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                 fallthrough;
  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         unsigned long nsid = 0;
 119         struct nvmet_ns *cur;
 120         unsigned long idx;
 121
 122         xa_for_each(&subsys->namespaces, idx, cur)
 123                 nsid = cur->nsid;
 124
 125         return nsid;
 126 }
 127
 128 static u32 nvmet_async_event_result(struct nvmet_async_event *aen)
 129 {
 130         return aen->event_type | (aen->event_info << 8) | (aen->log_page << 16);
 131 }
 132
 133 static void nvmet_async_events_failall(struct nvmet_ctrl *ctrl)
 134 {
 135         u16 status = NVME_SC_INTERNAL | NVME_SC_DNR;
 136         struct nvmet_req *req;
 137
 138         mutex_lock(&ctrl->lock);
 139         while (ctrl->nr_async_event_cmds) {
 140                 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
 141                 mutex_unlock(&ctrl->lock);
 142                 nvmet_req_complete(req, status);
 143                 mutex_lock(&ctrl->lock);
 144         }
 145         mutex_unlock(&ctrl->lock);
 146 }
 147
 148 static void nvmet_async_events_process(struct nvmet_ctrl *ctrl)
 149 {
 150         struct nvmet_async_event *aen;
 151         struct nvmet_req *req;
 152
 153         mutex_lock(&ctrl->lock);
 154         while (ctrl->nr_async_event_cmds && !list_empty(&ctrl->async_events)) {
 155                 aen = list_first_entry(&ctrl->async_events,
 156                                        struct nvmet_async_event, entry);
 157                 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
 158                 nvmet_set_result(req, nvmet_async_event_result(aen));
 159
 160                 list_del(&aen->entry);
 161                 kfree(aen);
 162
 163                 mutex_unlock(&ctrl->lock);
 164                 trace_nvmet_async_event(ctrl, req->cqe->result.u32);
 165                 nvmet_req_complete(req, 0);
 166                 mutex_lock(&ctrl->lock);
 167         }
 168         mutex_unlock(&ctrl->lock);
 169 }
 170
 171 static void nvmet_async_events_free(struct nvmet_ctrl *ctrl)
 172 {
 173         struct nvmet_async_event *aen, *tmp;
 174
 175         mutex_lock(&ctrl->lock);
 176         list_for_each_entry_safe(aen, tmp, &ctrl->async_events, entry) {
 177                 list_del(&aen->entry);
 178                 kfree(aen);
 179         }
 180         mutex_unlock(&ctrl->lock);
 181 }
 182
 183 static void nvmet_async_event_work(struct work_struct *work)
 184 {
 185         struct nvmet_ctrl *ctrl =
 186                 container_of(work, struct nvmet_ctrl, async_event_work);
 187
 188         nvmet_async_events_process(ctrl);
 189 }
 190
 191 void nvmet_add_async_event(struct nvmet_ctrl *ctrl, u8 event_type,
 192                 u8 event_info, u8 log_page)
 193 {
 194         struct nvmet_async_event *aen;
 195
 196         aen = kmalloc(sizeof(*aen), GFP_KERNEL);
 197         if (!aen)
 198                 return;
 199
 200         aen->event_type = event_type;
 201         aen->event_info = event_info;
 202         aen->log_page = log_page;
 203
 204         mutex_lock(&ctrl->lock);
 205         list_add_tail(&aen->entry, &ctrl->async_events);
 206         mutex_unlock(&ctrl->lock);
 207
 208         schedule_work(&ctrl->async_event_work);
 209 }
 210
 211 static void nvmet_add_to_changed_ns_log(struct nvmet_ctrl *ctrl, __le32 nsid)
 212 {
 213         u32 i;
 214
 215         mutex_lock(&ctrl->lock);
 216         if (ctrl->nr_changed_ns > NVME_MAX_CHANGED_NAMESPACES)
 217                 goto out_unlock;
 218
 219         for (i = 0; i < ctrl->nr_changed_ns; i++) {
 220                 if (ctrl->changed_ns_list[i] == nsid)
 221                         goto out_unlock;
 222         }
 223
 224         if (ctrl->nr_changed_ns == NVME_MAX_CHANGED_NAMESPACES) {
 225                 ctrl->changed_ns_list[0] = cpu_to_le32(0xffffffff);
 226                 ctrl->nr_changed_ns = U32_MAX;
 227                 goto out_unlock;
 228         }
 229
 230         ctrl->changed_ns_list[ctrl->nr_changed_ns++] = nsid;
 231 out_unlock:
 232         mutex_unlock(&ctrl->lock);
 233 }
 234
 235 void nvmet_ns_changed(struct nvmet_subsys *subsys, u32 nsid)
 236 {
 237         struct nvmet_ctrl *ctrl;
 238
 239         lockdep_assert_held(&subsys->lock);
 240
 241         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 242                 nvmet_add_to_changed_ns_log(ctrl, cpu_to_le32(nsid));
 243                 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_NS_ATTR))
 244                         continue;
 245                 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
 246                                 NVME_AER_NOTICE_NS_CHANGED,
 247                                 NVME_LOG_CHANGED_NS);
 248         }
 249 }
 250
 251 void nvmet_send_ana_event(struct nvmet_subsys *subsys,
 252                 struct nvmet_port *port)
 253 {
 254         struct nvmet_ctrl *ctrl;
 255
 256         mutex_lock(&subsys->lock);
 257         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 258                 if (port && ctrl->port != port)
 259                         continue;
 260                 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_ANA_CHANGE))
 261                         continue;
 262                 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
 263                                 NVME_AER_NOTICE_ANA, NVME_LOG_ANA);
 264         }
 265         mutex_unlock(&subsys->lock);
 266 }
 267
 268 void nvmet_port_send_ana_event(struct nvmet_port *port)
 269 {
 270         struct nvmet_subsys_link *p;
 271
 272         down_read(&nvmet_config_sem);
 273         list_for_each_entry(p, &port->subsystems, entry)
 274                 nvmet_send_ana_event(p->subsys, port);
 275         up_read(&nvmet_config_sem);
 276 }
 277
 278 int nvmet_register_transport(const struct nvmet_fabrics_ops *ops)
 279 {
 280         int ret = 0;
 281
 282         down_write(&nvmet_config_sem);
 283         if (nvmet_transports[ops->type])
 284                 ret = -EINVAL;
 285         else
 286                 nvmet_transports[ops->type] = ops;
 287         up_write(&nvmet_config_sem);
 288
 289         return ret;
 290 }
 291 EXPORT_SYMBOL_GPL(nvmet_register_transport);
 292
 293 void nvmet_unregister_transport(const struct nvmet_fabrics_ops *ops)
 294 {
 295         down_write(&nvmet_config_sem);
 296         nvmet_transports[ops->type] = NULL;
 297         up_write(&nvmet_config_sem);
 298 }
 299 EXPORT_SYMBOL_GPL(nvmet_unregister_transport);
 300
 301 void nvmet_port_del_ctrls(struct nvmet_port *port, struct nvmet_subsys *subsys)
 302 {
 303         struct nvmet_ctrl *ctrl;
 304
 305         mutex_lock(&subsys->lock);
 306         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 307                 if (ctrl->port == port)
 308                         ctrl->ops->delete_ctrl(ctrl);
 309         }
 310         mutex_unlock(&subsys->lock);
 311 }
 312
 313 int nvmet_enable_port(struct nvmet_port *port)
 314 {
 315         const struct nvmet_fabrics_ops *ops;
 316         int ret;
 317
 318         lockdep_assert_held(&nvmet_config_sem);
 319
 320         ops = nvmet_transports[port->disc_addr.trtype];
 321         if (!ops) {
 322                 up_write(&nvmet_config_sem);
 323                 request_module("nvmet-transport-%d", port->disc_addr.trtype);
 324                 down_write(&nvmet_config_sem);
 325                 ops = nvmet_transports[port->disc_addr.trtype];
 326                 if (!ops) {
 327                         pr_err("transport type %d not supported\n",
 328                                 port->disc_addr.trtype);
 329                         return -EINVAL;
 330                 }
 331         }
 332
 333         if (!try_module_get(ops->owner))
 334                 return -EINVAL;
 335
 336         /*
 337          * If the user requested PI support and the transport isn't pi capable,
 338          * don't enable the port.
