drivers/nvme/target/core.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Common code for the NVMe target.
   4  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
   5  */
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7 #include <linux/module.h>
   8 #include <linux/random.h>
   9 #include <linux/rculist.h>
  10 #include <linux/pci-p2pdma.h>
  11 #include <linux/scatterlist.h>
  12
  13 #define CREATE_TRACE_POINTS
  14 #include "trace.h"
  15
  16 #include "nvmet.h"
  17
  18 struct workqueue_struct *buffered_io_wq;
  19 static const struct nvmet_fabrics_ops *nvmet_transports[NVMF_TRTYPE_MAX];
  20 static DEFINE_IDA(cntlid_ida);
  21
  22 /*
  23  * This read/write semaphore is used to synchronize access to configuration
  24  * information on a target system that will result in discovery log page
  25  * information change for at least one host.
  26  * The full list of resources to protected by this semaphore is:
  27  *
  28  *  - subsystems list
  29  *  - per-subsystem allowed hosts list
  30  *  - allow_any_host subsystem attribute
  31  *  - nvmet_genctr
  32  *  - the nvmet_transports array
  33  *
  34  * When updating any of those lists/structures write lock should be obtained,
  35  * while when reading (popolating discovery log page or checking host-subsystem
  36  * link) read lock is obtained to allow concurrent reads.
  37  */
  38 DECLARE_RWSEM(nvmet_config_sem);
  39
  40 u32 nvmet_ana_group_enabled[NVMET_MAX_ANAGRPS + 1];
  41 u64 nvmet_ana_chgcnt;
  42 DECLARE_RWSEM(nvmet_ana_sem);
  43
  44 inline u16 errno_to_nvme_status(struct nvmet_req *req, int errno)
  45 {
  46         u16 status;
  47
  48         switch (errno) {
  49         case 0:
  50                 status = NVME_SC_SUCCESS;
  51                 break;
  52         case -ENOSPC:
  53                 req->error_loc = offsetof(struct nvme_rw_command, length);
  54                 status = NVME_SC_CAP_EXCEEDED | NVME_SC_DNR;
  55                 break;
  56         case -EREMOTEIO:
  57                 req->error_loc = offsetof(struct nvme_rw_command, slba);
  58                 status = NVME_SC_LBA_RANGE | NVME_SC_DNR;
  59                 break;
  60         case -EOPNOTSUPP:
  61                 req->error_loc = offsetof(struct nvme_common_command, opcode);
  62                 switch (req->cmd->common.opcode) {
  63                 case nvme_cmd_dsm:
  64                 case nvme_cmd_write_zeroes:
  65                         status = NVME_SC_ONCS_NOT_SUPPORTED | NVME_SC_DNR;
  66                         break;
  67                 default:
  68                         status = NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
  69                 }
  70                 break;
  71         case -ENODATA:
  72                 req->error_loc = offsetof(struct nvme_rw_command, nsid);
  73                 status = NVME_SC_ACCESS_DENIED;
  74                 break;
  75         case -EIO:
  76                 /* FALLTHRU */
  77         default:
  78                 req->error_loc = offsetof(struct nvme_common_command, opcode);
  79                 status = NVME_SC_INTERNAL | NVME_SC_DNR;
  80         }
  81
  82         return status;
  83 }
  84
  85 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
  86                 const char *subsysnqn);
  87
  88 u16 nvmet_copy_to_sgl(struct nvmet_req *req, off_t off, const void *buf,
  89                 size_t len)
  90 {
  91         if (sg_pcopy_from_buffer(req->sg, req->sg_cnt, buf, len, off) != len) {
  92                 req->error_loc = offsetof(struct nvme_common_command, dptr);
  93                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
  94         }
  95         return 0;
  96 }
  97
  98 u16 nvmet_copy_from_sgl(struct nvmet_req *req, off_t off, void *buf, size_t len)
  99 {
 100         if (sg_pcopy_to_buffer(req->sg, req->sg_cnt, buf, len, off) != len) {
 101                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 102                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
 103         }
 104         return 0;
 105 }
 106
 107 u16 nvmet_zero_sgl(struct nvmet_req *req, off_t off, size_t len)
 108 {
 109         if (sg_zero_buffer(req->sg, req->sg_cnt, len, off) != len) {
 110                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 111                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
 112         }
 113         return 0;
 114 }
 115
 116 static unsigned int nvmet_max_nsid(struct nvmet_subsys *subsys)
 117 {
 118         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 static void nvmet_start_keep_alive_timer(struct nvmet_ctrl *ctrl)
 399 {
 400         pr_debug("ctrl %d start keep-alive timer for %d secs\n",
 401                 ctrl->cntlid, ctrl->kato);
 402
 403         INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer);
 404         schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
 405 }
 406
 407 static void nvmet_stop_keep_alive_timer(struct nvmet_ctrl *ctrl)
 408 {
 409         pr_debug("ctrl %d stop keep-alive\n", ctrl->cntlid);
 410
 411         cancel_delayed_work_sync(&ctrl->ka_work);
 412 }
 413
 414 struct nvmet_ns *nvmet_find_namespace(struct nvmet_ctrl *ctrl, __le32 nsid)
 415 {
 416         struct nvmet_ns *ns;
 417
 418         ns = xa_load(&ctrl->subsys->namespaces, le32_to_cpu(nsid));
 419         if (ns)
 420                 percpu_ref_get(&ns->ref);
 421
 422         return ns;
 423 }
 424
 425 static void nvmet_destroy_namespace(struct percpu_ref *ref)
 426 {
 427         struct nvmet_ns *ns = container_of(ref, struct nvmet_ns, ref);
 428
 429         complete(&ns->disable_done);
 430 }
 431
 432 void nvmet_put_namespace(struct nvmet_ns *ns)
 433 {
 434         percpu_ref_put(&ns->ref);
 435 }
 436
 437 static void nvmet_ns_dev_disable(struct nvmet_ns *ns)
 438 {
 439         nvmet_bdev_ns_disable(ns);
 440         nvmet_file_ns_disable(ns);
 441 }
 442
 443 static int nvmet_p2pmem_ns_enable(struct nvmet_ns *ns)
 444 {
 445         int ret;
 446         struct pci_dev *p2p_dev;
 447
 448         if (!ns->use_p2pmem)
 449                 return 0;
 450
 451         if (!ns->bdev) {
 452                 pr_err("peer-to-peer DMA is not supported by non-block device namespaces\n");
 453                 return -EINVAL;
 454         }
 455
 456         if (!blk_queue_pci_p2pdma(ns->bdev->bd_disk->queue)) {
 457                 pr_err("peer-to-peer DMA is not supported by the driver of %s\n",
 458                        ns->device_path);
 459                 return -EINVAL;
 460         }
 461
 462         if (ns->p2p_dev) {
 463                 ret = pci_p2pdma_distance(ns->p2p_dev, nvmet_ns_dev(ns), true);
 464                 if (ret < 0)
 465                         return -EINVAL;
 466         } else {
 467                 /*
 468                  * Right now we just check that there is p2pmem available so
 469                  * we can report an error to the user right away if there
 470                  * is not. We'll find the actual device to use once we
 471                  * setup the controller when the port's device is available.
 472                  */
 473
 474                 p2p_dev = pci_p2pmem_find(nvmet_ns_dev(ns));
 475                 if (!p2p_dev) {
 476                         pr_err("no peer-to-peer memory is available for %s\n",
 477                                ns->device_path);
 478                         return -EINVAL;
 479                 }
 480
 481                 pci_dev_put(p2p_dev);
 482         }
 483
 484         return 0;
 485 }
 486
 487 /*
 488  * Note: ctrl->subsys->lock should be held when calling this function
 489  */
 490 static void nvmet_p2pmem_ns_add_p2p(struct nvmet_ctrl *ctrl,
 491                                     struct nvmet_ns *ns)
 492 {
 493         struct device *clients[2];
 494         struct pci_dev *p2p_dev;
 495         int ret;
 496
 497         if (!ctrl->p2p_client || !ns->use_p2pmem)
 498                 return;
 499
 500         if (ns->p2p_dev) {
 501                 ret = pci_p2pdma_distance(ns->p2p_dev, ctrl->p2p_client, true);
 502                 if (ret < 0)
 503                         return;
 504
 505                 p2p_dev = pci_dev_get(ns->p2p_dev);
 506         } else {
 507                 clients[0] = ctrl->p2p_client;
 508                 clients[1] = nvmet_ns_dev(ns);
 509
 510                 p2p_dev = pci_p2pmem_find_many(clients, ARRAY_SIZE(clients));
 511                 if (!p2p_dev) {
 512                         pr_err("no peer-to-peer memory is available that's supported by %s and %s\n",
 513                                dev_name(ctrl->p2p_client), ns->device_path);
 514                         return;
 515                 }
 516         }
 517
 518         ret = radix_tree_insert(&ctrl->p2p_ns_map, ns->nsid, p2p_dev);
 519         if (ret < 0)
 520                 pci_dev_put(p2p_dev);
 521
 522         pr_info("using p2pmem on %s for nsid %d\n", pci_name(p2p_dev),
 523                 ns->nsid);
 524 }
 525
 526 void nvmet_ns_revalidate(struct nvmet_ns *ns)
 527 {
 528         loff_t oldsize = ns->size;
 529
 530         if (ns->bdev)
 531                 nvmet_bdev_ns_revalidate(ns);
 532         else
 533                 nvmet_file_ns_revalidate(ns);
 534
 535         if (oldsize != ns->size)
 536                 nvmet_ns_changed(ns->subsys, ns->nsid);
 537 }
 538
 539 int nvmet_ns_enable(struct nvmet_ns *ns)
 540 {
 541         struct nvmet_subsys *subsys = ns->subsys;
 542         struct nvmet_ctrl *ctrl;
 543         int ret;
 544
 545         mutex_lock(&subsys->lock);
 546         ret = 0;
 547
 548         if (nvmet_passthru_ctrl(subsys)) {
 549                 pr_info("cannot enable both passthru and regular namespaces for a single subsystem");
 550                 goto out_unlock;
 551         }
 552
 553         if (ns->enabled)
 554                 goto out_unlock;
 555
 556         ret = -EMFILE;
 557         if (subsys->nr_namespaces == NVMET_MAX_NAMESPACES)
 558                 goto out_unlock;
 559
 560         ret = nvmet_bdev_ns_enable(ns);
 561         if (ret == -ENOTBLK)
 562                 ret = nvmet_file_ns_enable(ns);
 563         if (ret)
 564                 goto out_unlock;
 565
 566         ret = nvmet_p2pmem_ns_enable(ns);
 567         if (ret)
 568                 goto out_dev_disable;
 569
 570         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 571                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
 572
 573         ret = percpu_ref_init(&ns->ref, nvmet_destroy_namespace,
 574                                 0, GFP_KERNEL);
 575         if (ret)
 576                 goto out_dev_put;
 577
 578         if (ns->nsid > subsys->max_nsid)
 579                 subsys->max_nsid = ns->nsid;
 580
 581         ret = xa_insert(&subsys->namespaces, ns->nsid, ns, GFP_KERNEL);
 582         if (ret)
 583                 goto out_restore_subsys_maxnsid;
 584
 585         subsys->nr_namespaces++;
 586
 587         nvmet_ns_changed(subsys, ns->nsid);
 588         ns->enabled = true;
 589         ret = 0;
 590 out_unlock:
 591         mutex_unlock(&subsys->lock);
 592         return ret;
 593
 594 out_restore_subsys_maxnsid:
 595         subsys->max_nsid = nvmet_max_nsid(subsys);
 596         percpu_ref_exit(&ns->ref);
 597 out_dev_put:
 598         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 599                 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
 600 out_dev_disable:
 601         nvmet_ns_dev_disable(ns);
 602         goto out_unlock;
 603 }
 604
 605 void nvmet_ns_disable(struct nvmet_ns *ns)
 606 {
 607         struct nvmet_subsys *subsys = ns->subsys;
 608         struct nvmet_ctrl *ctrl;
 609
 610         mutex_lock(&subsys->lock);
 611         if (!ns->enabled)
 612                 goto out_unlock;
 613
 614         ns->enabled = false;
 615         xa_erase(&ns->subsys->namespaces, ns->nsid);
 616         if (ns->nsid == subsys->max_nsid)
 617                 subsys->max_nsid = nvmet_max_nsid(subsys);
 618
 619         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 620                 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
 621
 622         mutex_unlock(&subsys->lock);
 623
 624         /*
 625          * Now that we removed the namespaces from the lookup list, we
 626          * can kill the per_cpu ref and wait for any remaining references
 627          * to be dropped, as well as a RCU grace period for anyone only
 628          * using the namepace under rcu_read_lock().  Note that we can't
 629          * use call_rcu here as we need to ensure the namespaces have
 630          * been fully destroyed before unloading the module.
 631          */
 632         percpu_ref_kill(&ns->ref);
 633         synchronize_rcu();
 634         wait_for_completion(&ns->disable_done);
 635         percpu_ref_exit(&ns->ref);
 636
 637         mutex_lock(&subsys->lock);
 638
 639         subsys->nr_namespaces--;
 640         nvmet_ns_changed(subsys, ns->nsid);
 641         nvmet_ns_dev_disable(ns);
 642 out_unlock:
 643         mutex_unlock(&subsys->lock);
 644 }
 645
 646 void nvmet_ns_free(struct nvmet_ns *ns)
 647 {
 648         nvmet_ns_disable(ns);
 649
 650         down_write(&nvmet_ana_sem);
 651         nvmet_ana_group_enabled[ns->anagrpid]--;
 652         up_write(&nvmet_ana_sem);
 653
 654         kfree(ns->device_path);
 655         kfree(ns);
 656 }
 657
 658 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
 659 {
 660         struct nvmet_ns *ns;
 661
 662         ns = kzalloc(sizeof(*ns), GFP_KERNEL);
 663         if (!ns)
 664                 return NULL;
 665
 666         init_completion(&ns->disable_done);
 667
 668         ns->nsid = nsid;
 669         ns->subsys = subsys;
 670
 671         down_write(&nvmet_ana_sem);
 672         ns->anagrpid = NVMET_DEFAULT_ANA_GRPID;
 673         nvmet_ana_group_enabled[ns->anagrpid]++;
 674         up_write(&nvmet_ana_sem);
 675
 676         uuid_gen(&ns->uuid);
 677         ns->buffered_io = false;
 678
 679         return ns;
 680 }
 681
 682 static void nvmet_update_sq_head(struct nvmet_req *req)
 683 {
 684         if (req->sq->size) {
 685                 u32 old_sqhd, new_sqhd;
 686
 687                 do {
 688                         old_sqhd = req->sq->sqhd;
 689                         new_sqhd = (old_sqhd + 1) % req->sq->size;
 690                 } while (cmpxchg(&req->sq->sqhd, old_sqhd, new_sqhd) !=
 691                                         old_sqhd);
 692         }
 693         req->cqe->sq_head = cpu_to_le16(req->sq->sqhd & 0x0000FFFF);
 694 }
 695
 696 static void nvmet_set_error(struct nvmet_req *req, u16 status)
 697 {
 698         struct nvmet_ctrl *ctrl = req->sq->ctrl;
 699         struct nvme_error_slot *new_error_slot;
 700         unsigned long flags;
 701
 702         req->cqe->status = cpu_to_le16(status << 1);
 703
 704         if (!ctrl || req->error_loc == NVMET_NO_ERROR_LOC)
 705                 return;
 706
 707         spin_lock_irqsave(&ctrl->error_lock, flags);
 708         ctrl->err_counter++;
 709         new_error_slot =
 710                 &ctrl->slots[ctrl->err_counter % NVMET_ERROR_LOG_SLOTS];
 711
 712         new_error_slot->error_count = cpu_to_le64(ctrl->err_counter);
 713         new_error_slot->sqid = cpu_to_le16(req->sq->qid);
 714         new_error_slot->cmdid = cpu_to_le16(req->cmd->common.command_id);
 715         new_error_slot->status_field = cpu_to_le16(status << 1);
 716         new_error_slot->param_error_location = cpu_to_le16(req->error_loc);
 717         new_error_slot->lba = cpu_to_le64(req->error_slba);
 718         new_error_slot->nsid = req->cmd->common.nsid;
 719         spin_unlock_irqrestore(&ctrl->error_lock, flags);
 720
 721         /* set the more bit for this request */
 722         req->cqe->status |= cpu_to_le16(1 << 14);
 723 }
 724
 725 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
 726 {
 727         if (!req->sq->sqhd_disabled)
 728                 nvmet_update_sq_head(req);
 729         req->cqe->sq_id = cpu_to_le16(req->sq->qid);
 730         req->cqe->command_id = req->cmd->common.command_id;
 731
 732         if (unlikely(status))
 733                 nvmet_set_error(req, status);
 734
 735         trace_nvmet_req_complete(req);
 736
 737         if (req->ns)
 738                 nvmet_put_namespace(req->ns);
 739         req->ops->queue_response(req);
 740 }
 741
 742 void nvmet_req_complete(struct nvmet_req *req, u16 status)
 743 {
 744         __nvmet_req_complete(req, status);
 745         percpu_ref_put(&req->sq->ref);
 746 }
 747 EXPORT_SYMBOL_GPL(nvmet_req_complete);
 748
 749 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
 750                 u16 qid, u16 size)
 751 {
 752         cq->qid = qid;
 753         cq->size = size;
 754
 755         ctrl->cqs[qid] = cq;
 756 }
 757
 758 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
 759                 u16 qid, u16 size)
 760 {
 761         sq->sqhd = 0;
 762         sq->qid = qid;
 763         sq->size = size;
 764
 765         ctrl->sqs[qid] = sq;
 766 }
 767
 768 static void nvmet_confirm_sq(struct percpu_ref *ref)
 769 {
 770         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 771
 772         complete(&sq->confirm_done);
 773 }
 774
 775 void nvmet_sq_destroy(struct nvmet_sq *sq)
 776 {
 777         struct nvmet_ctrl *ctrl = sq->ctrl;
 778
 779         /*
 780          * If this is the admin queue, complete all AERs so that our
 781          * queue doesn't have outstanding requests on it.
