1 /* 2 * NVMe over Fabrics loopback device. 3 * Copyright (c) 2015-2016 HGST, a Western Digital Company. 4 * 5 * This program is free software; you can redistribute it and/or modify it 6 * under the terms and conditions of the GNU General Public License, 7 * version 2, as published by the Free Software Foundation. 8 * 9 * This program is distributed in the hope it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 12 * more details. 13 */ 14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 15 #include <linux/scatterlist.h> 16 #include <linux/blk-mq.h> 17 #include <linux/nvme.h> 18 #include <linux/module.h> 19 #include <linux/parser.h> 20 #include "nvmet.h" 21 #include "../host/nvme.h" 22 #include "../host/fabrics.h" 23 24 #define NVME_LOOP_MAX_SEGMENTS 256 25 26 /* 27 * We handle AEN commands ourselves and don't even let the 28 * block layer know about them. 29 */ 30 #define NVME_LOOP_NR_AEN_COMMANDS 1 31 #define NVME_LOOP_AQ_BLKMQ_DEPTH \ 32 (NVME_AQ_DEPTH - NVME_LOOP_NR_AEN_COMMANDS) 33 34 struct nvme_loop_iod { 35 struct nvme_request nvme_req; 36 struct nvme_command cmd; 37 struct nvme_completion rsp; 38 struct nvmet_req req; 39 struct nvme_loop_queue *queue; 40 struct work_struct work; 41 struct sg_table sg_table; 42 struct scatterlist first_sgl[]; 43 }; 44 45 struct nvme_loop_ctrl { 46 struct nvme_loop_queue *queues; 47 48 struct blk_mq_tag_set admin_tag_set; 49 50 struct list_head list; 51 struct blk_mq_tag_set tag_set; 52 struct nvme_loop_iod async_event_iod; 53 struct nvme_ctrl ctrl; 54 55 struct nvmet_ctrl *target_ctrl; 56 struct work_struct delete_work; 57 }; 58 59 static inline struct nvme_loop_ctrl *to_loop_ctrl(struct nvme_ctrl *ctrl) 60 { 61 return container_of(ctrl, struct nvme_loop_ctrl, ctrl); 62 } 63 64 struct nvme_loop_queue { 65 struct nvmet_cq nvme_cq; 66 struct nvmet_sq nvme_sq; 67 struct nvme_loop_ctrl *ctrl; 68 }; 69 70 static struct nvmet_port *nvmet_loop_port; 71 72 static LIST_HEAD(nvme_loop_ctrl_list); 73 static DEFINE_MUTEX(nvme_loop_ctrl_mutex); 74 75 static void nvme_loop_queue_response(struct nvmet_req *nvme_req); 76 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *ctrl); 77 78 static struct nvmet_fabrics_ops nvme_loop_ops; 79 80 static inline int nvme_loop_queue_idx(struct nvme_loop_queue *queue) 81 { 82 return queue - queue->ctrl->queues; 83 } 84 85 static void nvme_loop_complete_rq(struct request *req) 86 { 87 struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req); 88 89 nvme_cleanup_cmd(req); 90 sg_free_table_chained(&iod->sg_table, true); 91 nvme_complete_rq(req); 92 } 93 94 static struct blk_mq_tags *nvme_loop_tagset(struct nvme_loop_queue *queue) 95 { 96 u32 queue_idx = nvme_loop_queue_idx(queue); 97 98 if (queue_idx == 0) 99 return queue->ctrl->admin_tag_set.tags[queue_idx]; 100 return queue->ctrl->tag_set.tags[queue_idx - 1]; 101 } 102 103 static void nvme_loop_queue_response(struct nvmet_req *req) 104 { 105 struct nvme_loop_queue *queue = 106 container_of(req->sq, struct nvme_loop_queue, nvme_sq); 107 struct nvme_completion *cqe = req->rsp; 108 109 /* 110 * AEN requests are special as they don't time out and can 111 * survive any kind of queue freeze and often don't respond to 112 * aborts. We don't even bother to allocate a struct request 113 * for them but rather special case them here. 114 */ 115 if (unlikely(nvme_loop_queue_idx(queue) == 0 && 116 cqe->command_id >= NVME_LOOP_AQ_BLKMQ_DEPTH)) { 117 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status, 118 &cqe->result); 119 } else { 120 struct request *rq; 121 122 rq = blk_mq_tag_to_rq(nvme_loop_tagset(queue), cqe->command_id); 123 if (!rq) { 124 dev_err(queue->ctrl->ctrl.device, 125 "tag 0x%x on queue %d not found\n", 126 cqe->command_id, nvme_loop_queue_idx(queue)); 127 return; 128 } 129 130 nvme_end_request(rq, cqe->status, cqe->result); 131 } 132 } 133 134 static void nvme_loop_execute_work(struct work_struct *work) 135 { 136 struct nvme_loop_iod *iod = 137 container_of(work, struct nvme_loop_iod, work); 138 139 iod->req.execute(&iod->req); 140 } 141 142 static enum blk_eh_timer_return 143 nvme_loop_timeout(struct request *rq, bool reserved) 144 { 145 struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(rq); 146 147 /* queue error recovery */ 148 nvme_reset_ctrl(&iod->queue->ctrl->ctrl); 149 150 /* fail with DNR on admin cmd timeout */ 151 nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR; 152 153 return BLK_EH_HANDLED; 154 } 155 156 static blk_status_t nvme_loop_queue_rq(struct blk_mq_hw_ctx *hctx, 157 const struct blk_mq_queue_data *bd) 158 { 159 struct nvme_ns *ns = hctx->queue->queuedata; 160 struct nvme_loop_queue *queue = hctx->driver_data; 161 struct request *req = bd->rq; 162 struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req); 163 blk_status_t ret; 164 165 ret = nvme_setup_cmd(ns, req, &iod->cmd); 166 if (ret) 167 return ret; 168 169 iod->cmd.common.flags |= NVME_CMD_SGL_METABUF; 170 iod->req.port = nvmet_loop_port; 171 if (!nvmet_req_init(&iod->req, &queue->nvme_cq, 172 &queue->nvme_sq, &nvme_loop_ops)) { 173 nvme_cleanup_cmd(req); 174 blk_mq_start_request(req); 175 nvme_loop_queue_response(&iod->req); 176 return BLK_STS_OK; 177 } 178 179 if (blk_rq_bytes(req)) { 180 iod->sg_table.sgl = iod->first_sgl; 181 if (sg_alloc_table_chained(&iod->sg_table, 182 blk_rq_nr_phys_segments(req), 183 iod->sg_table.sgl)) 184 return BLK_STS_RESOURCE; 185 186 iod->req.sg = iod->sg_table.sgl; 187 iod->req.sg_cnt = blk_rq_map_sg(req->q, req, iod->sg_table.sgl); 188 } 189 190 blk_mq_start_request(req); 191 192 schedule_work(&iod->work); 193 return BLK_STS_OK; 194 } 195 196 static void nvme_loop_submit_async_event(struct nvme_ctrl *arg, int aer_idx) 197 { 198 struct nvme_loop_ctrl *ctrl = to_loop_ctrl(arg); 199 struct nvme_loop_queue *queue = &ctrl->queues[0]; 200 struct nvme_loop_iod *iod = &ctrl->async_event_iod; 201 202 memset(&iod->cmd, 0, sizeof(iod->cmd)); 203 iod->cmd.common.opcode = nvme_admin_async_event; 204 iod->cmd.common.command_id = NVME_LOOP_AQ_BLKMQ_DEPTH; 205 iod->cmd.common.flags |= NVME_CMD_SGL_METABUF; 206 207 if (!nvmet_req_init(&iod->req, &queue->nvme_cq, &queue->nvme_sq, 208 &nvme_loop_ops)) { 209 dev_err(ctrl->ctrl.device, "failed async event work\n"); 210 return; 211 } 212 213 schedule_work(&iod->work); 214 } 215 216 static int nvme_loop_init_iod(struct nvme_loop_ctrl *ctrl, 217 struct nvme_loop_iod *iod, unsigned int queue_idx) 218 { 219 iod->req.cmd = &iod->cmd; 220 iod->req.rsp = &iod->rsp; 221 iod->queue = &ctrl->queues[queue_idx]; 222 INIT_WORK(&iod->work, nvme_loop_execute_work); 223 return 0; 224 } 225 226 static int nvme_loop_init_request(struct blk_mq_tag_set *set, 227 struct request *req, unsigned int hctx_idx, 228 unsigned int numa_node) 229 { 230 struct nvme_loop_ctrl *ctrl = set->driver_data; 231 232 return nvme_loop_init_iod(ctrl, blk_mq_rq_to_pdu(req), 233 (set == &ctrl->tag_set) ? hctx_idx + 1 : 0); 234 } 235 236 static int nvme_loop_init_hctx(struct blk_mq_hw_ctx *hctx, void *data, 237 unsigned int hctx_idx) 238 { 239 struct nvme_loop_ctrl *ctrl = data; 240 struct nvme_loop_queue *queue = &ctrl->queues[hctx_idx + 1]; 241 242 BUG_ON(hctx_idx >= ctrl->ctrl.queue_count); 243 244 hctx->driver_data = queue; 245 return 0; 246 } 247 248 static int nvme_loop_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data, 249 unsigned int hctx_idx) 250 { 251 struct nvme_loop_ctrl *ctrl = data; 252 struct nvme_loop_queue *queue = &ctrl->queues[0]; 253 254 BUG_ON(hctx_idx != 0); 255 256 hctx->driver_data = queue; 257 return 0; 258 } 259 260 static const struct blk_mq_ops nvme_loop_mq_ops = { 261 .queue_rq = nvme_loop_queue_rq, 262 .complete = nvme_loop_complete_rq, 263 .init_request = nvme_loop_init_request, 264 .init_hctx = nvme_loop_init_hctx, 265 .timeout = nvme_loop_timeout, 266 }; 267 268 static const struct blk_mq_ops nvme_loop_admin_mq_ops = { 269 .queue_rq = nvme_loop_queue_rq, 270 .complete = nvme_loop_complete_rq, 271 .init_request = nvme_loop_init_request, 272 .init_hctx = nvme_loop_init_admin_hctx, 273 .timeout = nvme_loop_timeout, 274 }; 275 276 static void nvme_loop_destroy_admin_queue(struct nvme_loop_ctrl *ctrl) 277 { 278 nvmet_sq_destroy(&ctrl->queues[0].nvme_sq); 279 blk_cleanup_queue(ctrl->ctrl.admin_q); 280 blk_mq_free_tag_set(&ctrl->admin_tag_set); 281 } 282 283 static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl) 284 { 285 struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl); 286 287 if (list_empty(&ctrl->list)) 288 goto free_ctrl; 289 290 mutex_lock(&nvme_loop_ctrl_mutex); 291 list_del(&ctrl->list); 292 mutex_unlock(&nvme_loop_ctrl_mutex); 293 294 if (nctrl->tagset) { 295 blk_cleanup_queue(ctrl->ctrl.connect_q); 296 blk_mq_free_tag_set(&ctrl->tag_set); 297 } 298 kfree(ctrl->queues); 299 nvmf_free_options(nctrl->opts); 300 free_ctrl: 301 kfree(ctrl); 302 } 303 304 static void nvme_loop_destroy_io_queues(struct nvme_loop_ctrl *ctrl) 305 { 306 int i; 307 308 for (i = 1; i < ctrl->ctrl.queue_count; i++) 309 nvmet_sq_destroy(&ctrl->queues[i].nvme_sq); 310 } 311 312 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl) 313 { 314 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts; 315 unsigned int nr_io_queues; 316 int ret, i; 317 318 nr_io_queues = min(opts->nr_io_queues, num_online_cpus()); 319 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues); 320 if (ret || !nr_io_queues) 321 return ret; 322 323 dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues); 324 325 for (i = 1; i <= nr_io_queues; i++) { 326 ctrl->queues[i].ctrl = ctrl; 327 ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq); 328 if (ret) 329 goto out_destroy_queues; 330 331 ctrl->ctrl.queue_count++; 332 } 333 334 return 0; 335 336 out_destroy_queues: 337 nvme_loop_destroy_io_queues(ctrl); 338 return ret; 339 } 340 341 static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl) 342 { 343 int i, ret; 344 345 for (i = 1; i < ctrl->ctrl.queue_count; i++) { 346 ret = nvmf_connect_io_queue(&ctrl->ctrl, i); 347 if (ret) 348 return ret; 349 } 350 351 return 0; 352 } 353 354 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl) 355 { 356 int error; 357 358 memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set)); 359 ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops; 360 ctrl->admin_tag_set.queue_depth = NVME_LOOP_AQ_BLKMQ_DEPTH; 361 ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */ 362 ctrl->admin_tag_set.numa_node = NUMA_NO_NODE; 363 ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) + 364 SG_CHUNK_SIZE * sizeof(struct scatterlist); 365 ctrl->admin_tag_set.driver_data = ctrl; 366 ctrl->admin_tag_set.nr_hw_queues = 1; 367 ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT; 368 ctrl->admin_tag_set.flags = BLK_MQ_F_NO_SCHED; 369 370 ctrl->queues[0].ctrl = ctrl; 371 error = nvmet_sq_init(&ctrl->queues[0].nvme_sq); 372 if (error) 373 return error; 374 ctrl->ctrl.queue_count = 1; 375 376 error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set); 377 if (error) 378 goto out_free_sq; 379 ctrl->ctrl.admin_tagset = &ctrl->admin_tag_set; 380 381 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set); 382 if (IS_ERR(ctrl->ctrl.admin_q)) { 383 error = PTR_ERR(ctrl->ctrl.admin_q); 384 goto out_free_tagset; 385 } 386 387 error = nvmf_connect_admin_queue(&ctrl->ctrl); 388 if (error) 389 goto out_cleanup_queue; 390 391 error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->ctrl.cap); 392 if (error) { 393 dev_err(ctrl->ctrl.device, 394 "prop_get NVME_REG_CAP failed\n"); 395 goto out_cleanup_queue; 396 } 397 398 ctrl->ctrl.sqsize = 399 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize); 400 401 error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap); 402 if (error) 403 goto out_cleanup_queue; 404 405 ctrl->ctrl.max_hw_sectors = 406 (NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9); 407 408 error = nvme_init_identify(&ctrl->ctrl); 409 if (error) 410 goto out_cleanup_queue; 411 412 return 0; 413 414 out_cleanup_queue: 415 blk_cleanup_queue(ctrl->ctrl.admin_q); 416 out_free_tagset: 417 blk_mq_free_tag_set(&ctrl->admin_tag_set); 418 out_free_sq: 419 nvmet_sq_destroy(&ctrl->queues[0].nvme_sq); 420 return error; 421 } 422 423 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl) 424 { 425 if (ctrl->ctrl.queue_count > 1) { 426 nvme_stop_queues(&ctrl->ctrl); 427 blk_mq_tagset_busy_iter(&ctrl->tag_set, 428 nvme_cancel_request, &ctrl->ctrl); 429 nvme_loop_destroy_io_queues(ctrl); 430 } 431 432 if (ctrl->ctrl.state == NVME_CTRL_LIVE) 433 nvme_shutdown_ctrl(&ctrl->ctrl); 434 435 blk_mq_quiesce_queue(ctrl->ctrl.admin_q); 436 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set, 437 nvme_cancel_request, &ctrl->ctrl); 438 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q); 439 nvme_loop_destroy_admin_queue(ctrl); 440 } 441 442 static void nvme_loop_del_ctrl_work(struct work_struct *work) 443 { 444 struct nvme_loop_ctrl *ctrl = container_of(work, 445 struct nvme_loop_ctrl, delete_work); 446 447 nvme_stop_ctrl(&ctrl->ctrl); 448 nvme_remove_namespaces(&ctrl->ctrl); 449 nvme_loop_shutdown_ctrl(ctrl); 450 nvme_uninit_ctrl(&ctrl->ctrl); 451 nvme_put_ctrl(&ctrl->ctrl); 452 } 453 454 static int __nvme_loop_del_ctrl(struct nvme_loop_ctrl *ctrl) 455 { 456 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING)) 457 return -EBUSY; 458 459 if (!queue_work(nvme_wq, &ctrl->delete_work)) 460 return -EBUSY; 461 462 return 0; 463 } 464 465 static int nvme_loop_del_ctrl(struct nvme_ctrl *nctrl) 466 { 467 struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl); 468 int ret; 469 470 ret = __nvme_loop_del_ctrl(ctrl); 471 if (ret) 472 return ret; 473 474 flush_work(&ctrl->delete_work); 475 476 return 0; 477 } 478 479 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl) 480 { 481 struct nvme_loop_ctrl *ctrl; 482 483 mutex_lock(&nvme_loop_ctrl_mutex); 484 list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) { 485 if (ctrl->ctrl.cntlid == nctrl->cntlid) 486 __nvme_loop_del_ctrl(ctrl); 487 } 488 mutex_unlock(&nvme_loop_ctrl_mutex); 489 } 490 491 static void nvme_loop_reset_ctrl_work(struct work_struct *work) 492 { 493 struct nvme_loop_ctrl *ctrl = 494 container_of(work, struct nvme_loop_ctrl, ctrl.reset_work); 495 bool changed; 496 int ret; 497 498 nvme_stop_ctrl(&ctrl->ctrl); 499 nvme_loop_shutdown_ctrl(ctrl); 500 501 ret = nvme_loop_configure_admin_queue(ctrl); 502 if (ret) 503 goto out_disable; 504 505 ret = nvme_loop_init_io_queues(ctrl); 506 if (ret) 507 goto out_destroy_admin; 508 509 ret = nvme_loop_connect_io_queues(ctrl); 510 if (ret) 511 goto out_destroy_io; 512 513 blk_mq_update_nr_hw_queues(&ctrl->tag_set, 514 ctrl->ctrl.queue_count - 1); 515 516 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE); 517 WARN_ON_ONCE(!changed); 518 519 nvme_start_ctrl(&ctrl->ctrl); 520 521 return; 522 523 out_destroy_io: 524 nvme_loop_destroy_io_queues(ctrl); 525 out_destroy_admin: 526 nvme_loop_destroy_admin_queue(ctrl); 527 out_disable: 528 dev_warn(ctrl->ctrl.device, "Removing after reset failure\n"); 529 nvme_uninit_ctrl(&ctrl->ctrl); 530 nvme_put_ctrl(&ctrl->ctrl); 531 } 532 533 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = { 534 .name = "loop", 535 .module = THIS_MODULE, 536 .flags = NVME_F_FABRICS, 537 .reg_read32 = nvmf_reg_read32, 538 .reg_read64 = nvmf_reg_read64, 539 .reg_write32 = nvmf_reg_write32, 540 .free_ctrl = nvme_loop_free_ctrl, 541 .submit_async_event = nvme_loop_submit_async_event, 542 .delete_ctrl = nvme_loop_del_ctrl, 543 }; 544 545 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl) 546 { 547 int ret; 548 549 ret = nvme_loop_init_io_queues(ctrl); 550 if (ret) 551 return ret; 552 553 memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set)); 554 ctrl->tag_set.ops = &nvme_loop_mq_ops; 555 ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size; 556 ctrl->tag_set.reserved_tags = 1; /* fabric connect */ 557 ctrl->tag_set.numa_node = NUMA_NO_NODE; 558 ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE; 559 ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) + 560 SG_CHUNK_SIZE * sizeof(struct scatterlist); 561 ctrl->tag_set.driver_data = ctrl; 562 ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1; 563 ctrl->tag_set.timeout = NVME_IO_TIMEOUT; 564 ctrl->ctrl.tagset = &ctrl->tag_set; 565 566 ret = blk_mq_alloc_tag_set(&ctrl->tag_set); 567 if (ret) 568 goto out_destroy_queues; 569 570 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set); 571 if (IS_ERR(ctrl->ctrl.connect_q)) { 572 ret = PTR_ERR(ctrl->ctrl.connect_q); 573 goto out_free_tagset; 574 } 575 576 ret = nvme_loop_connect_io_queues(ctrl); 577 if (ret) 578 goto out_cleanup_connect_q; 579 580 return 0; 581 582 out_cleanup_connect_q: 583 blk_cleanup_queue(ctrl->ctrl.connect_q); 584 out_free_tagset: 585 blk_mq_free_tag_set(&ctrl->tag_set); 586 out_destroy_queues: 587 nvme_loop_destroy_io_queues(ctrl); 588 return ret; 589 } 590 591 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev, 592 struct nvmf_ctrl_options *opts) 593 { 594 struct nvme_loop_ctrl *ctrl; 595 bool changed; 596 int ret; 597 598 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL); 599 if (!ctrl) 600 return ERR_PTR(-ENOMEM); 601 ctrl->ctrl.opts = opts; 602 INIT_LIST_HEAD(&ctrl->list); 603 604 INIT_WORK(&ctrl->delete_work, nvme_loop_del_ctrl_work); 605 INIT_WORK(&ctrl->ctrl.reset_work, nvme_loop_reset_ctrl_work); 606 607 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops, 608 0 /* no quirks, we're perfect! */); 609 if (ret) 610 goto out_put_ctrl; 611 612 ret = -ENOMEM; 613 614 ctrl->ctrl.sqsize = opts->queue_size - 1; 615 ctrl->ctrl.kato = opts->kato; 616 617 ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues), 618 GFP_KERNEL); 619 if (!ctrl->queues) 620 goto out_uninit_ctrl; 621 622 ret = nvme_loop_configure_admin_queue(ctrl); 623 if (ret) 624 goto out_free_queues; 625 626 if (opts->queue_size > ctrl->ctrl.maxcmd) { 627 /* warn if maxcmd is lower than queue_size */ 628 dev_warn(ctrl->ctrl.device, 629 "queue_size %zu > ctrl maxcmd %u, clamping down\n", 630 opts->queue_size, ctrl->ctrl.maxcmd); 631 opts->queue_size = ctrl->ctrl.maxcmd; 632 } 633 634 if (opts->nr_io_queues) { 635 ret = nvme_loop_create_io_queues(ctrl); 636 if (ret) 637 goto out_remove_admin_queue; 638 } 639 640 nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0); 641 642 dev_info(ctrl->ctrl.device, 643 "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn); 644 645 nvme_get_ctrl(&ctrl->ctrl); 646 647 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE); 648 WARN_ON_ONCE(!changed); 649 650 mutex_lock(&nvme_loop_ctrl_mutex); 651 list_add_tail(&ctrl->list, &nvme_loop_ctrl_list); 652 mutex_unlock(&nvme_loop_ctrl_mutex); 653 654 nvme_start_ctrl(&ctrl->ctrl); 655 656 return &ctrl->ctrl; 657 658 out_remove_admin_queue: 659 nvme_loop_destroy_admin_queue(ctrl); 660 out_free_queues: 661 kfree(ctrl->queues); 662 out_uninit_ctrl: 663 nvme_uninit_ctrl(&ctrl->ctrl); 664 out_put_ctrl: 665 nvme_put_ctrl(&ctrl->ctrl); 666 if (ret > 0) 667 ret = -EIO; 668 return ERR_PTR(ret); 669 } 670 671 static int nvme_loop_add_port(struct nvmet_port *port) 672 { 673 /* 674 * XXX: disalow adding more than one port so 675 * there is no connection rejections when a 676 * a subsystem is assigned to a port for which 677 * loop doesn't have a pointer. 678 * This scenario would be possible if we allowed 679 * more than one port to be added and a subsystem 680 * was assigned to a port other than nvmet_loop_port. 681 */ 682 683 if (nvmet_loop_port) 684 return -EPERM; 685 686 nvmet_loop_port = port; 687 return 0; 688 } 689 690 static void nvme_loop_remove_port(struct nvmet_port *port) 691 { 692 if (port == nvmet_loop_port) 693 nvmet_loop_port = NULL; 694 } 695 696 static struct nvmet_fabrics_ops nvme_loop_ops = { 697 .owner = THIS_MODULE, 698 .type = NVMF_TRTYPE_LOOP, 699 .add_port = nvme_loop_add_port, 700 .remove_port = nvme_loop_remove_port, 701 .queue_response = nvme_loop_queue_response, 702 .delete_ctrl = nvme_loop_delete_ctrl, 703 }; 704 705 static struct nvmf_transport_ops nvme_loop_transport = { 706 .name = "loop", 707 .create_ctrl = nvme_loop_create_ctrl, 708 }; 709 710 static int __init nvme_loop_init_module(void) 711 { 712 int ret; 713 714 ret = nvmet_register_transport(&nvme_loop_ops); 715 if (ret) 716 return ret; 717 718 ret = nvmf_register_transport(&nvme_loop_transport); 719 if (ret) 720 nvmet_unregister_transport(&nvme_loop_ops); 721 722 return ret; 723 } 724 725 static void __exit nvme_loop_cleanup_module(void) 726 { 727 struct nvme_loop_ctrl *ctrl, *next; 728 729 nvmf_unregister_transport(&nvme_loop_transport); 730 nvmet_unregister_transport(&nvme_loop_ops); 731 732 mutex_lock(&nvme_loop_ctrl_mutex); 733 list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list) 734 __nvme_loop_del_ctrl(ctrl); 735 mutex_unlock(&nvme_loop_ctrl_mutex); 736 737 flush_workqueue(nvme_wq); 738 } 739 740 module_init(nvme_loop_init_module); 741 module_exit(nvme_loop_cleanup_module); 742 743 MODULE_LICENSE("GPL v2"); 744 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */ 745