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         xa_for_each(&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)) {
 445                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 446                 if (nvmet_subsys_nsid_exists(subsys, nsid))
 447                         return NVME_SC_INTERNAL_PATH_ERROR;
 448                 return NVME_SC_INVALID_NS | NVME_STATUS_DNR;
 449         }
 450
 451         percpu_ref_get(&req->ns->ref);
 452         return NVME_SC_SUCCESS;
 453 }
 454
 455 static void nvmet_destroy_namespace(struct percpu_ref *ref)
 456 {
 457         struct nvmet_ns *ns = container_of(ref, struct nvmet_ns, ref);
 458
 459         complete(&ns->disable_done);
 460 }
 461
 462 void nvmet_put_namespace(struct nvmet_ns *ns)
 463 {
 464         percpu_ref_put(&ns->ref);
 465 }
 466
 467 static void nvmet_ns_dev_disable(struct nvmet_ns *ns)
 468 {
 469         nvmet_bdev_ns_disable(ns);
 470         nvmet_file_ns_disable(ns);
 471 }
 472
 473 static int nvmet_p2pmem_ns_enable(struct nvmet_ns *ns)
 474 {
 475         int ret;
 476         struct pci_dev *p2p_dev;
 477
 478         if (!ns->use_p2pmem)
 479                 return 0;
 480
 481         if (!ns->bdev) {
 482                 pr_err("peer-to-peer DMA is not supported by non-block device namespaces\n");
 483                 return -EINVAL;
 484         }
 485
 486         if (!blk_queue_pci_p2pdma(ns->bdev->bd_disk->queue)) {
 487                 pr_err("peer-to-peer DMA is not supported by the driver of %s\n",
 488                        ns->device_path);
 489                 return -EINVAL;
 490         }
 491
 492         if (ns->p2p_dev) {
 493                 ret = pci_p2pdma_distance(ns->p2p_dev, nvmet_ns_dev(ns), true);
 494                 if (ret < 0)
 495                         return -EINVAL;
 496         } else {
 497                 /*
 498                  * Right now we just check that there is p2pmem available so
 499                  * we can report an error to the user right away if there
 500                  * is not. We'll find the actual device to use once we
 501                  * setup the controller when the port's device is available.
 502                  */
 503
 504                 p2p_dev = pci_p2pmem_find(nvmet_ns_dev(ns));
 505                 if (!p2p_dev) {
 506                         pr_err("no peer-to-peer memory is available for %s\n",
 507                                ns->device_path);
 508                         return -EINVAL;
 509                 }
 510
 511                 pci_dev_put(p2p_dev);
 512         }
 513
 514         return 0;
 515 }
 516
 517 /*
 518  * Note: ctrl->subsys->lock should be held when calling this function
 519  */
 520 static void nvmet_p2pmem_ns_add_p2p(struct nvmet_ctrl *ctrl,
 521                                     struct nvmet_ns *ns)
 522 {
 523         struct device *clients[2];
 524         struct pci_dev *p2p_dev;
 525         int ret;
 526
 527         if (!ctrl->p2p_client || !ns->use_p2pmem)
 528                 return;
 529
 530         if (ns->p2p_dev) {
 531                 ret = pci_p2pdma_distance(ns->p2p_dev, ctrl->p2p_client, true);
 532                 if (ret < 0)
 533                         return;
 534
 535                 p2p_dev = pci_dev_get(ns->p2p_dev);
 536         } else {
 537                 clients[0] = ctrl->p2p_client;
 538                 clients[1] = nvmet_ns_dev(ns);
 539
 540                 p2p_dev = pci_p2pmem_find_many(clients, ARRAY_SIZE(clients));
 541                 if (!p2p_dev) {
 542                         pr_err("no peer-to-peer memory is available that's supported by %s and %s\n",
 543                                dev_name(ctrl->p2p_client), ns->device_path);
 544                         return;
 545                 }
 546         }
 547
 548         ret = radix_tree_insert(&ctrl->p2p_ns_map, ns->nsid, p2p_dev);
 549         if (ret < 0)
 550                 pci_dev_put(p2p_dev);
 551
 552         pr_info("using p2pmem on %s for nsid %d\n", pci_name(p2p_dev),
 553                 ns->nsid);
 554 }
 555
 556 bool nvmet_ns_revalidate(struct nvmet_ns *ns)
 557 {
 558         loff_t oldsize = ns->size;
 559
 560         if (ns->bdev)
 561                 nvmet_bdev_ns_revalidate(ns);
 562         else
 563                 nvmet_file_ns_revalidate(ns);
 564
 565         return oldsize != ns->size;
 566 }
 567
 568 int nvmet_ns_enable(struct nvmet_ns *ns)
 569 {
 570         struct nvmet_subsys *subsys = ns->subsys;
 571         struct nvmet_ctrl *ctrl;
 572         int ret;
 573
 574         mutex_lock(&subsys->lock);
 575         ret = 0;
 576
 577         if (nvmet_is_passthru_subsys(subsys)) {
 578                 pr_info("cannot enable both passthru and regular namespaces for a single subsystem");
 579                 goto out_unlock;
 580         }
 581
 582         if (ns->enabled)
 583                 goto out_unlock;
 584
 585         ret = -EMFILE;
 586         if (subsys->nr_namespaces == NVMET_MAX_NAMESPACES)
 587                 goto out_unlock;
 588
 589         ret = nvmet_bdev_ns_enable(ns);
 590         if (ret == -ENOTBLK)
 591                 ret = nvmet_file_ns_enable(ns);
 592         if (ret)
 593                 goto out_unlock;
 594
 595         ret = nvmet_p2pmem_ns_enable(ns);
 596         if (ret)
 597                 goto out_dev_disable;
 598
 599         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 600                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
 601
 602         ret = percpu_ref_init(&ns->ref, nvmet_destroy_namespace,
 603                                 0, GFP_KERNEL);
 604         if (ret)
 605                 goto out_dev_put;
 606
 607         if (ns->nsid > subsys->max_nsid)
 608                 subsys->max_nsid = ns->nsid;
 609
 610         ret = xa_insert(&subsys->namespaces, ns->nsid, ns, GFP_KERNEL);
 611         if (ret)
 612                 goto out_restore_subsys_maxnsid;
 613
 614         subsys->nr_namespaces++;
 615
 616         nvmet_ns_changed(subsys, ns->nsid);
 617         ns->enabled = true;
 618         ret = 0;
 619 out_unlock:
 620         mutex_unlock(&subsys->lock);
 621         return ret;
 622
 623 out_restore_subsys_maxnsid:
 624         subsys->max_nsid = nvmet_max_nsid(subsys);
