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 #include <generated/utsrelease.h>
  14
  15 #define CREATE_TRACE_POINTS
  16 #include "trace.h"
  17
  18 #include "nvmet.h"
  19 #include "debugfs.h"
  20
  21 struct kmem_cache *nvmet_bvec_cache;
  22 struct workqueue_struct *buffered_io_wq;
  23 struct workqueue_struct *zbd_wq;
  24 static const struct nvmet_fabrics_ops *nvmet_transports[NVMF_TRTYPE_MAX];
  25 static DEFINE_IDA(cntlid_ida);
  26
  27 struct workqueue_struct *nvmet_wq;
  28 EXPORT_SYMBOL_GPL(nvmet_wq);
  29
  30 /*
  31  * This read/write semaphore is used to synchronize access to configuration
  32  * information on a target system that will result in discovery log page
  33  * information change for at least one host.
  34  * The full list of resources to protected by this semaphore is:
  35  *
  36  *  - subsystems list
  37  *  - per-subsystem allowed hosts list
  38  *  - allow_any_host subsystem attribute
  39  *  - nvmet_genctr
  40  *  - the nvmet_transports array
  41  *
  42  * When updating any of those lists/structures write lock should be obtained,
  43  * while when reading (popolating discovery log page or checking host-subsystem
  44  * link) read lock is obtained to allow concurrent reads.
  45  */
  46 DECLARE_RWSEM(nvmet_config_sem);
  47
  48 u32 nvmet_ana_group_enabled[NVMET_MAX_ANAGRPS + 1];
  49 u64 nvmet_ana_chgcnt;
  50 DECLARE_RWSEM(nvmet_ana_sem);
  51
  52 inline u16 errno_to_nvme_status(struct nvmet_req *req, int errno)
  53 {
  54         switch (errno) {
  55         case 0:
  56                 return NVME_SC_SUCCESS;
  57         case -ENOSPC:
  58                 req->error_loc = offsetof(struct nvme_rw_command, length);
  59                 return NVME_SC_CAP_EXCEEDED | NVME_STATUS_DNR;
  60         case -EREMOTEIO:
  61                 req->error_loc = offsetof(struct nvme_rw_command, slba);
  62                 return  NVME_SC_LBA_RANGE | NVME_STATUS_DNR;
  63         case -EOPNOTSUPP:
  64                 req->error_loc = offsetof(struct nvme_common_command, opcode);
  65                 switch (req->cmd->common.opcode) {
  66                 case nvme_cmd_dsm:
  67                 case nvme_cmd_write_zeroes:
  68                         return NVME_SC_ONCS_NOT_SUPPORTED | NVME_STATUS_DNR;
  69                 default:
  70                         return NVME_SC_INVALID_OPCODE | NVME_STATUS_DNR;
  71                 }
  72                 break;
  73         case -ENODATA:
  74                 req->error_loc = offsetof(struct nvme_rw_command, nsid);
  75                 return NVME_SC_ACCESS_DENIED;
  76         case -EIO:
  77                 fallthrough;
  78         default:
  79                 req->error_loc = offsetof(struct nvme_common_command, opcode);
  80                 return NVME_SC_INTERNAL | NVME_STATUS_DNR;
  81         }
  82 }
  83
  84 u16 nvmet_report_invalid_opcode(struct nvmet_req *req)
  85 {
  86         pr_debug("unhandled cmd %d on qid %d\n", req->cmd->common.opcode,
  87                  req->sq->qid);
  88
  89         req->error_loc = offsetof(struct nvme_common_command, opcode);
  90         return NVME_SC_INVALID_OPCODE | NVME_STATUS_DNR;
  91 }
  92
  93 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
  94                 const char *subsysnqn);
  95
  96 u16 nvmet_copy_to_sgl(struct nvmet_req *req, off_t off, const void *buf,
  97                 size_t len)
  98 {
  99         if (sg_pcopy_from_buffer(req->sg, req->sg_cnt, buf, len, off) != len) {
 100                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 101                 return NVME_SC_SGL_INVALID_DATA | NVME_STATUS_DNR;
 102         }
 103         return 0;
 104 }
 105
 106 u16 nvmet_copy_from_sgl(struct nvmet_req *req, off_t off, void *buf, size_t len)
 107 {
 108         if (sg_pcopy_to_buffer(req->sg, req->sg_cnt, buf, len, off) != len) {
 109                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 110                 return NVME_SC_SGL_INVALID_DATA | NVME_STATUS_DNR;
 111         }
 112         return 0;
 113 }
 114
 115 u16 nvmet_zero_sgl(struct nvmet_req *req, off_t off, size_t len)
 116 {
 117         if (sg_zero_buffer(req->sg, req->sg_cnt, len, off) != len) {
 118                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 119                 return NVME_SC_SGL_INVALID_DATA | NVME_STATUS_DNR;
 120         }
 121         return 0;
 122 }
 123
 124 static u32 nvmet_max_nsid(struct nvmet_subsys *subsys)
 125 {
 126         struct nvmet_ns *cur;
 127         unsigned long idx;
 128         u32 nsid = 0;
 129
 130         nvmet_for_each_enabled_ns(&subsys->namespaces, idx, cur)
 131                 nsid = cur->nsid;
 132
 133         return nsid;
 134 }
 135
 136 static u32 nvmet_async_event_result(struct nvmet_async_event *aen)
 137 {
 138         return aen->event_type | (aen->event_info << 8) | (aen->log_page << 16);
 139 }
 140
 141 static void nvmet_async_events_failall(struct nvmet_ctrl *ctrl)
 142 {
 143         struct nvmet_req *req;
 144
 145         mutex_lock(&ctrl->lock);
 146         while (ctrl->nr_async_event_cmds) {
 147                 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
 148                 mutex_unlock(&ctrl->lock);
 149                 nvmet_req_complete(req, NVME_SC_INTERNAL | NVME_STATUS_DNR);
 150                 mutex_lock(&ctrl->lock);
 151         }
 152         mutex_unlock(&ctrl->lock);
 153 }
 154
 155 static void nvmet_async_events_process(struct nvmet_ctrl *ctrl)
 156 {
 157         struct nvmet_async_event *aen;
 158         struct nvmet_req *req;
 159
 160         mutex_lock(&ctrl->lock);
 161         while (ctrl->nr_async_event_cmds && !list_empty(&ctrl->async_events)) {
 162                 aen = list_first_entry(&ctrl->async_events,
 163                                        struct nvmet_async_event, entry);
 164                 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
 165                 nvmet_set_result(req, nvmet_async_event_result(aen));
 166
 167                 list_del(&aen->entry);
 168                 kfree(aen);
 169
 170                 mutex_unlock(&ctrl->lock);
 171                 trace_nvmet_async_event(ctrl, req->cqe->result.u32);
 172                 nvmet_req_complete(req, 0);
 173                 mutex_lock(&ctrl->lock);
 174         }
 175         mutex_unlock(&ctrl->lock);
 176 }
 177
 178 static void nvmet_async_events_free(struct nvmet_ctrl *ctrl)
 179 {
 180         struct nvmet_async_event *aen, *tmp;
 181
 182         mutex_lock(&ctrl->lock);
 183         list_for_each_entry_safe(aen, tmp, &ctrl->async_events, entry) {
 184                 list_del(&aen->entry);
 185                 kfree(aen);
 186         }
 187         mutex_unlock(&ctrl->lock);
 188 }
 189
 190 static void nvmet_async_event_work(struct work_struct *work)
 191 {
 192         struct nvmet_ctrl *ctrl =
 193                 container_of(work, struct nvmet_ctrl, async_event_work);
 194
 195         nvmet_async_events_process(ctrl);
 196 }
 197
 198 void nvmet_add_async_event(struct nvmet_ctrl *ctrl, u8 event_type,
 199                 u8 event_info, u8 log_page)
 200 {
 201         struct nvmet_async_event *aen;
 202
 203         aen = kmalloc(sizeof(*aen), GFP_KERNEL);
 204         if (!aen)
 205                 return;
 206
 207         aen->event_type = event_type;
 208         aen->event_info = event_info;
 209         aen->log_page = log_page;
 210
 211         mutex_lock(&ctrl->lock);
 212         list_add_tail(&aen->entry, &ctrl->async_events);
 213         mutex_unlock(&ctrl->lock);
 214
 215         queue_work(nvmet_wq, &ctrl->async_event_work);
 216 }
 217
 218 static void nvmet_add_to_changed_ns_log(struct nvmet_ctrl *ctrl, __le32 nsid)
 219 {
 220         u32 i;
 221
 222         mutex_lock(&ctrl->lock);
 223         if (ctrl->nr_changed_ns > NVME_MAX_CHANGED_NAMESPACES)
 224                 goto out_unlock;
 225
 226         for (i = 0; i < ctrl->nr_changed_ns; i++) {
 227                 if (ctrl->changed_ns_list[i] == nsid)
 228                         goto out_unlock;
 229         }
 230
 231         if (ctrl->nr_changed_ns == NVME_MAX_CHANGED_NAMESPACES) {
 232                 ctrl->changed_ns_list[0] = cpu_to_le32(0xffffffff);
 233                 ctrl->nr_changed_ns = U32_MAX;
 234                 goto out_unlock;
 235         }
 236
 237         ctrl->changed_ns_list[ctrl->nr_changed_ns++] = nsid;
 238 out_unlock:
 239         mutex_unlock(&ctrl->lock);
 240 }
 241
 242 void nvmet_ns_changed(struct nvmet_subsys *subsys, u32 nsid)
 243 {
 244         struct nvmet_ctrl *ctrl;
 245
 246         lockdep_assert_held(&subsys->lock);
 247
 248         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 249                 nvmet_add_to_changed_ns_log(ctrl, cpu_to_le32(nsid));
 250                 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_NS_ATTR))
 251                         continue;
 252                 nvmet_add_async_event(ctrl, NVME_AER_NOTICE,
 253                                 NVME_AER_NOTICE_NS_CHANGED,
 254                                 NVME_LOG_CHANGED_NS);
 255         }
 256 }
 257
 258 void nvmet_send_ana_event(struct nvmet_subsys *subsys,
 259                 struct nvmet_port *port)
 260 {
 261         struct nvmet_ctrl *ctrl;
 262
 263         mutex_lock(&subsys->lock);
 264         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 265                 if (port && ctrl->port != port)
 266                         continue;
 267                 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_ANA_CHANGE))
 268                         continue;
 269                 nvmet_add_async_event(ctrl, NVME_AER_NOTICE,
 270                                 NVME_AER_NOTICE_ANA, NVME_LOG_ANA);
 271         }
 272         mutex_unlock(&subsys->lock);
 273 }
 274
 275 void nvmet_port_send_ana_event(struct nvmet_port *port)
 276 {
 277         struct nvmet_subsys_link *p;
 278
 279         down_read(&nvmet_config_sem);
 280         list_for_each_entry(p, &port->subsystems, entry)
 281                 nvmet_send_ana_event(p->subsys, port);
 282         up_read(&nvmet_config_sem);
 283 }
 284
 285 int nvmet_register_transport(const struct nvmet_fabrics_ops *ops)
 286 {
 287         int ret = 0;
 288
 289         down_write(&nvmet_config_sem);
 290         if (nvmet_transports[ops->type])
 291                 ret = -EINVAL;
 292         else
 293                 nvmet_transports[ops->type] = ops;
 294         up_write(&nvmet_config_sem);
 295
 296         return ret;
 297 }
 298 EXPORT_SYMBOL_GPL(nvmet_register_transport);
 299
 300 void nvmet_unregister_transport(const struct nvmet_fabrics_ops *ops)
 301 {
 302         down_write(&nvmet_config_sem);
 303         nvmet_transports[ops->type] = NULL;
 304         up_write(&nvmet_config_sem);
 305 }
 306 EXPORT_SYMBOL_GPL(nvmet_unregister_transport);
 307
 308 void nvmet_port_del_ctrls(struct nvmet_port *port, struct nvmet_subsys *subsys)
 309 {
 310         struct nvmet_ctrl *ctrl;
 311
 312         mutex_lock(&subsys->lock);
 313         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 314                 if (ctrl->port == port)
 315                         ctrl->ops->delete_ctrl(ctrl);
 316         }
 317         mutex_unlock(&subsys->lock);
 318 }
 319
 320 int nvmet_enable_port(struct nvmet_port *port)
 321 {
 322         const struct nvmet_fabrics_ops *ops;
 323         int ret;
 324
 325         lockdep_assert_held(&nvmet_config_sem);
 326
 327         ops = nvmet_transports[port->disc_addr.trtype];
 328         if (!ops) {
 329                 up_write(&nvmet_config_sem);
 330                 request_module("nvmet-transport-%d", port->disc_addr.trtype);
 331                 down_write(&nvmet_config_sem);
 332                 ops = nvmet_transports[port->disc_addr.trtype];
 333                 if (!ops) {
 334                         pr_err("transport type %d not supported\n",
 335                                 port->disc_addr.trtype);
 336                         return -EINVAL;
 337                 }
 338         }
 339
 340         if (!try_module_get(ops->owner))
 341                 return -EINVAL;
 342
 343         /*
 344          * If the user requested PI support and the transport isn't pi capable,
 345          * don't enable the port.
 346          */
 347         if (port->pi_enable && !(ops->flags & NVMF_METADATA_SUPPORTED)) {
 348                 pr_err("T10-PI is not supported by transport type %d\n",
 349                        port->disc_addr.trtype);
 350                 ret = -EINVAL;
 351                 goto out_put;
 352         }
 353
 354         ret = ops->add_port(port);
 355         if (ret)
 356                 goto out_put;
 357
 358         /* If the transport didn't set inline_data_size, then disable it. */
 359         if (port->inline_data_size < 0)
 360                 port->inline_data_size = 0;
 361
 362         /*
 363          * If the transport didn't set the max_queue_size properly, then clamp
 364          * it to the target limits. Also set default values in case the
 365          * transport didn't set it at all.
 366          */
 367         if (port->max_queue_size < 0)
 368                 port->max_queue_size = NVMET_MAX_QUEUE_SIZE;
 369         else
 370                 port->max_queue_size = clamp_t(int, port->max_queue_size,
 371                                                NVMET_MIN_QUEUE_SIZE,
 372                                                NVMET_MAX_QUEUE_SIZE);
 373
 374         port->enabled = true;
 375         port->tr_ops = ops;
 376         return 0;
 377
 378 out_put:
 379         module_put(ops->owner);
 380         return ret;
 381 }
 382
 383 void nvmet_disable_port(struct nvmet_port *port)
 384 {
 385         const struct nvmet_fabrics_ops *ops;
 386
 387         lockdep_assert_held(&nvmet_config_sem);
 388
 389         port->enabled = false;
 390         port->tr_ops = NULL;
 391
 392         ops = nvmet_transports[port->disc_addr.trtype];
 393         ops->remove_port(port);
 394         module_put(ops->owner);
 395 }
 396
 397 static void nvmet_keep_alive_timer(struct work_struct *work)
 398 {
 399         struct nvmet_ctrl *ctrl = container_of(to_delayed_work(work),
 400                         struct nvmet_ctrl, ka_work);
 401         bool reset_tbkas = ctrl->reset_tbkas;
 402
 403         ctrl->reset_tbkas = false;
 404         if (reset_tbkas) {
 405                 pr_debug("ctrl %d reschedule traffic based keep-alive timer\n",
 406                         ctrl->cntlid);
 407                 queue_delayed_work(nvmet_wq, &ctrl->ka_work, ctrl->kato * HZ);
 408                 return;
 409         }
 410
 411         pr_err("ctrl %d keep-alive timer (%d seconds) expired!\n",
 412                 ctrl->cntlid, ctrl->kato);
 413
 414         nvmet_ctrl_fatal_error(ctrl);
 415 }
 416
 417 void nvmet_start_keep_alive_timer(struct nvmet_ctrl *ctrl)
 418 {
 419         if (unlikely(ctrl->kato == 0))
 420                 return;
 421
 422         pr_debug("ctrl %d start keep-alive timer for %d secs\n",
 423                 ctrl->cntlid, ctrl->kato);
 424
 425         queue_delayed_work(nvmet_wq, &ctrl->ka_work, ctrl->kato * HZ);
 426 }
 427
 428 void nvmet_stop_keep_alive_timer(struct nvmet_ctrl *ctrl)
 429 {
 430         if (unlikely(ctrl->kato == 0))
 431                 return;
 432
 433         pr_debug("ctrl %d stop keep-alive\n", ctrl->cntlid);
 434
 435         cancel_delayed_work_sync(&ctrl->ka_work);
 436 }
 437
 438 u16 nvmet_req_find_ns(struct nvmet_req *req)
 439 {
 440         u32 nsid = le32_to_cpu(req->cmd->common.nsid);
 441         struct nvmet_subsys *subsys = nvmet_req_subsys(req);
 442
 443         req->ns = xa_load(&subsys->namespaces, nsid);
 444         if (unlikely(!req->ns || !req->ns->enabled)) {
 445                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 446                 if (!req->ns) /* ns doesn't exist! */
 447                         return NVME_SC_INVALID_NS | NVME_STATUS_DNR;
 448
 449                 /* ns exists but it's disabled */
 450                 req->ns = NULL;
 451                 return NVME_SC_INTERNAL_PATH_ERROR;
 452         }
 453
 454         percpu_ref_get(&req->ns->ref);
 455         return NVME_SC_SUCCESS;
 456 }
 457
 458 static void nvmet_destroy_namespace(struct percpu_ref *ref)
 459 {
 460         struct nvmet_ns *ns = container_of(ref, struct nvmet_ns, ref);
 461
 462         complete(&ns->disable_done);
 463 }
 464
 465 void nvmet_put_namespace(struct nvmet_ns *ns)
 466 {
 467         percpu_ref_put(&ns->ref);
 468 }
 469
 470 static void nvmet_ns_dev_disable(struct nvmet_ns *ns)
 471 {
 472         nvmet_bdev_ns_disable(ns);
 473         nvmet_file_ns_disable(ns);
 474 }
 475
 476 static int nvmet_p2pmem_ns_enable(struct nvmet_ns *ns)
 477 {
 478         int ret;
 479         struct pci_dev *p2p_dev;
 480
 481         if (!ns->use_p2pmem)
 482                 return 0;
 483
 484         if (!ns->bdev) {
 485                 pr_err("peer-to-peer DMA is not supported by non-block device namespaces\n");
 486                 return -EINVAL;
 487         }
 488
 489         if (!blk_queue_pci_p2pdma(ns->bdev->bd_disk->queue)) {
 490                 pr_err("peer-to-peer DMA is not supported by the driver of %s\n",
 491                        ns->device_path);
 492                 return -EINVAL;
 493         }
 494
 495         if (ns->p2p_dev) {
 496                 ret = pci_p2pdma_distance(ns->p2p_dev, nvmet_ns_dev(ns), true);
 497                 if (ret < 0)
 498                         return -EINVAL;
 499         } else {
 500                 /*
 501                  * Right now we just check that there is p2pmem available so
 502                  * we can report an error to the user right away if there
 503                  * is not. We'll find the actual device to use once we
 504                  * setup the controller when the port's device is available.
 505                  */
 506
 507                 p2p_dev = pci_p2pmem_find(nvmet_ns_dev(ns));
 508                 if (!p2p_dev) {
 509                         pr_err("no peer-to-peer memory is available for %s\n",
 510                                ns->device_path);
 511                         return -EINVAL;
 512                 }
 513
 514                 pci_dev_put(p2p_dev);
 515         }
 516
 517         return 0;
 518 }
 519
 520 /*
 521  * Note: ctrl->subsys->lock should be held when calling this function
 522  */
 523 static void nvmet_p2pmem_ns_add_p2p(struct nvmet_ctrl *ctrl,
 524                                     struct nvmet_ns *ns)
 525 {
 526         struct device *clients[2];
 527         struct pci_dev *p2p_dev;
 528         int ret;
 529
 530         if (!ctrl->p2p_client || !ns->use_p2pmem)
 531                 return;
 532
 533         if (ns->p2p_dev) {
 534                 ret = pci_p2pdma_distance(ns->p2p_dev, ctrl->p2p_client, true);
 535                 if (ret < 0)
 536                         return;
 537
 538                 p2p_dev = pci_dev_get(ns->p2p_dev);
 539         } else {
 540                 clients[0] = ctrl->p2p_client;
 541                 clients[1] = nvmet_ns_dev(ns);
 542
 543                 p2p_dev = pci_p2pmem_find_many(clients, ARRAY_SIZE(clients));
 544                 if (!p2p_dev) {
 545                         pr_err("no peer-to-peer memory is available that's supported by %s and %s\n",
 546                                dev_name(ctrl->p2p_client), ns->device_path);
 547                         return;
 548                 }
 549         }
 550
 551         ret = radix_tree_insert(&ctrl->p2p_ns_map, ns->nsid, p2p_dev);
 552         if (ret < 0)
 553                 pci_dev_put(p2p_dev);
 554
 555         pr_info("using p2pmem on %s for nsid %d\n", pci_name(p2p_dev),
 556                 ns->nsid);
 557 }
 558
 559 bool nvmet_ns_revalidate(struct nvmet_ns *ns)
 560 {
 561         loff_t oldsize = ns->size;
 562
 563         if (ns->bdev)
 564                 nvmet_bdev_ns_revalidate(ns);
 565         else
 566                 nvmet_file_ns_revalidate(ns);
 567
 568         return oldsize != ns->size;
 569 }
 570
 571 int nvmet_ns_enable(struct nvmet_ns *ns)
 572 {
 573         struct nvmet_subsys *subsys = ns->subsys;
 574         struct nvmet_ctrl *ctrl;
 575         int ret;
 576
 577         mutex_lock(&subsys->lock);
 578         ret = 0;
 579
 580         if (nvmet_is_passthru_subsys(subsys)) {
 581                 pr_info("cannot enable both passthru and regular namespaces for a single subsystem");
 582                 goto out_unlock;
 583         }
 584
 585         if (ns->enabled)
 586                 goto out_unlock;
 587
 588         ret = -EMFILE;
 589
 590         ret = nvmet_bdev_ns_enable(ns);
 591         if (ret == -ENOTBLK)
 592                 ret = nvmet_file_ns_enable(ns);
 593         if (ret)
 594                 goto out_unlock;
 595
 596         ret = nvmet_p2pmem_ns_enable(ns);
 597         if (ret)
 598                 goto out_dev_disable;
 599
 600         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 601                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
 602
 603         if (ns->pr.enable) {
 604                 ret = nvmet_pr_init_ns(ns);
 605                 if (ret)
 606                         goto out_dev_put;
 607         }
 608
 609         nvmet_ns_changed(subsys, ns->nsid);
 610         ns->enabled = true;
 611         xa_set_mark(&subsys->namespaces, ns->nsid, NVMET_NS_ENABLED);
 612         ret = 0;
 613 out_unlock:
 614         mutex_unlock(&subsys->lock);
 615         return ret;
 616 out_dev_put:
 617         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 618                 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
 619 out_dev_disable:
 620         nvmet_ns_dev_disable(ns);
 621         goto out_unlock;
 622 }
 623
 624 void nvmet_ns_disable(struct nvmet_ns *ns)
 625 {
 626         struct nvmet_subsys *subsys = ns->subsys;
 627         struct nvmet_ctrl *ctrl;
 628
 629         mutex_lock(&subsys->lock);
 630         if (!ns->enabled)
 631                 goto out_unlock;
 632
 633         ns->enabled = false;
 634         xa_clear_mark(&subsys->namespaces, ns->nsid, NVMET_NS_ENABLED);
 635
 636         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 637                 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
 638
 639         mutex_unlock(&subsys->lock);
 640
 641         if (ns->pr.enable)
 642                 nvmet_pr_exit_ns(ns);
 643
 644         mutex_lock(&subsys->lock);
 645         nvmet_ns_changed(subsys, ns->nsid);
 646         nvmet_ns_dev_disable(ns);
 647 out_unlock:
 648         mutex_unlock(&subsys->lock);
 649 }
 650
 651 void nvmet_ns_free(struct nvmet_ns *ns)
 652 {
 653         struct nvmet_subsys *subsys = ns->subsys;
 654
 655         nvmet_ns_disable(ns);
 656
 657         mutex_lock(&subsys->lock);
 658
 659         xa_erase(&subsys->namespaces, ns->nsid);
 660         if (ns->nsid == subsys->max_nsid)
 661                 subsys->max_nsid = nvmet_max_nsid(subsys);
 662
 663         mutex_unlock(&subsys->lock);
 664
 665         /*
 666          * Now that we removed the namespaces from the lookup list, we
 667          * can kill the per_cpu ref and wait for any remaining references
 668          * to be dropped, as well as a RCU grace period for anyone only
 669          * using the namepace under rcu_read_lock().  Note that we can't
 670          * use call_rcu here as we need to ensure the namespaces have
 671          * been fully destroyed before unloading the module.
 672          */
 673         percpu_ref_kill(&ns->ref);
 674         synchronize_rcu();
 675         wait_for_completion(&ns->disable_done);
 676         percpu_ref_exit(&ns->ref);
 677
 678         mutex_lock(&subsys->lock);
 679         subsys->nr_namespaces--;
 680         mutex_unlock(&subsys->lock);
 681
 682         down_write(&nvmet_ana_sem);
 683         nvmet_ana_group_enabled[ns->anagrpid]--;
 684         up_write(&nvmet_ana_sem);
 685
 686         kfree(ns->device_path);
 687         kfree(ns);
 688 }
 689
 690 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
 691 {
 692         struct nvmet_ns *ns;
 693
 694         mutex_lock(&subsys->lock);
 695
 696         if (subsys->nr_namespaces == NVMET_MAX_NAMESPACES)
 697                 goto out_unlock;
 698
 699         ns = kzalloc(sizeof(*ns), GFP_KERNEL);
 700         if (!ns)
 701                 goto out_unlock;
 702
 703         init_completion(&ns->disable_done);
 704
 705         ns->nsid = nsid;
 706         ns->subsys = subsys;
 707
 708         if (percpu_ref_init(&ns->ref, nvmet_destroy_namespace, 0, GFP_KERNEL))
 709                 goto out_free;
 710
 711         if (ns->nsid > subsys->max_nsid)
 712                 subsys->max_nsid = nsid;
 713
 714         if (xa_insert(&subsys->namespaces, ns->nsid, ns, GFP_KERNEL))
 715                 goto out_exit;
 716
 717         subsys->nr_namespaces++;
 718
 719         mutex_unlock(&subsys->lock);
 720
 721         down_write(&nvmet_ana_sem);
 722         ns->anagrpid = NVMET_DEFAULT_ANA_GRPID;
 723         nvmet_ana_group_enabled[ns->anagrpid]++;
 724         up_write(&nvmet_ana_sem);
 725
 726         uuid_gen(&ns->uuid);
 727         ns->buffered_io = false;
 728         ns->csi = NVME_CSI_NVM;
 729
 730         return ns;
 731 out_exit:
 732         subsys->max_nsid = nvmet_max_nsid(subsys);
 733         percpu_ref_exit(&ns->ref);
 734 out_free:
 735         kfree(ns);
 736 out_unlock:
 737         mutex_unlock(&subsys->lock);
 738         return NULL;
 739 }
 740
 741 static void nvmet_update_sq_head(struct nvmet_req *req)
 742 {
 743         if (req->sq->size) {
 744                 u32 old_sqhd, new_sqhd;
 745
 746                 old_sqhd = READ_ONCE(req->sq->sqhd);
 747                 do {
 748                         new_sqhd = (old_sqhd + 1) % req->sq->size;
 749                 } while (!try_cmpxchg(&req->sq->sqhd, &old_sqhd, new_sqhd));
 750         }
 751         req->cqe->sq_head = cpu_to_le16(req->sq->sqhd & 0x0000FFFF);
 752 }
 753
 754 static void nvmet_set_error(struct nvmet_req *req, u16 status)
 755 {
 756         struct nvmet_ctrl *ctrl = req->sq->ctrl;
 757         struct nvme_error_slot *new_error_slot;
 758         unsigned long flags;
 759
 760         req->cqe->status = cpu_to_le16(status << 1);
 761
 762         if (!ctrl || req->error_loc == NVMET_NO_ERROR_LOC)
 763                 return;
 764
 765         spin_lock_irqsave(&ctrl->error_lock, flags);
 766         ctrl->err_counter++;
 767         new_error_slot =
 768                 &ctrl->slots[ctrl->err_counter % NVMET_ERROR_LOG_SLOTS];
 769
 770         new_error_slot->error_count = cpu_to_le64(ctrl->err_counter);
 771         new_error_slot->sqid = cpu_to_le16(req->sq->qid);
 772         new_error_slot->cmdid = cpu_to_le16(req->cmd->common.command_id);
 773         new_error_slot->status_field = cpu_to_le16(status << 1);
 774         new_error_slot->param_error_location = cpu_to_le16(req->error_loc);
 775         new_error_slot->lba = cpu_to_le64(req->error_slba);
 776         new_error_slot->nsid = req->cmd->common.nsid;
 777         spin_unlock_irqrestore(&ctrl->error_lock, flags);
 778
 779         /* set the more bit for this request */
 780         req->cqe->status |= cpu_to_le16(1 << 14);
 781 }
 782
 783 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
 784 {
 785         struct nvmet_ns *ns = req->ns;
 786         struct nvmet_pr_per_ctrl_ref *pc_ref = req->pc_ref;
 787
 788         if (!req->sq->sqhd_disabled)
 789                 nvmet_update_sq_head(req);
 790         req->cqe->sq_id = cpu_to_le16(req->sq->qid);
 791         req->cqe->command_id = req->cmd->common.command_id;
 792
 793         if (unlikely(status))
 794                 nvmet_set_error(req, status);
 795
 796         trace_nvmet_req_complete(req);
 797
 798         req->ops->queue_response(req);
 799
 800         if (pc_ref)
 801                 nvmet_pr_put_ns_pc_ref(pc_ref);
 802         if (ns)
 803                 nvmet_put_namespace(ns);
 804 }
 805
 806 void nvmet_req_complete(struct nvmet_req *req, u16 status)
 807 {
 808         struct nvmet_sq *sq = req->sq;
 809
 810         __nvmet_req_complete(req, status);
 811         percpu_ref_put(&sq->ref);
 812 }
 813 EXPORT_SYMBOL_GPL(nvmet_req_complete);
 814
 815 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
 816                 u16 qid, u16 size)
 817 {
 818         cq->qid = qid;
 819         cq->size = size;
 820 }
 821
 822 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
 823                 u16 qid, u16 size)
 824 {
 825         sq->sqhd = 0;
 826         sq->qid = qid;
 827         sq->size = size;
 828
 829         ctrl->sqs[qid] = sq;
 830 }
 831
 832 static void nvmet_confirm_sq(struct percpu_ref *ref)
 833 {
 834         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 835
 836         complete(&sq->confirm_done);
 837 }
 838
 839 void nvmet_sq_destroy(struct nvmet_sq *sq)
 840 {
 841         struct nvmet_ctrl *ctrl = sq->ctrl;
 842
 843         /*
 844          * If this is the admin queue, complete all AERs so that our
 845          * queue doesn't have outstanding requests on it.
 846          */
 847         if (ctrl && ctrl->sqs && ctrl->sqs[0] == sq)
 848                 nvmet_async_events_failall(ctrl);
 849         percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
 850         wait_for_completion(&sq->confirm_done);
 851         wait_for_completion(&sq->free_done);
 852         percpu_ref_exit(&sq->ref);
 853         nvmet_auth_sq_free(sq);
 854
 855         /*
 856          * we must reference the ctrl again after waiting for inflight IO
 857          * to complete. Because admin connect may have sneaked in after we
 858          * store sq->ctrl locally, but before we killed the percpu_ref. the
 859          * admin connect allocates and assigns sq->ctrl, which now needs a
 860          * final ref put, as this ctrl is going away.
 861          */
 862         ctrl = sq->ctrl;
 863
 864         if (ctrl) {
 865                 /*
 866                  * The teardown flow may take some time, and the host may not
 867                  * send us keep-alive during this period, hence reset the
 868                  * traffic based keep-alive timer so we don't trigger a
 869                  * controller teardown as a result of a keep-alive expiration.
 870                  */
 871                 ctrl->reset_tbkas = true;
 872                 sq->ctrl->sqs[sq->qid] = NULL;
 873                 nvmet_ctrl_put(ctrl);
 874                 sq->ctrl = NULL; /* allows reusing the queue later */
 875         }
 876 }
 877 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
 878
 879 static void nvmet_sq_free(struct percpu_ref *ref)
 880 {
 881         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 882
 883         complete(&sq->free_done);
 884 }
 885
 886 int nvmet_sq_init(struct nvmet_sq *sq)
 887 {
 888         int ret;
 889
 890         ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
 891         if (ret) {
 892                 pr_err("percpu_ref init failed!\n");
 893                 return ret;
 894         }
 895         init_completion(&sq->free_done);
 896         init_completion(&sq->confirm_done);
 897         nvmet_auth_sq_init(sq);
 898
 899         return 0;
 900 }
 901 EXPORT_SYMBOL_GPL(nvmet_sq_init);
 902
 903 static inline u16 nvmet_check_ana_state(struct nvmet_port *port,
 904                 struct nvmet_ns *ns)
 905 {
 906         enum nvme_ana_state state = port->ana_state[ns->anagrpid];
 907
 908         if (unlikely(state == NVME_ANA_INACCESSIBLE))
 909                 return NVME_SC_ANA_INACCESSIBLE;
 910         if (unlikely(state == NVME_ANA_PERSISTENT_LOSS))
 911                 return NVME_SC_ANA_PERSISTENT_LOSS;
 912         if (unlikely(state == NVME_ANA_CHANGE))
 913                 return NVME_SC_ANA_TRANSITION;
 914         return 0;
 915 }
 916
 917 static inline u16 nvmet_io_cmd_check_access(struct nvmet_req *req)
 918 {
 919         if (unlikely(req->ns->readonly)) {
 920                 switch (req->cmd->common.opcode) {
 921                 case nvme_cmd_read:
 922                 case nvme_cmd_flush:
 923                         break;
 924                 default:
 925                         return NVME_SC_NS_WRITE_PROTECTED;
 926                 }
 927         }
 928
 929         return 0;
 930 }
 931
 932 static u16 nvmet_parse_io_cmd(struct nvmet_req *req)
 933 {
 934         struct nvme_command *cmd = req->cmd;
 935         u16 ret;
 936
 937         if (nvme_is_fabrics(cmd))
 938                 return nvmet_parse_fabrics_io_cmd(req);
 939
 940         if (unlikely(!nvmet_check_auth_status(req)))
 941                 return NVME_SC_AUTH_REQUIRED | NVME_STATUS_DNR;
 942
 943         ret = nvmet_check_ctrl_status(req);
 944         if (unlikely(ret))
 945                 return ret;
 946
 947         if (nvmet_is_passthru_req(req))
 948                 return nvmet_parse_passthru_io_cmd(req);
 949
 950         ret = nvmet_req_find_ns(req);
 951         if (unlikely(ret))
 952                 return ret;
 953
 954         ret = nvmet_check_ana_state(req->port, req->ns);
 955         if (unlikely(ret)) {
 956                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 957                 return ret;
 958         }
 959         ret = nvmet_io_cmd_check_access(req);
 960         if (unlikely(ret)) {
 961                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 962                 return ret;
 963         }
 964
 965         if (req->ns->pr.enable) {
 966                 ret = nvmet_parse_pr_cmd(req);
 967                 if (!ret)
 968                         return ret;
 969         }
 970
 971         switch (req->ns->csi) {
 972         case NVME_CSI_NVM:
 973                 if (req->ns->file)
 974                         ret = nvmet_file_parse_io_cmd(req);
 975                 else
 976                         ret = nvmet_bdev_parse_io_cmd(req);
 977                 break;
 978         case NVME_CSI_ZNS:
 979                 if (IS_ENABLED(CONFIG_BLK_DEV_ZONED))
 980                         ret = nvmet_bdev_zns_parse_io_cmd(req);
 981                 else
 982                         ret = NVME_SC_INVALID_IO_CMD_SET;
 983                 break;
 984         default:
 985                 ret = NVME_SC_INVALID_IO_CMD_SET;
 986         }
 987         if (ret)
 988                 return ret;
 989
 990         if (req->ns->pr.enable) {
 991                 ret = nvmet_pr_check_cmd_access(req);
 992                 if (ret)
 993                         return ret;
 994
 995                 ret = nvmet_pr_get_ns_pc_ref(req);
 996         }
 997         return ret;
 998 }
 999
1000 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
1001                 struct nvmet_sq *sq, const struct nvmet_fabrics_ops *ops)
1002 {
1003         u8 flags = req->cmd->common.flags;
1004         u16 status;
1005
1006         req->cq = cq;
1007         req->sq = sq;
1008         req->ops = ops;
1009         req->sg = NULL;
1010         req->metadata_sg = NULL;
1011         req->sg_cnt = 0;
1012         req->metadata_sg_cnt = 0;
1013         req->transfer_len = 0;
1014         req->metadata_len = 0;
1015         req->cqe->result.u64 = 0;
1016         req->cqe->status = 0;
1017         req->cqe->sq_head = 0;
1018         req->ns = NULL;
1019         req->error_loc = NVMET_NO_ERROR_LOC;
1020         req->error_slba = 0;
1021         req->pc_ref = NULL;
1022
1023         /* no support for fused commands yet */
1024         if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
1025                 req->error_loc = offsetof(struct nvme_common_command, flags);
1026                 status = NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
1027                 goto fail;
1028         }
1029
1030         /*
1031          * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
1032          * contains an address of a single contiguous physical buffer that is
1033          * byte aligned.
1034          */
1035         if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
1036                 req->error_loc = offsetof(struct nvme_common_command, flags);
1037                 status = NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
1038                 goto fail;
1039         }
1040
1041         if (unlikely(!req->sq->ctrl))
1042                 /* will return an error for any non-connect command: */
1043                 status = nvmet_parse_connect_cmd(req);
1044         else if (likely(req->sq->qid != 0))
1045                 status = nvmet_parse_io_cmd(req);
1046         else
1047                 status = nvmet_parse_admin_cmd(req);
1048
1049         if (status)
1050                 goto fail;
1051
1052         trace_nvmet_req_init(req, req->cmd);
1053
1054         if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
1055                 status = NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
1056                 goto fail;
1057         }
1058
1059         if (sq->ctrl)
1060                 sq->ctrl->reset_tbkas = true;
1061
1062         return true;
1063
1064 fail:
1065         __nvmet_req_complete(req, status);
1066         return false;
1067 }
1068 EXPORT_SYMBOL_GPL(nvmet_req_init);
1069
1070 void nvmet_req_uninit(struct nvmet_req *req)
1071 {
1072         percpu_ref_put(&req->sq->ref);
1073         if (req->pc_ref)
1074                 nvmet_pr_put_ns_pc_ref(req->pc_ref);
1075         if (req->ns)
1076                 nvmet_put_namespace(req->ns);
1077 }
1078 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
1079
1080 bool nvmet_check_transfer_len(struct nvmet_req *req, size_t len)
1081 {
1082         if (unlikely(len != req->transfer_len)) {
1083                 req->error_loc = offsetof(struct nvme_common_command, dptr);
1084                 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_STATUS_DNR);
1085                 return false;
1086         }
1087
1088         return true;
1089 }
1090 EXPORT_SYMBOL_GPL(nvmet_check_transfer_len);
1091
1092 bool nvmet_check_data_len_lte(struct nvmet_req *req, size_t data_len)
1093 {
1094         if (unlikely(data_len > req->transfer_len)) {
1095                 req->error_loc = offsetof(struct nvme_common_command, dptr);
1096                 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_STATUS_DNR);
1097                 return false;
1098         }
1099
1100         return true;
1101 }
1102
1103 static unsigned int nvmet_data_transfer_len(struct nvmet_req *req)
1104 {
1105         return req->transfer_len - req->metadata_len;
1106 }
1107
1108 static int nvmet_req_alloc_p2pmem_sgls(struct pci_dev *p2p_dev,
1109                 struct nvmet_req *req)
1110 {
1111         req->sg = pci_p2pmem_alloc_sgl(p2p_dev, &req->sg_cnt,
1112                         nvmet_data_transfer_len(req));
1113         if (!req->sg)
1114                 goto out_err;
1115
1116         if (req->metadata_len) {
1117                 req->metadata_sg = pci_p2pmem_alloc_sgl(p2p_dev,
1118                                 &req->metadata_sg_cnt, req->metadata_len);
1119                 if (!req->metadata_sg)
1120                         goto out_free_sg;
1121         }
1122
1123         req->p2p_dev = p2p_dev;
1124
1125         return 0;
1126 out_free_sg:
1127         pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1128 out_err:
1129         return -ENOMEM;
1130 }
1131
1132 static struct pci_dev *nvmet_req_find_p2p_dev(struct nvmet_req *req)
1133 {
1134         if (!IS_ENABLED(CONFIG_PCI_P2PDMA) ||
1135             !req->sq->ctrl || !req->sq->qid || !req->ns)
1136                 return NULL;
1137         return radix_tree_lookup(&req->sq->ctrl->p2p_ns_map, req->ns->nsid);
1138 }
1139
1140 int nvmet_req_alloc_sgls(struct nvmet_req *req)
1141 {
1142         struct pci_dev *p2p_dev = nvmet_req_find_p2p_dev(req);
1143
1144         if (p2p_dev && !nvmet_req_alloc_p2pmem_sgls(p2p_dev, req))
1145                 return 0;
1146
1147         req->sg = sgl_alloc(nvmet_data_transfer_len(req), GFP_KERNEL,
1148                             &req->sg_cnt);
1149         if (unlikely(!req->sg))
1150                 goto out;
1151
1152         if (req->metadata_len) {
1153                 req->metadata_sg = sgl_alloc(req->metadata_len, GFP_KERNEL,
1154                                              &req->metadata_sg_cnt);
1155                 if (unlikely(!req->metadata_sg))
1156                         goto out_free;
1157         }
1158
1159         return 0;
1160 out_free:
1161         sgl_free(req->sg);
1162 out:
1163         return -ENOMEM;
1164 }
1165 EXPORT_SYMBOL_GPL(nvmet_req_alloc_sgls);
1166
1167 void nvmet_req_free_sgls(struct nvmet_req *req)
1168 {
1169         if (req->p2p_dev) {
1170                 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1171                 if (req->metadata_sg)
1172                         pci_p2pmem_free_sgl(req->p2p_dev, req->metadata_sg);
1173                 req->p2p_dev = NULL;
1174         } else {
1175                 sgl_free(req->sg);
1176                 if (req->metadata_sg)
1177                         sgl_free(req->metadata_sg);
1178         }
1179
1180         req->sg = NULL;
1181         req->metadata_sg = NULL;
1182         req->sg_cnt = 0;
1183         req->metadata_sg_cnt = 0;
1184 }
1185 EXPORT_SYMBOL_GPL(nvmet_req_free_sgls);
1186
1187 static inline bool nvmet_cc_en(u32 cc)
1188 {
1189         return (cc >> NVME_CC_EN_SHIFT) & 0x1;
1190 }
1191
1192 static inline u8 nvmet_cc_css(u32 cc)
1193 {
1194         return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
1195 }
1196
1197 static inline u8 nvmet_cc_mps(u32 cc)
1198 {
1199         return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
1200 }
1201
1202 static inline u8 nvmet_cc_ams(u32 cc)
1203 {
1204         return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
1205 }
1206
1207 static inline u8 nvmet_cc_shn(u32 cc)
1208 {
1209         return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
1210 }
1211
1212 static inline u8 nvmet_cc_iosqes(u32 cc)
1213 {
1214         return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
1215 }
1216
1217 static inline u8 nvmet_cc_iocqes(u32 cc)
1218 {
1219         return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
1220 }
1221
1222 static inline bool nvmet_css_supported(u8 cc_css)
1223 {
1224         switch (cc_css << NVME_CC_CSS_SHIFT) {
1225         case NVME_CC_CSS_NVM:
1226         case NVME_CC_CSS_CSI:
1227                 return true;
1228         default:
1229                 return false;
1230         }
1231 }
1232
1233 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
1234 {
1235         lockdep_assert_held(&ctrl->lock);
1236
1237         /*
1238          * Only I/O controllers should verify iosqes,iocqes.
1239          * Strictly speaking, the spec says a discovery controller
1240          * should verify iosqes,iocqes are zeroed, however that
1241          * would break backwards compatibility, so don't enforce it.
1242          */
1243         if (!nvmet_is_disc_subsys(ctrl->subsys) &&
1244             (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
1245              nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES)) {
1246                 ctrl->csts = NVME_CSTS_CFS;
1247                 return;
1248         }
1249
1250         if (nvmet_cc_mps(ctrl->cc) != 0 ||
1251             nvmet_cc_ams(ctrl->cc) != 0 ||
1252             !nvmet_css_supported(nvmet_cc_css(ctrl->cc))) {
1253                 ctrl->csts = NVME_CSTS_CFS;
1254                 return;
1255         }
1256
1257         ctrl->csts = NVME_CSTS_RDY;
1258
1259         /*
1260          * Controllers that are not yet enabled should not really enforce the
1261          * keep alive timeout, but we still want to track a timeout and cleanup
1262          * in case a host died before it enabled the controller.  Hence, simply
1263          * reset the keep alive timer when the controller is enabled.
1264          */
1265         if (ctrl->kato)
1266                 mod_delayed_work(nvmet_wq, &ctrl->ka_work, ctrl->kato * HZ);
1267 }
1268
1269 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
1270 {
1271         lockdep_assert_held(&ctrl->lock);
1272
1273         /* XXX: tear down queues? */
1274         ctrl->csts &= ~NVME_CSTS_RDY;
1275         ctrl->cc = 0;
1276 }
1277
1278 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
1279 {
1280         u32 old;
1281
1282         mutex_lock(&ctrl->lock);
1283         old = ctrl->cc;
1284         ctrl->cc = new;
1285
1286         if (nvmet_cc_en(new) && !nvmet_cc_en(old))
1287                 nvmet_start_ctrl(ctrl);
1288         if (!nvmet_cc_en(new) && nvmet_cc_en(old))
1289                 nvmet_clear_ctrl(ctrl);
1290         if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
1291                 nvmet_clear_ctrl(ctrl);
1292                 ctrl->csts |= NVME_CSTS_SHST_CMPLT;
1293         }
1294         if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
1295                 ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
1296         mutex_unlock(&ctrl->lock);
1297 }
1298
1299 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
1300 {
1301         /* command sets supported: NVMe command set: */
1302         ctrl->cap = (1ULL << 37);
1303         /* Controller supports one or more I/O Command Sets */
1304         ctrl->cap |= (1ULL << 43);
1305         /* CC.EN timeout in 500msec units: */
1306         ctrl->cap |= (15ULL << 24);
1307         /* maximum queue entries supported: */
1308         if (ctrl->ops->get_max_queue_size)
1309                 ctrl->cap |= min_t(u16, ctrl->ops->get_max_queue_size(ctrl),
1310                                    ctrl->port->max_queue_size) - 1;
1311         else
1312                 ctrl->cap |= ctrl->port->max_queue_size - 1;
1313
1314         if (nvmet_is_passthru_subsys(ctrl->subsys))
1315                 nvmet_passthrough_override_cap(ctrl);
1316 }
1317
1318 struct nvmet_ctrl *nvmet_ctrl_find_get(const char *subsysnqn,
1319                                        const char *hostnqn, u16 cntlid,
1320                                        struct nvmet_req *req)
1321 {
1322         struct nvmet_ctrl *ctrl = NULL;
1323         struct nvmet_subsys *subsys;
1324
1325         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1326         if (!subsys) {
1327                 pr_warn("connect request for invalid subsystem %s!\n",
1328                         subsysnqn);
1329                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1330                 goto out;
1331         }
1332
1333         mutex_lock(&subsys->lock);
1334         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
1335                 if (ctrl->cntlid == cntlid) {
1336                         if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
1337                                 pr_warn("hostnqn mismatch.\n");
1338                                 continue;
1339                         }
1340                         if (!kref_get_unless_zero(&ctrl->ref))
1341                                 continue;
1342
1343                         /* ctrl found */
1344                         goto found;
1345                 }
1346         }
1347
1348         ctrl = NULL; /* ctrl not found */
1349         pr_warn("could not find controller %d for subsys %s / host %s\n",
1350                 cntlid, subsysnqn, hostnqn);
1351         req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
1352
1353 found:
1354         mutex_unlock(&subsys->lock);
1355         nvmet_subsys_put(subsys);
1356 out:
1357         return ctrl;
1358 }
1359
1360 u16 nvmet_check_ctrl_status(struct nvmet_req *req)
1361 {
1362         if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
1363                 pr_err("got cmd %d while CC.EN == 0 on qid = %d\n",
1364                        req->cmd->common.opcode, req->sq->qid);
1365                 return NVME_SC_CMD_SEQ_ERROR | NVME_STATUS_DNR;
1366         }
1367
1368         if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
1369                 pr_err("got cmd %d while CSTS.RDY == 0 on qid = %d\n",
1370                        req->cmd->common.opcode, req->sq->qid);
1371                 return NVME_SC_CMD_SEQ_ERROR | NVME_STATUS_DNR;
1372         }
1373
1374         if (unlikely(!nvmet_check_auth_status(req))) {
1375                 pr_warn("qid %d not authenticated\n", req->sq->qid);
1376                 return NVME_SC_AUTH_REQUIRED | NVME_STATUS_DNR;
1377         }
1378         return 0;
1379 }
1380
1381 bool nvmet_host_allowed(struct nvmet_subsys *subsys, const char *hostnqn)
1382 {
1383         struct nvmet_host_link *p;
1384
1385         lockdep_assert_held(&nvmet_config_sem);
1386
1387         if (subsys->allow_any_host)
1388                 return true;
1389
1390         if (nvmet_is_disc_subsys(subsys)) /* allow all access to disc subsys */
1391                 return true;
1392
1393         list_for_each_entry(p, &subsys->hosts, entry) {
1394                 if (!strcmp(nvmet_host_name(p->host), hostnqn))
1395                         return true;
1396         }
1397
1398         return false;
1399 }
1400
1401 /*
1402  * Note: ctrl->subsys->lock should be held when calling this function
1403  */
1404 static void nvmet_setup_p2p_ns_map(struct nvmet_ctrl *ctrl,
1405                 struct nvmet_req *req)
1406 {
1407         struct nvmet_ns *ns;
1408         unsigned long idx;
1409
1410         if (!req->p2p_client)
1411                 return;
1412
1413         ctrl->p2p_client = get_device(req->p2p_client);
1414
1415         nvmet_for_each_enabled_ns(&ctrl->subsys->namespaces, idx, ns)
1416                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
1417 }
1418
1419 /*
1420  * Note: ctrl->subsys->lock should be held when calling this function
1421  */
1422 static void nvmet_release_p2p_ns_map(struct nvmet_ctrl *ctrl)
1423 {
1424         struct radix_tree_iter iter;
1425         void __rcu **slot;
1426
1427         radix_tree_for_each_slot(slot, &ctrl->p2p_ns_map, &iter, 0)
1428                 pci_dev_put(radix_tree_deref_slot(slot));
1429
1430         put_device(ctrl->p2p_client);
1431 }
1432
1433 static void nvmet_fatal_error_handler(struct work_struct *work)
1434 {
1435         struct nvmet_ctrl *ctrl =
1436                         container_of(work, struct nvmet_ctrl, fatal_err_work);
1437
1438         pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
1439         ctrl->ops->delete_ctrl(ctrl);
1440 }
1441
1442 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
1443                 struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp,
1444                 uuid_t *hostid)
1445 {
1446         struct nvmet_subsys *subsys;
1447         struct nvmet_ctrl *ctrl;
1448         int ret;
1449         u16 status;
1450
1451         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_STATUS_DNR;
1452         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1453         if (!subsys) {
1454                 pr_warn("connect request for invalid subsystem %s!\n",
1455                         subsysnqn);
1456                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1457                 req->error_loc = offsetof(struct nvme_common_command, dptr);
1458                 goto out;
1459         }
1460
1461         down_read(&nvmet_config_sem);
1462         if (!nvmet_host_allowed(subsys, hostnqn)) {
1463                 pr_info("connect by host %s for subsystem %s not allowed\n",
1464                         hostnqn, subsysnqn);
1465                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
1466                 up_read(&nvmet_config_sem);
1467                 status = NVME_SC_CONNECT_INVALID_HOST | NVME_STATUS_DNR;
1468                 req->error_loc = offsetof(struct nvme_common_command, dptr);
1469                 goto out_put_subsystem;
1470         }
1471         up_read(&nvmet_config_sem);
1472
1473         status = NVME_SC_INTERNAL;
1474         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1475         if (!ctrl)
1476                 goto out_put_subsystem;
1477         mutex_init(&ctrl->lock);
1478
1479         ctrl->port = req->port;
1480         ctrl->ops = req->ops;
1481
1482 #ifdef CONFIG_NVME_TARGET_PASSTHRU
1483         /* By default, set loop targets to clear IDS by default */
1484         if (ctrl->port->disc_addr.trtype == NVMF_TRTYPE_LOOP)
1485                 subsys->clear_ids = 1;
1486 #endif
1487
1488         INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
1489         INIT_LIST_HEAD(&ctrl->async_events);
1490         INIT_RADIX_TREE(&ctrl->p2p_ns_map, GFP_KERNEL);
1491         INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
1492         INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer);
1493
1494         memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
1495         memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
1496
1497         kref_init(&ctrl->ref);
1498         ctrl->subsys = subsys;
1499         ctrl->pi_support = ctrl->port->pi_enable && ctrl->subsys->pi_support;
1500         nvmet_init_cap(ctrl);
1501         WRITE_ONCE(ctrl->aen_enabled, NVMET_AEN_CFG_OPTIONAL);
1502
1503         ctrl->changed_ns_list = kmalloc_array(NVME_MAX_CHANGED_NAMESPACES,
1504                         sizeof(__le32), GFP_KERNEL);
1505         if (!ctrl->changed_ns_list)
1506                 goto out_free_ctrl;
1507
1508         ctrl->sqs = kcalloc(subsys->max_qid + 1,
1509                         sizeof(struct nvmet_sq *),
1510                         GFP_KERNEL);
1511         if (!ctrl->sqs)
1512                 goto out_free_changed_ns_list;
1513
1514         ret = ida_alloc_range(&cntlid_ida,
1515                              subsys->cntlid_min, subsys->cntlid_max,
1516                              GFP_KERNEL);
1517         if (ret < 0) {
1518                 status = NVME_SC_CONNECT_CTRL_BUSY | NVME_STATUS_DNR;
1519                 goto out_free_sqs;
1520         }
1521         ctrl->cntlid = ret;
1522
1523         uuid_copy(&ctrl->hostid, hostid);
1524
1525         /*
1526          * Discovery controllers may use some arbitrary high value
1527          * in order to cleanup stale discovery sessions
1528          */
1529         if (nvmet_is_disc_subsys(ctrl->subsys) && !kato)
1530                 kato = NVMET_DISC_KATO_MS;
1531
1532         /* keep-alive timeout in seconds */
1533         ctrl->kato = DIV_ROUND_UP(kato, 1000);
1534
1535         ctrl->err_counter = 0;
1536         spin_lock_init(&ctrl->error_lock);
1537
1538         nvmet_start_keep_alive_timer(ctrl);
1539
1540         mutex_lock(&subsys->lock);
1541         ret = nvmet_ctrl_init_pr(ctrl);
1542         if (ret)
1543                 goto init_pr_fail;
1544         list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
1545         nvmet_setup_p2p_ns_map(ctrl, req);
1546         nvmet_debugfs_ctrl_setup(ctrl);
1547         mutex_unlock(&subsys->lock);
1548
1549         *ctrlp = ctrl;
1550         return 0;
1551
1552 init_pr_fail:
1553         mutex_unlock(&subsys->lock);
1554         nvmet_stop_keep_alive_timer(ctrl);
1555         ida_free(&cntlid_ida, ctrl->cntlid);
1556 out_free_sqs:
1557         kfree(ctrl->sqs);
1558 out_free_changed_ns_list:
1559         kfree(ctrl->changed_ns_list);
1560 out_free_ctrl:
1561         kfree(ctrl);
1562 out_put_subsystem:
1563         nvmet_subsys_put(subsys);
1564 out:
1565         return status;
1566 }
1567
1568 static void nvmet_ctrl_free(struct kref *ref)
1569 {
1570         struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
1571         struct nvmet_subsys *subsys = ctrl->subsys;
1572
1573         mutex_lock(&subsys->lock);
1574         nvmet_ctrl_destroy_pr(ctrl);
1575         nvmet_release_p2p_ns_map(ctrl);
1576         list_del(&ctrl->subsys_entry);
1577         mutex_unlock(&subsys->lock);
1578
1579         nvmet_stop_keep_alive_timer(ctrl);
1580
1581         flush_work(&ctrl->async_event_work);
1582         cancel_work_sync(&ctrl->fatal_err_work);
1583
1584         nvmet_destroy_auth(ctrl);
1585
1586         nvmet_debugfs_ctrl_free(ctrl);
1587
1588         ida_free(&cntlid_ida, ctrl->cntlid);
1589
1590         nvmet_async_events_free(ctrl);
1591         kfree(ctrl->sqs);
1592         kfree(ctrl->changed_ns_list);
1593         kfree(ctrl);
1594
1595         nvmet_subsys_put(subsys);
1596 }
1597
1598 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
1599 {
1600         kref_put(&ctrl->ref, nvmet_ctrl_free);
1601 }
1602
1603 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
1604 {
1605         mutex_lock(&ctrl->lock);
1606         if (!(ctrl->csts & NVME_CSTS_CFS)) {
1607                 ctrl->csts |= NVME_CSTS_CFS;
1608                 queue_work(nvmet_wq, &ctrl->fatal_err_work);
1609         }
1610         mutex_unlock(&ctrl->lock);
1611 }
1612 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
1613
1614 ssize_t nvmet_ctrl_host_traddr(struct nvmet_ctrl *ctrl,
1615                 char *traddr, size_t traddr_len)
1616 {
1617         if (!ctrl->ops->host_traddr)
1618                 return -EOPNOTSUPP;
1619         return ctrl->ops->host_traddr(ctrl, traddr, traddr_len);
1620 }
1621
1622 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
1623                 const char *subsysnqn)
1624 {
1625         struct nvmet_subsys_link *p;
1626
1627         if (!port)
1628                 return NULL;
1629
1630         if (!strcmp(NVME_DISC_SUBSYS_NAME, subsysnqn)) {
1631                 if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
1632                         return NULL;
1633                 return nvmet_disc_subsys;
1634         }
1635
1636         down_read(&nvmet_config_sem);
1637         if (!strncmp(nvmet_disc_subsys->subsysnqn, subsysnqn,
1638                                 NVMF_NQN_SIZE)) {
1639                 if (kref_get_unless_zero(&nvmet_disc_subsys->ref)) {
1640                         up_read(&nvmet_config_sem);
1641                         return nvmet_disc_subsys;
1642                 }
1643         }
1644         list_for_each_entry(p, &port->subsystems, entry) {
1645                 if (!strncmp(p->subsys->subsysnqn, subsysnqn,
1646                                 NVMF_NQN_SIZE)) {
1647                         if (!kref_get_unless_zero(&p->subsys->ref))
1648                                 break;
1649                         up_read(&nvmet_config_sem);
1650                         return p->subsys;
1651                 }
1652         }
1653         up_read(&nvmet_config_sem);
1654         return NULL;
1655 }
1656
1657 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
1658                 enum nvme_subsys_type type)
1659 {
1660         struct nvmet_subsys *subsys;
1661         char serial[NVMET_SN_MAX_SIZE / 2];
1662         int ret;
1663
1664         subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
1665         if (!subsys)
1666                 return ERR_PTR(-ENOMEM);
1667
1668         subsys->ver = NVMET_DEFAULT_VS;
1669         /* generate a random serial number as our controllers are ephemeral: */
1670         get_random_bytes(&serial, sizeof(serial));
1671         bin2hex(subsys->serial, &serial, sizeof(serial));
1672
1673         subsys->model_number = kstrdup(NVMET_DEFAULT_CTRL_MODEL, GFP_KERNEL);
1674         if (!subsys->model_number) {
1675                 ret = -ENOMEM;
1676                 goto free_subsys;
1677         }
1678
1679         subsys->ieee_oui = 0;
1680
1681         subsys->firmware_rev = kstrndup(UTS_RELEASE, NVMET_FR_MAX_SIZE, GFP_KERNEL);
1682         if (!subsys->firmware_rev) {
1683                 ret = -ENOMEM;
1684                 goto free_mn;
1685         }
1686
1687         switch (type) {
1688         case NVME_NQN_NVME:
1689                 subsys->max_qid = NVMET_NR_QUEUES;
1690                 break;
1691         case NVME_NQN_DISC:
1692         case NVME_NQN_CURR:
1693                 subsys->max_qid = 0;
1694                 break;
1695         default:
1696                 pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
1697                 ret = -EINVAL;
1698                 goto free_fr;
1699         }
1700         subsys->type = type;
1701         subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
1702                         GFP_KERNEL);
1703         if (!subsys->subsysnqn) {
1704                 ret = -ENOMEM;
1705                 goto free_fr;
1706         }
1707         subsys->cntlid_min = NVME_CNTLID_MIN;
1708         subsys->cntlid_max = NVME_CNTLID_MAX;
1709         kref_init(&subsys->ref);
1710
1711         mutex_init(&subsys->lock);
1712         xa_init(&subsys->namespaces);
1713         INIT_LIST_HEAD(&subsys->ctrls);
1714         INIT_LIST_HEAD(&subsys->hosts);
1715
1716         ret = nvmet_debugfs_subsys_setup(subsys);
1717         if (ret)
1718                 goto free_subsysnqn;
1719
1720         return subsys;
1721
1722 free_subsysnqn:
1723         kfree(subsys->subsysnqn);
1724 free_fr:
1725         kfree(subsys->firmware_rev);
1726 free_mn:
1727         kfree(subsys->model_number);
1728 free_subsys:
1729         kfree(subsys);
1730         return ERR_PTR(ret);
1731 }
1732
1733 static void nvmet_subsys_free(struct kref *ref)
1734 {
1735         struct nvmet_subsys *subsys =
1736                 container_of(ref, struct nvmet_subsys, ref);
1737
1738         WARN_ON_ONCE(!xa_empty(&subsys->namespaces));
1739
1740         nvmet_debugfs_subsys_free(subsys);
1741
1742         xa_destroy(&subsys->namespaces);
1743         nvmet_passthru_subsys_free(subsys);
1744
1745         kfree(subsys->subsysnqn);
1746         kfree(subsys->model_number);
1747         kfree(subsys->firmware_rev);
1748         kfree(subsys);
1749 }
1750
1751 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1752 {
1753         struct nvmet_ctrl *ctrl;
1754
1755         mutex_lock(&subsys->lock);
1756         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1757                 ctrl->ops->delete_ctrl(ctrl);
1758         mutex_unlock(&subsys->lock);
1759 }
1760
1761 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1762 {
1763         kref_put(&subsys->ref, nvmet_subsys_free);
1764 }
1765
1766 static int __init nvmet_init(void)
1767 {
1768         int error = -ENOMEM;
1769
1770         nvmet_ana_group_enabled[NVMET_DEFAULT_ANA_GRPID] = 1;
1771
1772         nvmet_bvec_cache = kmem_cache_create("nvmet-bvec",
1773                         NVMET_MAX_MPOOL_BVEC * sizeof(struct bio_vec), 0,
1774                         SLAB_HWCACHE_ALIGN, NULL);
1775         if (!nvmet_bvec_cache)
1776                 return -ENOMEM;
1777
1778         zbd_wq = alloc_workqueue("nvmet-zbd-wq", WQ_MEM_RECLAIM, 0);
1779         if (!zbd_wq)
1780                 goto out_destroy_bvec_cache;
1781
1782         buffered_io_wq = alloc_workqueue("nvmet-buffered-io-wq",
1783                         WQ_MEM_RECLAIM, 0);
1784         if (!buffered_io_wq)
1785                 goto out_free_zbd_work_queue;
1786
1787         nvmet_wq = alloc_workqueue("nvmet-wq",
1788                         WQ_MEM_RECLAIM | WQ_UNBOUND | WQ_SYSFS, 0);
1789         if (!nvmet_wq)
1790                 goto out_free_buffered_work_queue;
1791
1792         error = nvmet_init_discovery();
1793         if (error)
1794                 goto out_free_nvmet_work_queue;
1795
1796         error = nvmet_init_debugfs();
1797         if (error)
1798                 goto out_exit_discovery;
1799
1800         error = nvmet_init_configfs();
1801         if (error)
1802                 goto out_exit_debugfs;
1803
1804         return 0;
1805
1806 out_exit_debugfs:
1807         nvmet_exit_debugfs();
1808 out_exit_discovery:
1809         nvmet_exit_discovery();
1810 out_free_nvmet_work_queue:
1811         destroy_workqueue(nvmet_wq);
1812 out_free_buffered_work_queue:
1813         destroy_workqueue(buffered_io_wq);
1814 out_free_zbd_work_queue:
1815         destroy_workqueue(zbd_wq);
1816 out_destroy_bvec_cache:
1817         kmem_cache_destroy(nvmet_bvec_cache);
1818         return error;
1819 }
1820
1821 static void __exit nvmet_exit(void)
1822 {
1823         nvmet_exit_configfs();
1824         nvmet_exit_debugfs();
1825         nvmet_exit_discovery();
1826         ida_destroy(&cntlid_ida);
1827         destroy_workqueue(nvmet_wq);
1828         destroy_workqueue(buffered_io_wq);
1829         destroy_workqueue(zbd_wq);
1830         kmem_cache_destroy(nvmet_bvec_cache);
1831
1832         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1833         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1834 }
1835
1836 module_init(nvmet_init);
1837 module_exit(nvmet_exit);
1838
1839 MODULE_DESCRIPTION("NVMe target core framework");
1840 MODULE_LICENSE("GPL v2");