drivers/nvme/target/core.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Common code for the NVMe target.
   4  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
   5  */
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7 #include <linux/module.h>
   8 #include <linux/random.h>
   9 #include <linux/rculist.h>
  10 #include <linux/pci-p2pdma.h>
  11 #include <linux/scatterlist.h>
  12
  13 #define CREATE_TRACE_POINTS
  14 #include "trace.h"
  15
  16 #include "nvmet.h"
  17
  18 struct workqueue_struct *buffered_io_wq;
  19 struct workqueue_struct *zbd_wq;
  20 static const struct nvmet_fabrics_ops *nvmet_transports[NVMF_TRTYPE_MAX];
  21 static DEFINE_IDA(cntlid_ida);
  22
  23 /*
  24  * This read/write semaphore is used to synchronize access to configuration
  25  * information on a target system that will result in discovery log page
  26  * information change for at least one host.
  27  * The full list of resources to protected by this semaphore is:
  28  *
  29  *  - subsystems list
  30  *  - per-subsystem allowed hosts list
  31  *  - allow_any_host subsystem attribute
  32  *  - nvmet_genctr
  33  *  - the nvmet_transports array
  34  *
  35  * When updating any of those lists/structures write lock should be obtained,
  36  * while when reading (popolating discovery log page or checking host-subsystem
  37  * link) read lock is obtained to allow concurrent reads.
  38  */
  39 DECLARE_RWSEM(nvmet_config_sem);
  40
  41 u32 nvmet_ana_group_enabled[NVMET_MAX_ANAGRPS + 1];
  42 u64 nvmet_ana_chgcnt;
  43 DECLARE_RWSEM(nvmet_ana_sem);
  44
  45 inline u16 errno_to_nvme_status(struct nvmet_req *req, int errno)
  46 {
  47         switch (errno) {
  48         case 0:
  49                 return NVME_SC_SUCCESS;
  50         case -ENOSPC:
  51                 req->error_loc = offsetof(struct nvme_rw_command, length);
  52                 return NVME_SC_CAP_EXCEEDED | NVME_SC_DNR;
  53         case -EREMOTEIO:
  54                 req->error_loc = offsetof(struct nvme_rw_command, slba);
  55                 return  NVME_SC_LBA_RANGE | NVME_SC_DNR;
  56         case -EOPNOTSUPP:
  57                 req->error_loc = offsetof(struct nvme_common_command, opcode);
  58                 switch (req->cmd->common.opcode) {
  59                 case nvme_cmd_dsm:
  60                 case nvme_cmd_write_zeroes:
  61                         return NVME_SC_ONCS_NOT_SUPPORTED | NVME_SC_DNR;
  62                 default:
  63                         return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
  64                 }
  65                 break;
  66         case -ENODATA:
  67                 req->error_loc = offsetof(struct nvme_rw_command, nsid);
  68                 return NVME_SC_ACCESS_DENIED;
  69         case -EIO:
  70                 fallthrough;
  71         default:
  72                 req->error_loc = offsetof(struct nvme_common_command, opcode);
  73                 return NVME_SC_INTERNAL | NVME_SC_DNR;
  74         }
  75 }
  76
  77 u16 nvmet_report_invalid_opcode(struct nvmet_req *req)
  78 {
  79         pr_debug("unhandled cmd %d on qid %d\n", req->cmd->common.opcode,
  80                  req->sq->qid);
  81
  82         req->error_loc = offsetof(struct nvme_common_command, opcode);
  83         return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
  84 }
  85
  86 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
  87                 const char *subsysnqn);
  88
  89 u16 nvmet_copy_to_sgl(struct nvmet_req *req, off_t off, const void *buf,
  90                 size_t len)
  91 {
  92         if (sg_pcopy_from_buffer(req->sg, req->sg_cnt, buf, len, off) != len) {
  93                 req->error_loc = offsetof(struct nvme_common_command, dptr);
  94                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
  95         }
  96         return 0;
  97 }
  98
  99 u16 nvmet_copy_from_sgl(struct nvmet_req *req, off_t off, void *buf, size_t len)
 100 {
 101         if (sg_pcopy_to_buffer(req->sg, req->sg_cnt, buf, len, off) != len) {
 102                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 103                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
 104         }
 105         return 0;
 106 }
 107
 108 u16 nvmet_zero_sgl(struct nvmet_req *req, off_t off, size_t len)
 109 {
 110         if (sg_zero_buffer(req->sg, req->sg_cnt, len, off) != len) {
 111                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 112                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
 113         }
 114         return 0;
 115 }
 116
 117 static u32 nvmet_max_nsid(struct nvmet_subsys *subsys)
 118 {
 119         struct nvmet_ns *cur;
 120         unsigned long idx;
 121         u32 nsid = 0;
 122
 123         xa_for_each(&subsys->namespaces, idx, cur)
 124                 nsid = cur->nsid;
 125
 126         return nsid;
 127 }
 128
 129 static u32 nvmet_async_event_result(struct nvmet_async_event *aen)
 130 {
 131         return aen->event_type | (aen->event_info << 8) | (aen->log_page << 16);
 132 }
 133
 134 static void nvmet_async_events_failall(struct nvmet_ctrl *ctrl)
 135 {
 136         struct nvmet_req *req;
 137
 138         mutex_lock(&ctrl->lock);
 139         while (ctrl->nr_async_event_cmds) {
 140                 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
 141                 mutex_unlock(&ctrl->lock);
 142                 nvmet_req_complete(req, NVME_SC_INTERNAL | NVME_SC_DNR);
 143                 mutex_lock(&ctrl->lock);
 144         }
 145         mutex_unlock(&ctrl->lock);
 146 }
 147
 148 static void nvmet_async_events_process(struct nvmet_ctrl *ctrl)
 149 {
 150         struct nvmet_async_event *aen;
 151         struct nvmet_req *req;
 152
 153         mutex_lock(&ctrl->lock);
 154         while (ctrl->nr_async_event_cmds && !list_empty(&ctrl->async_events)) {
 155                 aen = list_first_entry(&ctrl->async_events,
 156                                        struct nvmet_async_event, entry);
 157                 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
 158                 nvmet_set_result(req, nvmet_async_event_result(aen));
 159
 160                 list_del(&aen->entry);
 161                 kfree(aen);