 339          */
 340         if (port->pi_enable && !(ops->flags & NVMF_METADATA_SUPPORTED)) {
 341                 pr_err("T10-PI is not supported by transport type %d\n",
 342                        port->disc_addr.trtype);
 343                 ret = -EINVAL;
 344                 goto out_put;
 345         }
 346
 347         ret = ops->add_port(port);
 348         if (ret)
 349                 goto out_put;
 350
 351         /* If the transport didn't set inline_data_size, then disable it. */
 352         if (port->inline_data_size < 0)
 353                 port->inline_data_size = 0;
 354
 355         port->enabled = true;
 356         port->tr_ops = ops;
 357         return 0;
 358
 359 out_put:
 360         module_put(ops->owner);
 361         return ret;
 362 }
 363
 364 void nvmet_disable_port(struct nvmet_port *port)
 365 {
 366         const struct nvmet_fabrics_ops *ops;
 367
 368         lockdep_assert_held(&nvmet_config_sem);
 369
 370         port->enabled = false;
 371         port->tr_ops = NULL;
 372
 373         ops = nvmet_transports[port->disc_addr.trtype];
 374         ops->remove_port(port);
 375         module_put(ops->owner);
 376 }
 377
 378 static void nvmet_keep_alive_timer(struct work_struct *work)
 379 {
 380         struct nvmet_ctrl *ctrl = container_of(to_delayed_work(work),
 381                         struct nvmet_ctrl, ka_work);
 382         bool reset_tbkas = ctrl->reset_tbkas;
 383
 384         ctrl->reset_tbkas = false;
 385         if (reset_tbkas) {
 386                 pr_debug("ctrl %d reschedule traffic based keep-alive timer\n",
 387                         ctrl->cntlid);
 388                 schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
 389                 return;
 390         }
 391
 392         pr_err("ctrl %d keep-alive timer (%d seconds) expired!\n",
 393                 ctrl->cntlid, ctrl->kato);
 394
 395         nvmet_ctrl_fatal_error(ctrl);
 396 }
 397
 398 void nvmet_start_keep_alive_timer(struct nvmet_ctrl *ctrl)
 399 {
 400         if (unlikely(ctrl->kato == 0))
 401                 return;
 402
 403         pr_debug("ctrl %d start keep-alive timer for %d secs\n",
 404                 ctrl->cntlid, ctrl->kato);
 405
 406         INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer);
 407         schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
 408 }
 409
 410 void nvmet_stop_keep_alive_timer(struct nvmet_ctrl *ctrl)
 411 {
 412         if (unlikely(ctrl->kato == 0))
 413                 return;
 414
 415         pr_debug("ctrl %d stop keep-alive\n", ctrl->cntlid);
 416
 417         cancel_delayed_work_sync(&ctrl->ka_work);
 418 }
 419
 420 struct nvmet_ns *nvmet_find_namespace(struct nvmet_ctrl *ctrl, __le32 nsid)
 421 {
 422         struct nvmet_ns *ns;
 423
 424         ns = xa_load(&ctrl->subsys->namespaces, le32_to_cpu(nsid));
 425         if (ns)
 426                 percpu_ref_get(&ns->ref);
 427
 428         return ns;
 429 }
 430
 431 static void nvmet_destroy_namespace(struct percpu_ref *ref)
 432 {
 433         struct nvmet_ns *ns = container_of(ref, struct nvmet_ns, ref);
 434
 435         complete(&ns->disable_done);
 436 }
 437
 438 void nvmet_put_namespace(struct nvmet_ns *ns)
 439 {
 440         percpu_ref_put(&ns->ref);
 441 }
 442
 443 static void nvmet_ns_dev_disable(struct nvmet_ns *ns)
 444 {
 445         nvmet_bdev_ns_disable(ns);
 446         nvmet_file_ns_disable(ns);
 447 }
 448
 449 static int nvmet_p2pmem_ns_enable(struct nvmet_ns *ns)
 450 {
 451         int ret;
 452         struct pci_dev *p2p_dev;
 453
 454         if (!ns->use_p2pmem)
 455                 return 0;
 456
 457         if (!ns->bdev) {
 458                 pr_err("peer-to-peer DMA is not supported by non-block device namespaces\n");
 459                 return -EINVAL;
 460         }
 461
 462         if (!blk_queue_pci_p2pdma(ns->bdev->bd_disk->queue)) {
 463                 pr_err("peer-to-peer DMA is not supported by the driver of %s\n",
 464                        ns->device_path);
 465                 return -EINVAL;
 466         }
 467
 468         if (ns->p2p_dev) {
 469                 ret = pci_p2pdma_distance(ns->p2p_dev, nvmet_ns_dev(ns), true);
 470                 if (ret < 0)
 471                         return -EINVAL;
 472         } else {
 473                 /*
 474                  * Right now we just check that there is p2pmem available so
 475                  * we can report an error to the user right away if there
 476                  * is not. We'll find the actual device to use once we
 477                  * setup the controller when the port's device is available.
 478                  */
 479
 480                 p2p_dev = pci_p2pmem_find(nvmet_ns_dev(ns));
 481                 if (!p2p_dev) {
 482                         pr_err("no peer-to-peer memory is available for %s\n",
 483                                ns->device_path);
 484                         return -EINVAL;
 485                 }
 486
 487                 pci_dev_put(p2p_dev);
 488         }
 489
 490         return 0;
 491 }
 492
 493 /*
 494  * Note: ctrl->subsys->lock should be held when calling this function
 495  */
 496 static void nvmet_p2pmem_ns_add_p2p(struct nvmet_ctrl *ctrl,
 497                                     struct nvmet_ns *ns)
 498 {
 499         struct device *clients[2];
 500         struct pci_dev *p2p_dev;
 501         int ret;
 502
 503         if (!ctrl->p2p_client || !ns->use_p2pmem)
 504                 return;
 505
 506         if (ns->p2p_dev) {
 507                 ret = pci_p2pdma_distance(ns->p2p_dev, ctrl->p2p_client, true);
 508                 if (ret < 0)
 509                         return;
 510
 511                 p2p_dev = pci_dev_get(ns->p2p_dev);
 512         } else {
 513                 clients[0] = ctrl->p2p_client;
 514                 clients[1] = nvmet_ns_dev(ns);
 515
 516                 p2p_dev = pci_p2pmem_find_many(clients, ARRAY_SIZE(clients));
 517                 if (!p2p_dev) {
 518                         pr_err("no peer-to-peer memory is available that's supported by %s and %s\n",
 519                                dev_name(ctrl->p2p_client), ns->device_path);
 520                         return;
 521                 }
 522         }
 523
 524         ret = radix_tree_insert(&ctrl->p2p_ns_map, ns->nsid, p2p_dev);
 525         if (ret < 0)
 526                 pci_dev_put(p2p_dev);
 527
 528         pr_info("using p2pmem on %s for nsid %d\n", pci_name(p2p_dev),
 529                 ns->nsid);
 530 }
 531
 532 void nvmet_ns_revalidate(struct nvmet_ns *ns)
 533 {
 534         loff_t oldsize = ns->size;
 535
 536         if (ns->bdev)
 537                 nvmet_bdev_ns_revalidate(ns);
 538         else
 539                 nvmet_file_ns_revalidate(ns);
 540
 541         if (oldsize != ns->size)
 542                 nvmet_ns_changed(ns->subsys, ns->nsid);
 543 }
 544
 545 int nvmet_ns_enable(struct nvmet_ns *ns)
 546 {
 547         struct nvmet_subsys *subsys = ns->subsys;
 548         struct nvmet_ctrl *ctrl;
 549         int ret;
 550
 551         mutex_lock(&subsys->lock);
 552         ret = 0;
 553
 554         if (nvmet_passthru_ctrl(subsys)) {
 555                 pr_info("cannot enable both passthru and regular namespaces for a single subsystem");
 556                 goto out_unlock;
 557         }
 558
 559         if (ns->enabled)
 560                 goto out_unlock;
 561
 562         ret = -EMFILE;
 563         if (subsys->nr_namespaces == NVMET_MAX_NAMESPACES)
 564                 goto out_unlock;
 565
 566         ret = nvmet_bdev_ns_enable(ns);
 567         if (ret == -ENOTBLK)
 568                 ret = nvmet_file_ns_enable(ns);
 569         if (ret)
 570                 goto out_unlock;
 571
 572         ret = nvmet_p2pmem_ns_enable(ns);
 573         if (ret)
 574                 goto out_dev_disable;
 575
 576         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 577                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
 578
 579         ret = percpu_ref_init(&ns->ref, nvmet_destroy_namespace,
 580                                 0, GFP_KERNEL);
 581         if (ret)
 582                 goto out_dev_put;
 583
 584         if (ns->nsid > subsys->max_nsid)
 585                 subsys->max_nsid = ns->nsid;
 586
 587         ret = xa_insert(&subsys->namespaces, ns->nsid, ns, GFP_KERNEL);
 588         if (ret)
 589                 goto out_restore_subsys_maxnsid;
 590
 591         subsys->nr_namespaces++;
 592
 593         nvmet_ns_changed(subsys, ns->nsid);
 594         ns->enabled = true;
 595         ret = 0;
 596 out_unlock:
 597         mutex_unlock(&subsys->lock);
 598         return ret;
 599
 600 out_restore_subsys_maxnsid:
 601         subsys->max_nsid = nvmet_max_nsid(subsys);
 602         percpu_ref_exit(&ns->ref);
 603 out_dev_put:
 604         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 605                 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
 606 out_dev_disable:
 607         nvmet_ns_dev_disable(ns);
 608         goto out_unlock;
 609 }
 610
 611 void nvmet_ns_disable(struct nvmet_ns *ns)
 612 {
 613         struct nvmet_subsys *subsys = ns->subsys;
 614         struct nvmet_ctrl *ctrl;
 615
 616         mutex_lock(&subsys->lock);
 617         if (!ns->enabled)
 618                 goto out_unlock;
 619
 620         ns->enabled = false;
 621         xa_erase(&ns->subsys->namespaces, ns->nsid);
 622         if (ns->nsid == subsys->max_nsid)
 623                 subsys->max_nsid = nvmet_max_nsid(subsys);
 624
 625         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 626                 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
 627
 628         mutex_unlock(&subsys->lock);
 629
 630         /*
 631          * Now that we removed the namespaces from the lookup list, we
 632          * can kill the per_cpu ref and wait for any remaining references
 633          * to be dropped, as well as a RCU grace period for anyone only
 634          * using the namepace under rcu_read_lock().  Note that we can't
 635          * use call_rcu here as we need to ensure the namespaces have
 636          * been fully destroyed before unloading the module.
