drivers/nvme/target/loop.c

   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/delay.h>
  17 #include <linux/blk-mq.h>
  18 #include <linux/nvme.h>
  19 #include <linux/module.h>
  20 #include <linux/parser.h>
  21 #include <linux/t10-pi.h>
  22 #include "nvmet.h"
  23 #include "../host/nvme.h"
  24 #include "../host/fabrics.h"
  25
  26 #define NVME_LOOP_AQ_DEPTH              256
  27
  28 #define NVME_LOOP_MAX_SEGMENTS          256
  29
  30 /*
  31  * We handle AEN commands ourselves and don't even let the
  32  * block layer know about them.
  33  */
  34 #define NVME_LOOP_NR_AEN_COMMANDS       1
  35 #define NVME_LOOP_AQ_BLKMQ_DEPTH        \
  36         (NVME_LOOP_AQ_DEPTH - NVME_LOOP_NR_AEN_COMMANDS)
  37
  38 struct nvme_loop_iod {
  39         struct nvme_command     cmd;
  40         struct nvme_completion  rsp;
  41         struct nvmet_req        req;
  42         struct nvme_loop_queue  *queue;
  43         struct work_struct      work;
  44         struct sg_table         sg_table;
  45         struct scatterlist      first_sgl[];
  46 };
  47
  48 struct nvme_loop_ctrl {
  49         spinlock_t              lock;
  50         struct nvme_loop_queue  *queues;
  51         u32                     queue_count;
  52
  53         struct blk_mq_tag_set   admin_tag_set;
  54
  55         struct list_head        list;
  56         u64                     cap;
  57         struct blk_mq_tag_set   tag_set;
  58         struct nvme_loop_iod    async_event_iod;
  59         struct nvme_ctrl        ctrl;
  60
  61         struct nvmet_ctrl       *target_ctrl;
  62         struct work_struct      delete_work;
  63         struct work_struct      reset_work;
  64 };
  65
  66 static inline struct nvme_loop_ctrl *to_loop_ctrl(struct nvme_ctrl *ctrl)
  67 {
  68         return container_of(ctrl, struct nvme_loop_ctrl, ctrl);
  69 }
  70
  71 struct nvme_loop_queue {
  72         struct nvmet_cq         nvme_cq;
  73         struct nvmet_sq         nvme_sq;
  74         struct nvme_loop_ctrl   *ctrl;
  75 };
  76
  77 static struct nvmet_port *nvmet_loop_port;
  78
  79 static LIST_HEAD(nvme_loop_ctrl_list);
  80 static DEFINE_MUTEX(nvme_loop_ctrl_mutex);
  81
  82 static void nvme_loop_queue_response(struct nvmet_req *nvme_req);
  83 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *ctrl);
  84
  85 static struct nvmet_fabrics_ops nvme_loop_ops;
  86
  87 static inline int nvme_loop_queue_idx(struct nvme_loop_queue *queue)
  88 {
  89         return queue - queue->ctrl->queues;
  90 }
  91
  92 static void nvme_loop_complete_rq(struct request *req)
  93 {
  94         struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
  95         int error = 0;
  96
  97         nvme_cleanup_cmd(req);
  98         sg_free_table_chained(&iod->sg_table, true);
  99
 100         if (unlikely(req->errors)) {
 101                 if (nvme_req_needs_retry(req, req->errors)) {
 102                         nvme_requeue_req(req);
 103                         return;
 104                 }
 105
 106                 if (req->cmd_type == REQ_TYPE_DRV_PRIV)
 107                         error = req->errors;
 108                 else
 109                         error = nvme_error_status(req->errors);
 110         }
 111
 112         blk_mq_end_request(req, error);
 113 }
 114
 115 static void nvme_loop_queue_response(struct nvmet_req *nvme_req)
 116 {
 117         struct nvme_loop_iod *iod =
 118                 container_of(nvme_req, struct nvme_loop_iod, req);
 119         struct nvme_completion *cqe = &iod->rsp;
 120
 121         /*
 122          * AEN requests are special as they don't time out and can
 123          * survive any kind of queue freeze and often don't respond to
 124          * aborts.  We don't even bother to allocate a struct request
 125          * for them but rather special case them here.