 162
 163                 mutex_unlock(&ctrl->lock);
 164                 trace_nvmet_async_event(ctrl, req->cqe->result.u32);
 165                 nvmet_req_complete(req, 0);
 166                 mutex_lock(&ctrl->lock);
 167         }
 168         mutex_unlock(&ctrl->lock);
 169 }
 170
 171 static void nvmet_async_events_free(struct nvmet_ctrl *ctrl)
 172 {
 173         struct nvmet_async_event *aen, *tmp;
 174
 175         mutex_lock(&ctrl->lock);
 176         list_for_each_entry_safe(aen, tmp, &ctrl->async_events, entry) {
 177                 list_del(&aen->entry);
 178                 kfree(aen);
 179         }
 180         mutex_unlock(&ctrl->lock);
 181 }
 182
 183 static void nvmet_async_event_work(struct work_struct *work)
 184 {
 185         struct nvmet_ctrl *ctrl =
 186                 container_of(work, struct nvmet_ctrl, async_event_work);
 187
 188         nvmet_async_events_process(ctrl);
 189 }
 190
 191 void nvmet_add_async_event(struct nvmet_ctrl *ctrl, u8 event_type,
 192                 u8 event_info, u8 log_page)
 193 {
 194         struct nvmet_async_event *aen;
 195
 196         aen = kmalloc(sizeof(*aen), GFP_KERNEL);
 197         if (!aen)
 198                 return;
 199
 200         aen->event_type = event_type;
 201         aen->event_info = event_info;
 202         aen->log_page = log_page;
 203
 204         mutex_lock(&ctrl->lock);
 205         list_add_tail(&aen->entry, &ctrl->async_events);
 206         mutex_unlock(&ctrl->lock);
 207
 208         schedule_work(&ctrl->async_event_work);
 209 }
 210
 211 static void nvmet_add_to_changed_ns_log(struct nvmet_ctrl *ctrl, __le32 nsid)
 212 {
 213         u32 i;
 214
 215         mutex_lock(&ctrl->lock);
 216         if (ctrl->nr_changed_ns > NVME_MAX_CHANGED_NAMESPACES)
 217                 goto out_unlock;
 218
 219         for (i = 0; i < ctrl->nr_changed_ns; i++) {
 220                 if (ctrl->changed_ns_list[i] == nsid)
 221                         goto out_unlock;
 222         }
 223
 224         if (ctrl->nr_changed_ns == NVME_MAX_CHANGED_NAMESPACES) {
 225                 ctrl->changed_ns_list[0] = cpu_to_le32(0xffffffff);
 226                 ctrl->nr_changed_ns = U32_MAX;
 227                 goto out_unlock;
 228         }
 229
 230         ctrl->changed_ns_list[ctrl->nr_changed_ns++] = nsid;
 231 out_unlock:
 232         mutex_unlock(&ctrl->lock);
 233 }
 234
 235 void nvmet_ns_changed(struct nvmet_subsys *subsys, u32 nsid)
 236 {
 237         struct nvmet_ctrl *ctrl;
 238
 239         lockdep_assert_held(&subsys->lock);
 240
 241         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 242                 nvmet_add_to_changed_ns_log(ctrl, cpu_to_le32(nsid));
 243                 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_NS_ATTR))
 244                         continue;
 245                 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
 246                                 NVME_AER_NOTICE_NS_CHANGED,
 247                                 NVME_LOG_CHANGED_NS);
 248         }
 249 }
 250
 251 void nvmet_send_ana_event(struct nvmet_subsys *subsys,
 252                 struct nvmet_port *port)
 253 {
 254         struct nvmet_ctrl *ctrl;
 255
 256         mutex_lock(&subsys->lock);
 257         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 258                 if (port && ctrl->port != port)
 259                         continue;
 260                 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_ANA_CHANGE))
 261                         continue;
 262                 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
 263                                 NVME_AER_NOTICE_ANA, NVME_LOG_ANA);
 264         }
 265         mutex_unlock(&subsys->lock);
 266 }
 267
 268 void nvmet_port_send_ana_event(struct nvmet_port *port)
 269 {
 270         struct nvmet_subsys_link *p;
 271
 272         down_read(&nvmet_config_sem);
 273         list_for_each_entry(p, &port->subsystems, entry)
 274                 nvmet_send_ana_event(p->subsys, port);
 275         up_read(&nvmet_config_sem);
 276 }
 277
 278 int nvmet_register_transport(const struct nvmet_fabrics_ops *ops)
 279 {
 280         int ret = 0;
 281
 282         down_write(&nvmet_config_sem);
 283         if (nvmet_transports[ops->type])
 284                 ret = -EINVAL;
 285         else
 286                 nvmet_transports[ops->type] = ops;
 287         up_write(&nvmet_config_sem);
 288
 289         return ret;
 290 }
 291 EXPORT_SYMBOL_GPL(nvmet_register_transport);
 292
 293 void nvmet_unregister_transport(const struct nvmet_fabrics_ops *ops)
 294 {
 295         down_write(&nvmet_config_sem);
 296         nvmet_transports[ops->type] = NULL;
 297         up_write(&nvmet_config_sem);
 298 }
 299 EXPORT_SYMBOL_GPL(nvmet_unregister_transport);
 300
 301 void nvmet_port_del_ctrls(struct nvmet_port *port, struct nvmet_subsys *subsys)
 302 {
 303         struct nvmet_ctrl *ctrl;
 304
 305         mutex_lock(&subsys->lock);
 306         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 307                 if (ctrl->port == port)
 308                         ctrl->ops->delete_ctrl(ctrl);
 309         }
 310         mutex_unlock(&subsys->lock);
 311 }
 312
 313 int nvmet_enable_port(struct nvmet_port *port)
 314 {
 315         const struct nvmet_fabrics_ops *ops;
 316         int ret;
 317
 318         lockdep_assert_held(&nvmet_config_sem);
 319
 320         ops = nvmet_transports[port->disc_addr.trtype];
 321         if (!ops) {
 322                 up_write(&nvmet_config_sem);
 323                 request_module("nvmet-transport-%d", port->disc_addr.trtype);
 324                 down_write(&nvmet_config_sem);
 325                 ops = nvmet_transports[port->disc_addr.trtype];
 326                 if (!ops) {
 327                         pr_err("transport type %d not supported\n",
 328                                 port->disc_addr.trtype);
 329                         return -EINVAL;
 330                 }
 331         }
 332
 333         if (!try_module_get(ops->owner))
 334                 return -EINVAL;
 335
 336         /*
 337          * If the user requested PI support and the transport isn't pi capable,
 338          * don't enable the port.