 637          */
 638         percpu_ref_kill(&ns->ref);
 639         synchronize_rcu();
 640         wait_for_completion(&ns->disable_done);
 641         percpu_ref_exit(&ns->ref);
 642
 643         mutex_lock(&subsys->lock);
 644
 645         subsys->nr_namespaces--;
 646         nvmet_ns_changed(subsys, ns->nsid);
 647         nvmet_ns_dev_disable(ns);
 648 out_unlock:
 649         mutex_unlock(&subsys->lock);
 650 }
 651
 652 void nvmet_ns_free(struct nvmet_ns *ns)
 653 {
 654         nvmet_ns_disable(ns);
 655
 656         down_write(&nvmet_ana_sem);
 657         nvmet_ana_group_enabled[ns->anagrpid]--;
 658         up_write(&nvmet_ana_sem);
 659
 660         kfree(ns->device_path);
 661         kfree(ns);
 662 }
 663
 664 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
 665 {
 666         struct nvmet_ns *ns;
 667
 668         ns = kzalloc(sizeof(*ns), GFP_KERNEL);
 669         if (!ns)
 670                 return NULL;
 671
 672         init_completion(&ns->disable_done);
 673
 674         ns->nsid = nsid;
 675         ns->subsys = subsys;
 676
 677         down_write(&nvmet_ana_sem);
 678         ns->anagrpid = NVMET_DEFAULT_ANA_GRPID;
 679         nvmet_ana_group_enabled[ns->anagrpid]++;
 680         up_write(&nvmet_ana_sem);
 681
 682         uuid_gen(&ns->uuid);
 683         ns->buffered_io = false;
 684
 685         return ns;
 686 }
 687
 688 static void nvmet_update_sq_head(struct nvmet_req *req)
 689 {
 690         if (req->sq->size) {
 691                 u32 old_sqhd, new_sqhd;
 692
 693                 do {
 694                         old_sqhd = req->sq->sqhd;
 695                         new_sqhd = (old_sqhd + 1) % req->sq->size;
 696                 } while (cmpxchg(&req->sq->sqhd, old_sqhd, new_sqhd) !=
 697                                         old_sqhd);
 698         }
 699         req->cqe->sq_head = cpu_to_le16(req->sq->sqhd & 0x0000FFFF);
 700 }
 701
 702 static void nvmet_set_error(struct nvmet_req *req, u16 status)
 703 {
 704         struct nvmet_ctrl *ctrl = req->sq->ctrl;
 705         struct nvme_error_slot *new_error_slot;
 706         unsigned long flags;
 707
 708         req->cqe->status = cpu_to_le16(status << 1);
 709
 710         if (!ctrl || req->error_loc == NVMET_NO_ERROR_LOC)
 711                 return;
 712
 713         spin_lock_irqsave(&ctrl->error_lock, flags);
 714         ctrl->err_counter++;
 715         new_error_slot =
 716                 &ctrl->slots[ctrl->err_counter % NVMET_ERROR_LOG_SLOTS];
 717
 718         new_error_slot->error_count = cpu_to_le64(ctrl->err_counter);
 719         new_error_slot->sqid = cpu_to_le16(req->sq->qid);
 720         new_error_slot->cmdid = cpu_to_le16(req->cmd->common.command_id);
 721         new_error_slot->status_field = cpu_to_le16(status << 1);
 722         new_error_slot->param_error_location = cpu_to_le16(req->error_loc);
 723         new_error_slot->lba = cpu_to_le64(req->error_slba);
 724         new_error_slot->nsid = req->cmd->common.nsid;
 725         spin_unlock_irqrestore(&ctrl->error_lock, flags);
 726
 727         /* set the more bit for this request */
 728         req->cqe->status |= cpu_to_le16(1 << 14);
 729 }
 730
 731 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
 732 {
 733         struct nvmet_ns *ns = req->ns;
 734
 735         if (!req->sq->sqhd_disabled)
 736                 nvmet_update_sq_head(req);
 737         req->cqe->sq_id = cpu_to_le16(req->sq->qid);
 738         req->cqe->command_id = req->cmd->common.command_id;
 739
 740         if (unlikely(status))
 741                 nvmet_set_error(req, status);
 742
 743         trace_nvmet_req_complete(req);
 744
 745         req->ops->queue_response(req);
 746         if (ns)
 747                 nvmet_put_namespace(ns);
 748 }
 749
 750 void nvmet_req_complete(struct nvmet_req *req, u16 status)
 751 {
 752         __nvmet_req_complete(req, status);
 753         percpu_ref_put(&req->sq->ref);
 754 }
 755 EXPORT_SYMBOL_GPL(nvmet_req_complete);
 756
 757 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
 758                 u16 qid, u16 size)
 759 {
 760         cq->qid = qid;
 761         cq->size = size;
 762 }
 763
 764 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
 765                 u16 qid, u16 size)
 766 {
 767         sq->sqhd = 0;
 768         sq->qid = qid;
 769         sq->size = size;
 770
 771         ctrl->sqs[qid] = sq;
 772 }
 773
 774 static void nvmet_confirm_sq(struct percpu_ref *ref)
 775 {
 776         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 777
 778         complete(&sq->confirm_done);
 779 }
 780
 781 void nvmet_sq_destroy(struct nvmet_sq *sq)
 782 {
 783         struct nvmet_ctrl *ctrl = sq->ctrl;
 784
 785         /*
 786          * If this is the admin queue, complete all AERs so that our
 787          * queue doesn't have outstanding requests on it.
 788          */
 789         if (ctrl && ctrl->sqs && ctrl->sqs[0] == sq)
 790                 nvmet_async_events_failall(ctrl);
 791         percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
 792         wait_for_completion(&sq->confirm_done);
 793         wait_for_completion(&sq->free_done);
 794         percpu_ref_exit(&sq->ref);
 795
 796         if (ctrl) {
 797                 /*
 798                  * The teardown flow may take some time, and the host may not
 799                  * send us keep-alive during this period, hence reset the
 800                  * traffic based keep-alive timer so we don't trigger a
 801                  * controller teardown as a result of a keep-alive expiration.