 782          */
 783         if (ctrl && ctrl->sqs && ctrl->sqs[0] == sq)
 784                 nvmet_async_events_failall(ctrl);
 785         percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
 786         wait_for_completion(&sq->confirm_done);
 787         wait_for_completion(&sq->free_done);
 788         percpu_ref_exit(&sq->ref);
 789
 790         if (ctrl) {
 791                 nvmet_ctrl_put(ctrl);
 792                 sq->ctrl = NULL; /* allows reusing the queue later */
 793         }
 794 }
 795 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
 796
 797 static void nvmet_sq_free(struct percpu_ref *ref)
 798 {
 799         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 800
 801         complete(&sq->free_done);
 802 }
 803
 804 int nvmet_sq_init(struct nvmet_sq *sq)
 805 {
 806         int ret;
 807
 808         ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
 809         if (ret) {
 810                 pr_err("percpu_ref init failed!\n");
 811                 return ret;
 812         }
 813         init_completion(&sq->free_done);
 814         init_completion(&sq->confirm_done);
 815
 816         return 0;
 817 }
 818 EXPORT_SYMBOL_GPL(nvmet_sq_init);
 819
 820 static inline u16 nvmet_check_ana_state(struct nvmet_port *port,
 821                 struct nvmet_ns *ns)
 822 {
 823         enum nvme_ana_state state = port->ana_state[ns->anagrpid];
 824
 825         if (unlikely(state == NVME_ANA_INACCESSIBLE))
 826                 return NVME_SC_ANA_INACCESSIBLE;
 827         if (unlikely(state == NVME_ANA_PERSISTENT_LOSS))
 828                 return NVME_SC_ANA_PERSISTENT_LOSS;
 829         if (unlikely(state == NVME_ANA_CHANGE))
 830                 return NVME_SC_ANA_TRANSITION;
 831         return 0;
 832 }
 833
 834 static inline u16 nvmet_io_cmd_check_access(struct nvmet_req *req)
 835 {
 836         if (unlikely(req->ns->readonly)) {
 837                 switch (req->cmd->common.opcode) {
 838                 case nvme_cmd_read:
 839                 case nvme_cmd_flush:
 840                         break;
 841                 default:
 842                         return NVME_SC_NS_WRITE_PROTECTED;
 843                 }
 844         }
 845
 846         return 0;
 847 }
 848
 849 static u16 nvmet_parse_io_cmd(struct nvmet_req *req)
 850 {
 851         struct nvme_command *cmd = req->cmd;
 852         u16 ret;
 853
 854         ret = nvmet_check_ctrl_status(req, cmd);
 855         if (unlikely(ret))
 856                 return ret;
 857
 858         if (nvmet_req_passthru_ctrl(req))
 859                 return nvmet_parse_passthru_io_cmd(req);
 860
 861         req->ns = nvmet_find_namespace(req->sq->ctrl, cmd->rw.nsid);
 862         if (unlikely(!req->ns)) {
 863                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 864                 return NVME_SC_INVALID_NS | NVME_SC_DNR;
 865         }
 866         ret = nvmet_check_ana_state(req->port, req->ns);
 867         if (unlikely(ret)) {
 868                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 869                 return ret;
 870         }
 871         ret = nvmet_io_cmd_check_access(req);
 872         if (unlikely(ret)) {
 873                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 874                 return ret;
 875         }
 876
 877         if (req->ns->file)
 878                 return nvmet_file_parse_io_cmd(req);
 879         else
 880                 return nvmet_bdev_parse_io_cmd(req);
 881 }
 882
 883 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
 884                 struct nvmet_sq *sq, const struct nvmet_fabrics_ops *ops)
 885 {
 886         u8 flags = req->cmd->common.flags;
 887         u16 status;
 888
 889         req->cq = cq;
 890         req->sq = sq;
 891         req->ops = ops;
 892         req->sg = NULL;
 893         req->metadata_sg = NULL;
 894         req->sg_cnt = 0;
 895         req->metadata_sg_cnt = 0;
 896         req->transfer_len = 0;
 897         req->metadata_len = 0;
 898         req->cqe->status = 0;
 899         req->cqe->sq_head = 0;
 900         req->ns = NULL;
 901         req->error_loc = NVMET_NO_ERROR_LOC;
 902         req->error_slba = 0;
 903
 904         trace_nvmet_req_init(req, req->cmd);
 905
 906         /* no support for fused commands yet */
 907         if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
 908                 req->error_loc = offsetof(struct nvme_common_command, flags);
 909                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 910                 goto fail;
 911         }
 912
 913         /*
 914          * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
 915          * contains an address of a single contiguous physical buffer that is
 916          * byte aligned.
 917          */
 918         if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
 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         if (unlikely(!req->sq->ctrl))
 925                 /* will return an error for any non-connect command: */
 926                 status = nvmet_parse_connect_cmd(req);
 927         else if (likely(req->sq->qid != 0))
 928                 status = nvmet_parse_io_cmd(req);
 929         else
 930                 status = nvmet_parse_admin_cmd(req);
 931
 932         if (status)
 933                 goto fail;
 934
 935         if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
 936                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 937                 goto fail;
 938         }
 939
 940         if (sq->ctrl)
 941                 sq->ctrl->cmd_seen = true;
 942
 943         return true;
 944
 945 fail:
 946         __nvmet_req_complete(req, status);
 947         return false;
 948 }
 949 EXPORT_SYMBOL_GPL(nvmet_req_init);
 950
 951 void nvmet_req_uninit(struct nvmet_req *req)
 952 {
 953         percpu_ref_put(&req->sq->ref);
 954         if (req->ns)
 955                 nvmet_put_namespace(req->ns);
 956 }
 957 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
 958
 959 bool nvmet_check_transfer_len(struct nvmet_req *req, size_t len)
 960 {
 961         if (unlikely(len != req->transfer_len)) {
 962                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 963                 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
 964                 return false;
 965         }
 966
 967         return true;
 968 }
 969 EXPORT_SYMBOL_GPL(nvmet_check_transfer_len);
 970
 971 bool nvmet_check_data_len_lte(struct nvmet_req *req, size_t data_len)
 972 {
 973         if (unlikely(data_len > req->transfer_len)) {
 974                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 975                 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
 976                 return false;
 977         }
 978
 979         return true;
 980 }
 981
 982 static unsigned int nvmet_data_transfer_len(struct nvmet_req *req)
 983 {
 984         return req->transfer_len - req->metadata_len;
 985 }
 986
 987 static int nvmet_req_alloc_p2pmem_sgls(struct nvmet_req *req)
 988 {
 989         req->sg = pci_p2pmem_alloc_sgl(req->p2p_dev, &req->sg_cnt,
 990                         nvmet_data_transfer_len(req));
 991         if (!req->sg)
 992                 goto out_err;
 993
 994         if (req->metadata_len) {
 995                 req->metadata_sg = pci_p2pmem_alloc_sgl(req->p2p_dev,
 996                                 &req->metadata_sg_cnt, req->metadata_len);
 997                 if (!req->metadata_sg)
 998                         goto out_free_sg;
 999         }
1000         return 0;
1001 out_free_sg:
1002         pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1003 out_err:
1004         return -ENOMEM;
1005 }
1006
1007 static bool nvmet_req_find_p2p_dev(struct nvmet_req *req)
1008 {
1009         if (!IS_ENABLED(CONFIG_PCI_P2PDMA))
1010                 return false;
1011
1012         if (req->sq->ctrl && req->sq->qid && req->ns) {
1013                 req->p2p_dev = radix_tree_lookup(&req->sq->ctrl->p2p_ns_map,
1014                                                  req->ns->nsid);
1015                 if (req->p2p_dev)
1016                         return true;
1017         }
1018
1019         req->p2p_dev = NULL;
1020         return false;
1021 }
1022
1023 int nvmet_req_alloc_sgls(struct nvmet_req *req)
1024 {
1025         if (nvmet_req_find_p2p_dev(req) && !nvmet_req_alloc_p2pmem_sgls(req))
1026                 return 0;
1027
1028         req->sg = sgl_alloc(nvmet_data_transfer_len(req), GFP_KERNEL,
1029                             &req->sg_cnt);
1030         if (unlikely(!req->sg))
1031                 goto out;
1032
1033         if (req->metadata_len) {
1034                 req->metadata_sg = sgl_alloc(req->metadata_len, GFP_KERNEL,
1035                                              &req->metadata_sg_cnt);
1036                 if (unlikely(!req->metadata_sg))
1037                         goto out_free;
1038         }
1039
1040         return 0;
1041 out_free:
1042         sgl_free(req->sg);
1043 out:
1044         return -ENOMEM;
1045 }
1046 EXPORT_SYMBOL_GPL(nvmet_req_alloc_sgls);
1047
1048 void nvmet_req_free_sgls(struct nvmet_req *req)
1049 {
1050         if (req->p2p_dev) {
1051                 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1052                 if (req->metadata_sg)
1053                         pci_p2pmem_free_sgl(req->p2p_dev, req->metadata_sg);
1054         } else {
1055                 sgl_free(req->sg);
1056                 if (req->metadata_sg)
1057                         sgl_free(req->metadata_sg);
1058         }
1059
1060         req->sg = NULL;
1061         req->metadata_sg = NULL;
1062         req->sg_cnt = 0;
1063         req->metadata_sg_cnt = 0;
1064 }
1065 EXPORT_SYMBOL_GPL(nvmet_req_free_sgls);
1066
1067 static inline bool nvmet_cc_en(u32 cc)
1068 {
1069         return (cc >> NVME_CC_EN_SHIFT) & 0x1;
1070 }
1071
1072 static inline u8 nvmet_cc_css(u32 cc)
1073 {
1074         return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
1075 }
1076
1077 static inline u8 nvmet_cc_mps(u32 cc)
1078 {
1079         return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
1080 }
1081
1082 static inline u8 nvmet_cc_ams(u32 cc)
1083 {
1084         return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
1085 }
1086
1087 static inline u8 nvmet_cc_shn(u32 cc)
1088 {
1089         return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
1090 }
1091
1092 static inline u8 nvmet_cc_iosqes(u32 cc)
1093 {
1094         return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
1095 }
1096
1097 static inline u8 nvmet_cc_iocqes(u32 cc)
1098 {
1099         return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
1100 }
1101
1102 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
1103 {
1104         lockdep_assert_held(&ctrl->lock);
1105
1106         if (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
1107             nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES ||
1108             nvmet_cc_mps(ctrl->cc) != 0 ||
1109             nvmet_cc_ams(ctrl->cc) != 0 ||
1110             nvmet_cc_css(ctrl->cc) != 0) {
1111                 ctrl->csts = NVME_CSTS_CFS;
1112                 return;
1113         }
1114
1115         ctrl->csts = NVME_CSTS_RDY;
1116
1117         /*
1118          * Controllers that are not yet enabled should not really enforce the
1119          * keep alive timeout, but we still want to track a timeout and cleanup
1120          * in case a host died before it enabled the controller.  Hence, simply
1121          * reset the keep alive timer when the controller is enabled.
1122          */
1123         mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ);
1124 }
1125
1126 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
1127 {
1128         lockdep_assert_held(&ctrl->lock);
1129
1130         /* XXX: tear down queues? */
1131         ctrl->csts &= ~NVME_CSTS_RDY;
1132         ctrl->cc = 0;
1133 }
1134
1135 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
1136 {
1137         u32 old;
1138
1139         mutex_lock(&ctrl->lock);
1140         old = ctrl->cc;
1141         ctrl->cc = new;
1142
1143         if (nvmet_cc_en(new) && !nvmet_cc_en(old))
1144                 nvmet_start_ctrl(ctrl);
1145         if (!nvmet_cc_en(new) && nvmet_cc_en(old))
1146                 nvmet_clear_ctrl(ctrl);
1147         if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
1148                 nvmet_clear_ctrl(ctrl);
1149                 ctrl->csts |= NVME_CSTS_SHST_CMPLT;
1150         }
1151         if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
1152                 ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
1153         mutex_unlock(&ctrl->lock);
1154 }
1155
1156 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
1157 {
1158         /* command sets supported: NVMe command set: */
1159         ctrl->cap = (1ULL << 37);
1160         /* CC.EN timeout in 500msec units: */
1161         ctrl->cap |= (15ULL << 24);
1162         /* maximum queue entries supported: */
1163         ctrl->cap |= NVMET_QUEUE_SIZE - 1;
1164 }
1165
1166 u16 nvmet_ctrl_find_get(const char *subsysnqn, const char *hostnqn, u16 cntlid,
1167                 struct nvmet_req *req, struct nvmet_ctrl **ret)
1168 {
1169         struct nvmet_subsys *subsys;
1170         struct nvmet_ctrl *ctrl;
1171         u16 status = 0;
1172
1173         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1174         if (!subsys) {
1175                 pr_warn("connect request for invalid subsystem %s!\n",
1176                         subsysnqn);
1177                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1178                 return NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1179         }
1180
1181         mutex_lock(&subsys->lock);
1182         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
1183                 if (ctrl->cntlid == cntlid) {
1184                         if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
1185                                 pr_warn("hostnqn mismatch.\n");
1186                                 continue;
1187                         }
1188                         if (!kref_get_unless_zero(&ctrl->ref))
1189                                 continue;
1190
1191                         *ret = ctrl;
1192                         goto out;
1193                 }
1194         }
1195
1196         pr_warn("could not find controller %d for subsys %s / host %s\n",
1197                 cntlid, subsysnqn, hostnqn);
1198         req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
1199         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1200
1201 out:
1202         mutex_unlock(&subsys->lock);
1203         nvmet_subsys_put(subsys);
1204         return status;
1205 }
1206
1207 u16 nvmet_check_ctrl_status(struct nvmet_req *req, struct nvme_command *cmd)
1208 {
1209         if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
1210                 pr_err("got cmd %d while CC.EN == 0 on qid = %d\n",
1211                        cmd->common.opcode, req->sq->qid);
1212                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1213         }
1214
1215         if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
1216                 pr_err("got cmd %d while CSTS.RDY == 0 on qid = %d\n",
1217                        cmd->common.opcode, req->sq->qid);
1218                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1219         }
1220         return 0;
1221 }
1222
1223 bool nvmet_host_allowed(struct nvmet_subsys *subsys, const char *hostnqn)
1224 {
1225         struct nvmet_host_link *p;
1226
1227         lockdep_assert_held(&nvmet_config_sem);
1228
1229         if (subsys->allow_any_host)
1230                 return true;
1231
1232         if (subsys->type == NVME_NQN_DISC) /* allow all access to disc subsys */
1233                 return true;
1234
1235         list_for_each_entry(p, &subsys->hosts, entry) {
1236                 if (!strcmp(nvmet_host_name(p->host), hostnqn))
1237                         return true;
1238         }
1239
1240         return false;
1241 }
1242
1243 /*
1244  * Note: ctrl->subsys->lock should be held when calling this function
1245  */
1246 static void nvmet_setup_p2p_ns_map(struct nvmet_ctrl *ctrl,
1247                 struct nvmet_req *req)
1248 {
1249         struct nvmet_ns *ns;
1250         unsigned long idx;
1251
1252         if (!req->p2p_client)
1253                 return;
1254
1255         ctrl->p2p_client = get_device(req->p2p_client);
1256
1257         xa_for_each(&ctrl->subsys->namespaces, idx, ns)
1258                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
1259 }
1260
1261 /*
1262  * Note: ctrl->subsys->lock should be held when calling this function
1263  */
1264 static void nvmet_release_p2p_ns_map(struct nvmet_ctrl *ctrl)
1265 {
1266         struct radix_tree_iter iter;
1267         void __rcu **slot;
1268
1269         radix_tree_for_each_slot(slot, &ctrl->p2p_ns_map, &iter, 0)
1270                 pci_dev_put(radix_tree_deref_slot(slot));
1271
1272         put_device(ctrl->p2p_client);
1273 }
1274
1275 static void nvmet_fatal_error_handler(struct work_struct *work)
1276 {
1277         struct nvmet_ctrl *ctrl =
1278                         container_of(work, struct nvmet_ctrl, fatal_err_work);
1279
1280         pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
1281         ctrl->ops->delete_ctrl(ctrl);
1282 }
1283
1284 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
1285                 struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
1286 {
1287         struct nvmet_subsys *subsys;
1288         struct nvmet_ctrl *ctrl;
1289         int ret;
1290         u16 status;
1291
1292         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1293         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1294         if (!subsys) {
1295                 pr_warn("connect request for invalid subsystem %s!\n",
1296                         subsysnqn);
1297                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1298                 goto out;
1299         }
1300
1301         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1302         down_read(&nvmet_config_sem);
1303         if (!nvmet_host_allowed(subsys, hostnqn)) {
1304                 pr_info("connect by host %s for subsystem %s not allowed\n",
1305                         hostnqn, subsysnqn);
1306                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
1307                 up_read(&nvmet_config_sem);
1308                 status = NVME_SC_CONNECT_INVALID_HOST | NVME_SC_DNR;
1309                 goto out_put_subsystem;
1310         }
1311         up_read(&nvmet_config_sem);
1312
1313         status = NVME_SC_INTERNAL;
1314         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1315         if (!ctrl)
1316                 goto out_put_subsystem;
1317         mutex_init(&ctrl->lock);
1318
1319         nvmet_init_cap(ctrl);
1320
1321         ctrl->port = req->port;
1322
1323         INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
1324         INIT_LIST_HEAD(&ctrl->async_events);
1325         INIT_RADIX_TREE(&ctrl->p2p_ns_map, GFP_KERNEL);
1326         INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
1327
1328         memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
1329         memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
1330
1331         kref_init(&ctrl->ref);
1332         ctrl->subsys = subsys;
1333         WRITE_ONCE(ctrl->aen_enabled, NVMET_AEN_CFG_OPTIONAL);
1334
1335         ctrl->changed_ns_list = kmalloc_array(NVME_MAX_CHANGED_NAMESPACES,
1336                         sizeof(__le32), GFP_KERNEL);
1337         if (!ctrl->changed_ns_list)
1338                 goto out_free_ctrl;
1339
1340         ctrl->cqs = kcalloc(subsys->max_qid + 1,
1341                         sizeof(struct nvmet_cq *),
1342                         GFP_KERNEL);
1343         if (!ctrl->cqs)
1344                 goto out_free_changed_ns_list;
1345
1346         ctrl->sqs = kcalloc(subsys->max_qid + 1,
1347                         sizeof(struct nvmet_sq *),
1348                         GFP_KERNEL);
1349         if (!ctrl->sqs)
1350                 goto out_free_cqs;
1351
1352         if (subsys->cntlid_min > subsys->cntlid_max)
1353                 goto out_free_cqs;
1354
1355         ret = ida_simple_get(&cntlid_ida,
1356                              subsys->cntlid_min, subsys->cntlid_max,
1357                              GFP_KERNEL);
1358         if (ret < 0) {
1359                 status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
1360                 goto out_free_sqs;
1361         }
1362         ctrl->cntlid = ret;
1363
1364         ctrl->ops = req->ops;
1365
1366         /*
1367          * Discovery controllers may use some arbitrary high value
1368          * in order to cleanup stale discovery sessions
1369          */
1370         if ((ctrl->subsys->type == NVME_NQN_DISC) && !kato)
1371                 kato = NVMET_DISC_KATO_MS;
1372
1373         /* keep-alive timeout in seconds */
1374         ctrl->kato = DIV_ROUND_UP(kato, 1000);
1375
1376         ctrl->err_counter = 0;
1377         spin_lock_init(&ctrl->error_lock);
1378
1379         nvmet_start_keep_alive_timer(ctrl);
1380
1381         mutex_lock(&subsys->lock);
1382         list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
1383         nvmet_setup_p2p_ns_map(ctrl, req);
1384         mutex_unlock(&subsys->lock);
1385
1386         *ctrlp = ctrl;
1387         return 0;
1388
1389 out_free_sqs:
1390         kfree(ctrl->sqs);
1391 out_free_cqs:
1392         kfree(ctrl->cqs);
1393 out_free_changed_ns_list:
1394         kfree(ctrl->changed_ns_list);
1395 out_free_ctrl:
1396         kfree(ctrl);
1397 out_put_subsystem:
1398         nvmet_subsys_put(subsys);
1399 out:
1400         return status;
1401 }
1402
1403 static void nvmet_ctrl_free(struct kref *ref)
1404 {
1405         struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
1406         struct nvmet_subsys *subsys = ctrl->subsys;
1407
1408         mutex_lock(&subsys->lock);
1409         nvmet_release_p2p_ns_map(ctrl);
1410         list_del(&ctrl->subsys_entry);
1411         mutex_unlock(&subsys->lock);
1412
1413         nvmet_stop_keep_alive_timer(ctrl);
1414
1415         flush_work(&ctrl->async_event_work);
1416         cancel_work_sync(&ctrl->fatal_err_work);
1417
1418         ida_simple_remove(&cntlid_ida, ctrl->cntlid);
1419
1420         nvmet_async_events_free(ctrl);
1421         kfree(ctrl->sqs);
1422         kfree(ctrl->cqs);
1423         kfree(ctrl->changed_ns_list);
1424         kfree(ctrl);
1425
1426         nvmet_subsys_put(subsys);
1427 }
1428
1429 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
1430 {
1431         kref_put(&ctrl->ref, nvmet_ctrl_free);
1432 }
1433
1434 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
1435 {
1436         mutex_lock(&ctrl->lock);
1437         if (!(ctrl->csts & NVME_CSTS_CFS)) {
1438                 ctrl->csts |= NVME_CSTS_CFS;
1439                 schedule_work(&ctrl->fatal_err_work);
1440         }
1441         mutex_unlock(&ctrl->lock);
1442 }
1443 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
1444
1445 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
1446                 const char *subsysnqn)
1447 {
1448         struct nvmet_subsys_link *p;
1449
1450         if (!port)
1451                 return NULL;
1452
1453         if (!strcmp(NVME_DISC_SUBSYS_NAME, subsysnqn)) {
1454                 if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
1455                         return NULL;
1456                 return nvmet_disc_subsys;
1457         }
1458
1459         down_read(&nvmet_config_sem);
1460         list_for_each_entry(p, &port->subsystems, entry) {
1461                 if (!strncmp(p->subsys->subsysnqn, subsysnqn,
1462                                 NVMF_NQN_SIZE)) {
1463                         if (!kref_get_unless_zero(&p->subsys->ref))
1464                                 break;
1465                         up_read(&nvmet_config_sem);
1466                         return p->subsys;
1467                 }
1468         }
1469         up_read(&nvmet_config_sem);
1470         return NULL;
1471 }
1472
1473 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
1474                 enum nvme_subsys_type type)
1475 {
1476         struct nvmet_subsys *subsys;
1477
1478         subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
1479         if (!subsys)
1480                 return ERR_PTR(-ENOMEM);
1481
1482         subsys->ver = NVMET_DEFAULT_VS;
1483         /* generate a random serial number as our controllers are ephemeral: */
1484         get_random_bytes(&subsys->serial, sizeof(subsys->serial));
1485
1486         switch (type) {
1487         case NVME_NQN_NVME:
1488                 subsys->max_qid = NVMET_NR_QUEUES;
1489                 break;
1490         case NVME_NQN_DISC:
1491                 subsys->max_qid = 0;
1492                 break;
1493         default:
1494                 pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
1495                 kfree(subsys);
1496                 return ERR_PTR(-EINVAL);
1497         }
1498         subsys->type = type;
1499         subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
1500                         GFP_KERNEL);
1501         if (!subsys->subsysnqn) {
1502                 kfree(subsys);
1503                 return ERR_PTR(-ENOMEM);
1504         }
1505         subsys->cntlid_min = NVME_CNTLID_MIN;
1506         subsys->cntlid_max = NVME_CNTLID_MAX;
1507         kref_init(&subsys->ref);
1508
1509         mutex_init(&subsys->lock);
1510         xa_init(&subsys->namespaces);
1511         INIT_LIST_HEAD(&subsys->ctrls);
1512         INIT_LIST_HEAD(&subsys->hosts);
1513
1514         return subsys;
1515 }
1516
1517 static void nvmet_subsys_free(struct kref *ref)
1518 {
1519         struct nvmet_subsys *subsys =
1520                 container_of(ref, struct nvmet_subsys, ref);
1521
1522         WARN_ON_ONCE(!xa_empty(&subsys->namespaces));
1523
1524         xa_destroy(&subsys->namespaces);
1525         nvmet_passthru_subsys_free(subsys);
1526
1527         kfree(subsys->subsysnqn);
1528         kfree_rcu(subsys->model, rcuhead);
1529         kfree(subsys);
1530 }
1531
1532 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1533 {
1534         struct nvmet_ctrl *ctrl;
1535
1536         mutex_lock(&subsys->lock);
1537         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1538                 ctrl->ops->delete_ctrl(ctrl);
1539         mutex_unlock(&subsys->lock);
1540 }
1541
1542 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1543 {
1544         kref_put(&subsys->ref, nvmet_subsys_free);
1545 }
1546
1547 static int __init nvmet_init(void)
1548 {
1549         int error;
1550
1551         nvmet_ana_group_enabled[NVMET_DEFAULT_ANA_GRPID] = 1;
1552
1553         buffered_io_wq = alloc_workqueue("nvmet-buffered-io-wq",
1554                         WQ_MEM_RECLAIM, 0);
1555         if (!buffered_io_wq) {
1556                 error = -ENOMEM;
1557                 goto out;
1558         }
1559
1560         error = nvmet_init_discovery();
1561         if (error)
1562                 goto out_free_work_queue;
1563
1564         error = nvmet_init_configfs();
1565         if (error)
1566                 goto out_exit_discovery;
1567         return 0;
1568
1569 out_exit_discovery:
1570         nvmet_exit_discovery();
1571 out_free_work_queue:
1572         destroy_workqueue(buffered_io_wq);
1573 out:
1574         return error;
1575 }
1576
1577 static void __exit nvmet_exit(void)
1578 {
1579         nvmet_exit_configfs();
1580         nvmet_exit_discovery();
1581         ida_destroy(&cntlid_ida);
1582         destroy_workqueue(buffered_io_wq);
1583
1584         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1585         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1586 }
1587
1588 module_init(nvmet_init);
1589 module_exit(nvmet_exit);
1590
1591 MODULE_LICENSE("GPL v2");