 339          */
 340         if (port->pi_enable && !(ops->flags & NVMF_METADATA_SUPPORTED)) {
 341                 pr_err("T10-PI is not supported by transport type %d\n",
 342                        port->disc_addr.trtype);
 343                 ret = -EINVAL;
 344                 goto out_put;
 345         }
 346
 347         ret = ops->add_port(port);
 348         if (ret)
 349                 goto out_put;
 350
 351         /* If the transport didn't set inline_data_size, then disable it. */
 352         if (port->inline_data_size < 0)
 353                 port->inline_data_size = 0;
 354
 355         port->enabled = true;
 356         port->tr_ops = ops;
 357         return 0;
 358
 359 out_put:
 360         module_put(ops->owner);
 361         return ret;
 362 }
 363
 364 void nvmet_disable_port(struct nvmet_port *port)
 365 {
 366         const struct nvmet_fabrics_ops *ops;
 367
 368         lockdep_assert_held(&nvmet_config_sem);
 369
 370         port->enabled = false;
 371         port->tr_ops = NULL;
 372
 373         ops = nvmet_transports[port->disc_addr.trtype];
 374         ops->remove_port(port);
 375         module_put(ops->owner);
 376 }
 377
 378 static void nvmet_keep_alive_timer(struct work_struct *work)
 379 {
 380         struct nvmet_ctrl *ctrl = container_of(to_delayed_work(work),
 381                         struct nvmet_ctrl, ka_work);
 382         bool reset_tbkas = ctrl->reset_tbkas;
 383
 384         ctrl->reset_tbkas = false;
 385         if (reset_tbkas) {
 386                 pr_debug("ctrl %d reschedule traffic based keep-alive timer\n",
 387                         ctrl->cntlid);
 388                 schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
 389                 return;
 390         }
 391
 392         pr_err("ctrl %d keep-alive timer (%d seconds) expired!\n",
 393                 ctrl->cntlid, ctrl->kato);
 394
 395         nvmet_ctrl_fatal_error(ctrl);
 396 }
 397
 398 void nvmet_start_keep_alive_timer(struct nvmet_ctrl *ctrl)
 399 {
 400         if (unlikely(ctrl->kato == 0))
 401                 return;
 402
 403         pr_debug("ctrl %d start keep-alive timer for %d secs\n",
 404                 ctrl->cntlid, ctrl->kato);
 405
 406         schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
 407 }
 408
 409 void nvmet_stop_keep_alive_timer(struct nvmet_ctrl *ctrl)
 410 {
 411         if (unlikely(ctrl->kato == 0))
 412                 return;
 413
 414         pr_debug("ctrl %d stop keep-alive\n", ctrl->cntlid);
 415
 416         cancel_delayed_work_sync(&ctrl->ka_work);
 417 }
 418
 419 u16 nvmet_req_find_ns(struct nvmet_req *req)
 420 {
 421         u32 nsid = le32_to_cpu(req->cmd->common.nsid);
 422
 423         req->ns = xa_load(&nvmet_req_subsys(req)->namespaces, nsid);
 424         if (unlikely(!req->ns)) {
 425                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 426                 return NVME_SC_INVALID_NS | NVME_SC_DNR;
 427         }
 428
 429         percpu_ref_get(&req->ns->ref);
 430         return NVME_SC_SUCCESS;
 431 }
 432
 433 static void nvmet_destroy_namespace(struct percpu_ref *ref)
 434 {
 435         struct nvmet_ns *ns = container_of(ref, struct nvmet_ns, ref);
 436
 437         complete(&ns->disable_done);
 438 }
 439
 440 void nvmet_put_namespace(struct nvmet_ns *ns)
 441 {
 442         percpu_ref_put(&ns->ref);
 443 }
 444
 445 static void nvmet_ns_dev_disable(struct nvmet_ns *ns)
 446 {
 447         nvmet_bdev_ns_disable(ns);
 448         nvmet_file_ns_disable(ns);
 449 }
 450
 451 static int nvmet_p2pmem_ns_enable(struct nvmet_ns *ns)
 452 {
 453         int ret;
 454         struct pci_dev *p2p_dev;
 455
 456         if (!ns->use_p2pmem)
 457                 return 0;
 458
 459         if (!ns->bdev) {
 460                 pr_err("peer-to-peer DMA is not supported by non-block device namespaces\n");
 461                 return -EINVAL;
 462         }
 463
 464         if (!blk_queue_pci_p2pdma(ns->bdev->bd_disk->queue)) {
 465                 pr_err("peer-to-peer DMA is not supported by the driver of %s\n",
 466                        ns->device_path);
 467                 return -EINVAL;
 468         }
 469
 470         if (ns->p2p_dev) {
 471                 ret = pci_p2pdma_distance(ns->p2p_dev, nvmet_ns_dev(ns), true);
 472                 if (ret < 0)
 473                         return -EINVAL;
 474         } else {
 475                 /*
 476                  * Right now we just check that there is p2pmem available so
 477                  * we can report an error to the user right away if there
 478                  * is not. We'll find the actual device to use once we
 479                  * setup the controller when the port's device is available.
 480                  */
 481
 482                 p2p_dev = pci_p2pmem_find(nvmet_ns_dev(ns));
 483                 if (!p2p_dev) {
 484                         pr_err("no peer-to-peer memory is available for %s\n",
 485                                ns->device_path);
 486                         return -EINVAL;
 487                 }
 488
 489                 pci_dev_put(p2p_dev);
 490         }
 491
 492         return 0;
 493 }
 494
 495 /*
 496  * Note: ctrl->subsys->lock should be held when calling this function
 497  */
 498 static void nvmet_p2pmem_ns_add_p2p(struct nvmet_ctrl *ctrl,
 499                                     struct nvmet_ns *ns)
 500 {
 501         struct device *clients[2];
 502         struct pci_dev *p2p_dev;
 503         int ret;
 504
 505         if (!ctrl->p2p_client || !ns->use_p2pmem)
 506                 return;
 507
 508         if (ns->p2p_dev) {
 509                 ret = pci_p2pdma_distance(ns->p2p_dev, ctrl->p2p_client, true);
 510                 if (ret < 0)
 511                         return;
 512
 513                 p2p_dev = pci_dev_get(ns->p2p_dev);
 514         } else {
 515                 clients[0] = ctrl->p2p_client;
 516                 clients[1] = nvmet_ns_dev(ns);
 517
 518                 p2p_dev = pci_p2pmem_find_many(clients, ARRAY_SIZE(clients));
 519                 if (!p2p_dev) {
 520                         pr_err("no peer-to-peer memory is available that's supported by %s and %s\n",
 521                                dev_name(ctrl->p2p_client), ns->device_path);
 522                         return;
 523                 }
 524         }
 525
 526         ret = radix_tree_insert(&ctrl->p2p_ns_map, ns->nsid, p2p_dev);
 527         if (ret < 0)
 528                 pci_dev_put(p2p_dev);
 529
 530         pr_info("using p2pmem on %s for nsid %d\n", pci_name(p2p_dev),
 531                 ns->nsid);
 532 }
 533
 534 bool nvmet_ns_revalidate(struct nvmet_ns *ns)
 535 {
 536         loff_t oldsize = ns->size;
 537
 538         if (ns->bdev)
 539                 nvmet_bdev_ns_revalidate(ns);
 540         else
 541                 nvmet_file_ns_revalidate(ns);
 542
 543         return oldsize != ns->size;
 544 }
 545
 546 int nvmet_ns_enable(struct nvmet_ns *ns)
 547 {
 548         struct nvmet_subsys *subsys = ns->subsys;
 549         struct nvmet_ctrl *ctrl;
 550         int ret;
 551
 552         mutex_lock(&subsys->lock);
 553         ret = 0;
 554
 555         if (nvmet_is_passthru_subsys(subsys)) {
 556                 pr_info("cannot enable both passthru and regular namespaces for a single subsystem");
 557                 goto out_unlock;
 558         }
 559
 560         if (ns->enabled)
 561                 goto out_unlock;
 562
 563         ret = -EMFILE;
 564         if (subsys->nr_namespaces == NVMET_MAX_NAMESPACES)
 565                 goto out_unlock;
 566
 567         ret = nvmet_bdev_ns_enable(ns);
 568         if (ret == -ENOTBLK)
 569                 ret = nvmet_file_ns_enable(ns);
 570         if (ret)
 571                 goto out_unlock;
 572
 573         ret = nvmet_p2pmem_ns_enable(ns);
 574         if (ret)
 575                 goto out_dev_disable;
 576
 577         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 578                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
 579
 580         ret = percpu_ref_init(&ns->ref, nvmet_destroy_namespace,
 581                                 0, GFP_KERNEL);
 582         if (ret)
 583                 goto out_dev_put;
 584
 585         if (ns->nsid > subsys->max_nsid)
 586                 subsys->max_nsid = ns->nsid;
 587
 588         ret = xa_insert(&subsys->namespaces, ns->nsid, ns, GFP_KERNEL);
 589         if (ret)
 590                 goto out_restore_subsys_maxnsid;
 591
 592         subsys->nr_namespaces++;
 593
 594         nvmet_ns_changed(subsys, ns->nsid);
 595         ns->enabled = true;
 596         ret = 0;
 597 out_unlock:
 598         mutex_unlock(&subsys->lock);
 599         return ret;
 600
 601 out_restore_subsys_maxnsid:
 602         subsys->max_nsid = nvmet_max_nsid(subsys);
 603         percpu_ref_exit(&ns->ref);
 604 out_dev_put:
 605         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 606                 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
 607 out_dev_disable:
 608         nvmet_ns_dev_disable(ns);
 609         goto out_unlock;
 610 }
 611
 612 void nvmet_ns_disable(struct nvmet_ns *ns)
 613 {
 614         struct nvmet_subsys *subsys = ns->subsys;
 615         struct nvmet_ctrl *ctrl;
 616
 617         mutex_lock(&subsys->lock);
 618         if (!ns->enabled)
 619                 goto out_unlock;
 620
 621         ns->enabled = false;
 622         xa_erase(&ns->subsys->namespaces, ns->nsid);
 623         if (ns->nsid == subsys->max_nsid)
 624                 subsys->max_nsid = nvmet_max_nsid(subsys);
 625
 626         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 627                 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
 628
 629         mutex_unlock(&subsys->lock);
 630
 631         /*
 632          * Now that we removed the namespaces from the lookup list, we
 633          * can kill the per_cpu ref and wait for any remaining references
 634          * to be dropped, as well as a RCU grace period for anyone only
 635          * using the namepace under rcu_read_lock().  Note that we can't
 636          * use call_rcu here as we need to ensure the namespaces have
 637          * been fully destroyed before unloading the module.
 638          */
 639         percpu_ref_kill(&ns->ref);
 640         synchronize_rcu();
 641         wait_for_completion(&ns->disable_done);
 642         percpu_ref_exit(&ns->ref);
 643
 644         mutex_lock(&subsys->lock);
 645
 646         subsys->nr_namespaces--;
 647         nvmet_ns_changed(subsys, ns->nsid);
 648         nvmet_ns_dev_disable(ns);
 649 out_unlock:
 650         mutex_unlock(&subsys->lock);
 651 }
 652
 653 void nvmet_ns_free(struct nvmet_ns *ns)
 654 {
 655         nvmet_ns_disable(ns);
 656
 657         down_write(&nvmet_ana_sem);
 658         nvmet_ana_group_enabled[ns->anagrpid]--;
 659         up_write(&nvmet_ana_sem);
 660
 661         kfree(ns->device_path);
 662         kfree(ns);
 663 }
 664
 665 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
 666 {
 667         struct nvmet_ns *ns;
 668
 669         ns = kzalloc(sizeof(*ns), GFP_KERNEL);
 670         if (!ns)
 671                 return NULL;
 672
 673         init_completion(&ns->disable_done);
 674
 675         ns->nsid = nsid;
 676         ns->subsys = subsys;
 677
 678         down_write(&nvmet_ana_sem);
 679         ns->anagrpid = NVMET_DEFAULT_ANA_GRPID;
 680         nvmet_ana_group_enabled[ns->anagrpid]++;
 681         up_write(&nvmet_ana_sem);
 682
 683         uuid_gen(&ns->uuid);
 684         ns->buffered_io = false;
 685         ns->csi = NVME_CSI_NVM;
 686
 687         return ns;
 688 }
 689
 690 static void nvmet_update_sq_head(struct nvmet_req *req)
 691 {
 692         if (req->sq->size) {
 693                 u32 old_sqhd, new_sqhd;
 694
 695                 do {
 696                         old_sqhd = req->sq->sqhd;
 697                         new_sqhd = (old_sqhd + 1) % req->sq->size;
 698                 } while (cmpxchg(&req->sq->sqhd, old_sqhd, new_sqhd) !=
 699                                         old_sqhd);
 700         }
 701         req->cqe->sq_head = cpu_to_le16(req->sq->sqhd & 0x0000FFFF);
 702 }
 703
 704 static void nvmet_set_error(struct nvmet_req *req, u16 status)
 705 {
 706         struct nvmet_ctrl *ctrl = req->sq->ctrl;
 707         struct nvme_error_slot *new_error_slot;
 708         unsigned long flags;
 709
 710         req->cqe->status = cpu_to_le16(status << 1);
 711
 712         if (!ctrl || req->error_loc == NVMET_NO_ERROR_LOC)
 713                 return;
 714
 715         spin_lock_irqsave(&ctrl->error_lock, flags);
 716         ctrl->err_counter++;
 717         new_error_slot =
 718                 &ctrl->slots[ctrl->err_counter % NVMET_ERROR_LOG_SLOTS];
 719
 720         new_error_slot->error_count = cpu_to_le64(ctrl->err_counter);
 721         new_error_slot->sqid = cpu_to_le16(req->sq->qid);
 722         new_error_slot->cmdid = cpu_to_le16(req->cmd->common.command_id);
 723         new_error_slot->status_field = cpu_to_le16(status << 1);
 724         new_error_slot->param_error_location = cpu_to_le16(req->error_loc);
 725         new_error_slot->lba = cpu_to_le64(req->error_slba);
 726         new_error_slot->nsid = req->cmd->common.nsid;
 727         spin_unlock_irqrestore(&ctrl->error_lock, flags);
 728
 729         /* set the more bit for this request */
 730         req->cqe->status |= cpu_to_le16(1 << 14);
 731 }
 732
 733 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
 734 {
 735         if (!req->sq->sqhd_disabled)
 736                 nvmet_update_sq_head(req);
 737         req->cqe->sq_id = cpu_to_le16(req->sq->qid);
 738         req->cqe->command_id = req->cmd->common.command_id;
 739
 740         if (unlikely(status))
 741                 nvmet_set_error(req, status);
 742
 743         trace_nvmet_req_complete(req);
 744
 745         if (req->ns)
 746                 nvmet_put_namespace(req->ns);
 747         req->ops->queue_response(req);
 748 }
 749
 750 void nvmet_req_complete(struct nvmet_req *req, u16 status)
 751 {
 752         __nvmet_req_complete(req, status);
 753         percpu_ref_put(&req->sq->ref);
 754 }
 755 EXPORT_SYMBOL_GPL(nvmet_req_complete);
 756
 757 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
 758                 u16 qid, u16 size)
 759 {
 760         cq->qid = qid;
 761         cq->size = size;
 762 }
 763
 764 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
 765                 u16 qid, u16 size)
 766 {
 767         sq->sqhd = 0;
 768         sq->qid = qid;
 769         sq->size = size;
 770
 771         ctrl->sqs[qid] = sq;
 772 }
 773
 774 static void nvmet_confirm_sq(struct percpu_ref *ref)
 775 {
 776         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 777
 778         complete(&sq->confirm_done);
 779 }
 780
 781 void nvmet_sq_destroy(struct nvmet_sq *sq)
 782 {
 783         struct nvmet_ctrl *ctrl = sq->ctrl;
 784
 785         /*
 786          * If this is the admin queue, complete all AERs so that our
 787          * queue doesn't have outstanding requests on it.