 802                  */
 803                 ctrl->reset_tbkas = true;
 804                 nvmet_ctrl_put(ctrl);
 805                 sq->ctrl = NULL; /* allows reusing the queue later */
 806         }
 807 }
 808 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
 809
 810 static void nvmet_sq_free(struct percpu_ref *ref)
 811 {
 812         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 813
 814         complete(&sq->free_done);
 815 }
 816
 817 int nvmet_sq_init(struct nvmet_sq *sq)
 818 {
 819         int ret;
 820
 821         ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
 822         if (ret) {
 823                 pr_err("percpu_ref init failed!\n");
 824                 return ret;
 825         }
 826         init_completion(&sq->free_done);
 827         init_completion(&sq->confirm_done);
 828
 829         return 0;
 830 }
 831 EXPORT_SYMBOL_GPL(nvmet_sq_init);
 832
 833 static inline u16 nvmet_check_ana_state(struct nvmet_port *port,
 834                 struct nvmet_ns *ns)
 835 {
 836         enum nvme_ana_state state = port->ana_state[ns->anagrpid];
 837
 838         if (unlikely(state == NVME_ANA_INACCESSIBLE))
 839                 return NVME_SC_ANA_INACCESSIBLE;
 840         if (unlikely(state == NVME_ANA_PERSISTENT_LOSS))
 841                 return NVME_SC_ANA_PERSISTENT_LOSS;
 842         if (unlikely(state == NVME_ANA_CHANGE))
 843                 return NVME_SC_ANA_TRANSITION;
 844         return 0;
 845 }
 846
 847 static inline u16 nvmet_io_cmd_check_access(struct nvmet_req *req)
 848 {
 849         if (unlikely(req->ns->readonly)) {
 850                 switch (req->cmd->common.opcode) {
 851                 case nvme_cmd_read:
 852                 case nvme_cmd_flush:
 853                         break;
 854                 default:
 855                         return NVME_SC_NS_WRITE_PROTECTED;
 856                 }
 857         }
 858
 859         return 0;
 860 }
 861
 862 static u16 nvmet_parse_io_cmd(struct nvmet_req *req)
 863 {
 864         struct nvme_command *cmd = req->cmd;
 865         u16 ret;
 866
 867         ret = nvmet_check_ctrl_status(req, cmd);
 868         if (unlikely(ret))
 869                 return ret;
 870
 871         if (nvmet_req_passthru_ctrl(req))
 872                 return nvmet_parse_passthru_io_cmd(req);
 873
 874         req->ns = nvmet_find_namespace(req->sq->ctrl, cmd->rw.nsid);
 875         if (unlikely(!req->ns)) {
 876                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 877                 return NVME_SC_INVALID_NS | NVME_SC_DNR;
 878         }
 879         ret = nvmet_check_ana_state(req->port, req->ns);
 880         if (unlikely(ret)) {
 881                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 882                 return ret;
 883         }
 884         ret = nvmet_io_cmd_check_access(req);
 885         if (unlikely(ret)) {
 886                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 887                 return ret;
 888         }
 889
 890         if (req->ns->file)
 891                 return nvmet_file_parse_io_cmd(req);
 892         else
 893                 return nvmet_bdev_parse_io_cmd(req);
 894 }
 895
 896 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
 897                 struct nvmet_sq *sq, const struct nvmet_fabrics_ops *ops)
 898 {
 899         u8 flags = req->cmd->common.flags;
 900         u16 status;
 901
 902         req->cq = cq;
 903         req->sq = sq;
 904         req->ops = ops;
 905         req->sg = NULL;
 906         req->metadata_sg = NULL;
 907         req->sg_cnt = 0;
 908         req->metadata_sg_cnt = 0;
 909         req->transfer_len = 0;
 910         req->metadata_len = 0;
 911         req->cqe->status = 0;
 912         req->cqe->sq_head = 0;
 913         req->ns = NULL;
 914         req->error_loc = NVMET_NO_ERROR_LOC;
 915         req->error_slba = 0;
 916
 917         /* no support for fused commands yet */
 918         if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
 919                 req->error_loc = offsetof(struct nvme_common_command, flags);
 920                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 921                 goto fail;
 922         }
 923
 924         /*
 925          * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
 926          * contains an address of a single contiguous physical buffer that is
 927          * byte aligned.
 928          */
 929         if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
 930                 req->error_loc = offsetof(struct nvme_common_command, flags);
 931                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 932                 goto fail;
 933         }
 934
 935         if (unlikely(!req->sq->ctrl))
 936                 /* will return an error for any non-connect command: */
 937                 status = nvmet_parse_connect_cmd(req);
 938         else if (likely(req->sq->qid != 0))
 939                 status = nvmet_parse_io_cmd(req);
 940         else
 941                 status = nvmet_parse_admin_cmd(req);
 942
 943         if (status)
 944                 goto fail;
 945
 946         trace_nvmet_req_init(req, req->cmd);
 947
 948         if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
 949                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 950                 goto fail;
 951         }
 952
 953         if (sq->ctrl)
 954                 sq->ctrl->reset_tbkas = true;
 955
 956         return true;
 957
 958 fail:
 959         __nvmet_req_complete(req, status);
 960         return false;
 961 }
 962 EXPORT_SYMBOL_GPL(nvmet_req_init);
 963
 964 void nvmet_req_uninit(struct nvmet_req *req)
 965 {
 966         percpu_ref_put(&req->sq->ref);
 967         if (req->ns)
 968                 nvmet_put_namespace(req->ns);
 969 }
 970 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
 971
 972 bool nvmet_check_transfer_len(struct nvmet_req *req, size_t len)
 973 {
 974         if (unlikely(len != req->transfer_len)) {
 975                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 976                 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
 977                 return false;
 978         }
 979
 980         return true;
 981 }
 982 EXPORT_SYMBOL_GPL(nvmet_check_transfer_len);
 983
 984 bool nvmet_check_data_len_lte(struct nvmet_req *req, size_t data_len)
 985 {
 986         if (unlikely(data_len > req->transfer_len)) {
 987                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 988                 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
 989                 return false;
 990         }
 991
 992         return true;
 993 }
 994
 995 static unsigned int nvmet_data_transfer_len(struct nvmet_req *req)
 996 {
 997         return req->transfer_len - req->metadata_len;
 998 }
 999
1000 static int nvmet_req_alloc_p2pmem_sgls(struct pci_dev *p2p_dev,
1001                 struct nvmet_req *req)
1002 {
1003         req->sg = pci_p2pmem_alloc_sgl(p2p_dev, &req->sg_cnt,
1004                         nvmet_data_transfer_len(req));
1005         if (!req->sg)
1006                 goto out_err;
1007
1008         if (req->metadata_len) {
1009                 req->metadata_sg = pci_p2pmem_alloc_sgl(p2p_dev,
1010                                 &req->metadata_sg_cnt, req->metadata_len);
1011                 if (!req->metadata_sg)
1012                         goto out_free_sg;
1013         }
1014
1015         req->p2p_dev = p2p_dev;
1016
1017         return 0;
1018 out_free_sg:
1019         pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1020 out_err:
1021         return -ENOMEM;
1022 }
1023
1024 static struct pci_dev *nvmet_req_find_p2p_dev(struct nvmet_req *req)
1025 {
1026         if (!IS_ENABLED(CONFIG_PCI_P2PDMA) ||
1027             !req->sq->ctrl || !req->sq->qid || !req->ns)
1028                 return NULL;
1029         return radix_tree_lookup(&req->sq->ctrl->p2p_ns_map, req->ns->nsid);
1030 }
1031
1032 int nvmet_req_alloc_sgls(struct nvmet_req *req)
1033 {
1034         struct pci_dev *p2p_dev = nvmet_req_find_p2p_dev(req);
1035
1036         if (p2p_dev && !nvmet_req_alloc_p2pmem_sgls(p2p_dev, req))
1037                 return 0;
1038
1039         req->sg = sgl_alloc(nvmet_data_transfer_len(req), GFP_KERNEL,
1040                             &req->sg_cnt);
1041         if (unlikely(!req->sg))
1042                 goto out;
1043
1044         if (req->metadata_len) {
1045                 req->metadata_sg = sgl_alloc(req->metadata_len, GFP_KERNEL,
1046                                              &req->metadata_sg_cnt);
1047                 if (unlikely(!req->metadata_sg))
1048                         goto out_free;
1049         }
1050
1051         return 0;
1052 out_free:
1053         sgl_free(req->sg);
1054 out:
1055         return -ENOMEM;
1056 }
1057 EXPORT_SYMBOL_GPL(nvmet_req_alloc_sgls);
1058
1059 void nvmet_req_free_sgls(struct nvmet_req *req)
1060 {
1061         if (req->p2p_dev) {
1062                 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1063                 if (req->metadata_sg)
1064                         pci_p2pmem_free_sgl(req->p2p_dev, req->metadata_sg);
1065                 req->p2p_dev = NULL;
1066         } else {
1067                 sgl_free(req->sg);
1068                 if (req->metadata_sg)
1069                         sgl_free(req->metadata_sg);
1070         }
1071
1072         req->sg = NULL;
1073         req->metadata_sg = NULL;
1074         req->sg_cnt = 0;
1075         req->metadata_sg_cnt = 0;
1076 }
1077 EXPORT_SYMBOL_GPL(nvmet_req_free_sgls);
1078
1079 static inline bool nvmet_cc_en(u32 cc)
1080 {
1081         return (cc >> NVME_CC_EN_SHIFT) & 0x1;
1082 }
1083
1084 static inline u8 nvmet_cc_css(u32 cc)
1085 {
1086         return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
1087 }
1088
1089 static inline u8 nvmet_cc_mps(u32 cc)
1090 {
1091         return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
1092 }
1093
1094 static inline u8 nvmet_cc_ams(u32 cc)
1095 {
1096         return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
1097 }
1098
1099 static inline u8 nvmet_cc_shn(u32 cc)
1100 {
1101         return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
1102 }
1103
1104 static inline u8 nvmet_cc_iosqes(u32 cc)
1105 {
1106         return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
1107 }
1108
1109 static inline u8 nvmet_cc_iocqes(u32 cc)
1110 {
1111         return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
1112 }
1113
1114 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
1115 {
1116         lockdep_assert_held(&ctrl->lock);
1117
1118         /*
1119          * Only I/O controllers should verify iosqes,iocqes.
1120          * Strictly speaking, the spec says a discovery controller
1121          * should verify iosqes,iocqes are zeroed, however that
1122          * would break backwards compatibility, so don't enforce it.
1123          */
1124         if (ctrl->subsys->type != NVME_NQN_DISC &&
1125             (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
1126              nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES)) {
1127                 ctrl->csts = NVME_CSTS_CFS;
1128                 return;
1129         }
1130
1131         if (nvmet_cc_mps(ctrl->cc) != 0 ||
1132             nvmet_cc_ams(ctrl->cc) != 0 ||
1133             nvmet_cc_css(ctrl->cc) != 0) {
1134                 ctrl->csts = NVME_CSTS_CFS;
1135                 return;
1136         }
1137
1138         ctrl->csts = NVME_CSTS_RDY;
1139
1140         /*
1141          * Controllers that are not yet enabled should not really enforce the
1142          * keep alive timeout, but we still want to track a timeout and cleanup
1143          * in case a host died before it enabled the controller.  Hence, simply
1144          * reset the keep alive timer when the controller is enabled.
1145          */
1146         if (ctrl->kato)
1147                 mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ);
1148 }
1149
1150 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
1151 {
1152         lockdep_assert_held(&ctrl->lock);
1153
1154         /* XXX: tear down queues? */
1155         ctrl->csts &= ~NVME_CSTS_RDY;
1156         ctrl->cc = 0;
1157 }
1158
1159 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
1160 {
1161         u32 old;
1162
1163         mutex_lock(&ctrl->lock);
1164         old = ctrl->cc;
1165         ctrl->cc = new;
1166
1167         if (nvmet_cc_en(new) && !nvmet_cc_en(old))
1168                 nvmet_start_ctrl(ctrl);
1169         if (!nvmet_cc_en(new) && nvmet_cc_en(old))
1170                 nvmet_clear_ctrl(ctrl);
1171         if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
1172                 nvmet_clear_ctrl(ctrl);
1173                 ctrl->csts |= NVME_CSTS_SHST_CMPLT;
1174         }
1175         if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
1176                 ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
1177         mutex_unlock(&ctrl->lock);
1178 }
1179
1180 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
1181 {
1182         /* command sets supported: NVMe command set: */
1183         ctrl->cap = (1ULL << 37);
1184         /* CC.EN timeout in 500msec units: */
1185         ctrl->cap |= (15ULL << 24);
1186         /* maximum queue entries supported: */
1187         ctrl->cap |= NVMET_QUEUE_SIZE - 1;
1188 }
1189
1190 u16 nvmet_ctrl_find_get(const char *subsysnqn, const char *hostnqn, u16 cntlid,
1191                 struct nvmet_req *req, struct nvmet_ctrl **ret)
1192 {
1193         struct nvmet_subsys *subsys;
1194         struct nvmet_ctrl *ctrl;
1195         u16 status = 0;
1196
1197         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1198         if (!subsys) {
1199                 pr_warn("connect request for invalid subsystem %s!\n",
1200                         subsysnqn);
1201                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1202                 return NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1203         }