 126          */
 127         if (unlikely(nvme_loop_queue_idx(iod->queue) == 0 &&
 128                         cqe->command_id >= NVME_LOOP_AQ_BLKMQ_DEPTH)) {
 129                 nvme_complete_async_event(&iod->queue->ctrl->ctrl, cqe);
 130         } else {
 131                 struct request *req = blk_mq_rq_from_pdu(iod);
 132
 133                 if (req->cmd_type == REQ_TYPE_DRV_PRIV && req->special)
 134                         memcpy(req->special, cqe, sizeof(*cqe));
 135                 blk_mq_complete_request(req, le16_to_cpu(cqe->status) >> 1);
 136         }
 137 }
 138
 139 static void nvme_loop_execute_work(struct work_struct *work)
 140 {
 141         struct nvme_loop_iod *iod =
 142                 container_of(work, struct nvme_loop_iod, work);
 143
 144         iod->req.execute(&iod->req);
 145 }
 146
 147 static enum blk_eh_timer_return
 148 nvme_loop_timeout(struct request *rq, bool reserved)
 149 {
 150         struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(rq);
 151
 152         /* queue error recovery */
 153         schedule_work(&iod->queue->ctrl->reset_work);
 154
 155         /* fail with DNR on admin cmd timeout */
 156         rq->errors = NVME_SC_ABORT_REQ | NVME_SC_DNR;
 157
 158         return BLK_EH_HANDLED;
 159 }
 160
 161 static int nvme_loop_queue_rq(struct blk_mq_hw_ctx *hctx,
 162                 const struct blk_mq_queue_data *bd)
 163 {
 164         struct nvme_ns *ns = hctx->queue->queuedata;
 165         struct nvme_loop_queue *queue = hctx->driver_data;
 166         struct request *req = bd->rq;
 167         struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
 168         int ret;
 169
 170         ret = nvme_setup_cmd(ns, req, &iod->cmd);
 171         if (ret)
 172                 return ret;
 173
 174         iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
 175         iod->req.port = nvmet_loop_port;
 176         if (!nvmet_req_init(&iod->req, &queue->nvme_cq,
 177                         &queue->nvme_sq, &nvme_loop_ops)) {
 178                 nvme_cleanup_cmd(req);
 179                 blk_mq_start_request(req);
 180                 nvme_loop_queue_response(&iod->req);
 181                 return 0;
 182         }
 183
 184         if (blk_rq_bytes(req)) {
 185                 iod->sg_table.sgl = iod->first_sgl;
 186                 ret = sg_alloc_table_chained(&iod->sg_table,
 187                         req->nr_phys_segments, iod->sg_table.sgl);
 188                 if (ret)
 189                         return BLK_MQ_RQ_QUEUE_BUSY;
 190
 191                 iod->req.sg = iod->sg_table.sgl;
 192                 iod->req.sg_cnt = blk_rq_map_sg(req->q, req, iod->sg_table.sgl);
 193                 BUG_ON(iod->req.sg_cnt > req->nr_phys_segments);
 194         }
 195
 196         iod->cmd.common.command_id = req->tag;
 197         blk_mq_start_request(req);
 198
 199         schedule_work(&iod->work);
 200         return 0;
 201 }
 202
 203 static void nvme_loop_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
 204 {
 205         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(arg);
 206         struct nvme_loop_queue *queue = &ctrl->queues[0];
 207         struct nvme_loop_iod *iod = &ctrl->async_event_iod;
 208
 209         memset(&iod->cmd, 0, sizeof(iod->cmd));
 210         iod->cmd.common.opcode = nvme_admin_async_event;
 211         iod->cmd.common.command_id = NVME_LOOP_AQ_BLKMQ_DEPTH;
 212         iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
 213
 214         if (!nvmet_req_init(&iod->req, &queue->nvme_cq, &queue->nvme_sq,
 215                         &nvme_loop_ops)) {
 216                 dev_err(ctrl->ctrl.device, "failed async event work\n");
 217                 return;
 218         }
 219
 220         schedule_work(&iod->work);
 221 }
 222
 223 static int nvme_loop_init_iod(struct nvme_loop_ctrl *ctrl,
 224                 struct nvme_loop_iod *iod, unsigned int queue_idx)
 225 {
 226         BUG_ON(queue_idx >= ctrl->queue_count);
 227
 228         iod->req.cmd = &iod->cmd;
 229         iod->req.rsp = &iod->rsp;
 230         iod->queue = &ctrl->queues[queue_idx];
 231         INIT_WORK(&iod->work, nvme_loop_execute_work);
 232         return 0;
 233 }
 234
 235 static int nvme_loop_init_request(void *data, struct request *req,
 236                                 unsigned int hctx_idx, unsigned int rq_idx,
 237                                 unsigned int numa_node)
 238 {
 239         return nvme_loop_init_iod(data, blk_mq_rq_to_pdu(req), hctx_idx + 1);
 240 }
 241
 242 static int nvme_loop_init_admin_request(void *data, struct request *req,
 243                                 unsigned int hctx_idx, unsigned int rq_idx,
 244                                 unsigned int numa_node)
 245 {
 246         return nvme_loop_init_iod(data, blk_mq_rq_to_pdu(req), 0);
 247 }
 248
 249 static int nvme_loop_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 250                 unsigned int hctx_idx)
 251 {
 252         struct nvme_loop_ctrl *ctrl = data;
 253         struct nvme_loop_queue *queue = &ctrl->queues[hctx_idx + 1];
 254
 255         BUG_ON(hctx_idx >= ctrl->queue_count);
 256
 257         hctx->driver_data = queue;
 258         return 0;
 259 }
 260
 261 static int nvme_loop_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 262                 unsigned int hctx_idx)
 263 {
 264         struct nvme_loop_ctrl *ctrl = data;
 265         struct nvme_loop_queue *queue = &ctrl->queues[0];
 266
 267         BUG_ON(hctx_idx != 0);
 268
 269         hctx->driver_data = queue;
 270         return 0;
 271 }
 272
 273 static struct blk_mq_ops nvme_loop_mq_ops = {
 274         .queue_rq       = nvme_loop_queue_rq,
 275         .complete       = nvme_loop_complete_rq,
 276         .map_queue      = blk_mq_map_queue,
 277         .init_request   = nvme_loop_init_request,
 278         .init_hctx      = nvme_loop_init_hctx,
 279         .timeout        = nvme_loop_timeout,
 280 };
 281
 282 static struct blk_mq_ops nvme_loop_admin_mq_ops = {
 283         .queue_rq       = nvme_loop_queue_rq,
 284         .complete       = nvme_loop_complete_rq,
 285         .map_queue      = blk_mq_map_queue,
 286         .init_request   = nvme_loop_init_admin_request,
 287         .init_hctx      = nvme_loop_init_admin_hctx,
 288         .timeout        = nvme_loop_timeout,
 289 };
 290
 291 static void nvme_loop_destroy_admin_queue(struct nvme_loop_ctrl *ctrl)
 292 {
 293         blk_cleanup_queue(ctrl->ctrl.admin_q);
 294         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 295         nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
 296 }
 297
 298 static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl)
 299 {
 300         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
 301
 302         if (list_empty(&ctrl->list))
 303                 goto free_ctrl;
 304
 305         mutex_lock(&nvme_loop_ctrl_mutex);
 306         list_del(&ctrl->list);
 307         mutex_unlock(&nvme_loop_ctrl_mutex);
 308
 309         if (nctrl->tagset) {
 310                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 311                 blk_mq_free_tag_set(&ctrl->tag_set);
 312         }
 313         kfree(ctrl->queues);
 314         nvmf_free_options(nctrl->opts);
 315 free_ctrl:
 316         kfree(ctrl);
 317 }
 318
 319 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
 320 {
 321         int error;
 322
 323         memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
 324         ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
 325         ctrl->admin_tag_set.queue_depth = NVME_LOOP_AQ_BLKMQ_DEPTH;
 326         ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
 327         ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
 328         ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
 329                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
 330         ctrl->admin_tag_set.driver_data = ctrl;
 331         ctrl->admin_tag_set.nr_hw_queues = 1;
 332         ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
 333
 334         ctrl->queues[0].ctrl = ctrl;
 335         error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
 336         if (error)
 337                 return error;
 338         ctrl->queue_count = 1;
 339
 340         error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
 341         if (error)
 342                 goto out_free_sq;
 343
 344         ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 345         if (IS_ERR(ctrl->ctrl.admin_q)) {
 346                 error = PTR_ERR(ctrl->ctrl.admin_q);
 347                 goto out_free_tagset;
 348         }
 349
 350         error = nvmf_connect_admin_queue(&ctrl->ctrl);
 351         if (error)
 352                 goto out_cleanup_queue;
 353
 354         error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->cap);
 355         if (error) {
 356                 dev_err(ctrl->ctrl.device,
 357                         "prop_get NVME_REG_CAP failed\n");
 358                 goto out_cleanup_queue;
 359         }
 360
 361         ctrl->ctrl.sqsize =
 362                 min_t(int, NVME_CAP_MQES(ctrl->cap) + 1, ctrl->ctrl.sqsize);
 363