 788          */
 789         if (ctrl && ctrl->sqs && ctrl->sqs[0] == sq)
 790                 nvmet_async_events_failall(ctrl);
 791         percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
 792         wait_for_completion(&sq->confirm_done);
 793         wait_for_completion(&sq->free_done);
 794         percpu_ref_exit(&sq->ref);
 795
 796         if (ctrl) {
 797                 /*
 798                  * The teardown flow may take some time, and the host may not
 799                  * send us keep-alive during this period, hence reset the
 800                  * traffic based keep-alive timer so we don't trigger a
 801                  * controller teardown as a result of a keep-alive expiration.
 802                  */
 803                 ctrl->reset_tbkas = true;
 804                 sq->ctrl->sqs[sq->qid] = NULL;
 805                 nvmet_ctrl_put(ctrl);
 806                 sq->ctrl = NULL; /* allows reusing the queue later */
 807         }
 808 }
 809 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
 810
 811 static void nvmet_sq_free(struct percpu_ref *ref)
 812 {
 813         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 814
 815         complete(&sq->free_done);
 816 }
 817
 818 int nvmet_sq_init(struct nvmet_sq *sq)
 819 {
 820         int ret;
 821
 822         ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
 823         if (ret) {
 824                 pr_err("percpu_ref init failed!\n");
 825                 return ret;
 826         }
 827         init_completion(&sq->free_done);
 828         init_completion(&sq->confirm_done);
 829
 830         return 0;
 831 }
 832 EXPORT_SYMBOL_GPL(nvmet_sq_init);
 833
 834 static inline u16 nvmet_check_ana_state(struct nvmet_port *port,
 835                 struct nvmet_ns *ns)
 836 {
 837         enum nvme_ana_state state = port->ana_state[ns->anagrpid];
 838
 839         if (unlikely(state == NVME_ANA_INACCESSIBLE))
 840                 return NVME_SC_ANA_INACCESSIBLE;
 841         if (unlikely(state == NVME_ANA_PERSISTENT_LOSS))
 842                 return NVME_SC_ANA_PERSISTENT_LOSS;
 843         if (unlikely(state == NVME_ANA_CHANGE))
 844                 return NVME_SC_ANA_TRANSITION;
 845         return 0;
 846 }
 847
 848 static inline u16 nvmet_io_cmd_check_access(struct nvmet_req *req)
 849 {
 850         if (unlikely(req->ns->readonly)) {
 851                 switch (req->cmd->common.opcode) {
 852                 case nvme_cmd_read:
 853                 case nvme_cmd_flush:
 854                         break;
 855                 default:
 856                         return NVME_SC_NS_WRITE_PROTECTED;
 857                 }
 858         }
 859
 860         return 0;
 861 }
 862
 863 static u16 nvmet_parse_io_cmd(struct nvmet_req *req)
 864 {
 865         u16 ret;
 866
 867         ret = nvmet_check_ctrl_status(req);
 868         if (unlikely(ret))
 869                 return ret;
 870
 871         if (nvmet_is_passthru_req(req))
 872                 return nvmet_parse_passthru_io_cmd(req);
 873
 874         ret = nvmet_req_find_ns(req);
 875         if (unlikely(ret))
 876                 return ret;
 877
 878         ret = nvmet_check_ana_state(req->port, req->ns);
 879         if (unlikely(ret)) {
 880                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 881                 return ret;
 882         }
 883         ret = nvmet_io_cmd_check_access(req);
 884         if (unlikely(ret)) {
 885                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 886                 return ret;
 887         }
 888
 889         switch (req->ns->csi) {
 890         case NVME_CSI_NVM:
 891                 if (req->ns->file)
 892                         return nvmet_file_parse_io_cmd(req);
 893                 return nvmet_bdev_parse_io_cmd(req);
 894         case NVME_CSI_ZNS:
 895                 if (IS_ENABLED(CONFIG_BLK_DEV_ZONED))
 896                         return nvmet_bdev_zns_parse_io_cmd(req);
 897                 return NVME_SC_INVALID_IO_CMD_SET;
 898         default:
 899                 return NVME_SC_INVALID_IO_CMD_SET;
 900         }
 901 }
 902
 903 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
 904                 struct nvmet_sq *sq, const struct nvmet_fabrics_ops *ops)
 905 {
 906         u8 flags = req->cmd->common.flags;
 907         u16 status;
 908
 909         req->cq = cq;
 910         req->sq = sq;
 911         req->ops = ops;
 912         req->sg = NULL;
 913         req->metadata_sg = NULL;
 914         req->sg_cnt = 0;
 915         req->metadata_sg_cnt = 0;
 916         req->transfer_len = 0;
 917         req->metadata_len = 0;
 918         req->cqe->status = 0;
 919         req->cqe->sq_head = 0;
 920         req->ns = NULL;
 921         req->error_loc = NVMET_NO_ERROR_LOC;
 922         req->error_slba = 0;
 923
 924         /* no support for fused commands yet */
 925         if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
 926                 req->error_loc = offsetof(struct nvme_common_command, flags);
 927                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 928                 goto fail;
 929         }
 930
 931         /*
 932          * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
 933          * contains an address of a single contiguous physical buffer that is
 934          * byte aligned.
 935          */
 936         if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
 937                 req->error_loc = offsetof(struct nvme_common_command, flags);
 938                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 939                 goto fail;
 940         }
 941
 942         if (unlikely(!req->sq->ctrl))
 943                 /* will return an error for any non-connect command: */
 944                 status = nvmet_parse_connect_cmd(req);
 945         else if (likely(req->sq->qid != 0))
 946                 status = nvmet_parse_io_cmd(req);
 947         else
 948                 status = nvmet_parse_admin_cmd(req);
 949
 950         if (status)
 951                 goto fail;
 952
 953         trace_nvmet_req_init(req, req->cmd);
 954
 955         if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
 956                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 957                 goto fail;
 958         }
 959
 960         if (sq->ctrl)
 961                 sq->ctrl->reset_tbkas = true;
 962
 963         return true;
 964
 965 fail:
 966         __nvmet_req_complete(req, status);
 967         return false;
 968 }
 969 EXPORT_SYMBOL_GPL(nvmet_req_init);
 970
 971 void nvmet_req_uninit(struct nvmet_req *req)
 972 {
 973         percpu_ref_put(&req->sq->ref);
 974         if (req->ns)
 975                 nvmet_put_namespace(req->ns);
 976 }
 977 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
 978
 979 bool nvmet_check_transfer_len(struct nvmet_req *req, size_t len)
 980 {
 981         if (unlikely(len != req->transfer_len)) {
 982                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 983                 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
 984                 return false;
 985         }
 986
 987         return true;
 988 }
 989 EXPORT_SYMBOL_GPL(nvmet_check_transfer_len);
 990
 991 bool nvmet_check_data_len_lte(struct nvmet_req *req, size_t data_len)
 992 {
 993         if (unlikely(data_len > req->transfer_len)) {
 994                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 995                 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
 996                 return false;
 997         }
 998
 999         return true;
1000 }
1001
1002 static unsigned int nvmet_data_transfer_len(struct nvmet_req *req)
1003 {
1004         return req->transfer_len - req->metadata_len;
1005 }
1006
1007 static int nvmet_req_alloc_p2pmem_sgls(struct pci_dev *p2p_dev,
1008                 struct nvmet_req *req)
1009 {
1010         req->sg = pci_p2pmem_alloc_sgl(p2p_dev, &req->sg_cnt,
1011                         nvmet_data_transfer_len(req));
1012         if (!req->sg)
1013                 goto out_err;
1014
1015         if (req->metadata_len) {
1016                 req->metadata_sg = pci_p2pmem_alloc_sgl(p2p_dev,
1017                                 &req->metadata_sg_cnt, req->metadata_len);
1018                 if (!req->metadata_sg)
1019                         goto out_free_sg;
1020         }
1021
1022         req->p2p_dev = p2p_dev;
1023
1024         return 0;
1025 out_free_sg:
1026         pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1027 out_err:
1028         return -ENOMEM;
1029 }
1030
1031 static struct pci_dev *nvmet_req_find_p2p_dev(struct nvmet_req *req)
1032 {
1033         if (!IS_ENABLED(CONFIG_PCI_P2PDMA) ||
1034             !req->sq->ctrl || !req->sq->qid || !req->ns)
1035                 return NULL;
1036         return radix_tree_lookup(&req->sq->ctrl->p2p_ns_map, req->ns->nsid);
1037 }
1038
1039 int nvmet_req_alloc_sgls(struct nvmet_req *req)
1040 {
1041         struct pci_dev *p2p_dev = nvmet_req_find_p2p_dev(req);
1042
1043         if (p2p_dev && !nvmet_req_alloc_p2pmem_sgls(p2p_dev, req))
1044                 return 0;
1045
1046         req->sg = sgl_alloc(nvmet_data_transfer_len(req), GFP_KERNEL,
1047                             &req->sg_cnt);
1048         if (unlikely(!req->sg))
1049                 goto out;
1050
1051         if (req->metadata_len) {
1052                 req->metadata_sg = sgl_alloc(req->metadata_len, GFP_KERNEL,
1053                                              &req->metadata_sg_cnt);
1054                 if (unlikely(!req->metadata_sg))
1055                         goto out_free;
1056         }
1057
1058         return 0;
1059 out_free:
1060         sgl_free(req->sg);
1061 out:
1062         return -ENOMEM;
1063 }
1064 EXPORT_SYMBOL_GPL(nvmet_req_alloc_sgls);
1065
1066 void nvmet_req_free_sgls(struct nvmet_req *req)
1067 {
1068         if (req->p2p_dev) {
1069                 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1070                 if (req->metadata_sg)
1071                         pci_p2pmem_free_sgl(req->p2p_dev, req->metadata_sg);
1072                 req->p2p_dev = NULL;
1073         } else {
1074                 sgl_free(req->sg);
1075                 if (req->metadata_sg)
1076                         sgl_free(req->metadata_sg);
1077         }
1078
1079         req->sg = NULL;
1080         req->metadata_sg = NULL;
1081         req->sg_cnt = 0;
1082         req->metadata_sg_cnt = 0;
1083 }
1084 EXPORT_SYMBOL_GPL(nvmet_req_free_sgls);
1085
1086 static inline bool nvmet_cc_en(u32 cc)
1087 {
1088         return (cc >> NVME_CC_EN_SHIFT) & 0x1;
1089 }
1090
1091 static inline u8 nvmet_cc_css(u32 cc)
1092 {
1093         return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
1094 }
1095
1096 static inline u8 nvmet_cc_mps(u32 cc)
1097 {
1098         return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
1099 }
1100
1101 static inline u8 nvmet_cc_ams(u32 cc)
1102 {
1103         return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
1104 }
1105
1106 static inline u8 nvmet_cc_shn(u32 cc)
1107 {
1108         return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
1109 }
1110
1111 static inline u8 nvmet_cc_iosqes(u32 cc)
1112 {
1113         return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
1114 }
1115
1116 static inline u8 nvmet_cc_iocqes(u32 cc)
1117 {
1118         return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
1119 }
1120
1121 static inline bool nvmet_css_supported(u8 cc_css)
1122 {
1123         switch (cc_css <<= NVME_CC_CSS_SHIFT) {
1124         case NVME_CC_CSS_NVM:
1125         case NVME_CC_CSS_CSI:
1126                 return true;
1127         default:
1128                 return false;
1129         }
1130 }
1131
1132 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
1133 {
1134         lockdep_assert_held(&ctrl->lock);
1135
1136         /*
1137          * Only I/O controllers should verify iosqes,iocqes.
1138          * Strictly speaking, the spec says a discovery controller
1139          * should verify iosqes,iocqes are zeroed, however that
1140          * would break backwards compatibility, so don't enforce it.
1141          */
1142         if (!nvmet_is_disc_subsys(ctrl->subsys) &&
1143             (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
1144              nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES)) {
1145                 ctrl->csts = NVME_CSTS_CFS;
1146                 return;
1147         }
1148
1149         if (nvmet_cc_mps(ctrl->cc) != 0 ||
1150             nvmet_cc_ams(ctrl->cc) != 0 ||
1151             !nvmet_css_supported(nvmet_cc_css(ctrl->cc))) {
1152                 ctrl->csts = NVME_CSTS_CFS;
1153                 return;
1154         }
1155
1156         ctrl->csts = NVME_CSTS_RDY;
1157
1158         /*
1159          * Controllers that are not yet enabled should not really enforce the
1160          * keep alive timeout, but we still want to track a timeout and cleanup
1161          * in case a host died before it enabled the controller.  Hence, simply
1162          * reset the keep alive timer when the controller is enabled.
1163          */
1164         if (ctrl->kato)
1165                 mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ);
1166 }
1167
1168 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
1169 {
1170         lockdep_assert_held(&ctrl->lock);
1171
1172         /* XXX: tear down queues? */
1173         ctrl->csts &= ~NVME_CSTS_RDY;
1174         ctrl->cc = 0;
1175 }
1176
1177 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
1178 {
1179         u32 old;
1180
1181         mutex_lock(&ctrl->lock);
1182         old = ctrl->cc;
1183         ctrl->cc = new;
1184
1185         if (nvmet_cc_en(new) && !nvmet_cc_en(old))
1186                 nvmet_start_ctrl(ctrl);
1187         if (!nvmet_cc_en(new) && nvmet_cc_en(old))
1188                 nvmet_clear_ctrl(ctrl);
1189         if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
1190                 nvmet_clear_ctrl(ctrl);
1191                 ctrl->csts |= NVME_CSTS_SHST_CMPLT;
1192         }
1193         if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
1194                 ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
1195         mutex_unlock(&ctrl->lock);
1196 }
1197
1198 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
1199 {
1200         /* command sets supported: NVMe command set: */
1201         ctrl->cap = (1ULL << 37);
1202         /* Controller supports one or more I/O Command Sets */
1203         ctrl->cap |= (1ULL << 43);
1204         /* CC.EN timeout in 500msec units: */
1205         ctrl->cap |= (15ULL << 24);
1206         /* maximum queue entries supported: */
1207         if (ctrl->ops->get_max_queue_size)
1208                 ctrl->cap |= ctrl->ops->get_max_queue_size(ctrl) - 1;
1209         else
1210                 ctrl->cap |= NVMET_QUEUE_SIZE - 1;
1211
1212         if (nvmet_is_passthru_subsys(ctrl->subsys))
1213                 nvmet_passthrough_override_cap(ctrl);
1214 }
1215
1216 struct nvmet_ctrl *nvmet_ctrl_find_get(const char *subsysnqn,
1217                                        const char *hostnqn, u16 cntlid,
1218                                        struct nvmet_req *req)
1219 {
1220         struct nvmet_ctrl *ctrl = NULL;
1221         struct nvmet_subsys *subsys;
1222
1223         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1224         if (!subsys) {
1225                 pr_warn("connect request for invalid subsystem %s!\n",
1226                         subsysnqn);
1227                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1228                 goto out;
1229         }
1230
1231         mutex_lock(&subsys->lock);
1232         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
1233                 if (ctrl->cntlid == cntlid) {
1234                         if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
1235                                 pr_warn("hostnqn mismatch.\n");
1236                                 continue;
1237                         }
1238                         if (!kref_get_unless_zero(&ctrl->ref))
1239                                 continue;
1240
1241                         /* ctrl found */
1242                         goto found;
1243                 }
1244         }
1245
1246         ctrl = NULL; /* ctrl not found */
1247         pr_warn("could not find controller %d for subsys %s / host %s\n",
1248                 cntlid, subsysnqn, hostnqn);
1249         req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
1250
1251 found:
1252         mutex_unlock(&subsys->lock);
1253         nvmet_subsys_put(subsys);
1254 out:
1255         return ctrl;
1256 }
1257
1258 u16 nvmet_check_ctrl_status(struct nvmet_req *req)
1259 {
1260         if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
1261                 pr_err("got cmd %d while CC.EN == 0 on qid = %d\n",
1262                        req->cmd->common.opcode, req->sq->qid);
1263                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1264         }
1265
1266         if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
1267                 pr_err("got cmd %d while CSTS.RDY == 0 on qid = %d\n",
1268                        req->cmd->common.opcode, req->sq->qid);
1269                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1270         }
1271         return 0;
1272 }
1273
1274 bool nvmet_host_allowed(struct nvmet_subsys *subsys, const char *hostnqn)
1275 {
1276         struct nvmet_host_link *p;
1277
1278         lockdep_assert_held(&nvmet_config_sem);
1279
1280         if (subsys->allow_any_host)
1281                 return true;
1282
1283         if (nvmet_is_disc_subsys(subsys)) /* allow all access to disc subsys */
1284                 return true;
1285
1286         list_for_each_entry(p, &subsys->hosts, entry) {
1287                 if (!strcmp(nvmet_host_name(p->host), hostnqn))
1288                         return true;
1289         }
1290
1291         return false;
1292 }
1293
1294 /*
1295  * Note: ctrl->subsys->lock should be held when calling this function
1296  */
1297 static void nvmet_setup_p2p_ns_map(struct nvmet_ctrl *ctrl,
1298                 struct nvmet_req *req)
1299 {
1300         struct nvmet_ns *ns;
1301         unsigned long idx;
1302
1303         if (!req->p2p_client)
1304                 return;
1305
1306         ctrl->p2p_client = get_device(req->p2p_client);
1307
1308         xa_for_each(&ctrl->subsys->namespaces, idx, ns)
1309                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
1310 }
1311
1312 /*
1313  * Note: ctrl->subsys->lock should be held when calling this function
1314  */
1315 static void nvmet_release_p2p_ns_map(struct nvmet_ctrl *ctrl)
1316 {
1317         struct radix_tree_iter iter;
1318         void __rcu **slot;
1319
1320         radix_tree_for_each_slot(slot, &ctrl->p2p_ns_map, &iter, 0)
1321                 pci_dev_put(radix_tree_deref_slot(slot));
1322
1323         put_device(ctrl->p2p_client);
1324 }
1325
1326 static void nvmet_fatal_error_handler(struct work_struct *work)
1327 {
1328         struct nvmet_ctrl *ctrl =
1329                         container_of(work, struct nvmet_ctrl, fatal_err_work);
1330
1331         pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
1332         ctrl->ops->delete_ctrl(ctrl);
1333 }
1334
1335 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
1336                 struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
1337 {
1338         struct nvmet_subsys *subsys;
1339         struct nvmet_ctrl *ctrl;
1340         int ret;
1341         u16 status;
1342
1343         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1344         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1345         if (!subsys) {
1346                 pr_warn("connect request for invalid subsystem %s!\n",
1347                         subsysnqn);
1348                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1349                 req->error_loc = offsetof(struct nvme_common_command, dptr);
1350                 goto out;
1351         }
1352
1353         down_read(&nvmet_config_sem);
1354         if (!nvmet_host_allowed(subsys, hostnqn)) {
1355                 pr_info("connect by host %s for subsystem %s not allowed\n",
1356                         hostnqn, subsysnqn);
1357                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
1358                 up_read(&nvmet_config_sem);
1359                 status = NVME_SC_CONNECT_INVALID_HOST | NVME_SC_DNR;
1360                 req->error_loc = offsetof(struct nvme_common_command, dptr);
1361                 goto out_put_subsystem;
1362         }
1363         up_read(&nvmet_config_sem);
1364
1365         status = NVME_SC_INTERNAL;
1366         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1367         if (!ctrl)
1368                 goto out_put_subsystem;
1369         mutex_init(&ctrl->lock);
1370
1371         ctrl->port = req->port;
1372         ctrl->ops = req->ops;
1373
1374         INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
1375         INIT_LIST_HEAD(&ctrl->async_events);
1376         INIT_RADIX_TREE(&ctrl->p2p_ns_map, GFP_KERNEL);
1377         INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
1378         INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer);
1379
1380         memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
1381         memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
1382
1383         kref_init(&ctrl->ref);
1384         ctrl->subsys = subsys;
1385         nvmet_init_cap(ctrl);
1386         WRITE_ONCE(ctrl->aen_enabled, NVMET_AEN_CFG_OPTIONAL);
1387
1388         ctrl->changed_ns_list = kmalloc_array(NVME_MAX_CHANGED_NAMESPACES,
1389                         sizeof(__le32), GFP_KERNEL);
1390         if (!ctrl->changed_ns_list)
1391                 goto out_free_ctrl;
1392
1393         ctrl->sqs = kcalloc(subsys->max_qid + 1,
1394                         sizeof(struct nvmet_sq *),
1395                         GFP_KERNEL);
1396         if (!ctrl->sqs)
1397                 goto out_free_changed_ns_list;
1398
1399         if (subsys->cntlid_min > subsys->cntlid_max)
1400                 goto out_free_sqs;
1401
1402         ret = ida_alloc_range(&cntlid_ida,
1403                              subsys->cntlid_min, subsys->cntlid_max,
1404                              GFP_KERNEL);
1405         if (ret < 0) {
1406                 status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
1407                 goto out_free_sqs;
1408         }
1409         ctrl->cntlid = ret;
1410
1411         /*
1412          * Discovery controllers may use some arbitrary high value
1413          * in order to cleanup stale discovery sessions
1414          */
1415         if (nvmet_is_disc_subsys(ctrl->subsys) && !kato)
1416                 kato = NVMET_DISC_KATO_MS;
1417
1418         /* keep-alive timeout in seconds */
1419         ctrl->kato = DIV_ROUND_UP(kato, 1000);
1420
1421         ctrl->err_counter = 0;
1422         spin_lock_init(&ctrl->error_lock);
1423
1424         nvmet_start_keep_alive_timer(ctrl);
1425
1426         mutex_lock(&subsys->lock);
1427         list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
1428         nvmet_setup_p2p_ns_map(ctrl, req);
1429         mutex_unlock(&subsys->lock);
1430
1431         *ctrlp = ctrl;
1432         return 0;
1433
1434 out_free_sqs:
1435         kfree(ctrl->sqs);
1436 out_free_changed_ns_list:
1437         kfree(ctrl->changed_ns_list);
1438 out_free_ctrl:
1439         kfree(ctrl);
1440 out_put_subsystem:
1441         nvmet_subsys_put(subsys);
1442 out:
1443         return status;
1444 }
1445
1446 static void nvmet_ctrl_free(struct kref *ref)
1447 {
1448         struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
1449         struct nvmet_subsys *subsys = ctrl->subsys;
1450
1451         mutex_lock(&subsys->lock);
1452         nvmet_release_p2p_ns_map(ctrl);
1453         list_del(&ctrl->subsys_entry);
1454         mutex_unlock(&subsys->lock);
1455
1456         nvmet_stop_keep_alive_timer(ctrl);
1457
1458         flush_work(&ctrl->async_event_work);
1459         cancel_work_sync(&ctrl->fatal_err_work);
1460
1461         ida_free(&cntlid_ida, ctrl->cntlid);
1462
1463         nvmet_async_events_free(ctrl);
1464         kfree(ctrl->sqs);
1465         kfree(ctrl->changed_ns_list);
1466         kfree(ctrl);
1467
1468         nvmet_subsys_put(subsys);
1469 }
1470
1471 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
1472 {
1473         kref_put(&ctrl->ref, nvmet_ctrl_free);
1474 }
1475
1476 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
1477 {
1478         mutex_lock(&ctrl->lock);
1479         if (!(ctrl->csts & NVME_CSTS_CFS)) {
1480                 ctrl->csts |= NVME_CSTS_CFS;
1481                 schedule_work(&ctrl->fatal_err_work);
1482         }
1483         mutex_unlock(&ctrl->lock);
1484 }
1485 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
1486
1487 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
1488                 const char *subsysnqn)
1489 {
1490         struct nvmet_subsys_link *p;
1491
1492         if (!port)
1493                 return NULL;
1494
1495         if (!strcmp(NVME_DISC_SUBSYS_NAME, subsysnqn)) {
1496                 if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
1497                         return NULL;
1498                 return nvmet_disc_subsys;
1499         }
1500
1501         down_read(&nvmet_config_sem);
1502         list_for_each_entry(p, &port->subsystems, entry) {
1503                 if (!strncmp(p->subsys->subsysnqn, subsysnqn,
1504                                 NVMF_NQN_SIZE)) {
1505                         if (!kref_get_unless_zero(&p->subsys->ref))
1506                                 break;
1507                         up_read(&nvmet_config_sem);
1508                         return p->subsys;
1509                 }
1510         }
1511         up_read(&nvmet_config_sem);
1512         return NULL;
1513 }
1514
1515 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
1516                 enum nvme_subsys_type type)
1517 {
1518         struct nvmet_subsys *subsys;
1519         char serial[NVMET_SN_MAX_SIZE / 2];
1520         int ret;
1521
1522         subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
1523         if (!subsys)
1524                 return ERR_PTR(-ENOMEM);
1525
1526         subsys->ver = NVMET_DEFAULT_VS;
1527         /* generate a random serial number as our controllers are ephemeral: */
1528         get_random_bytes(&serial, sizeof(serial));
1529         bin2hex(subsys->serial, &serial, sizeof(serial));
1530
1531         subsys->model_number = kstrdup(NVMET_DEFAULT_CTRL_MODEL, GFP_KERNEL);
1532         if (!subsys->model_number) {
1533                 ret = -ENOMEM;
1534                 goto free_subsys;
1535         }
1536
1537         switch (type) {
1538         case NVME_NQN_NVME:
1539                 subsys->max_qid = NVMET_NR_QUEUES;
1540                 break;
1541         case NVME_NQN_DISC:
1542         case NVME_NQN_CURR:
1543                 subsys->max_qid = 0;
1544                 break;
1545         default:
1546                 pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
1547                 ret = -EINVAL;
1548                 goto free_mn;
1549         }
1550         subsys->type = type;
1551         subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
1552                         GFP_KERNEL);
1553         if (!subsys->subsysnqn) {
1554                 ret = -ENOMEM;
1555                 goto free_mn;
1556         }
1557         subsys->cntlid_min = NVME_CNTLID_MIN;
1558         subsys->cntlid_max = NVME_CNTLID_MAX;
1559         kref_init(&subsys->ref);
1560
1561         mutex_init(&subsys->lock);
1562         xa_init(&subsys->namespaces);
1563         INIT_LIST_HEAD(&subsys->ctrls);
1564         INIT_LIST_HEAD(&subsys->hosts);
1565
1566         return subsys;
1567
1568 free_mn:
1569         kfree(subsys->model_number);
1570 free_subsys:
1571         kfree(subsys);
1572         return ERR_PTR(ret);
1573 }
1574
1575 static void nvmet_subsys_free(struct kref *ref)
1576 {
1577         struct nvmet_subsys *subsys =
1578                 container_of(ref, struct nvmet_subsys, ref);
1579
1580         WARN_ON_ONCE(!xa_empty(&subsys->namespaces));
1581
1582         xa_destroy(&subsys->namespaces);
1583         nvmet_passthru_subsys_free(subsys);
1584
1585         kfree(subsys->subsysnqn);
1586         kfree(subsys->model_number);
1587         kfree(subsys);
1588 }
1589
1590 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1591 {
1592         struct nvmet_ctrl *ctrl;
1593
1594         mutex_lock(&subsys->lock);
1595         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1596                 ctrl->ops->delete_ctrl(ctrl);
1597         mutex_unlock(&subsys->lock);
1598 }
1599
1600 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1601 {
1602         kref_put(&subsys->ref, nvmet_subsys_free);
1603 }
1604
1605 static int __init nvmet_init(void)
1606 {
1607         int error;
1608
1609         nvmet_ana_group_enabled[NVMET_DEFAULT_ANA_GRPID] = 1;
1610
1611         zbd_wq = alloc_workqueue("nvmet-zbd-wq", WQ_MEM_RECLAIM, 0);
1612         if (!zbd_wq)
1613                 return -ENOMEM;
1614
1615         buffered_io_wq = alloc_workqueue("nvmet-buffered-io-wq",
1616                         WQ_MEM_RECLAIM, 0);
1617         if (!buffered_io_wq) {
1618                 error = -ENOMEM;
1619                 goto out_free_zbd_work_queue;
1620         }
1621
1622         error = nvmet_init_discovery();
1623         if (error)
1624                 goto out_free_work_queue;
1625
1626         error = nvmet_init_configfs();
1627         if (error)
1628                 goto out_exit_discovery;
1629         return 0;
1630
1631 out_exit_discovery:
1632         nvmet_exit_discovery();
1633 out_free_work_queue:
1634         destroy_workqueue(buffered_io_wq);
1635 out_free_zbd_work_queue:
1636         destroy_workqueue(zbd_wq);
1637         return error;
1638 }
1639
1640 static void __exit nvmet_exit(void)
1641 {
1642         nvmet_exit_configfs();
1643         nvmet_exit_discovery();
1644         ida_destroy(&cntlid_ida);
1645         destroy_workqueue(buffered_io_wq);
1646         destroy_workqueue(zbd_wq);
1647
1648         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1649         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1650 }
1651
1652 module_init(nvmet_init);
1653 module_exit(nvmet_exit);
1654
1655 MODULE_LICENSE("GPL v2");