1204
1205         mutex_lock(&subsys->lock);
1206         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
1207                 if (ctrl->cntlid == cntlid) {
1208                         if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
1209                                 pr_warn("hostnqn mismatch.\n");
1210                                 continue;
1211                         }
1212                         if (!kref_get_unless_zero(&ctrl->ref))
1213                                 continue;
1214
1215                         *ret = ctrl;
1216                         goto out;
1217                 }
1218         }
1219
1220         pr_warn("could not find controller %d for subsys %s / host %s\n",
1221                 cntlid, subsysnqn, hostnqn);
1222         req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
1223         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1224
1225 out:
1226         mutex_unlock(&subsys->lock);
1227         nvmet_subsys_put(subsys);
1228         return status;
1229 }
1230
1231 u16 nvmet_check_ctrl_status(struct nvmet_req *req, struct nvme_command *cmd)
1232 {
1233         if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
1234                 pr_err("got cmd %d while CC.EN == 0 on qid = %d\n",
1235                        cmd->common.opcode, req->sq->qid);
1236                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1237         }
1238
1239         if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
1240                 pr_err("got cmd %d while CSTS.RDY == 0 on qid = %d\n",
1241                        cmd->common.opcode, req->sq->qid);
1242                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1243         }
1244         return 0;
1245 }
1246
1247 bool nvmet_host_allowed(struct nvmet_subsys *subsys, const char *hostnqn)
1248 {
1249         struct nvmet_host_link *p;
1250
1251         lockdep_assert_held(&nvmet_config_sem);
1252
1253         if (subsys->allow_any_host)
1254                 return true;
1255
1256         if (subsys->type == NVME_NQN_DISC) /* allow all access to disc subsys */
1257                 return true;
1258
1259         list_for_each_entry(p, &subsys->hosts, entry) {
1260                 if (!strcmp(nvmet_host_name(p->host), hostnqn))
1261                         return true;
1262         }
1263
1264         return false;
1265 }
1266
1267 /*
1268  * Note: ctrl->subsys->lock should be held when calling this function
1269  */
1270 static void nvmet_setup_p2p_ns_map(struct nvmet_ctrl *ctrl,
1271                 struct nvmet_req *req)
1272 {
1273         struct nvmet_ns *ns;
1274         unsigned long idx;
1275
1276         if (!req->p2p_client)
1277                 return;
1278
1279         ctrl->p2p_client = get_device(req->p2p_client);
1280
1281         xa_for_each(&ctrl->subsys->namespaces, idx, ns)
1282                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
1283 }
1284
1285 /*
1286  * Note: ctrl->subsys->lock should be held when calling this function
1287  */
1288 static void nvmet_release_p2p_ns_map(struct nvmet_ctrl *ctrl)
1289 {
1290         struct radix_tree_iter iter;
1291         void __rcu **slot;
1292
1293         radix_tree_for_each_slot(slot, &ctrl->p2p_ns_map, &iter, 0)
1294                 pci_dev_put(radix_tree_deref_slot(slot));
1295
1296         put_device(ctrl->p2p_client);
1297 }
1298
1299 static void nvmet_fatal_error_handler(struct work_struct *work)
1300 {
1301         struct nvmet_ctrl *ctrl =
1302                         container_of(work, struct nvmet_ctrl, fatal_err_work);
1303
1304         pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
1305         ctrl->ops->delete_ctrl(ctrl);
1306 }
1307
1308 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
1309                 struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
1310 {
1311         struct nvmet_subsys *subsys;
1312         struct nvmet_ctrl *ctrl;
1313         int ret;
1314         u16 status;
1315
1316         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1317         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1318         if (!subsys) {
1319                 pr_warn("connect request for invalid subsystem %s!\n",
1320                         subsysnqn);
1321                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1322                 goto out;
1323         }
1324
1325         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1326         down_read(&nvmet_config_sem);
1327         if (!nvmet_host_allowed(subsys, hostnqn)) {
1328                 pr_info("connect by host %s for subsystem %s not allowed\n",
1329                         hostnqn, subsysnqn);
1330                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
1331                 up_read(&nvmet_config_sem);
1332                 status = NVME_SC_CONNECT_INVALID_HOST | NVME_SC_DNR;
1333                 goto out_put_subsystem;
1334         }
1335         up_read(&nvmet_config_sem);
1336
1337         status = NVME_SC_INTERNAL;
1338         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1339         if (!ctrl)
1340                 goto out_put_subsystem;
1341         mutex_init(&ctrl->lock);
1342
1343         nvmet_init_cap(ctrl);
1344
1345         ctrl->port = req->port;
1346
1347         INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
1348         INIT_LIST_HEAD(&ctrl->async_events);
1349         INIT_RADIX_TREE(&ctrl->p2p_ns_map, GFP_KERNEL);
1350         INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
1351
1352         memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
1353         memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
1354
1355         kref_init(&ctrl->ref);
1356         ctrl->subsys = subsys;
1357         WRITE_ONCE(ctrl->aen_enabled, NVMET_AEN_CFG_OPTIONAL);
1358
1359         ctrl->changed_ns_list = kmalloc_array(NVME_MAX_CHANGED_NAMESPACES,
1360                         sizeof(__le32), GFP_KERNEL);
1361         if (!ctrl->changed_ns_list)
1362                 goto out_free_ctrl;
1363
1364         ctrl->sqs = kcalloc(subsys->max_qid + 1,
1365                         sizeof(struct nvmet_sq *),
1366                         GFP_KERNEL);
1367         if (!ctrl->sqs)
1368                 goto out_free_changed_ns_list;
1369
1370         if (subsys->cntlid_min > subsys->cntlid_max)
1371                 goto out_free_sqs;
1372
1373         ret = ida_simple_get(&cntlid_ida,
1374                              subsys->cntlid_min, subsys->cntlid_max,
1375                              GFP_KERNEL);
1376         if (ret < 0) {
1377                 status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
1378                 goto out_free_sqs;
1379         }
1380         ctrl->cntlid = ret;
1381
1382         ctrl->ops = req->ops;
1383
1384         /*
1385          * Discovery controllers may use some arbitrary high value
1386          * in order to cleanup stale discovery sessions
1387          */
1388         if ((ctrl->subsys->type == NVME_NQN_DISC) && !kato)
1389                 kato = NVMET_DISC_KATO_MS;
1390
1391         /* keep-alive timeout in seconds */
1392         ctrl->kato = DIV_ROUND_UP(kato, 1000);
1393
1394         ctrl->err_counter = 0;
1395         spin_lock_init(&ctrl->error_lock);
1396
1397         nvmet_start_keep_alive_timer(ctrl);
1398
1399         mutex_lock(&subsys->lock);
1400         list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
1401         nvmet_setup_p2p_ns_map(ctrl, req);
1402         mutex_unlock(&subsys->lock);
1403
1404         *ctrlp = ctrl;
1405         return 0;
1406
1407 out_free_sqs:
1408         kfree(ctrl->sqs);
1409 out_free_changed_ns_list:
1410         kfree(ctrl->changed_ns_list);
1411 out_free_ctrl:
1412         kfree(ctrl);
1413 out_put_subsystem:
1414         nvmet_subsys_put(subsys);
1415 out:
1416         return status;
1417 }
1418
1419 static void nvmet_ctrl_free(struct kref *ref)
1420 {
1421         struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
1422         struct nvmet_subsys *subsys = ctrl->subsys;
1423
1424         mutex_lock(&subsys->lock);
1425         nvmet_release_p2p_ns_map(ctrl);
1426         list_del(&ctrl->subsys_entry);
1427         mutex_unlock(&subsys->lock);
1428
1429         nvmet_stop_keep_alive_timer(ctrl);
1430
1431         flush_work(&ctrl->async_event_work);
1432         cancel_work_sync(&ctrl->fatal_err_work);
1433
1434         ida_simple_remove(&cntlid_ida, ctrl->cntlid);
1435
1436         nvmet_async_events_free(ctrl);
1437         kfree(ctrl->sqs);
1438         kfree(ctrl->changed_ns_list);
1439         kfree(ctrl);
1440
1441         nvmet_subsys_put(subsys);
1442 }
1443
1444 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
1445 {
1446         kref_put(&ctrl->ref, nvmet_ctrl_free);
1447 }
1448
1449 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
1450 {
1451         mutex_lock(&ctrl->lock);
1452         if (!(ctrl->csts & NVME_CSTS_CFS)) {
1453                 ctrl->csts |= NVME_CSTS_CFS;
1454                 schedule_work(&ctrl->fatal_err_work);
1455         }
1456         mutex_unlock(&ctrl->lock);
1457 }
1458 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
1459
1460 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
1461                 const char *subsysnqn)
1462 {
1463         struct nvmet_subsys_link *p;
1464
1465         if (!port)
1466                 return NULL;
1467
1468         if (!strcmp(NVME_DISC_SUBSYS_NAME, subsysnqn)) {
1469                 if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
1470                         return NULL;
1471                 return nvmet_disc_subsys;
1472         }
1473
1474         down_read(&nvmet_config_sem);
1475         list_for_each_entry(p, &port->subsystems, entry) {
1476                 if (!strncmp(p->subsys->subsysnqn, subsysnqn,
1477                                 NVMF_NQN_SIZE)) {
1478                         if (!kref_get_unless_zero(&p->subsys->ref))
1479                                 break;
1480                         up_read(&nvmet_config_sem);
1481                         return p->subsys;
1482                 }
1483         }
1484         up_read(&nvmet_config_sem);
1485         return NULL;
1486 }
1487
1488 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
1489                 enum nvme_subsys_type type)
1490 {
1491         struct nvmet_subsys *subsys;
1492
1493         subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
1494         if (!subsys)
1495                 return ERR_PTR(-ENOMEM);
1496
1497         subsys->ver = NVMET_DEFAULT_VS;
1498         /* generate a random serial number as our controllers are ephemeral: */
1499         get_random_bytes(&subsys->serial, sizeof(subsys->serial));
1500
1501         switch (type) {
1502         case NVME_NQN_NVME:
1503                 subsys->max_qid = NVMET_NR_QUEUES;
1504                 break;
1505         case NVME_NQN_DISC:
1506                 subsys->max_qid = 0;
1507                 break;
1508         default:
1509                 pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
1510                 kfree(subsys);
1511                 return ERR_PTR(-EINVAL);
1512         }
1513         subsys->type = type;
1514         subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
1515                         GFP_KERNEL);
1516         if (!subsys->subsysnqn) {
1517                 kfree(subsys);
1518                 return ERR_PTR(-ENOMEM);
1519         }
1520         subsys->cntlid_min = NVME_CNTLID_MIN;
1521         subsys->cntlid_max = NVME_CNTLID_MAX;
1522         kref_init(&subsys->ref);
1523
1524         mutex_init(&subsys->lock);
1525         xa_init(&subsys->namespaces);
1526         INIT_LIST_HEAD(&subsys->ctrls);
1527         INIT_LIST_HEAD(&subsys->hosts);
1528
1529         return subsys;
1530 }
1531
1532 static void nvmet_subsys_free(struct kref *ref)
1533 {
1534         struct nvmet_subsys *subsys =
1535                 container_of(ref, struct nvmet_subsys, ref);
1536
1537         WARN_ON_ONCE(!xa_empty(&subsys->namespaces));
1538
1539         xa_destroy(&subsys->namespaces);
1540         nvmet_passthru_subsys_free(subsys);
1541
1542         kfree(subsys->subsysnqn);
1543         kfree_rcu(subsys->model, rcuhead);
1544         kfree(subsys);
1545 }
1546
1547 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1548 {
1549         struct nvmet_ctrl *ctrl;
1550
1551         mutex_lock(&subsys->lock);
1552         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1553                 ctrl->ops->delete_ctrl(ctrl);
1554         mutex_unlock(&subsys->lock);
1555 }
1556
1557 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1558 {
1559         kref_put(&subsys->ref, nvmet_subsys_free);
1560 }
1561
1562 static int __init nvmet_init(void)
1563 {
1564         int error;
1565
1566         nvmet_ana_group_enabled[NVMET_DEFAULT_ANA_GRPID] = 1;
1567
1568         buffered_io_wq = alloc_workqueue("nvmet-buffered-io-wq",
1569                         WQ_MEM_RECLAIM, 0);
1570         if (!buffered_io_wq) {
1571                 error = -ENOMEM;
1572                 goto out;
1573         }
1574
1575         error = nvmet_init_discovery();
1576         if (error)
1577                 goto out_free_work_queue;
1578
1579         error = nvmet_init_configfs();
1580         if (error)
1581                 goto out_exit_discovery;
1582         return 0;
1583
1584 out_exit_discovery:
1585         nvmet_exit_discovery();
1586 out_free_work_queue:
1587         destroy_workqueue(buffered_io_wq);
1588 out:
1589         return error;
1590 }
1591
1592 static void __exit nvmet_exit(void)
1593 {
1594         nvmet_exit_configfs();
1595         nvmet_exit_discovery();
1596         ida_destroy(&cntlid_ida);
1597         destroy_workqueue(buffered_io_wq);
1598
1599         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1600         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1601 }
1602
1603 module_init(nvmet_init);
1604 module_exit(nvmet_exit);
1605
1606 MODULE_LICENSE("GPL v2");