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 struct workqueue_struct *zbd_wq;
  20 static const struct nvmet_fabrics_ops *nvmet_transports[NVMF_TRTYPE_MAX];
  21 static DEFINE_IDA(cntlid_ida);
  22
  23 struct workqueue_struct *nvmet_wq;
  24 EXPORT_SYMBOL_GPL(nvmet_wq);
  25
  26 /*
  27  * This read/write semaphore is used to synchronize access to configuration
  28  * information on a target system that will result in discovery log page
  29  * information change for at least one host.
  30  * The full list of resources to protected by this semaphore is:
  31  *
  32  *  - subsystems list
  33  *  - per-subsystem allowed hosts list
  34  *  - allow_any_host subsystem attribute
  35  *  - nvmet_genctr
  36  *  - the nvmet_transports array
  37  *
  38  * When updating any of those lists/structures write lock should be obtained,
  39  * while when reading (popolating discovery log page or checking host-subsystem
  40  * link) read lock is obtained to allow concurrent reads.
  41  */
  42 DECLARE_RWSEM(nvmet_config_sem);
  43
  44 u32 nvmet_ana_group_enabled[NVMET_MAX_ANAGRPS + 1];
  45 u64 nvmet_ana_chgcnt;
  46 DECLARE_RWSEM(nvmet_ana_sem);
  47
  48 inline u16 errno_to_nvme_status(struct nvmet_req *req, int errno)
  49 {
  50         switch (errno) {
  51         case 0:
  52                 return NVME_SC_SUCCESS;
  53         case -ENOSPC:
  54                 req->error_loc = offsetof(struct nvme_rw_command, length);
  55                 return NVME_SC_CAP_EXCEEDED | NVME_SC_DNR;
  56         case -EREMOTEIO:
  57                 req->error_loc = offsetof(struct nvme_rw_command, slba);
  58                 return  NVME_SC_LBA_RANGE | NVME_SC_DNR;
  59         case -EOPNOTSUPP:
  60                 req->error_loc = offsetof(struct nvme_common_command, opcode);
  61                 switch (req->cmd->common.opcode) {
  62                 case nvme_cmd_dsm:
  63                 case nvme_cmd_write_zeroes:
  64                         return NVME_SC_ONCS_NOT_SUPPORTED | NVME_SC_DNR;
  65                 default:
  66                         return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
  67                 }
  68                 break;
  69         case -ENODATA:
  70                 req->error_loc = offsetof(struct nvme_rw_command, nsid);
  71                 return NVME_SC_ACCESS_DENIED;
  72         case -EIO:
  73                 fallthrough;
  74         default:
  75                 req->error_loc = offsetof(struct nvme_common_command, opcode);
  76                 return NVME_SC_INTERNAL | NVME_SC_DNR;
  77         }
  78 }
  79
  80 u16 nvmet_report_invalid_opcode(struct nvmet_req *req)
  81 {
  82         pr_debug("unhandled cmd %d on qid %d\n", req->cmd->common.opcode,
  83                  req->sq->qid);
  84
  85         req->error_loc = offsetof(struct nvme_common_command, opcode);
  86         return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
  87 }
  88
  89 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
  90                 const char *subsysnqn);
  91
  92 u16 nvmet_copy_to_sgl(struct nvmet_req *req, off_t off, const void *buf,
  93                 size_t len)
  94 {
  95         if (sg_pcopy_from_buffer(req->sg, req->sg_cnt, buf, len, off) != len) {
  96                 req->error_loc = offsetof(struct nvme_common_command, dptr);
  97                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
  98         }
  99         return 0;
 100 }
 101
 102 u16 nvmet_copy_from_sgl(struct nvmet_req *req, off_t off, void *buf, size_t len)
 103 {
 104         if (sg_pcopy_to_buffer(req->sg, req->sg_cnt, buf, len, off) != len) {
 105                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 106                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
 107         }
 108         return 0;
 109 }
 110
 111 u16 nvmet_zero_sgl(struct nvmet_req *req, off_t off, size_t len)
 112 {
 113         if (sg_zero_buffer(req->sg, req->sg_cnt, len, off) != len) {
 114                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 115                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
 116         }
 117         return 0;
 118 }
 119
 120 static u32 nvmet_max_nsid(struct nvmet_subsys *subsys)
 121 {
 122         struct nvmet_ns *cur;
 123         unsigned long idx;
 124         u32 nsid = 0;
 125
 126         xa_for_each(&subsys->namespaces, idx, cur)
 127                 nsid = cur->nsid;
 128
 129         return nsid;
 130 }
 131
 132 static u32 nvmet_async_event_result(struct nvmet_async_event *aen)
 133 {
 134         return aen->event_type | (aen->event_info << 8) | (aen->log_page << 16);
 135 }
 136
 137 static void nvmet_async_events_failall(struct nvmet_ctrl *ctrl)
 138 {
 139         struct nvmet_req *req;
 140
 141         mutex_lock(&ctrl->lock);
 142         while (ctrl->nr_async_event_cmds) {
 143                 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
 144                 mutex_unlock(&ctrl->lock);
 145                 nvmet_req_complete(req, NVME_SC_INTERNAL | NVME_SC_DNR);
 146                 mutex_lock(&ctrl->lock);
 147         }
 148         mutex_unlock(&ctrl->lock);
 149 }
 150
 151 static void nvmet_async_events_process(struct nvmet_ctrl *ctrl)
 152 {
 153         struct nvmet_async_event *aen;
 154         struct nvmet_req *req;
 155
 156         mutex_lock(&ctrl->lock);
 157         while (ctrl->nr_async_event_cmds && !list_empty(&ctrl->async_events)) {
 158                 aen = list_first_entry(&ctrl->async_events,
 159                                        struct nvmet_async_event, entry);
 160                 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
 161                 nvmet_set_result(req, nvmet_async_event_result(aen));
 162
 163                 list_del(&aen->entry);
 164                 kfree(aen);
 165
 166                 mutex_unlock(&ctrl->lock);
 167                 trace_nvmet_async_event(ctrl, req->cqe->result.u32);
 168                 nvmet_req_complete(req, 0);
 169                 mutex_lock(&ctrl->lock);
 170         }
 171         mutex_unlock(&ctrl->lock);
 172 }
 173
 174 static void nvmet_async_events_free(struct nvmet_ctrl *ctrl)
 175 {
 176         struct nvmet_async_event *aen, *tmp;
 177
 178         mutex_lock(&ctrl->lock);
 179         list_for_each_entry_safe(aen, tmp, &ctrl->async_events, entry) {
 180                 list_del(&aen->entry);
 181                 kfree(aen);
 182         }
 183         mutex_unlock(&ctrl->lock);
 184 }
 185
 186 static void nvmet_async_event_work(struct work_struct *work)
 187 {
 188         struct nvmet_ctrl *ctrl =
 189                 container_of(work, struct nvmet_ctrl, async_event_work);
 190
 191         nvmet_async_events_process(ctrl);
 192 }
 193
 194 void nvmet_add_async_event(struct nvmet_ctrl *ctrl, u8 event_type,
 195                 u8 event_info, u8 log_page)
 196 {
 197         struct nvmet_async_event *aen;
 198
 199         aen = kmalloc(sizeof(*aen), GFP_KERNEL);
 200         if (!aen)
 201                 return;
 202
 203         aen->event_type = event_type;
 204         aen->event_info = event_info;
 205         aen->log_page = log_page;
 206
 207         mutex_lock(&ctrl->lock);
 208         list_add_tail(&aen->entry, &ctrl->async_events);
 209         mutex_unlock(&ctrl->lock);
 210
 211         queue_work(nvmet_wq, &ctrl->async_event_work);
 212 }
 213
 214 static void nvmet_add_to_changed_ns_log(struct nvmet_ctrl *ctrl, __le32 nsid)
 215 {
 216         u32 i;
 217
 218         mutex_lock(&ctrl->lock);
 219         if (ctrl->nr_changed_ns > NVME_MAX_CHANGED_NAMESPACES)
 220                 goto out_unlock;
 221
 222         for (i = 0; i < ctrl->nr_changed_ns; i++) {
 223                 if (ctrl->changed_ns_list[i] == nsid)
 224                         goto out_unlock;
 225         }
 226
 227         if (ctrl->nr_changed_ns == NVME_MAX_CHANGED_NAMESPACES) {
 228                 ctrl->changed_ns_list[0] = cpu_to_le32(0xffffffff);
 229                 ctrl->nr_changed_ns = U32_MAX;
 230                 goto out_unlock;
 231         }
 232
 233         ctrl->changed_ns_list[ctrl->nr_changed_ns++] = nsid;
 234 out_unlock:
 235         mutex_unlock(&ctrl->lock);
 236 }
 237
 238 void nvmet_ns_changed(struct nvmet_subsys *subsys, u32 nsid)
 239 {
 240         struct nvmet_ctrl *ctrl;
 241
 242         lockdep_assert_held(&subsys->lock);
 243
 244         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 245                 nvmet_add_to_changed_ns_log(ctrl, cpu_to_le32(nsid));
 246                 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_NS_ATTR))
 247                         continue;
 248                 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
 249                                 NVME_AER_NOTICE_NS_CHANGED,
 250                                 NVME_LOG_CHANGED_NS);
 251         }
 252 }
 253
 254 void nvmet_send_ana_event(struct nvmet_subsys *subsys,
 255                 struct nvmet_port *port)
 256 {
 257         struct nvmet_ctrl *ctrl;
 258
 259         mutex_lock(&subsys->lock);
 260         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 261                 if (port && ctrl->port != port)
 262                         continue;
 263                 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_ANA_CHANGE))
 264                         continue;
 265                 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
 266                                 NVME_AER_NOTICE_ANA, NVME_LOG_ANA);
 267         }
 268         mutex_unlock(&subsys->lock);
 269 }
 270
 271 void nvmet_port_send_ana_event(struct nvmet_port *port)
 272 {
 273         struct nvmet_subsys_link *p;
 274
 275         down_read(&nvmet_config_sem);
 276         list_for_each_entry(p, &port->subsystems, entry)
 277                 nvmet_send_ana_event(p->subsys, port);
 278         up_read(&nvmet_config_sem);
 279 }
 280
 281 int nvmet_register_transport(const struct nvmet_fabrics_ops *ops)
 282 {
 283         int ret = 0;
 284
 285         down_write(&nvmet_config_sem);
 286         if (nvmet_transports[ops->type])
 287                 ret = -EINVAL;
 288         else
 289                 nvmet_transports[ops->type] = ops;
 290         up_write(&nvmet_config_sem);
 291
 292         return ret;
 293 }
 294 EXPORT_SYMBOL_GPL(nvmet_register_transport);
 295
 296 void nvmet_unregister_transport(const struct nvmet_fabrics_ops *ops)
 297 {
 298         down_write(&nvmet_config_sem);
 299         nvmet_transports[ops->type] = NULL;
 300         up_write(&nvmet_config_sem);
 301 }
 302 EXPORT_SYMBOL_GPL(nvmet_unregister_transport);
 303
 304 void nvmet_port_del_ctrls(struct nvmet_port *port, struct nvmet_subsys *subsys)
 305 {
 306         struct nvmet_ctrl *ctrl;
 307
 308         mutex_lock(&subsys->lock);
 309         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 310                 if (ctrl->port == port)
 311                         ctrl->ops->delete_ctrl(ctrl);
 312         }
 313         mutex_unlock(&subsys->lock);
 314 }
 315
 316 int nvmet_enable_port(struct nvmet_port *port)
 317 {
 318         const struct nvmet_fabrics_ops *ops;
 319         int ret;
 320
 321         lockdep_assert_held(&nvmet_config_sem);
 322
 323         ops = nvmet_transports[port->disc_addr.trtype];
 324         if (!ops) {
 325                 up_write(&nvmet_config_sem);
 326                 request_module("nvmet-transport-%d", port->disc_addr.trtype);
 327                 down_write(&nvmet_config_sem);
 328                 ops = nvmet_transports[port->disc_addr.trtype];
 329                 if (!ops) {
 330                         pr_err("transport type %d not supported\n",
 331                                 port->disc_addr.trtype);
 332                         return -EINVAL;
 333                 }
 334         }
 335
 336         if (!try_module_get(ops->owner))
 337                 return -EINVAL;
 338
 339         /*
 340          * If the user requested PI support and the transport isn't pi capable,
 341          * don't enable the port.
 342          */
 343         if (port->pi_enable && !(ops->flags & NVMF_METADATA_SUPPORTED)) {
 344                 pr_err("T10-PI is not supported by transport type %d\n",
 345                        port->disc_addr.trtype);
 346                 ret = -EINVAL;
 347                 goto out_put;
 348         }
 349
 350         ret = ops->add_port(port);
 351         if (ret)
 352                 goto out_put;
 353
 354         /* If the transport didn't set inline_data_size, then disable it. */
 355         if (port->inline_data_size < 0)
 356                 port->inline_data_size = 0;
 357
 358         port->enabled = true;
 359         port->tr_ops = ops;
 360         return 0;
 361
 362 out_put:
 363         module_put(ops->owner);
 364         return ret;
 365 }
 366
 367 void nvmet_disable_port(struct nvmet_port *port)
 368 {
 369         const struct nvmet_fabrics_ops *ops;
 370
 371         lockdep_assert_held(&nvmet_config_sem);
 372
 373         port->enabled = false;
 374         port->tr_ops = NULL;
 375
 376         ops = nvmet_transports[port->disc_addr.trtype];
 377         ops->remove_port(port);
 378         module_put(ops->owner);
 379 }
 380
 381 static void nvmet_keep_alive_timer(struct work_struct *work)
 382 {
 383         struct nvmet_ctrl *ctrl = container_of(to_delayed_work(work),
 384                         struct nvmet_ctrl, ka_work);
 385         bool reset_tbkas = ctrl->reset_tbkas;
 386
 387         ctrl->reset_tbkas = false;
 388         if (reset_tbkas) {
 389                 pr_debug("ctrl %d reschedule traffic based keep-alive timer\n",
 390                         ctrl->cntlid);
 391                 queue_delayed_work(nvmet_wq, &ctrl->ka_work, ctrl->kato * HZ);
 392                 return;
 393         }
 394
 395         pr_err("ctrl %d keep-alive timer (%d seconds) expired!\n",
 396                 ctrl->cntlid, ctrl->kato);
 397
 398         nvmet_ctrl_fatal_error(ctrl);
 399 }
 400
 401 void nvmet_start_keep_alive_timer(struct nvmet_ctrl *ctrl)
 402 {
 403         if (unlikely(ctrl->kato == 0))
 404                 return;
 405
 406         pr_debug("ctrl %d start keep-alive timer for %d secs\n",
 407                 ctrl->cntlid, ctrl->kato);
 408
 409         queue_delayed_work(nvmet_wq, &ctrl->ka_work, ctrl->kato * HZ);
 410 }
 411
 412 void nvmet_stop_keep_alive_timer(struct nvmet_ctrl *ctrl)
 413 {
 414         if (unlikely(ctrl->kato == 0))
 415                 return;
 416
 417         pr_debug("ctrl %d stop keep-alive\n", ctrl->cntlid);
 418
 419         cancel_delayed_work_sync(&ctrl->ka_work);
 420 }
 421
 422 u16 nvmet_req_find_ns(struct nvmet_req *req)
 423 {
 424         u32 nsid = le32_to_cpu(req->cmd->common.nsid);
 425
 426         req->ns = xa_load(&nvmet_req_subsys(req)->namespaces, nsid);
 427         if (unlikely(!req->ns)) {
 428                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 429                 return NVME_SC_INVALID_NS | NVME_SC_DNR;
 430         }
 431
 432         percpu_ref_get(&req->ns->ref);
 433         return NVME_SC_SUCCESS;
 434 }
 435
 436 static void nvmet_destroy_namespace(struct percpu_ref *ref)
 437 {
 438         struct nvmet_ns *ns = container_of(ref, struct nvmet_ns, ref);
 439
 440         complete(&ns->disable_done);
 441 }
 442
 443 void nvmet_put_namespace(struct nvmet_ns *ns)
 444 {
 445         percpu_ref_put(&ns->ref);
 446 }
 447
 448 static void nvmet_ns_dev_disable(struct nvmet_ns *ns)
 449 {
 450         nvmet_bdev_ns_disable(ns);
 451         nvmet_file_ns_disable(ns);
 452 }
 453
 454 static int nvmet_p2pmem_ns_enable(struct nvmet_ns *ns)
 455 {
 456         int ret;
 457         struct pci_dev *p2p_dev;
 458
 459         if (!ns->use_p2pmem)
 460                 return 0;
 461
 462         if (!ns->bdev) {
 463                 pr_err("peer-to-peer DMA is not supported by non-block device namespaces\n");
 464                 return -EINVAL;
 465         }
 466
 467         if (!blk_queue_pci_p2pdma(ns->bdev->bd_disk->queue)) {
 468                 pr_err("peer-to-peer DMA is not supported by the driver of %s\n",
 469                        ns->device_path);
 470                 return -EINVAL;
 471         }
 472
 473         if (ns->p2p_dev) {
 474                 ret = pci_p2pdma_distance(ns->p2p_dev, nvmet_ns_dev(ns), true);
 475                 if (ret < 0)
 476                         return -EINVAL;
 477         } else {
 478                 /*
 479                  * Right now we just check that there is p2pmem available so
 480                  * we can report an error to the user right away if there
 481                  * is not. We'll find the actual device to use once we
 482                  * setup the controller when the port's device is available.
 483                  */
 484
 485                 p2p_dev = pci_p2pmem_find(nvmet_ns_dev(ns));
 486                 if (!p2p_dev) {
 487                         pr_err("no peer-to-peer memory is available for %s\n",
 488                                ns->device_path);
 489                         return -EINVAL;
 490                 }
 491
 492                 pci_dev_put(p2p_dev);
 493         }
 494
 495         return 0;
 496 }
 497
 498 /*
 499  * Note: ctrl->subsys->lock should be held when calling this function
 500  */
 501 static void nvmet_p2pmem_ns_add_p2p(struct nvmet_ctrl *ctrl,
 502                                     struct nvmet_ns *ns)
 503 {
 504         struct device *clients[2];
 505         struct pci_dev *p2p_dev;
 506         int ret;
 507
 508         if (!ctrl->p2p_client || !ns->use_p2pmem)
 509                 return;
 510
 511         if (ns->p2p_dev) {
 512                 ret = pci_p2pdma_distance(ns->p2p_dev, ctrl->p2p_client, true);
 513                 if (ret < 0)
 514                         return;
 515
 516                 p2p_dev = pci_dev_get(ns->p2p_dev);
 517         } else {
 518                 clients[0] = ctrl->p2p_client;
 519                 clients[1] = nvmet_ns_dev(ns);
 520
 521                 p2p_dev = pci_p2pmem_find_many(clients, ARRAY_SIZE(clients));
 522                 if (!p2p_dev) {
 523                         pr_err("no peer-to-peer memory is available that's supported by %s and %s\n",
 524                                dev_name(ctrl->p2p_client), ns->device_path);
 525                         return;
 526                 }
 527         }
 528
 529         ret = radix_tree_insert(&ctrl->p2p_ns_map, ns->nsid, p2p_dev);
 530         if (ret < 0)
 531                 pci_dev_put(p2p_dev);
 532
 533         pr_info("using p2pmem on %s for nsid %d\n", pci_name(p2p_dev),
 534                 ns->nsid);
 535 }
 536
 537 bool nvmet_ns_revalidate(struct nvmet_ns *ns)
 538 {
 539         loff_t oldsize = ns->size;
 540
 541         if (ns->bdev)
 542                 nvmet_bdev_ns_revalidate(ns);
 543         else
 544                 nvmet_file_ns_revalidate(ns);
 545
 546         return oldsize != ns->size;
 547 }
 548
 549 int nvmet_ns_enable(struct nvmet_ns *ns)
 550 {
 551         struct nvmet_subsys *subsys = ns->subsys;
 552         struct nvmet_ctrl *ctrl;
 553         int ret;
 554
 555         mutex_lock(&subsys->lock);
 556         ret = 0;
 557
 558         if (nvmet_is_passthru_subsys(subsys)) {
 559                 pr_info("cannot enable both passthru and regular namespaces for a single subsystem");
 560                 goto out_unlock;
 561         }
 562
 563         if (ns->enabled)
 564                 goto out_unlock;
 565
 566         ret = -EMFILE;
 567         if (subsys->nr_namespaces == NVMET_MAX_NAMESPACES)
 568                 goto out_unlock;
 569
 570         ret = nvmet_bdev_ns_enable(ns);
 571         if (ret == -ENOTBLK)
 572                 ret = nvmet_file_ns_enable(ns);
 573         if (ret)
 574                 goto out_unlock;
 575
 576         ret = nvmet_p2pmem_ns_enable(ns);
 577         if (ret)
 578                 goto out_dev_disable;
 579
 580         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 581                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
 582
 583         ret = percpu_ref_init(&ns->ref, nvmet_destroy_namespace,
 584                                 0, GFP_KERNEL);
 585         if (ret)
 586                 goto out_dev_put;
 587
 588         if (ns->nsid > subsys->max_nsid)
 589                 subsys->max_nsid = ns->nsid;
 590
 591         ret = xa_insert(&subsys->namespaces, ns->nsid, ns, GFP_KERNEL);
 592         if (ret)
 593                 goto out_restore_subsys_maxnsid;
 594
 595         subsys->nr_namespaces++;
 596
 597         nvmet_ns_changed(subsys, ns->nsid);
 598         ns->enabled = true;
 599         ret = 0;
 600 out_unlock:
 601         mutex_unlock(&subsys->lock);
 602         return ret;
 603
 604 out_restore_subsys_maxnsid:
 605         subsys->max_nsid = nvmet_max_nsid(subsys);
 606         percpu_ref_exit(&ns->ref);
 607 out_dev_put:
 608         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 609                 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
 610 out_dev_disable:
 611         nvmet_ns_dev_disable(ns);
 612         goto out_unlock;
 613 }
 614
 615 void nvmet_ns_disable(struct nvmet_ns *ns)
 616 {
 617         struct nvmet_subsys *subsys = ns->subsys;
 618         struct nvmet_ctrl *ctrl;
 619
 620         mutex_lock(&subsys->lock);
 621         if (!ns->enabled)
 622                 goto out_unlock;
 623
 624         ns->enabled = false;
 625         xa_erase(&ns->subsys->namespaces, ns->nsid);
 626         if (ns->nsid == subsys->max_nsid)
 627                 subsys->max_nsid = nvmet_max_nsid(subsys);
 628
 629         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 630                 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
 631
 632         mutex_unlock(&subsys->lock);
 633
 634         /*
 635          * Now that we removed the namespaces from the lookup list, we
 636          * can kill the per_cpu ref and wait for any remaining references
 637          * to be dropped, as well as a RCU grace period for anyone only
 638          * using the namepace under rcu_read_lock().  Note that we can't
 639          * use call_rcu here as we need to ensure the namespaces have
 640          * been fully destroyed before unloading the module.
 641          */
 642         percpu_ref_kill(&ns->ref);
 643         synchronize_rcu();
 644         wait_for_completion(&ns->disable_done);
 645         percpu_ref_exit(&ns->ref);
 646
 647         mutex_lock(&subsys->lock);
 648
 649         subsys->nr_namespaces--;
 650         nvmet_ns_changed(subsys, ns->nsid);
 651         nvmet_ns_dev_disable(ns);
 652 out_unlock:
 653         mutex_unlock(&subsys->lock);
 654 }
 655
 656 void nvmet_ns_free(struct nvmet_ns *ns)
 657 {
 658         nvmet_ns_disable(ns);
 659
 660         down_write(&nvmet_ana_sem);
 661         nvmet_ana_group_enabled[ns->anagrpid]--;
 662         up_write(&nvmet_ana_sem);
 663
 664         kfree(ns->device_path);
 665         kfree(ns);
 666 }
 667
 668 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
 669 {
 670         struct nvmet_ns *ns;
 671
 672         ns = kzalloc(sizeof(*ns), GFP_KERNEL);
 673         if (!ns)
 674                 return NULL;
 675
 676         init_completion(&ns->disable_done);
 677
 678         ns->nsid = nsid;
 679         ns->subsys = subsys;
 680
 681         down_write(&nvmet_ana_sem);
 682         ns->anagrpid = NVMET_DEFAULT_ANA_GRPID;
 683         nvmet_ana_group_enabled[ns->anagrpid]++;
 684         up_write(&nvmet_ana_sem);
 685
 686         uuid_gen(&ns->uuid);
 687         ns->buffered_io = false;
 688         ns->csi = NVME_CSI_NVM;
 689
 690         return ns;
 691 }
 692
 693 static void nvmet_update_sq_head(struct nvmet_req *req)
 694 {
 695         if (req->sq->size) {
 696                 u32 old_sqhd, new_sqhd;
 697
 698                 do {
 699                         old_sqhd = req->sq->sqhd;
 700                         new_sqhd = (old_sqhd + 1) % req->sq->size;
 701                 } while (cmpxchg(&req->sq->sqhd, old_sqhd, new_sqhd) !=
 702                                         old_sqhd);
 703         }
 704         req->cqe->sq_head = cpu_to_le16(req->sq->sqhd & 0x0000FFFF);
 705 }
 706
 707 static void nvmet_set_error(struct nvmet_req *req, u16 status)
 708 {
 709         struct nvmet_ctrl *ctrl = req->sq->ctrl;
 710         struct nvme_error_slot *new_error_slot;
 711         unsigned long flags;
 712
 713         req->cqe->status = cpu_to_le16(status << 1);
 714
 715         if (!ctrl || req->error_loc == NVMET_NO_ERROR_LOC)
 716                 return;
 717
 718         spin_lock_irqsave(&ctrl->error_lock, flags);
 719         ctrl->err_counter++;
 720         new_error_slot =
 721                 &ctrl->slots[ctrl->err_counter % NVMET_ERROR_LOG_SLOTS];
 722
 723         new_error_slot->error_count = cpu_to_le64(ctrl->err_counter);
 724         new_error_slot->sqid = cpu_to_le16(req->sq->qid);
 725         new_error_slot->cmdid = cpu_to_le16(req->cmd->common.command_id);
 726         new_error_slot->status_field = cpu_to_le16(status << 1);
 727         new_error_slot->param_error_location = cpu_to_le16(req->error_loc);
 728         new_error_slot->lba = cpu_to_le64(req->error_slba);
 729         new_error_slot->nsid = req->cmd->common.nsid;
 730         spin_unlock_irqrestore(&ctrl->error_lock, flags);
 731
 732         /* set the more bit for this request */
 733         req->cqe->status |= cpu_to_le16(1 << 14);
 734 }
 735
 736 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
 737 {
 738         if (!req->sq->sqhd_disabled)
 739                 nvmet_update_sq_head(req);
 740         req->cqe->sq_id = cpu_to_le16(req->sq->qid);
 741         req->cqe->command_id = req->cmd->common.command_id;
 742
 743         if (unlikely(status))
 744                 nvmet_set_error(req, status);
 745
 746         trace_nvmet_req_complete(req);
 747
 748         if (req->ns)
 749                 nvmet_put_namespace(req->ns);
 750         req->ops->queue_response(req);
 751 }
 752
 753 void nvmet_req_complete(struct nvmet_req *req, u16 status)
 754 {
 755         __nvmet_req_complete(req, status);
 756         percpu_ref_put(&req->sq->ref);
 757 }
 758 EXPORT_SYMBOL_GPL(nvmet_req_complete);
 759
 760 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
 761                 u16 qid, u16 size)
 762 {
 763         cq->qid = qid;
 764         cq->size = size;
 765 }
 766
 767 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
 768                 u16 qid, u16 size)
 769 {
 770         sq->sqhd = 0;
 771         sq->qid = qid;
 772         sq->size = size;
 773
 774         ctrl->sqs[qid] = sq;
 775 }
 776
 777 static void nvmet_confirm_sq(struct percpu_ref *ref)
 778 {
 779         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 780
 781         complete(&sq->confirm_done);
 782 }
 783
 784 void nvmet_sq_destroy(struct nvmet_sq *sq)
 785 {
 786         struct nvmet_ctrl *ctrl = sq->ctrl;
 787
 788         /*
 789          * If this is the admin queue, complete all AERs so that our
 790          * queue doesn't have outstanding requests on it.
 791          */
 792         if (ctrl && ctrl->sqs && ctrl->sqs[0] == sq)
 793                 nvmet_async_events_failall(ctrl);
 794         percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
 795         wait_for_completion(&sq->confirm_done);
 796         wait_for_completion(&sq->free_done);
 797         percpu_ref_exit(&sq->ref);
 798
 799         if (ctrl) {
 800                 /*
 801                  * The teardown flow may take some time, and the host may not
 802                  * send us keep-alive during this period, hence reset the
 803                  * traffic based keep-alive timer so we don't trigger a
 804                  * controller teardown as a result of a keep-alive expiration.
 805                  */
 806                 ctrl->reset_tbkas = true;
 807                 sq->ctrl->sqs[sq->qid] = NULL;
 808                 nvmet_ctrl_put(ctrl);
 809                 sq->ctrl = NULL; /* allows reusing the queue later */
 810         }
 811 }
 812 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
 813
 814 static void nvmet_sq_free(struct percpu_ref *ref)
 815 {
 816         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 817
 818         complete(&sq->free_done);
 819 }
 820
 821 int nvmet_sq_init(struct nvmet_sq *sq)
 822 {
 823         int ret;
 824
 825         ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
 826         if (ret) {
 827                 pr_err("percpu_ref init failed!\n");
 828                 return ret;
 829         }
 830         init_completion(&sq->free_done);
 831         init_completion(&sq->confirm_done);
 832
 833         return 0;
 834 }
 835 EXPORT_SYMBOL_GPL(nvmet_sq_init);
 836
 837 static inline u16 nvmet_check_ana_state(struct nvmet_port *port,
 838                 struct nvmet_ns *ns)
 839 {
 840         enum nvme_ana_state state = port->ana_state[ns->anagrpid];
 841
 842         if (unlikely(state == NVME_ANA_INACCESSIBLE))
 843                 return NVME_SC_ANA_INACCESSIBLE;
 844         if (unlikely(state == NVME_ANA_PERSISTENT_LOSS))
 845                 return NVME_SC_ANA_PERSISTENT_LOSS;
 846         if (unlikely(state == NVME_ANA_CHANGE))
 847                 return NVME_SC_ANA_TRANSITION;
 848         return 0;
 849 }
 850
 851 static inline u16 nvmet_io_cmd_check_access(struct nvmet_req *req)
 852 {
 853         if (unlikely(req->ns->readonly)) {
 854                 switch (req->cmd->common.opcode) {
 855                 case nvme_cmd_read:
 856                 case nvme_cmd_flush:
 857                         break;
 858                 default:
 859                         return NVME_SC_NS_WRITE_PROTECTED;
 860                 }
 861         }
 862
 863         return 0;
 864 }
 865
 866 static u16 nvmet_parse_io_cmd(struct nvmet_req *req)
 867 {
 868         u16 ret;
 869
 870         ret = nvmet_check_ctrl_status(req);
 871         if (unlikely(ret))
 872                 return ret;
 873
 874         if (nvmet_is_passthru_req(req))
 875                 return nvmet_parse_passthru_io_cmd(req);
 876
 877         ret = nvmet_req_find_ns(req);
 878         if (unlikely(ret))
 879                 return ret;
 880
 881         ret = nvmet_check_ana_state(req->port, req->ns);
 882         if (unlikely(ret)) {
 883                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 884                 return ret;
 885         }
 886         ret = nvmet_io_cmd_check_access(req);
 887         if (unlikely(ret)) {
 888                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 889                 return ret;
 890         }
 891
 892         switch (req->ns->csi) {
 893         case NVME_CSI_NVM:
 894                 if (req->ns->file)
 895                         return nvmet_file_parse_io_cmd(req);
 896                 return nvmet_bdev_parse_io_cmd(req);
 897         case NVME_CSI_ZNS:
 898                 if (IS_ENABLED(CONFIG_BLK_DEV_ZONED))
 899                         return nvmet_bdev_zns_parse_io_cmd(req);
 900                 return NVME_SC_INVALID_IO_CMD_SET;
 901         default:
 902                 return NVME_SC_INVALID_IO_CMD_SET;
 903         }
 904 }
 905
 906 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
 907                 struct nvmet_sq *sq, const struct nvmet_fabrics_ops *ops)
 908 {
 909         u8 flags = req->cmd->common.flags;
 910         u16 status;
 911
 912         req->cq = cq;
 913         req->sq = sq;
 914         req->ops = ops;
 915         req->sg = NULL;
 916         req->metadata_sg = NULL;
 917         req->sg_cnt = 0;
 918         req->metadata_sg_cnt = 0;
 919         req->transfer_len = 0;
 920         req->metadata_len = 0;
 921         req->cqe->status = 0;
 922         req->cqe->sq_head = 0;
 923         req->ns = NULL;
 924         req->error_loc = NVMET_NO_ERROR_LOC;
 925         req->error_slba = 0;
 926
 927         /* no support for fused commands yet */
 928         if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
 929                 req->error_loc = offsetof(struct nvme_common_command, flags);
 930                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 931                 goto fail;
 932         }
 933
 934         /*
 935          * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
 936          * contains an address of a single contiguous physical buffer that is
 937          * byte aligned.
 938          */
 939         if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
 940                 req->error_loc = offsetof(struct nvme_common_command, flags);
 941                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 942                 goto fail;
 943         }
 944
 945         if (unlikely(!req->sq->ctrl))
 946                 /* will return an error for any non-connect command: */
 947                 status = nvmet_parse_connect_cmd(req);
 948         else if (likely(req->sq->qid != 0))
 949                 status = nvmet_parse_io_cmd(req);
 950         else
 951                 status = nvmet_parse_admin_cmd(req);
 952
 953         if (status)
 954                 goto fail;
 955
 956         trace_nvmet_req_init(req, req->cmd);
 957
 958         if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
 959                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 960                 goto fail;
 961         }
 962
 963         if (sq->ctrl)
 964                 sq->ctrl->reset_tbkas = true;
 965
 966         return true;
 967
 968 fail:
 969         __nvmet_req_complete(req, status);
 970         return false;
 971 }
 972 EXPORT_SYMBOL_GPL(nvmet_req_init);
 973
 974 void nvmet_req_uninit(struct nvmet_req *req)
 975 {
 976         percpu_ref_put(&req->sq->ref);
 977         if (req->ns)
 978                 nvmet_put_namespace(req->ns);
 979 }
 980 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
 981
 982 bool nvmet_check_transfer_len(struct nvmet_req *req, size_t len)
 983 {
 984         if (unlikely(len != req->transfer_len)) {
 985                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 986                 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
 987                 return false;
 988         }
 989
 990         return true;
 991 }
 992 EXPORT_SYMBOL_GPL(nvmet_check_transfer_len);
 993
 994 bool nvmet_check_data_len_lte(struct nvmet_req *req, size_t data_len)
 995 {
 996         if (unlikely(data_len > req->transfer_len)) {
 997                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 998                 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
 999                 return false;
1000         }
1001
1002         return true;
1003 }
1004
1005 static unsigned int nvmet_data_transfer_len(struct nvmet_req *req)
1006 {
1007         return req->transfer_len - req->metadata_len;
1008 }
1009
1010 static int nvmet_req_alloc_p2pmem_sgls(struct pci_dev *p2p_dev,
1011                 struct nvmet_req *req)
1012 {
1013         req->sg = pci_p2pmem_alloc_sgl(p2p_dev, &req->sg_cnt,
1014                         nvmet_data_transfer_len(req));
1015         if (!req->sg)
1016                 goto out_err;
1017
1018         if (req->metadata_len) {
1019                 req->metadata_sg = pci_p2pmem_alloc_sgl(p2p_dev,
1020                                 &req->metadata_sg_cnt, req->metadata_len);
1021                 if (!req->metadata_sg)
1022                         goto out_free_sg;
1023         }
1024
1025         req->p2p_dev = p2p_dev;
1026
1027         return 0;
1028 out_free_sg:
1029         pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1030 out_err:
1031         return -ENOMEM;
1032 }
1033
1034 static struct pci_dev *nvmet_req_find_p2p_dev(struct nvmet_req *req)
1035 {
1036         if (!IS_ENABLED(CONFIG_PCI_P2PDMA) ||
1037             !req->sq->ctrl || !req->sq->qid || !req->ns)
1038                 return NULL;
1039         return radix_tree_lookup(&req->sq->ctrl->p2p_ns_map, req->ns->nsid);
1040 }
1041
1042 int nvmet_req_alloc_sgls(struct nvmet_req *req)
1043 {
1044         struct pci_dev *p2p_dev = nvmet_req_find_p2p_dev(req);
1045
1046         if (p2p_dev && !nvmet_req_alloc_p2pmem_sgls(p2p_dev, req))
1047                 return 0;
1048
1049         req->sg = sgl_alloc(nvmet_data_transfer_len(req), GFP_KERNEL,
1050                             &req->sg_cnt);
1051         if (unlikely(!req->sg))
1052                 goto out;
1053
1054         if (req->metadata_len) {
1055                 req->metadata_sg = sgl_alloc(req->metadata_len, GFP_KERNEL,
1056                                              &req->metadata_sg_cnt);
1057                 if (unlikely(!req->metadata_sg))
1058                         goto out_free;
1059         }
1060
1061         return 0;
1062 out_free:
1063         sgl_free(req->sg);
1064 out:
1065         return -ENOMEM;
1066 }
1067 EXPORT_SYMBOL_GPL(nvmet_req_alloc_sgls);
1068
1069 void nvmet_req_free_sgls(struct nvmet_req *req)
1070 {
1071         if (req->p2p_dev) {
1072                 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1073                 if (req->metadata_sg)
1074                         pci_p2pmem_free_sgl(req->p2p_dev, req->metadata_sg);
1075                 req->p2p_dev = NULL;
1076         } else {
1077                 sgl_free(req->sg);
1078                 if (req->metadata_sg)
1079                         sgl_free(req->metadata_sg);
1080         }
1081
1082         req->sg = NULL;
1083         req->metadata_sg = NULL;
1084         req->sg_cnt = 0;
1085         req->metadata_sg_cnt = 0;
1086 }
1087 EXPORT_SYMBOL_GPL(nvmet_req_free_sgls);
1088
1089 static inline bool nvmet_cc_en(u32 cc)
1090 {
1091         return (cc >> NVME_CC_EN_SHIFT) & 0x1;
1092 }
1093
1094 static inline u8 nvmet_cc_css(u32 cc)
1095 {
1096         return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
1097 }
1098
1099 static inline u8 nvmet_cc_mps(u32 cc)
1100 {
1101         return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
1102 }
1103
1104 static inline u8 nvmet_cc_ams(u32 cc)
1105 {
1106         return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
1107 }
1108
1109 static inline u8 nvmet_cc_shn(u32 cc)
1110 {
1111         return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
1112 }
1113
1114 static inline u8 nvmet_cc_iosqes(u32 cc)
1115 {
1116         return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
1117 }
1118
1119 static inline u8 nvmet_cc_iocqes(u32 cc)
1120 {
1121         return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
1122 }
1123
1124 static inline bool nvmet_css_supported(u8 cc_css)
1125 {
1126         switch (cc_css << NVME_CC_CSS_SHIFT) {
1127         case NVME_CC_CSS_NVM:
1128         case NVME_CC_CSS_CSI:
1129                 return true;
1130         default:
1131                 return false;
1132         }
1133 }
1134
1135 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
1136 {
1137         lockdep_assert_held(&ctrl->lock);
1138
1139         /*
1140          * Only I/O controllers should verify iosqes,iocqes.
1141          * Strictly speaking, the spec says a discovery controller
1142          * should verify iosqes,iocqes are zeroed, however that
1143          * would break backwards compatibility, so don't enforce it.
1144          */
1145         if (!nvmet_is_disc_subsys(ctrl->subsys) &&
1146             (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
1147              nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES)) {
1148                 ctrl->csts = NVME_CSTS_CFS;
1149                 return;
1150         }
1151
1152         if (nvmet_cc_mps(ctrl->cc) != 0 ||
1153             nvmet_cc_ams(ctrl->cc) != 0 ||
1154             !nvmet_css_supported(nvmet_cc_css(ctrl->cc))) {
1155                 ctrl->csts = NVME_CSTS_CFS;
1156                 return;
1157         }
1158
1159         ctrl->csts = NVME_CSTS_RDY;
1160
1161         /*
1162          * Controllers that are not yet enabled should not really enforce the
1163          * keep alive timeout, but we still want to track a timeout and cleanup
1164          * in case a host died before it enabled the controller.  Hence, simply
1165          * reset the keep alive timer when the controller is enabled.
1166          */
1167         if (ctrl->kato)
1168                 mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ);
1169 }
1170
1171 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
1172 {
1173         lockdep_assert_held(&ctrl->lock);
1174
1175         /* XXX: tear down queues? */
1176         ctrl->csts &= ~NVME_CSTS_RDY;
1177         ctrl->cc = 0;
1178 }
1179
1180 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
1181 {
1182         u32 old;
1183
1184         mutex_lock(&ctrl->lock);
1185         old = ctrl->cc;
1186         ctrl->cc = new;
1187
1188         if (nvmet_cc_en(new) && !nvmet_cc_en(old))
1189                 nvmet_start_ctrl(ctrl);
1190         if (!nvmet_cc_en(new) && nvmet_cc_en(old))
1191                 nvmet_clear_ctrl(ctrl);
1192         if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
1193                 nvmet_clear_ctrl(ctrl);
1194                 ctrl->csts |= NVME_CSTS_SHST_CMPLT;
1195         }
1196         if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
1197                 ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
1198         mutex_unlock(&ctrl->lock);
1199 }
1200
1201 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
1202 {
1203         /* command sets supported: NVMe command set: */
1204         ctrl->cap = (1ULL << 37);
1205         /* Controller supports one or more I/O Command Sets */
1206         ctrl->cap |= (1ULL << 43);
1207         /* CC.EN timeout in 500msec units: */
1208         ctrl->cap |= (15ULL << 24);
1209         /* maximum queue entries supported: */
1210         if (ctrl->ops->get_max_queue_size)
1211                 ctrl->cap |= ctrl->ops->get_max_queue_size(ctrl) - 1;
1212         else
1213                 ctrl->cap |= NVMET_QUEUE_SIZE - 1;
1214
1215         if (nvmet_is_passthru_subsys(ctrl->subsys))
1216                 nvmet_passthrough_override_cap(ctrl);
1217 }
1218
1219 struct nvmet_ctrl *nvmet_ctrl_find_get(const char *subsysnqn,
1220                                        const char *hostnqn, u16 cntlid,
1221                                        struct nvmet_req *req)
1222 {
1223         struct nvmet_ctrl *ctrl = NULL;
1224         struct nvmet_subsys *subsys;
1225
1226         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1227         if (!subsys) {
1228                 pr_warn("connect request for invalid subsystem %s!\n",
1229                         subsysnqn);
1230                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1231                 goto out;
1232         }
1233
1234         mutex_lock(&subsys->lock);
1235         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
1236                 if (ctrl->cntlid == cntlid) {
1237                         if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
1238                                 pr_warn("hostnqn mismatch.\n");
1239                                 continue;
1240                         }
1241                         if (!kref_get_unless_zero(&ctrl->ref))
1242                                 continue;
1243
1244                         /* ctrl found */
1245                         goto found;
1246                 }
1247         }
1248
1249         ctrl = NULL; /* ctrl not found */
1250         pr_warn("could not find controller %d for subsys %s / host %s\n",
1251                 cntlid, subsysnqn, hostnqn);
1252         req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
1253
1254 found:
1255         mutex_unlock(&subsys->lock);
1256         nvmet_subsys_put(subsys);
1257 out:
1258         return ctrl;
1259 }
1260
1261 u16 nvmet_check_ctrl_status(struct nvmet_req *req)
1262 {
1263         if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
1264                 pr_err("got cmd %d while CC.EN == 0 on qid = %d\n",
1265                        req->cmd->common.opcode, req->sq->qid);
1266                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1267         }
1268
1269         if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
1270                 pr_err("got cmd %d while CSTS.RDY == 0 on qid = %d\n",
1271                        req->cmd->common.opcode, req->sq->qid);
1272                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1273         }
1274         return 0;
1275 }
1276
1277 bool nvmet_host_allowed(struct nvmet_subsys *subsys, const char *hostnqn)
1278 {
1279         struct nvmet_host_link *p;
1280
1281         lockdep_assert_held(&nvmet_config_sem);
1282
1283         if (subsys->allow_any_host)
1284                 return true;
1285
1286         if (nvmet_is_disc_subsys(subsys)) /* allow all access to disc subsys */
1287                 return true;
1288
1289         list_for_each_entry(p, &subsys->hosts, entry) {
1290                 if (!strcmp(nvmet_host_name(p->host), hostnqn))
1291                         return true;
1292         }
1293
1294         return false;
1295 }
1296
1297 /*
1298  * Note: ctrl->subsys->lock should be held when calling this function
1299  */
1300 static void nvmet_setup_p2p_ns_map(struct nvmet_ctrl *ctrl,
1301                 struct nvmet_req *req)
1302 {
1303         struct nvmet_ns *ns;
1304         unsigned long idx;
1305
1306         if (!req->p2p_client)
1307                 return;
1308
1309         ctrl->p2p_client = get_device(req->p2p_client);
1310
1311         xa_for_each(&ctrl->subsys->namespaces, idx, ns)
1312                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
1313 }
1314
1315 /*
1316  * Note: ctrl->subsys->lock should be held when calling this function
1317  */
1318 static void nvmet_release_p2p_ns_map(struct nvmet_ctrl *ctrl)
1319 {
1320         struct radix_tree_iter iter;
1321         void __rcu **slot;
1322
1323         radix_tree_for_each_slot(slot, &ctrl->p2p_ns_map, &iter, 0)
1324                 pci_dev_put(radix_tree_deref_slot(slot));
1325
1326         put_device(ctrl->p2p_client);
1327 }
1328
1329 static void nvmet_fatal_error_handler(struct work_struct *work)
1330 {
1331         struct nvmet_ctrl *ctrl =
1332                         container_of(work, struct nvmet_ctrl, fatal_err_work);
1333
1334         pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
1335         ctrl->ops->delete_ctrl(ctrl);
1336 }
1337
1338 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
1339                 struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
1340 {
1341         struct nvmet_subsys *subsys;
1342         struct nvmet_ctrl *ctrl;
1343         int ret;
1344         u16 status;
1345
1346         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1347         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1348         if (!subsys) {
1349                 pr_warn("connect request for invalid subsystem %s!\n",
1350                         subsysnqn);
1351                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1352                 req->error_loc = offsetof(struct nvme_common_command, dptr);
1353                 goto out;
1354         }
1355
1356         down_read(&nvmet_config_sem);
1357         if (!nvmet_host_allowed(subsys, hostnqn)) {
1358                 pr_info("connect by host %s for subsystem %s not allowed\n",
1359                         hostnqn, subsysnqn);
1360                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
1361                 up_read(&nvmet_config_sem);
1362                 status = NVME_SC_CONNECT_INVALID_HOST | NVME_SC_DNR;
1363                 req->error_loc = offsetof(struct nvme_common_command, dptr);
1364                 goto out_put_subsystem;
1365         }
1366         up_read(&nvmet_config_sem);
1367
1368         status = NVME_SC_INTERNAL;
1369         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1370         if (!ctrl)
1371                 goto out_put_subsystem;
1372         mutex_init(&ctrl->lock);
1373
1374         ctrl->port = req->port;
1375         ctrl->ops = req->ops;
1376
1377         INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
1378         INIT_LIST_HEAD(&ctrl->async_events);
1379         INIT_RADIX_TREE(&ctrl->p2p_ns_map, GFP_KERNEL);
1380         INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
1381         INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer);
1382
1383         memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
1384         memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
1385
1386         kref_init(&ctrl->ref);
1387         ctrl->subsys = subsys;
1388         nvmet_init_cap(ctrl);
1389         WRITE_ONCE(ctrl->aen_enabled, NVMET_AEN_CFG_OPTIONAL);
1390
1391         ctrl->changed_ns_list = kmalloc_array(NVME_MAX_CHANGED_NAMESPACES,
1392                         sizeof(__le32), GFP_KERNEL);
1393         if (!ctrl->changed_ns_list)
1394                 goto out_free_ctrl;
1395
1396         ctrl->sqs = kcalloc(subsys->max_qid + 1,
1397                         sizeof(struct nvmet_sq *),
1398                         GFP_KERNEL);
1399         if (!ctrl->sqs)
1400                 goto out_free_changed_ns_list;
1401
1402         if (subsys->cntlid_min > subsys->cntlid_max)
1403                 goto out_free_sqs;
1404
1405         ret = ida_alloc_range(&cntlid_ida,
1406                              subsys->cntlid_min, subsys->cntlid_max,
1407                              GFP_KERNEL);
1408         if (ret < 0) {
1409                 status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
1410                 goto out_free_sqs;
1411         }
1412         ctrl->cntlid = ret;
1413
1414         /*
1415          * Discovery controllers may use some arbitrary high value
1416          * in order to cleanup stale discovery sessions
1417          */
1418         if (nvmet_is_disc_subsys(ctrl->subsys) && !kato)
1419                 kato = NVMET_DISC_KATO_MS;
1420
1421         /* keep-alive timeout in seconds */
1422         ctrl->kato = DIV_ROUND_UP(kato, 1000);
1423
1424         ctrl->err_counter = 0;
1425         spin_lock_init(&ctrl->error_lock);
1426
1427         nvmet_start_keep_alive_timer(ctrl);
1428
1429         mutex_lock(&subsys->lock);
1430         list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
1431         nvmet_setup_p2p_ns_map(ctrl, req);
1432         mutex_unlock(&subsys->lock);
1433
1434         *ctrlp = ctrl;
1435         return 0;
1436
1437 out_free_sqs:
1438         kfree(ctrl->sqs);
1439 out_free_changed_ns_list:
1440         kfree(ctrl->changed_ns_list);
1441 out_free_ctrl:
1442         kfree(ctrl);
1443 out_put_subsystem:
1444         nvmet_subsys_put(subsys);
1445 out:
1446         return status;
1447 }
1448
1449 static void nvmet_ctrl_free(struct kref *ref)
1450 {
1451         struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
1452         struct nvmet_subsys *subsys = ctrl->subsys;
1453
1454         mutex_lock(&subsys->lock);
1455         nvmet_release_p2p_ns_map(ctrl);
1456         list_del(&ctrl->subsys_entry);
1457         mutex_unlock(&subsys->lock);
1458
1459         nvmet_stop_keep_alive_timer(ctrl);
1460
1461         flush_work(&ctrl->async_event_work);
1462         cancel_work_sync(&ctrl->fatal_err_work);
1463
1464         ida_free(&cntlid_ida, ctrl->cntlid);
1465
1466         nvmet_async_events_free(ctrl);
1467         kfree(ctrl->sqs);
1468         kfree(ctrl->changed_ns_list);
1469         kfree(ctrl);
1470
1471         nvmet_subsys_put(subsys);
1472 }
1473
1474 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
1475 {
1476         kref_put(&ctrl->ref, nvmet_ctrl_free);
1477 }
1478
1479 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
1480 {
1481         mutex_lock(&ctrl->lock);
1482         if (!(ctrl->csts & NVME_CSTS_CFS)) {
1483                 ctrl->csts |= NVME_CSTS_CFS;
1484                 queue_work(nvmet_wq, &ctrl->fatal_err_work);
1485         }
1486         mutex_unlock(&ctrl->lock);
1487 }
1488 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
1489
1490 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
1491                 const char *subsysnqn)
1492 {
1493         struct nvmet_subsys_link *p;
1494
1495         if (!port)
1496                 return NULL;
1497
1498         if (!strcmp(NVME_DISC_SUBSYS_NAME, subsysnqn)) {
1499                 if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
1500                         return NULL;
1501                 return nvmet_disc_subsys;
1502         }
1503
1504         down_read(&nvmet_config_sem);
1505         list_for_each_entry(p, &port->subsystems, entry) {
1506                 if (!strncmp(p->subsys->subsysnqn, subsysnqn,
1507                                 NVMF_NQN_SIZE)) {
1508                         if (!kref_get_unless_zero(&p->subsys->ref))
1509                                 break;
1510                         up_read(&nvmet_config_sem);
1511                         return p->subsys;
1512                 }
1513         }
1514         up_read(&nvmet_config_sem);
1515         return NULL;
1516 }
1517
1518 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
1519                 enum nvme_subsys_type type)
1520 {
1521         struct nvmet_subsys *subsys;
1522         char serial[NVMET_SN_MAX_SIZE / 2];
1523         int ret;
1524
1525         subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
1526         if (!subsys)
1527                 return ERR_PTR(-ENOMEM);
1528
1529         subsys->ver = NVMET_DEFAULT_VS;
1530         /* generate a random serial number as our controllers are ephemeral: */
1531         get_random_bytes(&serial, sizeof(serial));
1532         bin2hex(subsys->serial, &serial, sizeof(serial));
1533
1534         subsys->model_number = kstrdup(NVMET_DEFAULT_CTRL_MODEL, GFP_KERNEL);
1535         if (!subsys->model_number) {
1536                 ret = -ENOMEM;
1537                 goto free_subsys;
1538         }
1539
1540         switch (type) {
1541         case NVME_NQN_NVME:
1542                 subsys->max_qid = NVMET_NR_QUEUES;
1543                 break;
1544         case NVME_NQN_DISC:
1545         case NVME_NQN_CURR:
1546                 subsys->max_qid = 0;
1547                 break;
1548         default:
1549                 pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
1550                 ret = -EINVAL;
1551                 goto free_mn;
1552         }
1553         subsys->type = type;
1554         subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
1555                         GFP_KERNEL);
1556         if (!subsys->subsysnqn) {
1557                 ret = -ENOMEM;
1558                 goto free_mn;
1559         }
1560         subsys->cntlid_min = NVME_CNTLID_MIN;
1561         subsys->cntlid_max = NVME_CNTLID_MAX;
1562         kref_init(&subsys->ref);
1563
1564         mutex_init(&subsys->lock);
1565         xa_init(&subsys->namespaces);
1566         INIT_LIST_HEAD(&subsys->ctrls);
1567         INIT_LIST_HEAD(&subsys->hosts);
1568
1569         return subsys;
1570
1571 free_mn:
1572         kfree(subsys->model_number);
1573 free_subsys:
1574         kfree(subsys);
1575         return ERR_PTR(ret);
1576 }
1577
1578 static void nvmet_subsys_free(struct kref *ref)
1579 {
1580         struct nvmet_subsys *subsys =
1581                 container_of(ref, struct nvmet_subsys, ref);
1582
1583         WARN_ON_ONCE(!xa_empty(&subsys->namespaces));
1584
1585         xa_destroy(&subsys->namespaces);
1586         nvmet_passthru_subsys_free(subsys);
1587
1588         kfree(subsys->subsysnqn);
1589         kfree(subsys->model_number);
1590         kfree(subsys);
1591 }
1592
1593 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1594 {
1595         struct nvmet_ctrl *ctrl;
1596
1597         mutex_lock(&subsys->lock);
1598         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1599                 ctrl->ops->delete_ctrl(ctrl);
1600         mutex_unlock(&subsys->lock);
1601 }
1602
1603 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1604 {
1605         kref_put(&subsys->ref, nvmet_subsys_free);
1606 }
1607
1608 static int __init nvmet_init(void)
1609 {
1610         int error;
1611
1612         nvmet_ana_group_enabled[NVMET_DEFAULT_ANA_GRPID] = 1;
1613
1614         zbd_wq = alloc_workqueue("nvmet-zbd-wq", WQ_MEM_RECLAIM, 0);
1615         if (!zbd_wq)
1616                 return -ENOMEM;
1617
1618         buffered_io_wq = alloc_workqueue("nvmet-buffered-io-wq",
1619                         WQ_MEM_RECLAIM, 0);
1620         if (!buffered_io_wq) {
1621                 error = -ENOMEM;
1622                 goto out_free_zbd_work_queue;
1623         }
1624
1625         nvmet_wq = alloc_workqueue("nvmet-wq", WQ_MEM_RECLAIM, 0);
1626         if (!nvmet_wq) {
1627                 error = -ENOMEM;
1628                 goto out_free_buffered_work_queue;
1629         }
1630
1631         error = nvmet_init_discovery();
1632         if (error)
1633                 goto out_free_nvmet_work_queue;
1634
1635         error = nvmet_init_configfs();
1636         if (error)
1637                 goto out_exit_discovery;
1638         return 0;
1639
1640 out_exit_discovery:
1641         nvmet_exit_discovery();
1642 out_free_nvmet_work_queue:
1643         destroy_workqueue(nvmet_wq);
1644 out_free_buffered_work_queue:
1645         destroy_workqueue(buffered_io_wq);
1646 out_free_zbd_work_queue:
1647         destroy_workqueue(zbd_wq);
1648         return error;
1649 }
1650
1651 static void __exit nvmet_exit(void)
1652 {
1653         nvmet_exit_configfs();
1654         nvmet_exit_discovery();
1655         ida_destroy(&cntlid_ida);
1656         destroy_workqueue(nvmet_wq);
1657         destroy_workqueue(buffered_io_wq);
1658         destroy_workqueue(zbd_wq);
1659
1660         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1661         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1662 }
1663
1664 module_init(nvmet_init);
1665 module_exit(nvmet_exit);
1666
1667 MODULE_LICENSE("GPL v2");