 364         error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
 365         if (error)
 366                 goto out_cleanup_queue;
 367
 368         ctrl->ctrl.max_hw_sectors =
 369                 (NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
 370
 371         error = nvme_init_identify(&ctrl->ctrl);
 372         if (error)
 373                 goto out_cleanup_queue;
 374
 375         nvme_start_keep_alive(&ctrl->ctrl);
 376
 377         return 0;
 378
 379 out_cleanup_queue:
 380         blk_cleanup_queue(ctrl->ctrl.admin_q);
 381 out_free_tagset:
 382         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 383 out_free_sq:
 384         nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
 385         return error;
 386 }
 387
 388 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
 389 {
 390         int i;
 391
 392         nvme_stop_keep_alive(&ctrl->ctrl);
 393
 394         if (ctrl->queue_count > 1) {
 395                 nvme_stop_queues(&ctrl->ctrl);
 396                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
 397                                         nvme_cancel_request, &ctrl->ctrl);
 398
 399                 for (i = 1; i < ctrl->queue_count; i++)
 400                         nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
 401         }
 402
 403         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
 404                 nvme_shutdown_ctrl(&ctrl->ctrl);
 405
 406         blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
 407         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
 408                                 nvme_cancel_request, &ctrl->ctrl);
 409         nvme_loop_destroy_admin_queue(ctrl);
 410 }
 411
 412 static void nvme_loop_del_ctrl_work(struct work_struct *work)
 413 {
 414         struct nvme_loop_ctrl *ctrl = container_of(work,
 415                                 struct nvme_loop_ctrl, delete_work);
 416
 417         nvme_uninit_ctrl(&ctrl->ctrl);
 418         nvme_loop_shutdown_ctrl(ctrl);
 419         nvme_put_ctrl(&ctrl->ctrl);
 420 }
 421
 422 static int __nvme_loop_del_ctrl(struct nvme_loop_ctrl *ctrl)
 423 {
 424         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
 425                 return -EBUSY;
 426
 427         if (!schedule_work(&ctrl->delete_work))
 428                 return -EBUSY;
 429
 430         return 0;
 431 }
 432
 433 static int nvme_loop_del_ctrl(struct nvme_ctrl *nctrl)
 434 {
 435         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
 436         int ret;
 437
 438         ret = __nvme_loop_del_ctrl(ctrl);
 439         if (ret)
 440                 return ret;
 441
 442         flush_work(&ctrl->delete_work);
 443
 444         return 0;
 445 }
 446
 447 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
 448 {
 449         struct nvme_loop_ctrl *ctrl;
 450
 451         mutex_lock(&nvme_loop_ctrl_mutex);
 452         list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
 453                 if (ctrl->ctrl.cntlid == nctrl->cntlid)
 454                         __nvme_loop_del_ctrl(ctrl);
 455         }
 456         mutex_unlock(&nvme_loop_ctrl_mutex);
 457 }
 458
 459 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
 460 {
 461         struct nvme_loop_ctrl *ctrl = container_of(work,
 462                                         struct nvme_loop_ctrl, reset_work);
 463         bool changed;
 464         int i, ret;
 465
 466         nvme_loop_shutdown_ctrl(ctrl);
 467
 468         ret = nvme_loop_configure_admin_queue(ctrl);
 469         if (ret)
 470                 goto out_disable;
 471
 472         for (i = 1; i <= ctrl->ctrl.opts->nr_io_queues; i++) {
 473                 ctrl->queues[i].ctrl = ctrl;
 474                 ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
 475                 if (ret)
 476                         goto out_free_queues;
 477
 478                 ctrl->queue_count++;
 479         }
 480
 481         for (i = 1; i <= ctrl->ctrl.opts->nr_io_queues; i++) {
 482                 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
 483                 if (ret)
 484                         goto out_free_queues;
 485         }
 486
 487         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 488         WARN_ON_ONCE(!changed);
 489
 490         nvme_queue_scan(&ctrl->ctrl);
 491         nvme_queue_async_events(&ctrl->ctrl);
 492
 493         nvme_start_queues(&ctrl->ctrl);
 494
 495         return;
 496
 497 out_free_queues:
 498         for (i = 1; i < ctrl->queue_count; i++)
 499                 nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