 625         percpu_ref_exit(&ns->ref);
 626 out_dev_put:
 627         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 628                 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
 629 out_dev_disable:
 630         nvmet_ns_dev_disable(ns);
 631         goto out_unlock;
 632 }
 633
 634 void nvmet_ns_disable(struct nvmet_ns *ns)
 635 {
 636         struct nvmet_subsys *subsys = ns->subsys;
 637         struct nvmet_ctrl *ctrl;
 638
 639         mutex_lock(&subsys->lock);
 640         if (!ns->enabled)
 641                 goto out_unlock;
 642
 643         ns->enabled = false;
 644         xa_erase(&ns->subsys->namespaces, ns->nsid);
 645         if (ns->nsid == subsys->max_nsid)
 646                 subsys->max_nsid = nvmet_max_nsid(subsys);
 647
 648         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 649                 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
 650
 651         mutex_unlock(&subsys->lock);
 652
 653         /*
 654          * Now that we removed the namespaces from the lookup list, we
 655          * can kill the per_cpu ref and wait for any remaining references
 656          * to be dropped, as well as a RCU grace period for anyone only
 657          * using the namepace under rcu_read_lock().  Note that we can't
 658          * use call_rcu here as we need to ensure the namespaces have
 659          * been fully destroyed before unloading the module.
 660          */
 661         percpu_ref_kill(&ns->ref);
 662         synchronize_rcu();
 663         wait_for_completion(&ns->disable_done);
 664         percpu_ref_exit(&ns->ref);
 665
 666         mutex_lock(&subsys->lock);
 667
 668         subsys->nr_namespaces--;
 669         nvmet_ns_changed(subsys, ns->nsid);
 670         nvmet_ns_dev_disable(ns);
 671 out_unlock:
 672         mutex_unlock(&subsys->lock);
 673 }
 674
 675 void nvmet_ns_free(struct nvmet_ns *ns)
 676 {
 677         nvmet_ns_disable(ns);
 678
 679         down_write(&nvmet_ana_sem);
 680         nvmet_ana_group_enabled[ns->anagrpid]--;
 681         up_write(&nvmet_ana_sem);
 682
 683         kfree(ns->device_path);
 684         kfree(ns);
 685 }
 686
 687 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
 688 {
 689         struct nvmet_ns *ns;
 690
 691         ns = kzalloc(sizeof(*ns), GFP_KERNEL);
 692         if (!ns)
 693                 return NULL;
 694
 695         init_completion(&ns->disable_done);
 696
 697         ns->nsid = nsid;
 698         ns->subsys = subsys;
 699
 700         down_write(&nvmet_ana_sem);
 701         ns->anagrpid = NVMET_DEFAULT_ANA_GRPID;
 702         nvmet_ana_group_enabled[ns->anagrpid]++;
 703         up_write(&nvmet_ana_sem);
 704
 705         uuid_gen(&ns->uuid);
 706         ns->buffered_io = false;
 707         ns->csi = NVME_CSI_NVM;
 708
 709         return ns;
 710 }
 711
 712 static void nvmet_update_sq_head(struct nvmet_req *req)
 713 {
 714         if (req->sq->size) {
 715                 u32 old_sqhd, new_sqhd;
 716
 717                 old_sqhd = READ_ONCE(req->sq->sqhd);
 718                 do {
 719                         new_sqhd = (old_sqhd + 1) % req->sq->size;
 720                 } while (!try_cmpxchg(&req->sq->sqhd, &old_sqhd, new_sqhd));
 721         }
 722         req->cqe->sq_head = cpu_to_le16(req->sq->sqhd & 0x0000FFFF);
 723 }
 724
 725 static void nvmet_set_error(struct nvmet_req *req, u16 status)
 726 {
 727         struct nvmet_ctrl *ctrl = req->sq->ctrl;
 728         struct nvme_error_slot *new_error_slot;
 729         unsigned long flags;
 730
 731         req->cqe->status = cpu_to_le16(status << 1);
 732
 733         if (!ctrl || req->error_loc == NVMET_NO_ERROR_LOC)
 734                 return;
 735
 736         spin_lock_irqsave(&ctrl->error_lock, flags);
 737         ctrl->err_counter++;
 738         new_error_slot =
 739                 &ctrl->slots[ctrl->err_counter % NVMET_ERROR_LOG_SLOTS];
 740
 741         new_error_slot->error_count = cpu_to_le64(ctrl->err_counter);
 742         new_error_slot->sqid = cpu_to_le16(req->sq->qid);
 743         new_error_slot->cmdid = cpu_to_le16(req->cmd->common.command_id);
 744         new_error_slot->status_field = cpu_to_le16(status << 1);
 745         new_error_slot->param_error_location = cpu_to_le16(req->error_loc);
 746         new_error_slot->lba = cpu_to_le64(req->error_slba);
 747         new_error_slot->nsid = req->cmd->common.nsid;
 748         spin_unlock_irqrestore(&ctrl->error_lock, flags);
 749
 750         /* set the more bit for this request */
 751         req->cqe->status |= cpu_to_le16(1 << 14);
 752 }
 753
 754 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
 755 {
 756         struct nvmet_ns *ns = req->ns;
 757
 758         if (!req->sq->sqhd_disabled)
 759                 nvmet_update_sq_head(req);
 760         req->cqe->sq_id = cpu_to_le16(req->sq->qid);
 761         req->cqe->command_id = req->cmd->common.command_id;
 762
 763         if (unlikely(status))
 764                 nvmet_set_error(req, status);
 765
 766         trace_nvmet_req_complete(req);
 767
 768         req->ops->queue_response(req);
 769         if (ns)
 770                 nvmet_put_namespace(ns);
 771 }
 772
 773 void nvmet_req_complete(struct nvmet_req *req, u16 status)
 774 {
 775         struct nvmet_sq *sq = req->sq;
 776
 777         __nvmet_req_complete(req, status);
 778         percpu_ref_put(&sq->ref);
 779 }
 780 EXPORT_SYMBOL_GPL(nvmet_req_complete);
 781
 782 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
 783                 u16 qid, u16 size)
 784 {
 785         cq->qid = qid;
 786         cq->size = size;
 787 }
 788
 789 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
 790                 u16 qid, u16 size)
 791 {
 792         sq->sqhd = 0;
 793         sq->qid = qid;
 794         sq->size = size;
 795
 796         ctrl->sqs[qid] = sq;
 797 }
 798
 799 static void nvmet_confirm_sq(struct percpu_ref *ref)
 800 {
 801         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 802
 803         complete(&sq->confirm_done);
 804 }
 805
 806 void nvmet_sq_destroy(struct nvmet_sq *sq)
 807 {
 808         struct nvmet_ctrl *ctrl = sq->ctrl;
 809
 810         /*
 811          * If this is the admin queue, complete all AERs so that our
 812          * queue doesn't have outstanding requests on it.
 813          */
 814         if (ctrl && ctrl->sqs && ctrl->sqs[0] == sq)
 815                 nvmet_async_events_failall(ctrl);
 816         percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
 817         wait_for_completion(&sq->confirm_done);
 818         wait_for_completion(&sq->free_done);
 819         percpu_ref_exit(&sq->ref);
 820         nvmet_auth_sq_free(sq);
 821
 822         /*
 823          * we must reference the ctrl again after waiting for inflight IO
 824          * to complete. Because admin connect may have sneaked in after we
 825          * store sq->ctrl locally, but before we killed the percpu_ref. the
 826          * admin connect allocates and assigns sq->ctrl, which now needs a
 827          * final ref put, as this ctrl is going away.
 828          */
 829         ctrl = sq->ctrl;
 830
 831         if (ctrl) {
 832                 /*
 833                  * The teardown flow may take some time, and the host may not
 834                  * send us keep-alive during this period, hence reset the
 835                  * traffic based keep-alive timer so we don't trigger a
 836                  * controller teardown as a result of a keep-alive expiration.
 837                  */
 838                 ctrl->reset_tbkas = true;
 839                 sq->ctrl->sqs[sq->qid] = NULL;
 840                 nvmet_ctrl_put(ctrl);
 841                 sq->ctrl = NULL; /* allows reusing the queue later */
 842         }
 843 }
 844 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
 845
 846 static void nvmet_sq_free(struct percpu_ref *ref)
 847 {
 848         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 849
 850         complete(&sq->free_done);
 851 }
 852
 853 int nvmet_sq_init(struct nvmet_sq *sq)
 854 {
 855         int ret;
 856
 857         ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
 858         if (ret) {
 859                 pr_err("percpu_ref init failed!\n");
 860                 return ret;
 861         }
 862         init_completion(&sq->free_done);
 863         init_completion(&sq->confirm_done);
 864         nvmet_auth_sq_init(sq);
 865
 866         return 0;
 867 }
 868 EXPORT_SYMBOL_GPL(nvmet_sq_init);
 869
 870 static inline u16 nvmet_check_ana_state(struct nvmet_port *port,
 871                 struct nvmet_ns *ns)
 872 {
 873         enum nvme_ana_state state = port->ana_state[ns->anagrpid];
 874
 875         if (unlikely(state == NVME_ANA_INACCESSIBLE))
 876                 return NVME_SC_ANA_INACCESSIBLE;
 877         if (unlikely(state == NVME_ANA_PERSISTENT_LOSS))
 878                 return NVME_SC_ANA_PERSISTENT_LOSS;
 879         if (unlikely(state == NVME_ANA_CHANGE))
 880                 return NVME_SC_ANA_TRANSITION;
 881         return 0;
 882 }
 883
 884 static inline u16 nvmet_io_cmd_check_access(struct nvmet_req *req)
 885 {
 886         if (unlikely(req->ns->readonly)) {
 887                 switch (req->cmd->common.opcode) {
 888                 case nvme_cmd_read:
 889                 case nvme_cmd_flush:
 890                         break;
 891                 default:
 892                         return NVME_SC_NS_WRITE_PROTECTED;
 893                 }
 894         }
 895
 896         return 0;
 897 }
 898
 899 static u16 nvmet_parse_io_cmd(struct nvmet_req *req)
 900 {
 901         struct nvme_command *cmd = req->cmd;
 902         u16 ret;
 903
 904         if (nvme_is_fabrics(cmd))
 905                 return nvmet_parse_fabrics_io_cmd(req);
 906
 907         if (unlikely(!nvmet_check_auth_status(req)))
 908                 return NVME_SC_AUTH_REQUIRED | NVME_STATUS_DNR;
 909
 910         ret = nvmet_check_ctrl_status(req);
 911         if (unlikely(ret))
 912                 return ret;
 913
 914         if (nvmet_is_passthru_req(req))
 915                 return nvmet_parse_passthru_io_cmd(req);
 916
 917         ret = nvmet_req_find_ns(req);
 918         if (unlikely(ret))
 919                 return ret;
 920
 921         ret = nvmet_check_ana_state(req->port, req->ns);
 922         if (unlikely(ret)) {
 923                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 924                 return ret;
 925         }
 926         ret = nvmet_io_cmd_check_access(req);
 927         if (unlikely(ret)) {
 928                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 929                 return ret;
 930         }
 931
 932         switch (req->ns->csi) {
 933         case NVME_CSI_NVM:
 934                 if (req->ns->file)
 935                         return nvmet_file_parse_io_cmd(req);
 936                 return nvmet_bdev_parse_io_cmd(req);
 937         case NVME_CSI_ZNS:
 938                 if (IS_ENABLED(CONFIG_BLK_DEV_ZONED))
 939                         return nvmet_bdev_zns_parse_io_cmd(req);
 940                 return NVME_SC_INVALID_IO_CMD_SET;
 941         default:
 942                 return NVME_SC_INVALID_IO_CMD_SET;
 943         }
 944 }
 945
 946 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
 947                 struct nvmet_sq *sq, const struct nvmet_fabrics_ops *ops)
 948 {
 949         u8 flags = req->cmd->common.flags;
 950         u16 status;
 951
 952         req->cq = cq;
 953         req->sq = sq;
 954         req->ops = ops;
 955         req->sg = NULL;
 956         req->metadata_sg = NULL;
 957         req->sg_cnt = 0;
 958         req->metadata_sg_cnt = 0;
 959         req->transfer_len = 0;
 960         req->metadata_len = 0;
 961         req->cqe->result.u64 = 0;
 962         req->cqe->status = 0;
 963         req->cqe->sq_head = 0;
 964         req->ns = NULL;
 965         req->error_loc = NVMET_NO_ERROR_LOC;
 966         req->error_slba = 0;
 967
 968         /* no support for fused commands yet */
 969         if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
 970                 req->error_loc = offsetof(struct nvme_common_command, flags);
 971                 status = NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
 972                 goto fail;
 973         }
 974
 975         /*
 976          * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
 977          * contains an address of a single contiguous physical buffer that is
 978          * byte aligned.
 979          */
 980         if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
 981                 req->error_loc = offsetof(struct nvme_common_command, flags);
 982                 status = NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
 983                 goto fail;
 984         }
 985
 986         if (unlikely(!req->sq->ctrl))
 987                 /* will return an error for any non-connect command: */
 988                 status = nvmet_parse_connect_cmd(req);
 989         else if (likely(req->sq->qid != 0))
 990                 status = nvmet_parse_io_cmd(req);
 991         else
 992                 status = nvmet_parse_admin_cmd(req);
 993
 994         if (status)
 995                 goto fail;
 996
 997         trace_nvmet_req_init(req, req->cmd);
 998
 999         if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
1000                 status = NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
1001                 goto fail;
1002         }
1003
1004         if (sq->ctrl)
1005                 sq->ctrl->reset_tbkas = true;
1006
1007         return true;
1008
1009 fail:
1010         __nvmet_req_complete(req, status);
1011         return false;
1012 }
1013 EXPORT_SYMBOL_GPL(nvmet_req_init);
1014
1015 void nvmet_req_uninit(struct nvmet_req *req)
1016 {
1017         percpu_ref_put(&req->sq->ref);
1018         if (req->ns)
1019                 nvmet_put_namespace(req->ns);
1020 }
1021 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
1022
1023 bool nvmet_check_transfer_len(struct nvmet_req *req, size_t len)
1024 {
1025         if (unlikely(len != req->transfer_len)) {
1026                 req->error_loc = offsetof(struct nvme_common_command, dptr);
1027                 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_STATUS_DNR);
1028                 return false;
1029         }
1030
1031         return true;
1032 }
1033 EXPORT_SYMBOL_GPL(nvmet_check_transfer_len);
1034
1035 bool nvmet_check_data_len_lte(struct nvmet_req *req, size_t data_len)
1036 {
1037         if (unlikely(data_len > req->transfer_len)) {
1038                 req->error_loc = offsetof(struct nvme_common_command, dptr);
1039                 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_STATUS_DNR);
1040                 return false;
1041         }
1042
1043         return true;
1044 }
1045
1046 static unsigned int nvmet_data_transfer_len(struct nvmet_req *req)
1047 {
1048         return req->transfer_len - req->metadata_len;
1049 }
1050
1051 static int nvmet_req_alloc_p2pmem_sgls(struct pci_dev *p2p_dev,
1052                 struct nvmet_req *req)
1053 {
1054         req->sg = pci_p2pmem_alloc_sgl(p2p_dev, &req->sg_cnt,
1055                         nvmet_data_transfer_len(req));
1056         if (!req->sg)
1057                 goto out_err;
1058
1059         if (req->metadata_len) {
1060                 req->metadata_sg = pci_p2pmem_alloc_sgl(p2p_dev,
1061                                 &req->metadata_sg_cnt, req->metadata_len);
1062                 if (!req->metadata_sg)
1063                         goto out_free_sg;
1064         }
1065
1066         req->p2p_dev = p2p_dev;
1067
1068         return 0;
1069 out_free_sg:
1070         pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1071 out_err:
1072         return -ENOMEM;
1073 }
1074
1075 static struct pci_dev *nvmet_req_find_p2p_dev(struct nvmet_req *req)
1076 {
1077         if (!IS_ENABLED(CONFIG_PCI_P2PDMA) ||
1078             !req->sq->ctrl || !req->sq->qid || !req->ns)
1079                 return NULL;
1080         return radix_tree_lookup(&req->sq->ctrl->p2p_ns_map, req->ns->nsid);
1081 }
1082
1083 int nvmet_req_alloc_sgls(struct nvmet_req *req)
1084 {
1085         struct pci_dev *p2p_dev = nvmet_req_find_p2p_dev(req);
1086
1087         if (p2p_dev && !nvmet_req_alloc_p2pmem_sgls(p2p_dev, req))
1088                 return 0;
1089
1090         req->sg = sgl_alloc(nvmet_data_transfer_len(req), GFP_KERNEL,
1091                             &req->sg_cnt);
1092         if (unlikely(!req->sg))
1093                 goto out;
1094
1095         if (req->metadata_len) {
1096                 req->metadata_sg = sgl_alloc(req->metadata_len, GFP_KERNEL,
1097                                              &req->metadata_sg_cnt);
1098                 if (unlikely(!req->metadata_sg))
1099                         goto out_free;
1100         }
1101
1102         return 0;
1103 out_free:
1104         sgl_free(req->sg);
1105 out:
1106         return -ENOMEM;
1107 }
1108 EXPORT_SYMBOL_GPL(nvmet_req_alloc_sgls);
1109
1110 void nvmet_req_free_sgls(struct nvmet_req *req)
1111 {
1112         if (req->p2p_dev) {
1113                 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1114                 if (req->metadata_sg)
1115                         pci_p2pmem_free_sgl(req->p2p_dev, req->metadata_sg);
1116                 req->p2p_dev = NULL;
1117         } else {
1118                 sgl_free(req->sg);
1119                 if (req->metadata_sg)
1120                         sgl_free(req->metadata_sg);
1121         }
1122
1123         req->sg = NULL;
1124         req->metadata_sg = NULL;
1125         req->sg_cnt = 0;
1126         req->metadata_sg_cnt = 0;
1127 }
1128 EXPORT_SYMBOL_GPL(nvmet_req_free_sgls);
1129
1130 static inline bool nvmet_cc_en(u32 cc)
1131 {
1132         return (cc >> NVME_CC_EN_SHIFT) & 0x1;
1133 }
1134
1135 static inline u8 nvmet_cc_css(u32 cc)
1136 {
1137         return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
1138 }
1139
1140 static inline u8 nvmet_cc_mps(u32 cc)
1141 {
1142         return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
1143 }
1144
1145 static inline u8 nvmet_cc_ams(u32 cc)
1146 {
1147         return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
1148 }
1149
1150 static inline u8 nvmet_cc_shn(u32 cc)
1151 {
1152         return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
1153 }
1154
1155 static inline u8 nvmet_cc_iosqes(u32 cc)
1156 {
1157         return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
1158 }
1159
1160 static inline u8 nvmet_cc_iocqes(u32 cc)
1161 {
1162         return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
1163 }
1164
1165 static inline bool nvmet_css_supported(u8 cc_css)
1166 {
1167         switch (cc_css << NVME_CC_CSS_SHIFT) {
1168         case NVME_CC_CSS_NVM:
1169         case NVME_CC_CSS_CSI:
1170                 return true;
1171         default:
1172                 return false;
1173         }
1174 }
1175
1176 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
1177 {
1178         lockdep_assert_held(&ctrl->lock);
1179
1180         /*
1181          * Only I/O controllers should verify iosqes,iocqes.
1182          * Strictly speaking, the spec says a discovery controller
1183          * should verify iosqes,iocqes are zeroed, however that
1184          * would break backwards compatibility, so don't enforce it.
1185          */
1186         if (!nvmet_is_disc_subsys(ctrl->subsys) &&
1187             (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
1188              nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES)) {
1189                 ctrl->csts = NVME_CSTS_CFS;
1190                 return;
1191         }
1192
1193         if (nvmet_cc_mps(ctrl->cc) != 0 ||
1194             nvmet_cc_ams(ctrl->cc) != 0 ||
1195             !nvmet_css_supported(nvmet_cc_css(ctrl->cc))) {
1196                 ctrl->csts = NVME_CSTS_CFS;
1197                 return;
1198         }
1199
1200         ctrl->csts = NVME_CSTS_RDY;
1201
1202         /*
1203          * Controllers that are not yet enabled should not really enforce the
1204          * keep alive timeout, but we still want to track a timeout and cleanup
1205          * in case a host died before it enabled the controller.  Hence, simply
1206          * reset the keep alive timer when the controller is enabled.
1207          */
1208         if (ctrl->kato)
1209                 mod_delayed_work(nvmet_wq, &ctrl->ka_work, ctrl->kato * HZ);
1210 }
1211
1212 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
1213 {
1214         lockdep_assert_held(&ctrl->lock);
1215
1216         /* XXX: tear down queues? */
1217         ctrl->csts &= ~NVME_CSTS_RDY;
1218         ctrl->cc = 0;
1219 }
1220
1221 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
1222 {
1223         u32 old;
1224
1225         mutex_lock(&ctrl->lock);
1226         old = ctrl->cc;
1227         ctrl->cc = new;
1228
1229         if (nvmet_cc_en(new) && !nvmet_cc_en(old))
1230                 nvmet_start_ctrl(ctrl);
1231         if (!nvmet_cc_en(new) && nvmet_cc_en(old))
1232                 nvmet_clear_ctrl(ctrl);
1233         if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
1234                 nvmet_clear_ctrl(ctrl);
1235                 ctrl->csts |= NVME_CSTS_SHST_CMPLT;
1236         }
1237         if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
1238                 ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
1239         mutex_unlock(&ctrl->lock);
1240 }
1241
1242 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
1243 {
1244         /* command sets supported: NVMe command set: */
1245         ctrl->cap = (1ULL << 37);
1246         /* Controller supports one or more I/O Command Sets */
1247         ctrl->cap |= (1ULL << 43);
1248         /* CC.EN timeout in 500msec units: */
1249         ctrl->cap |= (15ULL << 24);
1250         /* maximum queue entries supported: */
1251         if (ctrl->ops->get_max_queue_size)
1252                 ctrl->cap |= min_t(u16, ctrl->ops->get_max_queue_size(ctrl),
1253                                    ctrl->port->max_queue_size) - 1;
1254         else
1255                 ctrl->cap |= ctrl->port->max_queue_size - 1;
1256
1257         if (nvmet_is_passthru_subsys(ctrl->subsys))
1258                 nvmet_passthrough_override_cap(ctrl);
1259 }
1260
1261 struct nvmet_ctrl *nvmet_ctrl_find_get(const char *subsysnqn,
1262                                        const char *hostnqn, u16 cntlid,
1263                                        struct nvmet_req *req)
1264 {
1265         struct nvmet_ctrl *ctrl = NULL;
1266         struct nvmet_subsys *subsys;
1267
1268         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1269         if (!subsys) {
1270                 pr_warn("connect request for invalid subsystem %s!\n",
1271                         subsysnqn);
1272                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1273                 goto out;
1274         }
1275
1276         mutex_lock(&subsys->lock);
1277         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
1278                 if (ctrl->cntlid == cntlid) {
1279                         if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
1280                                 pr_warn("hostnqn mismatch.\n");
1281                                 continue;
1282                         }
1283                         if (!kref_get_unless_zero(&ctrl->ref))
1284                                 continue;
1285
1286                         /* ctrl found */
1287                         goto found;
1288                 }
1289         }
1290
1291         ctrl = NULL; /* ctrl not found */
1292         pr_warn("could not find controller %d for subsys %s / host %s\n",
1293                 cntlid, subsysnqn, hostnqn);
1294         req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
1295
1296 found:
1297         mutex_unlock(&subsys->lock);
1298         nvmet_subsys_put(subsys);
1299 out:
1300         return ctrl;
1301 }
1302
1303 u16 nvmet_check_ctrl_status(struct nvmet_req *req)
1304 {
1305         if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
1306                 pr_err("got cmd %d while CC.EN == 0 on qid = %d\n",
1307                        req->cmd->common.opcode, req->sq->qid);
1308                 return NVME_SC_CMD_SEQ_ERROR | NVME_STATUS_DNR;
1309         }
1310
1311         if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
1312                 pr_err("got cmd %d while CSTS.RDY == 0 on qid = %d\n",
1313                        req->cmd->common.opcode, req->sq->qid);
1314                 return NVME_SC_CMD_SEQ_ERROR | NVME_STATUS_DNR;
1315         }
1316
1317         if (unlikely(!nvmet_check_auth_status(req))) {
1318                 pr_warn("qid %d not authenticated\n", req->sq->qid);
1319                 return NVME_SC_AUTH_REQUIRED | NVME_STATUS_DNR;
1320         }
1321         return 0;
1322 }
1323
1324 bool nvmet_host_allowed(struct nvmet_subsys *subsys, const char *hostnqn)
1325 {
1326         struct nvmet_host_link *p;
1327
1328         lockdep_assert_held(&nvmet_config_sem);
1329
1330         if (subsys->allow_any_host)
1331                 return true;
1332
1333         if (nvmet_is_disc_subsys(subsys)) /* allow all access to disc subsys */
1334                 return true;
1335
1336         list_for_each_entry(p, &subsys->hosts, entry) {
1337                 if (!strcmp(nvmet_host_name(p->host), hostnqn))
1338                         return true;
1339         }
1340
1341         return false;
1342 }
1343
1344 /*
1345  * Note: ctrl->subsys->lock should be held when calling this function
1346  */
1347 static void nvmet_setup_p2p_ns_map(struct nvmet_ctrl *ctrl,
1348                 struct nvmet_req *req)
1349 {
1350         struct nvmet_ns *ns;
1351         unsigned long idx;
1352
1353         if (!req->p2p_client)
1354                 return;
1355
1356         ctrl->p2p_client = get_device(req->p2p_client);
1357
1358         xa_for_each(&ctrl->subsys->namespaces, idx, ns)
1359                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
1360 }
1361
1362 /*
1363  * Note: ctrl->subsys->lock should be held when calling this function
1364  */
1365 static void nvmet_release_p2p_ns_map(struct nvmet_ctrl *ctrl)
1366 {
1367         struct radix_tree_iter iter;
1368         void __rcu **slot;
1369
1370         radix_tree_for_each_slot(slot, &ctrl->p2p_ns_map, &iter, 0)
1371                 pci_dev_put(radix_tree_deref_slot(slot));
1372
1373         put_device(ctrl->p2p_client);
1374 }
1375
1376 static void nvmet_fatal_error_handler(struct work_struct *work)
1377 {
1378         struct nvmet_ctrl *ctrl =
1379                         container_of(work, struct nvmet_ctrl, fatal_err_work);
1380
1381         pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
1382         ctrl->ops->delete_ctrl(ctrl);
1383 }
1384
1385 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
1386                 struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
1387 {
1388         struct nvmet_subsys *subsys;
1389         struct nvmet_ctrl *ctrl;
1390         int ret;
1391         u16 status;
1392
1393         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_STATUS_DNR;
1394         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1395         if (!subsys) {
1396                 pr_warn("connect request for invalid subsystem %s!\n",
1397                         subsysnqn);
1398                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1399                 req->error_loc = offsetof(struct nvme_common_command, dptr);
1400                 goto out;
1401         }
1402
1403         down_read(&nvmet_config_sem);
1404         if (!nvmet_host_allowed(subsys, hostnqn)) {
1405                 pr_info("connect by host %s for subsystem %s not allowed\n",
1406                         hostnqn, subsysnqn);
1407                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
1408                 up_read(&nvmet_config_sem);
1409                 status = NVME_SC_CONNECT_INVALID_HOST | NVME_STATUS_DNR;
1410                 req->error_loc = offsetof(struct nvme_common_command, dptr);
1411                 goto out_put_subsystem;
1412         }
1413         up_read(&nvmet_config_sem);
1414
1415         status = NVME_SC_INTERNAL;
1416         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1417         if (!ctrl)
1418                 goto out_put_subsystem;
1419         mutex_init(&ctrl->lock);
1420
1421         ctrl->port = req->port;
1422         ctrl->ops = req->ops;
1423
1424 #ifdef CONFIG_NVME_TARGET_PASSTHRU
1425         /* By default, set loop targets to clear IDS by default */
1426         if (ctrl->port->disc_addr.trtype == NVMF_TRTYPE_LOOP)
1427                 subsys->clear_ids = 1;
1428 #endif
1429
1430         INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
1431         INIT_LIST_HEAD(&ctrl->async_events);
1432         INIT_RADIX_TREE(&ctrl->p2p_ns_map, GFP_KERNEL);
1433         INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
1434         INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer);
1435
1436         memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
1437         memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
1438
1439         kref_init(&ctrl->ref);
1440         ctrl->subsys = subsys;
1441         ctrl->pi_support = ctrl->port->pi_enable && ctrl->subsys->pi_support;
1442         nvmet_init_cap(ctrl);
1443         WRITE_ONCE(ctrl->aen_enabled, NVMET_AEN_CFG_OPTIONAL);
1444
1445         ctrl->changed_ns_list = kmalloc_array(NVME_MAX_CHANGED_NAMESPACES,
1446                         sizeof(__le32), GFP_KERNEL);
1447         if (!ctrl->changed_ns_list)
1448                 goto out_free_ctrl;
1449
1450         ctrl->sqs = kcalloc(subsys->max_qid + 1,
1451                         sizeof(struct nvmet_sq *),
1452                         GFP_KERNEL);
1453         if (!ctrl->sqs)
1454                 goto out_free_changed_ns_list;
1455
1456         ret = ida_alloc_range(&cntlid_ida,
1457                              subsys->cntlid_min, subsys->cntlid_max,
1458                              GFP_KERNEL);
1459         if (ret < 0) {
1460                 status = NVME_SC_CONNECT_CTRL_BUSY | NVME_STATUS_DNR;
1461                 goto out_free_sqs;
1462         }
1463         ctrl->cntlid = ret;
1464
1465         /*
1466          * Discovery controllers may use some arbitrary high value
1467          * in order to cleanup stale discovery sessions
1468          */
1469         if (nvmet_is_disc_subsys(ctrl->subsys) && !kato)
1470                 kato = NVMET_DISC_KATO_MS;
1471
1472         /* keep-alive timeout in seconds */
1473         ctrl->kato = DIV_ROUND_UP(kato, 1000);
1474
1475         ctrl->err_counter = 0;
1476         spin_lock_init(&ctrl->error_lock);
1477
1478         nvmet_start_keep_alive_timer(ctrl);
1479
1480         mutex_lock(&subsys->lock);
1481         list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
1482         nvmet_setup_p2p_ns_map(ctrl, req);
1483         nvmet_debugfs_ctrl_setup(ctrl);
1484         mutex_unlock(&subsys->lock);
1485
1486         *ctrlp = ctrl;
1487         return 0;
1488
1489 out_free_sqs:
1490         kfree(ctrl->sqs);
1491 out_free_changed_ns_list:
1492         kfree(ctrl->changed_ns_list);
1493 out_free_ctrl:
1494         kfree(ctrl);
1495 out_put_subsystem:
1496         nvmet_subsys_put(subsys);
1497 out:
1498         return status;
1499 }
1500
1501 static void nvmet_ctrl_free(struct kref *ref)
1502 {
1503         struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
1504         struct nvmet_subsys *subsys = ctrl->subsys;
1505
1506         mutex_lock(&subsys->lock);
1507         nvmet_release_p2p_ns_map(ctrl);
1508         list_del(&ctrl->subsys_entry);
1509         mutex_unlock(&subsys->lock);
1510
1511         nvmet_stop_keep_alive_timer(ctrl);
1512
1513         flush_work(&ctrl->async_event_work);
1514         cancel_work_sync(&ctrl->fatal_err_work);
1515
1516         nvmet_destroy_auth(ctrl);
1517
1518         nvmet_debugfs_ctrl_free(ctrl);
1519
1520         ida_free(&cntlid_ida, ctrl->cntlid);
1521
1522         nvmet_async_events_free(ctrl);
1523         kfree(ctrl->sqs);
1524         kfree(ctrl->changed_ns_list);
1525         kfree(ctrl);
1526
1527         nvmet_subsys_put(subsys);
1528 }
1529
1530 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
1531 {
1532         kref_put(&ctrl->ref, nvmet_ctrl_free);
1533 }
1534
1535 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
1536 {
1537         mutex_lock(&ctrl->lock);
1538         if (!(ctrl->csts & NVME_CSTS_CFS)) {
1539                 ctrl->csts |= NVME_CSTS_CFS;
1540                 queue_work(nvmet_wq, &ctrl->fatal_err_work);
1541         }
1542         mutex_unlock(&ctrl->lock);
1543 }
1544 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
1545
1546 ssize_t nvmet_ctrl_host_traddr(struct nvmet_ctrl *ctrl,
1547                 char *traddr, size_t traddr_len)
1548 {
1549         if (!ctrl->ops->host_traddr)
1550                 return -EOPNOTSUPP;
1551         return ctrl->ops->host_traddr(ctrl, traddr, traddr_len);
1552 }
1553
1554 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
1555                 const char *subsysnqn)
1556 {
1557         struct nvmet_subsys_link *p;
1558
1559         if (!port)
1560                 return NULL;
1561
1562         if (!strcmp(NVME_DISC_SUBSYS_NAME, subsysnqn)) {
1563                 if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
1564                         return NULL;
1565                 return nvmet_disc_subsys;
1566         }
1567
1568         down_read(&nvmet_config_sem);
1569         if (!strncmp(nvmet_disc_subsys->subsysnqn, subsysnqn,
1570                                 NVMF_NQN_SIZE)) {
1571                 if (kref_get_unless_zero(&nvmet_disc_subsys->ref)) {
1572                         up_read(&nvmet_config_sem);
1573                         return nvmet_disc_subsys;
1574                 }
1575         }
1576         list_for_each_entry(p, &port->subsystems, entry) {
1577                 if (!strncmp(p->subsys->subsysnqn, subsysnqn,
1578                                 NVMF_NQN_SIZE)) {
1579                         if (!kref_get_unless_zero(&p->subsys->ref))
1580                                 break;
1581                         up_read(&nvmet_config_sem);
1582                         return p->subsys;
1583                 }
1584         }
1585         up_read(&nvmet_config_sem);
1586         return NULL;
1587 }
1588
1589 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
1590                 enum nvme_subsys_type type)
1591 {
1592         struct nvmet_subsys *subsys;
1593         char serial[NVMET_SN_MAX_SIZE / 2];
1594         int ret;
1595
1596         subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
1597         if (!subsys)
1598                 return ERR_PTR(-ENOMEM);
1599
1600         subsys->ver = NVMET_DEFAULT_VS;
1601         /* generate a random serial number as our controllers are ephemeral: */
1602         get_random_bytes(&serial, sizeof(serial));
1603         bin2hex(subsys->serial, &serial, sizeof(serial));
1604
1605         subsys->model_number = kstrdup(NVMET_DEFAULT_CTRL_MODEL, GFP_KERNEL);
1606         if (!subsys->model_number) {
1607                 ret = -ENOMEM;
1608                 goto free_subsys;
1609         }
1610
1611         subsys->ieee_oui = 0;
1612
1613         subsys->firmware_rev = kstrndup(UTS_RELEASE, NVMET_FR_MAX_SIZE, GFP_KERNEL);
1614         if (!subsys->firmware_rev) {
1615                 ret = -ENOMEM;
1616                 goto free_mn;
1617         }
1618
1619         switch (type) {
1620         case NVME_NQN_NVME:
1621                 subsys->max_qid = NVMET_NR_QUEUES;
1622                 break;
1623         case NVME_NQN_DISC:
1624         case NVME_NQN_CURR:
1625                 subsys->max_qid = 0;
1626                 break;
1627         default:
1628                 pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
1629                 ret = -EINVAL;
1630                 goto free_fr;
1631         }
1632         subsys->type = type;
1633         subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
1634                         GFP_KERNEL);
1635         if (!subsys->subsysnqn) {
1636                 ret = -ENOMEM;
1637                 goto free_fr;
1638         }
1639         subsys->cntlid_min = NVME_CNTLID_MIN;
1640         subsys->cntlid_max = NVME_CNTLID_MAX;
1641         kref_init(&subsys->ref);
1642
1643         mutex_init(&subsys->lock);
1644         xa_init(&subsys->namespaces);
1645         INIT_LIST_HEAD(&subsys->ctrls);
1646         INIT_LIST_HEAD(&subsys->hosts);
1647
1648         ret = nvmet_debugfs_subsys_setup(subsys);
1649         if (ret)
1650                 goto free_subsysnqn;
1651
1652         return subsys;
1653
1654 free_subsysnqn:
1655         kfree(subsys->subsysnqn);
1656 free_fr:
1657         kfree(subsys->firmware_rev);
1658 free_mn:
1659         kfree(subsys->model_number);
1660 free_subsys:
1661         kfree(subsys);
1662         return ERR_PTR(ret);
1663 }
1664
1665 static void nvmet_subsys_free(struct kref *ref)
1666 {
1667         struct nvmet_subsys *subsys =
1668                 container_of(ref, struct nvmet_subsys, ref);
1669
1670         WARN_ON_ONCE(!xa_empty(&subsys->namespaces));
1671
1672         nvmet_debugfs_subsys_free(subsys);
1673
1674         xa_destroy(&subsys->namespaces);
1675         nvmet_passthru_subsys_free(subsys);
1676
1677         kfree(subsys->subsysnqn);
1678         kfree(subsys->model_number);
1679         kfree(subsys->firmware_rev);
1680         kfree(subsys);
1681 }
1682
1683 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1684 {
1685         struct nvmet_ctrl *ctrl;
1686
1687         mutex_lock(&subsys->lock);
1688         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1689                 ctrl->ops->delete_ctrl(ctrl);
1690         mutex_unlock(&subsys->lock);
1691 }
1692
1693 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1694 {
1695         kref_put(&subsys->ref, nvmet_subsys_free);
1696 }
1697
1698 static int __init nvmet_init(void)
1699 {
1700         int error = -ENOMEM;
1701
1702         nvmet_ana_group_enabled[NVMET_DEFAULT_ANA_GRPID] = 1;
1703
1704         nvmet_bvec_cache = kmem_cache_create("nvmet-bvec",
1705                         NVMET_MAX_MPOOL_BVEC * sizeof(struct bio_vec), 0,
1706                         SLAB_HWCACHE_ALIGN, NULL);
1707         if (!nvmet_bvec_cache)
1708                 return -ENOMEM;
1709
1710         zbd_wq = alloc_workqueue("nvmet-zbd-wq", WQ_MEM_RECLAIM, 0);
1711         if (!zbd_wq)
1712                 goto out_destroy_bvec_cache;
1713
1714         buffered_io_wq = alloc_workqueue("nvmet-buffered-io-wq",
1715                         WQ_MEM_RECLAIM, 0);
1716         if (!buffered_io_wq)
1717                 goto out_free_zbd_work_queue;
1718
1719         nvmet_wq = alloc_workqueue("nvmet-wq",
1720                         WQ_MEM_RECLAIM | WQ_UNBOUND, 0);
1721         if (!nvmet_wq)
1722                 goto out_free_buffered_work_queue;
1723
1724         error = nvmet_init_discovery();
1725         if (error)
1726                 goto out_free_nvmet_work_queue;
1727
1728         error = nvmet_init_debugfs();
1729         if (error)
1730                 goto out_exit_discovery;
1731
1732         error = nvmet_init_configfs();
1733         if (error)
1734                 goto out_exit_debugfs;
1735
1736         return 0;
1737
1738 out_exit_debugfs:
1739         nvmet_exit_debugfs();
1740 out_exit_discovery:
1741         nvmet_exit_discovery();
1742 out_free_nvmet_work_queue:
1743         destroy_workqueue(nvmet_wq);
1744 out_free_buffered_work_queue:
1745         destroy_workqueue(buffered_io_wq);
1746 out_free_zbd_work_queue:
1747         destroy_workqueue(zbd_wq);
1748 out_destroy_bvec_cache:
1749         kmem_cache_destroy(nvmet_bvec_cache);
1750         return error;
1751 }
1752
1753 static void __exit nvmet_exit(void)
1754 {
1755         nvmet_exit_configfs();
1756         nvmet_exit_debugfs();
1757         nvmet_exit_discovery();
1758         ida_destroy(&cntlid_ida);
1759         destroy_workqueue(nvmet_wq);
1760         destroy_workqueue(buffered_io_wq);
1761         destroy_workqueue(zbd_wq);
1762         kmem_cache_destroy(nvmet_bvec_cache);
1763
1764         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1765         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1766 }
1767
1768 module_init(nvmet_init);
1769 module_exit(nvmet_exit);
1770
1771 MODULE_DESCRIPTION("NVMe target core framework");
1772 MODULE_LICENSE("GPL v2");