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