 500         nvme_loop_destroy_admin_queue(ctrl);
 501 out_disable:
 502         dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
 503         nvme_uninit_ctrl(&ctrl->ctrl);
 504         nvme_put_ctrl(&ctrl->ctrl);
 505 }
 506
 507 static int nvme_loop_reset_ctrl(struct nvme_ctrl *nctrl)
 508 {
 509         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
 510
 511         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
 512                 return -EBUSY;
 513
 514         if (!schedule_work(&ctrl->reset_work))
 515                 return -EBUSY;
 516
 517         flush_work(&ctrl->reset_work);
 518
 519         return 0;
 520 }
 521
 522 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
 523         .name                   = "loop",
 524         .module                 = THIS_MODULE,
 525         .is_fabrics             = true,
 526         .reg_read32             = nvmf_reg_read32,
 527         .reg_read64             = nvmf_reg_read64,
 528         .reg_write32            = nvmf_reg_write32,
 529         .reset_ctrl             = nvme_loop_reset_ctrl,
 530         .free_ctrl              = nvme_loop_free_ctrl,
 531         .submit_async_event     = nvme_loop_submit_async_event,
 532         .delete_ctrl            = nvme_loop_del_ctrl,
 533         .get_subsysnqn          = nvmf_get_subsysnqn,
 534 };
 535
 536 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
 537 {
 538         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 539         int ret, i;
 540
 541         ret = nvme_set_queue_count(&ctrl->ctrl, &opts->nr_io_queues);
 542         if (ret || !opts->nr_io_queues)
 543                 return ret;
 544
 545         dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n",
 546                 opts->nr_io_queues);
 547
 548         for (i = 1; i <= opts->nr_io_queues; i++) {
 549                 ctrl->queues[i].ctrl = ctrl;
 550                 ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
 551                 if (ret)
 552                         goto out_destroy_queues;
 553
 554                 ctrl->queue_count++;
 555         }
 556
 557         memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
 558         ctrl->tag_set.ops = &nvme_loop_mq_ops;
 559         ctrl->tag_set.queue_depth = ctrl->ctrl.sqsize;
 560         ctrl->tag_set.reserved_tags = 1; /* fabric connect */
 561         ctrl->tag_set.numa_node = NUMA_NO_NODE;
 562         ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
 563         ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
 564                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
 565         ctrl->tag_set.driver_data = ctrl;
 566         ctrl->tag_set.nr_hw_queues = ctrl->queue_count - 1;
 567         ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
 568         ctrl->ctrl.tagset = &ctrl->tag_set;
 569
 570         ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
 571         if (ret)
 572                 goto out_destroy_queues;
 573
 574         ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 575         if (IS_ERR(ctrl->ctrl.connect_q)) {
 576                 ret = PTR_ERR(ctrl->ctrl.connect_q);
 577                 goto out_free_tagset;
 578         }
 579
 580         for (i = 1; i <= opts->nr_io_queues; i++) {
 581                 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
 582                 if (ret)
 583                         goto out_cleanup_connect_q;
 584         }
 585
 586         return 0;
 587
 588 out_cleanup_connect_q:
 589         blk_cleanup_queue(ctrl->ctrl.connect_q);
 590 out_free_tagset:
 591         blk_mq_free_tag_set(&ctrl->tag_set);
 592 out_destroy_queues:
 593         for (i = 1; i < ctrl->queue_count; i++)
 594                 nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
 595         return ret;
 596 }
 597
 598 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
 599                 struct nvmf_ctrl_options *opts)
 600 {
 601         struct nvme_loop_ctrl *ctrl;
 602         bool changed;
 603         int ret;
 604
 605         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
 606         if (!ctrl)
 607                 return ERR_PTR(-ENOMEM);
 608         ctrl->ctrl.opts = opts;
 609         INIT_LIST_HEAD(&ctrl->list);
 610
 611         INIT_WORK(&ctrl->delete_work, nvme_loop_del_ctrl_work);
 612         INIT_WORK(&ctrl->reset_work, nvme_loop_reset_ctrl_work);
 613
 614         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
 615                                 0 /* no quirks, we're perfect! */);
 616         if (ret)
 617                 goto out_put_ctrl;
 618
 619         spin_lock_init(&ctrl->lock);
 620
 621         ret = -ENOMEM;
 622
 623         ctrl->ctrl.sqsize = opts->queue_size;
 624         ctrl->ctrl.kato = opts->kato;
 625
 626         ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
 627                         GFP_KERNEL);
 628         if (!ctrl->queues)
 629                 goto out_uninit_ctrl;
 630
 631         ret = nvme_loop_configure_admin_queue(ctrl);
 632         if (ret)
 633                 goto out_free_queues;
 634
 635         if (opts->queue_size > ctrl->ctrl.maxcmd) {
 636                 /* warn if maxcmd is lower than queue_size */
 637                 dev_warn(ctrl->ctrl.device,
 638                         "queue_size %zu > ctrl maxcmd %u, clamping down\n",
 639                         opts->queue_size, ctrl->ctrl.maxcmd);
 640                 opts->queue_size = ctrl->ctrl.maxcmd;
 641         }
 642
 643         if (opts->nr_io_queues) {
 644                 ret = nvme_loop_create_io_queues(ctrl);
 645                 if (ret)
 646                         goto out_remove_admin_queue;
 647         }
 648
 649         nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
 650
 651         dev_info(ctrl->ctrl.device,
 652                  "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
 653
 654         kref_get(&ctrl->ctrl.kref);
 655
 656         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 657         WARN_ON_ONCE(!changed);
 658
 659         mutex_lock(&nvme_loop_ctrl_mutex);
 660         list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
 661         mutex_unlock(&nvme_loop_ctrl_mutex);
 662
 663         if (opts->nr_io_queues) {
 664                 nvme_queue_scan(&ctrl->ctrl);
 665                 nvme_queue_async_events(&ctrl->ctrl);
 666         }
 667
 668         return &ctrl->ctrl;
 669
 670 out_remove_admin_queue:
 671         nvme_loop_destroy_admin_queue(ctrl);
 672 out_free_queues:
 673         kfree(ctrl->queues);
 674 out_uninit_ctrl:
 675         nvme_uninit_ctrl(&ctrl->ctrl);
 676 out_put_ctrl:
 677         nvme_put_ctrl(&ctrl->ctrl);
 678         if (ret > 0)
 679                 ret = -EIO;
 680         return ERR_PTR(ret);
 681 }
 682
 683 static int nvme_loop_add_port(struct nvmet_port *port)
 684 {
 685         /*
 686          * XXX: disalow adding more than one port so
 687          * there is no connection rejections when a
 688          * a subsystem is assigned to a port for which
 689          * loop doesn't have a pointer.
 690          * This scenario would be possible if we allowed
 691          * more than one port to be added and a subsystem
 692          * was assigned to a port other than nvmet_loop_port.
 693          */
 694
 695         if (nvmet_loop_port)
 696                 return -EPERM;
 697
 698         nvmet_loop_port = port;
 699         return 0;
 700 }
 701
 702 static void nvme_loop_remove_port(struct nvmet_port *port)
 703 {
 704         if (port == nvmet_loop_port)
 705                 nvmet_loop_port = NULL;
 706 }
 707
 708 static struct nvmet_fabrics_ops nvme_loop_ops = {
 709         .owner          = THIS_MODULE,
 710         .type           = NVMF_TRTYPE_LOOP,
 711         .add_port       = nvme_loop_add_port,
 712         .remove_port    = nvme_loop_remove_port,
 713         .queue_response = nvme_loop_queue_response,
 714         .delete_ctrl    = nvme_loop_delete_ctrl,
 715 };
 716
 717 static struct nvmf_transport_ops nvme_loop_transport = {
 718         .name           = "loop",
 719         .create_ctrl    = nvme_loop_create_ctrl,
 720 };
 721
 722 static int __init nvme_loop_init_module(void)
 723 {
 724         int ret;
 725
 726         ret = nvmet_register_transport(&nvme_loop_ops);
 727         if (ret)
 728                 return ret;
 729         nvmf_register_transport(&nvme_loop_transport);
 730         return 0;
 731 }
 732
 733 static void __exit nvme_loop_cleanup_module(void)
 734 {
 735         struct nvme_loop_ctrl *ctrl, *next;
 736
 737         nvmf_unregister_transport(&nvme_loop_transport);
 738         nvmet_unregister_transport(&nvme_loop_ops);
 739
 740         mutex_lock(&nvme_loop_ctrl_mutex);
 741         list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
 742                 __nvme_loop_del_ctrl(ctrl);
 743         mutex_unlock(&nvme_loop_ctrl_mutex);
 744
 745         flush_scheduled_work();
 746 }
 747
 748 module_init(nvme_loop_init_module);
 749 module_exit(nvme_loop_cleanup_module);
 750
 751 MODULE_LICENSE("GPL v2");
 752 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */