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