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         iod->req.cmd = &iod->cmd;
 227         iod->req.rsp = &iod->rsp;
 228         iod->queue = &ctrl->queues[queue_idx];
 229         INIT_WORK(&iod->work, nvme_loop_execute_work);
 230         return 0;
 231 }
 232
 233 static int nvme_loop_init_request(void *data, struct request *req,
 234                                 unsigned int hctx_idx, unsigned int rq_idx,
 235                                 unsigned int numa_node)
 236 {
 237         return nvme_loop_init_iod(data, blk_mq_rq_to_pdu(req), hctx_idx + 1);
 238 }
 239
 240 static int nvme_loop_init_admin_request(void *data, struct request *req,
 241                                 unsigned int hctx_idx, unsigned int rq_idx,
 242                                 unsigned int numa_node)
 243 {
 244         return nvme_loop_init_iod(data, blk_mq_rq_to_pdu(req), 0);
 245 }
 246
 247 static int nvme_loop_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 248                 unsigned int hctx_idx)
 249 {
 250         struct nvme_loop_ctrl *ctrl = data;
 251         struct nvme_loop_queue *queue = &ctrl->queues[hctx_idx + 1];
 252
 253         BUG_ON(hctx_idx >= ctrl->queue_count);
 254
 255         hctx->driver_data = queue;
 256         return 0;
 257 }
 258
 259 static int nvme_loop_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 260                 unsigned int hctx_idx)
 261 {
 262         struct nvme_loop_ctrl *ctrl = data;
 263         struct nvme_loop_queue *queue = &ctrl->queues[0];
 264
 265         BUG_ON(hctx_idx != 0);
 266
 267         hctx->driver_data = queue;
 268         return 0;
 269 }
 270
 271 static struct blk_mq_ops nvme_loop_mq_ops = {
 272         .queue_rq       = nvme_loop_queue_rq,
 273         .complete       = nvme_loop_complete_rq,
 274         .init_request   = nvme_loop_init_request,
 275         .init_hctx      = nvme_loop_init_hctx,
 276         .timeout        = nvme_loop_timeout,
 277 };
 278
 279 static struct blk_mq_ops nvme_loop_admin_mq_ops = {
 280         .queue_rq       = nvme_loop_queue_rq,
 281         .complete       = nvme_loop_complete_rq,
 282         .init_request   = nvme_loop_init_admin_request,
 283         .init_hctx      = nvme_loop_init_admin_hctx,
 284         .timeout        = nvme_loop_timeout,
 285 };
 286
 287 static void nvme_loop_destroy_admin_queue(struct nvme_loop_ctrl *ctrl)
 288 {
 289         nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
 290         blk_cleanup_queue(ctrl->ctrl.admin_q);
 291         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 292 }
 293
 294 static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl)
 295 {
 296         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
 297
 298         if (list_empty(&ctrl->list))
 299                 goto free_ctrl;
 300
 301         mutex_lock(&nvme_loop_ctrl_mutex);
 302         list_del(&ctrl->list);
 303         mutex_unlock(&nvme_loop_ctrl_mutex);
 304
 305         if (nctrl->tagset) {
 306                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 307                 blk_mq_free_tag_set(&ctrl->tag_set);
 308         }
 309         kfree(ctrl->queues);
 310         nvmf_free_options(nctrl->opts);
 311 free_ctrl:
 312         kfree(ctrl);
 313 }
 314
 315 static void nvme_loop_destroy_io_queues(struct nvme_loop_ctrl *ctrl)
 316 {
 317         int i;
 318
 319         for (i = 1; i < ctrl->queue_count; i++)
 320                 nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
 321 }
 322
 323 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl)
 324 {
 325         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 326         unsigned int nr_io_queues;
 327         int ret, i;
 328
 329         nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
 330         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 331         if (ret || !nr_io_queues)
 332                 return ret;
 333
 334         dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues);
 335
 336         for (i = 1; i <= nr_io_queues; i++) {
 337                 ctrl->queues[i].ctrl = ctrl;
 338                 ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
 339                 if (ret)
 340                         goto out_destroy_queues;
 341
 342                 ctrl->queue_count++;
 343         }
 344
 345         return 0;
 346
 347 out_destroy_queues:
 348         nvme_loop_destroy_io_queues(ctrl);
 349         return ret;
 350 }
 351
 352 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
 353 {
 354         int error;
 355
 356         memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
 357         ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
 358         ctrl->admin_tag_set.queue_depth = NVME_LOOP_AQ_BLKMQ_DEPTH;
 359         ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
 360         ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
 361         ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
 362                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
 363         ctrl->admin_tag_set.driver_data = ctrl;
 364         ctrl->admin_tag_set.nr_hw_queues = 1;
 365         ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
 366
 367         ctrl->queues[0].ctrl = ctrl;
 368         error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
 369         if (error)
 370                 return error;
 371         ctrl->queue_count = 1;
 372
 373         error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
 374         if (error)
 375                 goto out_free_sq;
 376
 377         ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 378         if (IS_ERR(ctrl->ctrl.admin_q)) {
 379                 error = PTR_ERR(ctrl->ctrl.admin_q);
 380                 goto out_free_tagset;
 381         }
 382
 383         error = nvmf_connect_admin_queue(&ctrl->ctrl);
 384         if (error)
 385                 goto out_cleanup_queue;
 386
 387         error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->cap);
 388         if (error) {
 389                 dev_err(ctrl->ctrl.device,
 390                         "prop_get NVME_REG_CAP failed\n");
 391                 goto out_cleanup_queue;
 392         }
 393
 394         ctrl->ctrl.sqsize =
 395                 min_t(int, NVME_CAP_MQES(ctrl->cap) + 1, ctrl->ctrl.sqsize);
 396
 397         error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
 398         if (error)
 399                 goto out_cleanup_queue;
 400
 401         ctrl->ctrl.max_hw_sectors =
 402                 (NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
 403
 404         error = nvme_init_identify(&ctrl->ctrl);
 405         if (error)
 406                 goto out_cleanup_queue;
 407
 408         nvme_start_keep_alive(&ctrl->ctrl);
 409
 410         return 0;
 411
 412 out_cleanup_queue:
 413         blk_cleanup_queue(ctrl->ctrl.admin_q);
 414 out_free_tagset:
 415         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 416 out_free_sq:
 417         nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
 418         return error;
 419 }
 420
 421 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
 422 {
 423         nvme_stop_keep_alive(&ctrl->ctrl);
 424
 425         if (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_stop_hw_queues(ctrl->ctrl.admin_q);
 436         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
 437                                 nvme_cancel_request, &ctrl->ctrl);
 438         nvme_loop_destroy_admin_queue(ctrl);
 439 }
 440
 441 static void nvme_loop_del_ctrl_work(struct work_struct *work)
 442 {
 443         struct nvme_loop_ctrl *ctrl = container_of(work,
 444                                 struct nvme_loop_ctrl, delete_work);
 445
 446         nvme_uninit_ctrl(&ctrl->ctrl);
 447         nvme_loop_shutdown_ctrl(ctrl);
 448         nvme_put_ctrl(&ctrl->ctrl);
 449 }
 450
 451 static int __nvme_loop_del_ctrl(struct nvme_loop_ctrl *ctrl)
 452 {
 453         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
 454                 return -EBUSY;
 455
 456         if (!schedule_work(&ctrl->delete_work))
 457                 return -EBUSY;
 458
 459         return 0;
 460 }
 461
 462 static int nvme_loop_del_ctrl(struct nvme_ctrl *nctrl)
 463 {
 464         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
 465         int ret;
 466
 467         ret = __nvme_loop_del_ctrl(ctrl);
 468         if (ret)
 469                 return ret;
 470
 471         flush_work(&ctrl->delete_work);
 472
 473         return 0;
 474 }
 475
 476 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
 477 {
 478         struct nvme_loop_ctrl *ctrl;
 479
 480         mutex_lock(&nvme_loop_ctrl_mutex);
 481         list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
 482                 if (ctrl->ctrl.cntlid == nctrl->cntlid)
 483                         __nvme_loop_del_ctrl(ctrl);
 484         }
 485         mutex_unlock(&nvme_loop_ctrl_mutex);
 486 }
 487
 488 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
 489 {
 490         struct nvme_loop_ctrl *ctrl = container_of(work,
 491                                         struct nvme_loop_ctrl, reset_work);
 492         bool changed;
 493         int i, ret;
 494
 495         nvme_loop_shutdown_ctrl(ctrl);
 496
 497         ret = nvme_loop_configure_admin_queue(ctrl);
 498         if (ret)
 499                 goto out_disable;
 500
 501         ret = nvme_loop_init_io_queues(ctrl);
 502         if (ret)
 503                 goto out_destroy_admin;
 504
 505         for (i = 1; i < ctrl->queue_count; i++) {
 506                 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
 507                 if (ret)
 508                         goto out_destroy_io;
 509         }
 510
 511         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 512         WARN_ON_ONCE(!changed);
 513
 514         nvme_queue_scan(&ctrl->ctrl);
 515         nvme_queue_async_events(&ctrl->ctrl);
 516
 517         nvme_start_queues(&ctrl->ctrl);
 518
 519         return;
 520
 521 out_destroy_io:
 522         nvme_loop_destroy_io_queues(ctrl);
 523 out_destroy_admin:
 524         nvme_loop_destroy_admin_queue(ctrl);
 525 out_disable:
 526         dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
 527         nvme_uninit_ctrl(&ctrl->ctrl);
 528         nvme_put_ctrl(&ctrl->ctrl);
 529 }
 530
 531 static int nvme_loop_reset_ctrl(struct nvme_ctrl *nctrl)
 532 {
 533         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
 534
 535         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
 536                 return -EBUSY;
 537
 538         if (!schedule_work(&ctrl->reset_work))
 539                 return -EBUSY;
 540
 541         flush_work(&ctrl->reset_work);
 542
 543         return 0;
 544 }
 545
 546 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
 547         .name                   = "loop",
 548         .module                 = THIS_MODULE,
 549         .is_fabrics             = true,
 550         .reg_read32             = nvmf_reg_read32,
 551         .reg_read64             = nvmf_reg_read64,
 552         .reg_write32            = nvmf_reg_write32,
 553         .reset_ctrl             = nvme_loop_reset_ctrl,
 554         .free_ctrl              = nvme_loop_free_ctrl,
 555         .submit_async_event     = nvme_loop_submit_async_event,
 556         .delete_ctrl            = nvme_loop_del_ctrl,
 557         .get_subsysnqn          = nvmf_get_subsysnqn,
 558 };
 559
 560 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
 561 {
 562         int ret, i;
 563
 564         ret = nvme_loop_init_io_queues(ctrl);
 565         if (ret)
 566                 return ret;
 567
 568         memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
 569         ctrl->tag_set.ops = &nvme_loop_mq_ops;
 570         ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
 571         ctrl->tag_set.reserved_tags = 1; /* fabric connect */
 572         ctrl->tag_set.numa_node = NUMA_NO_NODE;
 573         ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
 574         ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
 575                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
 576         ctrl->tag_set.driver_data = ctrl;
 577         ctrl->tag_set.nr_hw_queues = ctrl->queue_count - 1;
 578         ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
 579         ctrl->ctrl.tagset = &ctrl->tag_set;
 580
 581         ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
 582         if (ret)
 583                 goto out_destroy_queues;
 584
 585         ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 586         if (IS_ERR(ctrl->ctrl.connect_q)) {
 587                 ret = PTR_ERR(ctrl->ctrl.connect_q);
 588                 goto out_free_tagset;
 589         }
 590
 591         for (i = 1; i < ctrl->queue_count; i++) {
 592                 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
 593                 if (ret)
 594                         goto out_cleanup_connect_q;
 595         }
 596
 597         return 0;
 598
 599 out_cleanup_connect_q:
 600         blk_cleanup_queue(ctrl->ctrl.connect_q);
 601 out_free_tagset:
 602         blk_mq_free_tag_set(&ctrl->tag_set);
 603 out_destroy_queues:
 604         nvme_loop_destroy_io_queues(ctrl);
 605         return ret;
 606 }
 607
 608 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
 609                 struct nvmf_ctrl_options *opts)
 610 {
 611         struct nvme_loop_ctrl *ctrl;
 612         bool changed;
 613         int ret;
 614
 615         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
 616         if (!ctrl)
 617                 return ERR_PTR(-ENOMEM);
 618         ctrl->ctrl.opts = opts;
 619         INIT_LIST_HEAD(&ctrl->list);
 620
 621         INIT_WORK(&ctrl->delete_work, nvme_loop_del_ctrl_work);
 622         INIT_WORK(&ctrl->reset_work, nvme_loop_reset_ctrl_work);
 623
 624         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
 625                                 0 /* no quirks, we're perfect! */);
 626         if (ret)
 627                 goto out_put_ctrl;
 628
 629         spin_lock_init(&ctrl->lock);
 630
 631         ret = -ENOMEM;
 632
 633         ctrl->ctrl.sqsize = opts->queue_size - 1;
 634         ctrl->ctrl.kato = opts->kato;
 635
 636         ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
 637                         GFP_KERNEL);
 638         if (!ctrl->queues)
 639                 goto out_uninit_ctrl;
 640
 641         ret = nvme_loop_configure_admin_queue(ctrl);
 642         if (ret)
 643                 goto out_free_queues;
 644
 645         if (opts->queue_size > ctrl->ctrl.maxcmd) {
 646                 /* warn if maxcmd is lower than queue_size */
 647                 dev_warn(ctrl->ctrl.device,
 648                         "queue_size %zu > ctrl maxcmd %u, clamping down\n",
 649                         opts->queue_size, ctrl->ctrl.maxcmd);
 650                 opts->queue_size = ctrl->ctrl.maxcmd;
 651         }
 652
 653         if (opts->nr_io_queues) {
 654                 ret = nvme_loop_create_io_queues(ctrl);
 655                 if (ret)
 656                         goto out_remove_admin_queue;
 657         }
 658
 659         nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
 660
 661         dev_info(ctrl->ctrl.device,
 662                  "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
 663
 664         kref_get(&ctrl->ctrl.kref);
 665
 666         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 667         WARN_ON_ONCE(!changed);
 668
 669         mutex_lock(&nvme_loop_ctrl_mutex);
 670         list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
 671         mutex_unlock(&nvme_loop_ctrl_mutex);
 672
 673         if (opts->nr_io_queues) {
 674                 nvme_queue_scan(&ctrl->ctrl);
 675                 nvme_queue_async_events(&ctrl->ctrl);
 676         }
 677
 678         return &ctrl->ctrl;
 679
 680 out_remove_admin_queue:
 681         nvme_loop_destroy_admin_queue(ctrl);
 682 out_free_queues:
 683         kfree(ctrl->queues);
 684 out_uninit_ctrl:
 685         nvme_uninit_ctrl(&ctrl->ctrl);
 686 out_put_ctrl:
 687         nvme_put_ctrl(&ctrl->ctrl);
 688         if (ret > 0)
 689                 ret = -EIO;
 690         return ERR_PTR(ret);
 691 }
 692
 693 static int nvme_loop_add_port(struct nvmet_port *port)
 694 {
 695         /*
 696          * XXX: disalow adding more than one port so
 697          * there is no connection rejections when a
 698          * a subsystem is assigned to a port for which
 699          * loop doesn't have a pointer.
 700          * This scenario would be possible if we allowed
 701          * more than one port to be added and a subsystem
 702          * was assigned to a port other than nvmet_loop_port.
 703          */
 704
 705         if (nvmet_loop_port)
 706                 return -EPERM;
 707
 708         nvmet_loop_port = port;
 709         return 0;
 710 }
 711
 712 static void nvme_loop_remove_port(struct nvmet_port *port)
 713 {
 714         if (port == nvmet_loop_port)
 715                 nvmet_loop_port = NULL;
 716 }
 717
 718 static struct nvmet_fabrics_ops nvme_loop_ops = {
 719         .owner          = THIS_MODULE,
 720         .type           = NVMF_TRTYPE_LOOP,
 721         .add_port       = nvme_loop_add_port,
 722         .remove_port    = nvme_loop_remove_port,
 723         .queue_response = nvme_loop_queue_response,
 724         .delete_ctrl    = nvme_loop_delete_ctrl,
 725 };
 726
 727 static struct nvmf_transport_ops nvme_loop_transport = {
 728         .name           = "loop",
 729         .create_ctrl    = nvme_loop_create_ctrl,
 730 };
 731
 732 static int __init nvme_loop_init_module(void)
 733 {
 734         int ret;
 735
 736         ret = nvmet_register_transport(&nvme_loop_ops);
 737         if (ret)
 738                 return ret;
 739         nvmf_register_transport(&nvme_loop_transport);
 740         return 0;
 741 }
 742
 743 static void __exit nvme_loop_cleanup_module(void)
 744 {
 745         struct nvme_loop_ctrl *ctrl, *next;
 746
 747         nvmf_unregister_transport(&nvme_loop_transport);
 748         nvmet_unregister_transport(&nvme_loop_ops);
 749
 750         mutex_lock(&nvme_loop_ctrl_mutex);
 751         list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
 752                 __nvme_loop_del_ctrl(ctrl);
 753         mutex_unlock(&nvme_loop_ctrl_mutex);
 754
 755         flush_scheduled_work();
 756 }
 757
 758 module_init(nvme_loop_init_module);
 759 module_exit(nvme_loop_cleanup_module);
 760
 761 MODULE_LICENSE("GPL v2");
 762 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */