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