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 cmd_seen = ctrl->cmd_seen;
 383
 384         ctrl->cmd_seen = false;
 385         if (cmd_seen) {
 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         if (!req->sq->sqhd_disabled)
 734                 nvmet_update_sq_head(req);
 735         req->cqe->sq_id = cpu_to_le16(req->sq->qid);
 736         req->cqe->command_id = req->cmd->common.command_id;
 737
 738         if (unlikely(status))
 739                 nvmet_set_error(req, status);
 740
 741         trace_nvmet_req_complete(req);
 742
 743         if (req->ns)
 744                 nvmet_put_namespace(req->ns);
 745         req->ops->queue_response(req);
 746 }
 747
 748 void nvmet_req_complete(struct nvmet_req *req, u16 status)
 749 {
 750         __nvmet_req_complete(req, status);
 751         percpu_ref_put(&req->sq->ref);
 752 }
 753 EXPORT_SYMBOL_GPL(nvmet_req_complete);
 754
 755 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
 756                 u16 qid, u16 size)
 757 {
 758         cq->qid = qid;
 759         cq->size = size;
 760 }
 761
 762 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
 763                 u16 qid, u16 size)
 764 {
 765         sq->sqhd = 0;
 766         sq->qid = qid;
 767         sq->size = size;
 768
 769         ctrl->sqs[qid] = sq;
 770 }
 771
 772 static void nvmet_confirm_sq(struct percpu_ref *ref)
 773 {
 774         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 775
 776         complete(&sq->confirm_done);
 777 }
 778
 779 void nvmet_sq_destroy(struct nvmet_sq *sq)
 780 {
 781         struct nvmet_ctrl *ctrl = sq->ctrl;
 782
 783         /*
 784          * If this is the admin queue, complete all AERs so that our
 785          * queue doesn't have outstanding requests on it.
 786          */
 787         if (ctrl && ctrl->sqs && ctrl->sqs[0] == sq)
 788                 nvmet_async_events_failall(ctrl);
 789         percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
 790         wait_for_completion(&sq->confirm_done);
 791         wait_for_completion(&sq->free_done);
 792         percpu_ref_exit(&sq->ref);
 793
 794         if (ctrl) {
 795                 nvmet_ctrl_put(ctrl);
 796                 sq->ctrl = NULL; /* allows reusing the queue later */
 797         }
 798 }
 799 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
 800
 801 static void nvmet_sq_free(struct percpu_ref *ref)
 802 {
 803         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 804
 805         complete(&sq->free_done);
 806 }
 807
 808 int nvmet_sq_init(struct nvmet_sq *sq)
 809 {
 810         int ret;
 811
 812         ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
 813         if (ret) {
 814                 pr_err("percpu_ref init failed!\n");
 815                 return ret;
 816         }
 817         init_completion(&sq->free_done);
 818         init_completion(&sq->confirm_done);
 819
 820         return 0;
 821 }
 822 EXPORT_SYMBOL_GPL(nvmet_sq_init);
 823
 824 static inline u16 nvmet_check_ana_state(struct nvmet_port *port,
 825                 struct nvmet_ns *ns)
 826 {
 827         enum nvme_ana_state state = port->ana_state[ns->anagrpid];
 828
 829         if (unlikely(state == NVME_ANA_INACCESSIBLE))
 830                 return NVME_SC_ANA_INACCESSIBLE;
 831         if (unlikely(state == NVME_ANA_PERSISTENT_LOSS))
 832                 return NVME_SC_ANA_PERSISTENT_LOSS;
 833         if (unlikely(state == NVME_ANA_CHANGE))
 834                 return NVME_SC_ANA_TRANSITION;
 835         return 0;
 836 }
 837
 838 static inline u16 nvmet_io_cmd_check_access(struct nvmet_req *req)
 839 {
 840         if (unlikely(req->ns->readonly)) {
 841                 switch (req->cmd->common.opcode) {
 842                 case nvme_cmd_read:
 843                 case nvme_cmd_flush:
 844                         break;
 845                 default:
 846                         return NVME_SC_NS_WRITE_PROTECTED;
 847                 }
 848         }
 849
 850         return 0;
 851 }
 852
 853 static u16 nvmet_parse_io_cmd(struct nvmet_req *req)
 854 {
 855         struct nvme_command *cmd = req->cmd;
 856         u16 ret;
 857
 858         ret = nvmet_check_ctrl_status(req, cmd);
 859         if (unlikely(ret))
 860                 return ret;
 861
 862         if (nvmet_req_passthru_ctrl(req))
 863                 return nvmet_parse_passthru_io_cmd(req);
 864
 865         req->ns = nvmet_find_namespace(req->sq->ctrl, cmd->rw.nsid);
 866         if (unlikely(!req->ns)) {
 867                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 868                 return NVME_SC_INVALID_NS | NVME_SC_DNR;
 869         }
 870         ret = nvmet_check_ana_state(req->port, req->ns);
 871         if (unlikely(ret)) {
 872                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 873                 return ret;
 874         }
 875         ret = nvmet_io_cmd_check_access(req);
 876         if (unlikely(ret)) {
 877                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 878                 return ret;
 879         }
 880
 881         if (req->ns->file)
 882                 return nvmet_file_parse_io_cmd(req);
 883         else
 884                 return nvmet_bdev_parse_io_cmd(req);
 885 }
 886
 887 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
 888                 struct nvmet_sq *sq, const struct nvmet_fabrics_ops *ops)
 889 {
 890         u8 flags = req->cmd->common.flags;
 891         u16 status;
 892
 893         req->cq = cq;
 894         req->sq = sq;
 895         req->ops = ops;
 896         req->sg = NULL;
 897         req->metadata_sg = NULL;
 898         req->sg_cnt = 0;
 899         req->metadata_sg_cnt = 0;
 900         req->transfer_len = 0;
 901         req->metadata_len = 0;
 902         req->cqe->status = 0;
 903         req->cqe->sq_head = 0;
 904         req->ns = NULL;
 905         req->error_loc = NVMET_NO_ERROR_LOC;
 906         req->error_slba = 0;
 907
 908         /* no support for fused commands yet */
 909         if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
 910                 req->error_loc = offsetof(struct nvme_common_command, flags);
 911                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 912                 goto fail;
 913         }
 914
 915         /*
 916          * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
 917          * contains an address of a single contiguous physical buffer that is
 918          * byte aligned.
 919          */
 920         if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
 921                 req->error_loc = offsetof(struct nvme_common_command, flags);
 922                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 923                 goto fail;
 924         }
 925
 926         if (unlikely(!req->sq->ctrl))
 927                 /* will return an error for any non-connect command: */
 928                 status = nvmet_parse_connect_cmd(req);
 929         else if (likely(req->sq->qid != 0))
 930                 status = nvmet_parse_io_cmd(req);
 931         else
 932                 status = nvmet_parse_admin_cmd(req);
 933
 934         if (status)
 935                 goto fail;
 936
 937         trace_nvmet_req_init(req, req->cmd);
 938
 939         if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
 940                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 941                 goto fail;
 942         }
 943
 944         if (sq->ctrl)
 945                 sq->ctrl->cmd_seen = true;
 946
 947         return true;
 948
 949 fail:
 950         __nvmet_req_complete(req, status);
 951         return false;
 952 }
 953 EXPORT_SYMBOL_GPL(nvmet_req_init);
 954
 955 void nvmet_req_uninit(struct nvmet_req *req)
 956 {
 957         percpu_ref_put(&req->sq->ref);
 958         if (req->ns)
 959                 nvmet_put_namespace(req->ns);
 960 }
 961 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
 962
 963 bool nvmet_check_transfer_len(struct nvmet_req *req, size_t len)
 964 {
 965         if (unlikely(len != req->transfer_len)) {
 966                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 967                 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
 968                 return false;
 969         }
 970
 971         return true;
 972 }
 973 EXPORT_SYMBOL_GPL(nvmet_check_transfer_len);
 974
 975 bool nvmet_check_data_len_lte(struct nvmet_req *req, size_t data_len)
 976 {
 977         if (unlikely(data_len > req->transfer_len)) {
 978                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 979                 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
 980                 return false;
 981         }
 982
 983         return true;
 984 }
 985
 986 static unsigned int nvmet_data_transfer_len(struct nvmet_req *req)
 987 {
 988         return req->transfer_len - req->metadata_len;
 989 }
 990
 991 static int nvmet_req_alloc_p2pmem_sgls(struct nvmet_req *req)
 992 {
 993         req->sg = pci_p2pmem_alloc_sgl(req->p2p_dev, &req->sg_cnt,
 994                         nvmet_data_transfer_len(req));
 995         if (!req->sg)
 996                 goto out_err;
 997
 998         if (req->metadata_len) {
 999                 req->metadata_sg = pci_p2pmem_alloc_sgl(req->p2p_dev,
1000                                 &req->metadata_sg_cnt, req->metadata_len);
1001                 if (!req->metadata_sg)
1002                         goto out_free_sg;
1003         }
1004         return 0;
1005 out_free_sg:
1006         pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1007 out_err:
1008         return -ENOMEM;
1009 }
1010
1011 static bool nvmet_req_find_p2p_dev(struct nvmet_req *req)
1012 {
1013         if (!IS_ENABLED(CONFIG_PCI_P2PDMA))
1014                 return false;
1015
1016         if (req->sq->ctrl && req->sq->qid && req->ns) {
1017                 req->p2p_dev = radix_tree_lookup(&req->sq->ctrl->p2p_ns_map,
1018                                                  req->ns->nsid);
1019                 if (req->p2p_dev)
1020                         return true;
1021         }
1022
1023         req->p2p_dev = NULL;
1024         return false;
1025 }
1026
1027 int nvmet_req_alloc_sgls(struct nvmet_req *req)
1028 {
1029         if (nvmet_req_find_p2p_dev(req) && !nvmet_req_alloc_p2pmem_sgls(req))
1030                 return 0;
1031
1032         req->sg = sgl_alloc(nvmet_data_transfer_len(req), GFP_KERNEL,
1033                             &req->sg_cnt);
1034         if (unlikely(!req->sg))
1035                 goto out;
1036
1037         if (req->metadata_len) {
1038                 req->metadata_sg = sgl_alloc(req->metadata_len, GFP_KERNEL,
1039                                              &req->metadata_sg_cnt);
1040                 if (unlikely(!req->metadata_sg))
1041                         goto out_free;
1042         }
1043
1044         return 0;
1045 out_free:
1046         sgl_free(req->sg);
1047 out:
1048         return -ENOMEM;
1049 }
1050 EXPORT_SYMBOL_GPL(nvmet_req_alloc_sgls);
1051
1052 void nvmet_req_free_sgls(struct nvmet_req *req)
1053 {
1054         if (req->p2p_dev) {
1055                 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1056                 if (req->metadata_sg)
1057                         pci_p2pmem_free_sgl(req->p2p_dev, req->metadata_sg);
1058         } else {
1059                 sgl_free(req->sg);
1060                 if (req->metadata_sg)
1061                         sgl_free(req->metadata_sg);
1062         }
1063
1064         req->sg = NULL;
1065         req->metadata_sg = NULL;
1066         req->sg_cnt = 0;
1067         req->metadata_sg_cnt = 0;
1068 }
1069 EXPORT_SYMBOL_GPL(nvmet_req_free_sgls);
1070
1071 static inline bool nvmet_cc_en(u32 cc)
1072 {
1073         return (cc >> NVME_CC_EN_SHIFT) & 0x1;
1074 }
1075
1076 static inline u8 nvmet_cc_css(u32 cc)
1077 {
1078         return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
1079 }
1080
1081 static inline u8 nvmet_cc_mps(u32 cc)
1082 {
1083         return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
1084 }
1085
1086 static inline u8 nvmet_cc_ams(u32 cc)
1087 {
1088         return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
1089 }
1090
1091 static inline u8 nvmet_cc_shn(u32 cc)
1092 {
1093         return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
1094 }
1095
1096 static inline u8 nvmet_cc_iosqes(u32 cc)
1097 {
1098         return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
1099 }
1100
1101 static inline u8 nvmet_cc_iocqes(u32 cc)
1102 {
1103         return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
1104 }
1105
1106 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
1107 {
1108         lockdep_assert_held(&ctrl->lock);
1109
1110         if (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
1111             nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES ||
1112             nvmet_cc_mps(ctrl->cc) != 0 ||
1113             nvmet_cc_ams(ctrl->cc) != 0 ||
1114             nvmet_cc_css(ctrl->cc) != 0) {
1115                 ctrl->csts = NVME_CSTS_CFS;
1116                 return;
1117         }
1118
1119         ctrl->csts = NVME_CSTS_RDY;
1120
1121         /*
1122          * Controllers that are not yet enabled should not really enforce the
1123          * keep alive timeout, but we still want to track a timeout and cleanup
1124          * in case a host died before it enabled the controller.  Hence, simply
1125          * reset the keep alive timer when the controller is enabled.
1126          */
1127         if (ctrl->kato)
1128                 mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ);
1129 }
1130
1131 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
1132 {
1133         lockdep_assert_held(&ctrl->lock);
1134
1135         /* XXX: tear down queues? */
1136         ctrl->csts &= ~NVME_CSTS_RDY;
1137         ctrl->cc = 0;
1138 }
1139
1140 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
1141 {
1142         u32 old;
1143
1144         mutex_lock(&ctrl->lock);
1145         old = ctrl->cc;
1146         ctrl->cc = new;
1147
1148         if (nvmet_cc_en(new) && !nvmet_cc_en(old))
1149                 nvmet_start_ctrl(ctrl);
1150         if (!nvmet_cc_en(new) && nvmet_cc_en(old))
1151                 nvmet_clear_ctrl(ctrl);
1152         if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
1153                 nvmet_clear_ctrl(ctrl);
1154                 ctrl->csts |= NVME_CSTS_SHST_CMPLT;
1155         }
1156         if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
1157                 ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
1158         mutex_unlock(&ctrl->lock);
1159 }
1160
1161 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
1162 {
1163         /* command sets supported: NVMe command set: */
1164         ctrl->cap = (1ULL << 37);
1165         /* CC.EN timeout in 500msec units: */
1166         ctrl->cap |= (15ULL << 24);
1167         /* maximum queue entries supported: */
1168         ctrl->cap |= NVMET_QUEUE_SIZE - 1;
1169 }
1170
1171 u16 nvmet_ctrl_find_get(const char *subsysnqn, const char *hostnqn, u16 cntlid,
1172                 struct nvmet_req *req, struct nvmet_ctrl **ret)
1173 {
1174         struct nvmet_subsys *subsys;
1175         struct nvmet_ctrl *ctrl;
1176         u16 status = 0;
1177
1178         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1179         if (!subsys) {
1180                 pr_warn("connect request for invalid subsystem %s!\n",
1181                         subsysnqn);
1182                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1183                 return NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1184         }
1185
1186         mutex_lock(&subsys->lock);
1187         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
1188                 if (ctrl->cntlid == cntlid) {
1189                         if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
1190                                 pr_warn("hostnqn mismatch.\n");
1191                                 continue;
1192                         }
1193                         if (!kref_get_unless_zero(&ctrl->ref))
1194                                 continue;
1195
1196                         *ret = ctrl;
1197                         goto out;
1198                 }
1199         }
1200
1201         pr_warn("could not find controller %d for subsys %s / host %s\n",
1202                 cntlid, subsysnqn, hostnqn);
1203         req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
1204         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1205
1206 out:
1207         mutex_unlock(&subsys->lock);
1208         nvmet_subsys_put(subsys);
1209         return status;
1210 }
1211
1212 u16 nvmet_check_ctrl_status(struct nvmet_req *req, struct nvme_command *cmd)
1213 {
1214         if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
1215                 pr_err("got cmd %d while CC.EN == 0 on qid = %d\n",
1216                        cmd->common.opcode, req->sq->qid);
1217                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1218         }
1219
1220         if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
1221                 pr_err("got cmd %d while CSTS.RDY == 0 on qid = %d\n",
1222                        cmd->common.opcode, req->sq->qid);
1223                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1224         }
1225         return 0;
1226 }
1227
1228 bool nvmet_host_allowed(struct nvmet_subsys *subsys, const char *hostnqn)
1229 {
1230         struct nvmet_host_link *p;
1231
1232         lockdep_assert_held(&nvmet_config_sem);
1233
1234         if (subsys->allow_any_host)
1235                 return true;
1236
1237         if (subsys->type == NVME_NQN_DISC) /* allow all access to disc subsys */
1238                 return true;
1239
1240         list_for_each_entry(p, &subsys->hosts, entry) {
1241                 if (!strcmp(nvmet_host_name(p->host), hostnqn))
1242                         return true;
1243         }
1244
1245         return false;
1246 }
1247
1248 /*
1249  * Note: ctrl->subsys->lock should be held when calling this function
1250  */
1251 static void nvmet_setup_p2p_ns_map(struct nvmet_ctrl *ctrl,
1252                 struct nvmet_req *req)
1253 {
1254         struct nvmet_ns *ns;
1255         unsigned long idx;
1256
1257         if (!req->p2p_client)
1258                 return;
1259
1260         ctrl->p2p_client = get_device(req->p2p_client);
1261
1262         xa_for_each(&ctrl->subsys->namespaces, idx, ns)
1263                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
1264 }
1265
1266 /*
1267  * Note: ctrl->subsys->lock should be held when calling this function
1268  */
1269 static void nvmet_release_p2p_ns_map(struct nvmet_ctrl *ctrl)
1270 {
1271         struct radix_tree_iter iter;
1272         void __rcu **slot;
1273
1274         radix_tree_for_each_slot(slot, &ctrl->p2p_ns_map, &iter, 0)
1275                 pci_dev_put(radix_tree_deref_slot(slot));
1276
1277         put_device(ctrl->p2p_client);
1278 }
1279
1280 static void nvmet_fatal_error_handler(struct work_struct *work)
1281 {
1282         struct nvmet_ctrl *ctrl =
1283                         container_of(work, struct nvmet_ctrl, fatal_err_work);
1284
1285         pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
1286         ctrl->ops->delete_ctrl(ctrl);
1287 }
1288
1289 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
1290                 struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
1291 {
1292         struct nvmet_subsys *subsys;
1293         struct nvmet_ctrl *ctrl;
1294         int ret;
1295         u16 status;
1296
1297         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1298         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1299         if (!subsys) {
1300                 pr_warn("connect request for invalid subsystem %s!\n",
1301                         subsysnqn);
1302                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1303                 goto out;
1304         }
1305
1306         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1307         down_read(&nvmet_config_sem);
1308         if (!nvmet_host_allowed(subsys, hostnqn)) {
1309                 pr_info("connect by host %s for subsystem %s not allowed\n",
1310                         hostnqn, subsysnqn);
1311                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
1312                 up_read(&nvmet_config_sem);
1313                 status = NVME_SC_CONNECT_INVALID_HOST | NVME_SC_DNR;
1314                 goto out_put_subsystem;
1315         }
1316         up_read(&nvmet_config_sem);
1317
1318         status = NVME_SC_INTERNAL;
1319         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1320         if (!ctrl)
1321                 goto out_put_subsystem;
1322         mutex_init(&ctrl->lock);
1323
1324         nvmet_init_cap(ctrl);
1325
1326         ctrl->port = req->port;
1327
1328         INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
1329         INIT_LIST_HEAD(&ctrl->async_events);
1330         INIT_RADIX_TREE(&ctrl->p2p_ns_map, GFP_KERNEL);
1331         INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
1332
1333         memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
1334         memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
1335
1336         kref_init(&ctrl->ref);
1337         ctrl->subsys = subsys;
1338         WRITE_ONCE(ctrl->aen_enabled, NVMET_AEN_CFG_OPTIONAL);
1339
1340         ctrl->changed_ns_list = kmalloc_array(NVME_MAX_CHANGED_NAMESPACES,
1341                         sizeof(__le32), GFP_KERNEL);
1342         if (!ctrl->changed_ns_list)
1343                 goto out_free_ctrl;
1344
1345         ctrl->sqs = kcalloc(subsys->max_qid + 1,
1346                         sizeof(struct nvmet_sq *),
1347                         GFP_KERNEL);
1348         if (!ctrl->sqs)
1349                 goto out_free_changed_ns_list;
1350
1351         if (subsys->cntlid_min > subsys->cntlid_max)
1352                 goto out_free_changed_ns_list;
1353
1354         ret = ida_simple_get(&cntlid_ida,
1355                              subsys->cntlid_min, subsys->cntlid_max,
1356                              GFP_KERNEL);
1357         if (ret < 0) {
1358                 status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
1359                 goto out_free_sqs;
1360         }
1361         ctrl->cntlid = ret;
1362
1363         ctrl->ops = req->ops;
1364
1365         /*
1366          * Discovery controllers may use some arbitrary high value
1367          * in order to cleanup stale discovery sessions
1368          */
1369         if ((ctrl->subsys->type == NVME_NQN_DISC) && !kato)
1370                 kato = NVMET_DISC_KATO_MS;
1371
1372         /* keep-alive timeout in seconds */
1373         ctrl->kato = DIV_ROUND_UP(kato, 1000);
1374
1375         ctrl->err_counter = 0;
1376         spin_lock_init(&ctrl->error_lock);
1377
1378         nvmet_start_keep_alive_timer(ctrl);
1379
1380         mutex_lock(&subsys->lock);
1381         list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
1382         nvmet_setup_p2p_ns_map(ctrl, req);
1383         mutex_unlock(&subsys->lock);
1384
1385         *ctrlp = ctrl;
1386         return 0;
1387
1388 out_free_sqs:
1389         kfree(ctrl->sqs);
1390 out_free_changed_ns_list:
1391         kfree(ctrl->changed_ns_list);
1392 out_free_ctrl:
1393         kfree(ctrl);
1394 out_put_subsystem:
1395         nvmet_subsys_put(subsys);
1396 out:
1397         return status;
1398 }
1399
1400 static void nvmet_ctrl_free(struct kref *ref)
1401 {
1402         struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
1403         struct nvmet_subsys *subsys = ctrl->subsys;
1404
1405         mutex_lock(&subsys->lock);
1406         nvmet_release_p2p_ns_map(ctrl);
1407         list_del(&ctrl->subsys_entry);
1408         mutex_unlock(&subsys->lock);
1409
1410         nvmet_stop_keep_alive_timer(ctrl);
1411
1412         flush_work(&ctrl->async_event_work);
1413         cancel_work_sync(&ctrl->fatal_err_work);
1414
1415         ida_simple_remove(&cntlid_ida, ctrl->cntlid);
1416
1417         nvmet_async_events_free(ctrl);
1418         kfree(ctrl->sqs);
1419         kfree(ctrl->changed_ns_list);
1420         kfree(ctrl);
1421
1422         nvmet_subsys_put(subsys);
1423 }
1424
1425 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
1426 {
1427         kref_put(&ctrl->ref, nvmet_ctrl_free);
1428 }
1429
1430 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
1431 {
1432         mutex_lock(&ctrl->lock);
1433         if (!(ctrl->csts & NVME_CSTS_CFS)) {
1434                 ctrl->csts |= NVME_CSTS_CFS;
1435                 schedule_work(&ctrl->fatal_err_work);
1436         }
1437         mutex_unlock(&ctrl->lock);
1438 }
1439 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
1440
1441 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
1442                 const char *subsysnqn)
1443 {
1444         struct nvmet_subsys_link *p;
1445
1446         if (!port)
1447                 return NULL;
1448
1449         if (!strcmp(NVME_DISC_SUBSYS_NAME, subsysnqn)) {
1450                 if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
1451                         return NULL;
1452                 return nvmet_disc_subsys;
1453         }
1454
1455         down_read(&nvmet_config_sem);
1456         list_for_each_entry(p, &port->subsystems, entry) {
1457                 if (!strncmp(p->subsys->subsysnqn, subsysnqn,
1458                                 NVMF_NQN_SIZE)) {
1459                         if (!kref_get_unless_zero(&p->subsys->ref))
1460                                 break;
1461                         up_read(&nvmet_config_sem);
1462                         return p->subsys;
1463                 }
1464         }
1465         up_read(&nvmet_config_sem);
1466         return NULL;
1467 }
1468
1469 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
1470                 enum nvme_subsys_type type)
1471 {
1472         struct nvmet_subsys *subsys;
1473
1474         subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
1475         if (!subsys)
1476                 return ERR_PTR(-ENOMEM);
1477
1478         subsys->ver = NVMET_DEFAULT_VS;
1479         /* generate a random serial number as our controllers are ephemeral: */
1480         get_random_bytes(&subsys->serial, sizeof(subsys->serial));
1481
1482         switch (type) {
1483         case NVME_NQN_NVME:
1484                 subsys->max_qid = NVMET_NR_QUEUES;
1485                 break;
1486         case NVME_NQN_DISC:
1487                 subsys->max_qid = 0;
1488                 break;
1489         default:
1490                 pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
1491                 kfree(subsys);
1492                 return ERR_PTR(-EINVAL);
1493         }
1494         subsys->type = type;
1495         subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
1496                         GFP_KERNEL);
1497         if (!subsys->subsysnqn) {
1498                 kfree(subsys);
1499                 return ERR_PTR(-ENOMEM);
1500         }
1501         subsys->cntlid_min = NVME_CNTLID_MIN;
1502         subsys->cntlid_max = NVME_CNTLID_MAX;
1503         kref_init(&subsys->ref);
1504
1505         mutex_init(&subsys->lock);
1506         xa_init(&subsys->namespaces);
1507         INIT_LIST_HEAD(&subsys->ctrls);
1508         INIT_LIST_HEAD(&subsys->hosts);
1509
1510         return subsys;
1511 }
1512
1513 static void nvmet_subsys_free(struct kref *ref)
1514 {
1515         struct nvmet_subsys *subsys =
1516                 container_of(ref, struct nvmet_subsys, ref);
1517
1518         WARN_ON_ONCE(!xa_empty(&subsys->namespaces));
1519
1520         xa_destroy(&subsys->namespaces);
1521         nvmet_passthru_subsys_free(subsys);
1522
1523         kfree(subsys->subsysnqn);
1524         kfree_rcu(subsys->model, rcuhead);
1525         kfree(subsys);
1526 }
1527
1528 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1529 {
1530         struct nvmet_ctrl *ctrl;
1531
1532         mutex_lock(&subsys->lock);
1533         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1534                 ctrl->ops->delete_ctrl(ctrl);
1535         mutex_unlock(&subsys->lock);
1536 }
1537
1538 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1539 {
1540         kref_put(&subsys->ref, nvmet_subsys_free);
1541 }
1542
1543 static int __init nvmet_init(void)
1544 {
1545         int error;
1546
1547         nvmet_ana_group_enabled[NVMET_DEFAULT_ANA_GRPID] = 1;
1548
1549         buffered_io_wq = alloc_workqueue("nvmet-buffered-io-wq",
1550                         WQ_MEM_RECLAIM, 0);
1551         if (!buffered_io_wq) {
1552                 error = -ENOMEM;
1553                 goto out;
1554         }
1555
1556         error = nvmet_init_discovery();
1557         if (error)
1558                 goto out_free_work_queue;
1559
1560         error = nvmet_init_configfs();
1561         if (error)
1562                 goto out_exit_discovery;
1563         return 0;
1564
1565 out_exit_discovery:
1566         nvmet_exit_discovery();
1567 out_free_work_queue:
1568         destroy_workqueue(buffered_io_wq);
1569 out:
1570         return error;
1571 }
1572
1573 static void __exit nvmet_exit(void)
1574 {
1575         nvmet_exit_configfs();
1576         nvmet_exit_discovery();
1577         ida_destroy(&cntlid_ida);
1578         destroy_workqueue(buffered_io_wq);
1579
1580         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1581         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1582 }
1583
1584 module_init(nvmet_init);
1585 module_exit(nvmet_exit);
1586
1587 MODULE_LICENSE("GPL v2");