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/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_AQ_DEPTH              256
  25
  26 #define NVME_LOOP_MAX_SEGMENTS          256
  27
  28 /*
  29  * We handle AEN commands ourselves and don't even let the
  30  * block layer know about them.
  31  */
  32 #define NVME_LOOP_NR_AEN_COMMANDS       1
  33 #define NVME_LOOP_AQ_BLKMQ_DEPTH        \
  34         (NVME_LOOP_AQ_DEPTH - NVME_LOOP_NR_AEN_COMMANDS)
  35
  36 struct nvme_loop_iod {
  37         struct nvme_request     nvme_req;
  38         struct nvme_command     cmd;
  39         struct nvme_completion  rsp;
  40         struct nvmet_req        req;
  41         struct nvme_loop_queue  *queue;
  42         struct work_struct      work;
  43         struct sg_table         sg_table;
  44         struct scatterlist      first_sgl[];
  45 };
  46
  47 struct nvme_loop_ctrl {
  48         spinlock_t              lock;
  49         struct nvme_loop_queue  *queues;
  50         u32                     queue_count;
  51
  52         struct blk_mq_tag_set   admin_tag_set;
  53
  54         struct list_head        list;
  55         u64                     cap;
  56         struct blk_mq_tag_set   tag_set;
  57         struct nvme_loop_iod    async_event_iod;
  58         struct nvme_ctrl        ctrl;
  59
  60         struct nvmet_ctrl       *target_ctrl;
  61         struct work_struct      delete_work;
  62         struct work_struct      reset_work;
  63 };
  64
  65 static inline struct nvme_loop_ctrl *to_loop_ctrl(struct nvme_ctrl *ctrl)
  66 {
  67         return container_of(ctrl, struct nvme_loop_ctrl, ctrl);
  68 }
  69
  70 struct nvme_loop_queue {
  71         struct nvmet_cq         nvme_cq;
  72         struct nvmet_sq         nvme_sq;
  73         struct nvme_loop_ctrl   *ctrl;
  74 };
  75
  76 static struct nvmet_port *nvmet_loop_port;
  77
  78 static LIST_HEAD(nvme_loop_ctrl_list);
  79 static DEFINE_MUTEX(nvme_loop_ctrl_mutex);
  80
  81 static void nvme_loop_queue_response(struct nvmet_req *nvme_req);
  82 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *ctrl);
  83
  84 static struct nvmet_fabrics_ops nvme_loop_ops;
  85
  86 static inline int nvme_loop_queue_idx(struct nvme_loop_queue *queue)
  87 {
  88         return queue - queue->ctrl->queues;
  89 }
  90
  91 static void nvme_loop_complete_rq(struct request *req)
  92 {
  93         struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
  94         int error = 0;
  95
  96         nvme_cleanup_cmd(req);
  97         sg_free_table_chained(&iod->sg_table, true);
  98
  99         if (unlikely(req->errors)) {
 100                 if (nvme_req_needs_retry(req, req->errors)) {
 101                         req->retries++;
 102                         nvme_requeue_req(req);
 103                         return;
 104                 }
 105
 106                 if (blk_rq_is_passthrough(req))
 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 struct blk_mq_tags *nvme_loop_tagset(struct nvme_loop_queue *queue)
 116 {
 117         u32 queue_idx = nvme_loop_queue_idx(queue);
 118
 119         if (queue_idx == 0)
 120                 return queue->ctrl->admin_tag_set.tags[queue_idx];
 121         return queue->ctrl->tag_set.tags[queue_idx - 1];
 122 }
 123
 124 static void nvme_loop_queue_response(struct nvmet_req *req)
 125 {
 126         struct nvme_loop_queue *queue =
 127                 container_of(req->sq, struct nvme_loop_queue, nvme_sq);
 128         struct nvme_completion *cqe = req->rsp;
 129
 130         /*
 131          * AEN requests are special as they don't time out and can
 132          * survive any kind of queue freeze and often don't respond to
 133          * aborts.  We don't even bother to allocate a struct request
 134          * for them but rather special case them here.
 135          */
 136         if (unlikely(nvme_loop_queue_idx(queue) == 0 &&
 137                         cqe->command_id >= NVME_LOOP_AQ_BLKMQ_DEPTH)) {
 138                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
 139                                 &cqe->result);
 140         } else {
 141                 struct request *rq;
 142                 struct nvme_loop_iod *iod;
 143
 144                 rq = blk_mq_tag_to_rq(nvme_loop_tagset(queue), cqe->command_id);
 145                 if (!rq) {
 146                         dev_err(queue->ctrl->ctrl.device,
 147                                 "tag 0x%x on queue %d not found\n",
 148                                 cqe->command_id, nvme_loop_queue_idx(queue));
 149                         return;
 150                 }
 151
 152                 iod = blk_mq_rq_to_pdu(rq);
 153                 iod->nvme_req.result = cqe->result;
 154                 blk_mq_complete_request(rq, le16_to_cpu(cqe->status) >> 1);
 155         }
 156 }
 157
 158 static void nvme_loop_execute_work(struct work_struct *work)
 159 {
 160         struct nvme_loop_iod *iod =
 161                 container_of(work, struct nvme_loop_iod, work);
 162
 163         iod->req.execute(&iod->req);
 164 }
 165
 166 static enum blk_eh_timer_return
 167 nvme_loop_timeout(struct request *rq, bool reserved)
 168 {
 169         struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(rq);
 170
 171         /* queue error recovery */
 172         schedule_work(&iod->queue->ctrl->reset_work);
 173
 174         /* fail with DNR on admin cmd timeout */
 175         rq->errors = NVME_SC_ABORT_REQ | NVME_SC_DNR;
 176
 177         return BLK_EH_HANDLED;
 178 }
 179
 180 static int nvme_loop_queue_rq(struct blk_mq_hw_ctx *hctx,
 181                 const struct blk_mq_queue_data *bd)
 182 {
 183         struct nvme_ns *ns = hctx->queue->queuedata;
 184         struct nvme_loop_queue *queue = hctx->driver_data;
 185         struct request *req = bd->rq;
 186         struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
 187         int ret;
 188
 189         ret = nvme_setup_cmd(ns, req, &iod->cmd);
 190         if (ret != BLK_MQ_RQ_QUEUE_OK)
 191                 return ret;
 192
 193         iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
 194         iod->req.port = nvmet_loop_port;
 195         if (!nvmet_req_init(&iod->req, &queue->nvme_cq,
 196                         &queue->nvme_sq, &nvme_loop_ops)) {
 197                 nvme_cleanup_cmd(req);
 198                 blk_mq_start_request(req);
 199                 nvme_loop_queue_response(&iod->req);
 200                 return BLK_MQ_RQ_QUEUE_OK;
 201         }
 202
 203         if (blk_rq_bytes(req)) {
 204                 iod->sg_table.sgl = iod->first_sgl;
 205                 ret = sg_alloc_table_chained(&iod->sg_table,
 206                                 blk_rq_nr_phys_segments(req),
 207                                 iod->sg_table.sgl);
 208                 if (ret)
 209                         return BLK_MQ_RQ_QUEUE_BUSY;
 210
 211                 iod->req.sg = iod->sg_table.sgl;
 212                 iod->req.sg_cnt = blk_rq_map_sg(req->q, req, iod->sg_table.sgl);
 213         }
 214
 215         blk_mq_start_request(req);
 216
 217         schedule_work(&iod->work);
 218         return BLK_MQ_RQ_QUEUE_OK;
 219 }
 220
 221 static void nvme_loop_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
 222 {
 223         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(arg);
 224         struct nvme_loop_queue *queue = &ctrl->queues[0];
 225         struct nvme_loop_iod *iod = &ctrl->async_event_iod;
 226
 227         memset(&iod->cmd, 0, sizeof(iod->cmd));
 228         iod->cmd.common.opcode = nvme_admin_async_event;
 229         iod->cmd.common.command_id = NVME_LOOP_AQ_BLKMQ_DEPTH;
 230         iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
 231
 232         if (!nvmet_req_init(&iod->req, &queue->nvme_cq, &queue->nvme_sq,
 233                         &nvme_loop_ops)) {
 234                 dev_err(ctrl->ctrl.device, "failed async event work\n");
 235                 return;
 236         }
 237
 238         schedule_work(&iod->work);
 239 }
 240
 241 static int nvme_loop_init_iod(struct nvme_loop_ctrl *ctrl,
 242                 struct nvme_loop_iod *iod, unsigned int queue_idx)
 243 {
 244         iod->req.cmd = &iod->cmd;
 245         iod->req.rsp = &iod->rsp;
 246         iod->queue = &ctrl->queues[queue_idx];
 247         INIT_WORK(&iod->work, nvme_loop_execute_work);
 248         return 0;
 249 }
 250
 251 static int nvme_loop_init_request(void *data, struct request *req,
 252                                 unsigned int hctx_idx, unsigned int rq_idx,
 253                                 unsigned int numa_node)
 254 {
 255         return nvme_loop_init_iod(data, blk_mq_rq_to_pdu(req), hctx_idx + 1);
 256 }
 257
 258 static int nvme_loop_init_admin_request(void *data, struct request *req,
 259                                 unsigned int hctx_idx, unsigned int rq_idx,
 260                                 unsigned int numa_node)
 261 {
 262         return nvme_loop_init_iod(data, blk_mq_rq_to_pdu(req), 0);
 263 }
 264
 265 static int nvme_loop_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 266                 unsigned int hctx_idx)
 267 {
 268         struct nvme_loop_ctrl *ctrl = data;
 269         struct nvme_loop_queue *queue = &ctrl->queues[hctx_idx + 1];
 270
 271         BUG_ON(hctx_idx >= ctrl->queue_count);
 272
 273         hctx->driver_data = queue;
 274         return 0;
 275 }
 276
 277 static int nvme_loop_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 278                 unsigned int hctx_idx)
 279 {
 280         struct nvme_loop_ctrl *ctrl = data;
 281         struct nvme_loop_queue *queue = &ctrl->queues[0];
 282
 283         BUG_ON(hctx_idx != 0);
 284
 285         hctx->driver_data = queue;
 286         return 0;
 287 }
 288
 289 static const struct blk_mq_ops nvme_loop_mq_ops = {
 290         .queue_rq       = nvme_loop_queue_rq,
 291         .complete       = nvme_loop_complete_rq,
 292         .init_request   = nvme_loop_init_request,
 293         .init_hctx      = nvme_loop_init_hctx,
 294         .timeout        = nvme_loop_timeout,
 295 };
 296
 297 static const struct blk_mq_ops nvme_loop_admin_mq_ops = {
 298         .queue_rq       = nvme_loop_queue_rq,
 299         .complete       = nvme_loop_complete_rq,
 300         .init_request   = nvme_loop_init_admin_request,
 301         .init_hctx      = nvme_loop_init_admin_hctx,
 302         .timeout        = nvme_loop_timeout,
 303 };
 304
 305 static void nvme_loop_destroy_admin_queue(struct nvme_loop_ctrl *ctrl)
 306 {
 307         nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
 308         blk_cleanup_queue(ctrl->ctrl.admin_q);
 309         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 310 }
 311
 312 static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl)
 313 {
 314         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
 315
 316         if (list_empty(&ctrl->list))
 317                 goto free_ctrl;
 318
 319         mutex_lock(&nvme_loop_ctrl_mutex);
 320         list_del(&ctrl->list);
 321         mutex_unlock(&nvme_loop_ctrl_mutex);
 322
 323         if (nctrl->tagset) {
 324                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 325                 blk_mq_free_tag_set(&ctrl->tag_set);
 326         }
 327         kfree(ctrl->queues);
 328         nvmf_free_options(nctrl->opts);
 329 free_ctrl:
 330         kfree(ctrl);
 331 }
 332
 333 static void nvme_loop_destroy_io_queues(struct nvme_loop_ctrl *ctrl)
 334 {
 335         int i;
 336
 337         for (i = 1; i < ctrl->queue_count; i++)
 338                 nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
 339 }
 340
 341 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl)
 342 {
 343         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 344         unsigned int nr_io_queues;
 345         int ret, i;
 346
 347         nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
 348         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 349         if (ret || !nr_io_queues)
 350                 return ret;
 351
 352         dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues);
 353
 354         for (i = 1; i <= nr_io_queues; i++) {
 355                 ctrl->queues[i].ctrl = ctrl;
 356                 ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
 357                 if (ret)
 358                         goto out_destroy_queues;
 359
 360                 ctrl->queue_count++;
 361         }
 362
 363         return 0;
 364
 365 out_destroy_queues:
 366         nvme_loop_destroy_io_queues(ctrl);
 367         return ret;
 368 }
 369
 370 static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl)
 371 {
 372         int i, ret;
 373
 374         for (i = 1; i < ctrl->queue_count; i++) {
 375                 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
 376                 if (ret)
 377                         return ret;
 378         }
 379
 380         return 0;
 381 }
 382
 383 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
 384 {
 385         int error;
 386
 387         memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
 388         ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
 389         ctrl->admin_tag_set.queue_depth = NVME_LOOP_AQ_BLKMQ_DEPTH;
 390         ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
 391         ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
 392         ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
 393                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
 394         ctrl->admin_tag_set.driver_data = ctrl;
 395         ctrl->admin_tag_set.nr_hw_queues = 1;
 396         ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
 397
 398         ctrl->queues[0].ctrl = ctrl;
 399         error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
 400         if (error)
 401                 return error;
 402         ctrl->queue_count = 1;
 403
 404         error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
 405         if (error)
 406                 goto out_free_sq;
 407
 408         ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 409         if (IS_ERR(ctrl->ctrl.admin_q)) {
 410                 error = PTR_ERR(ctrl->ctrl.admin_q);
 411                 goto out_free_tagset;
 412         }
 413
 414         error = nvmf_connect_admin_queue(&ctrl->ctrl);
 415         if (error)
 416                 goto out_cleanup_queue;
 417
 418         error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->cap);
 419         if (error) {
 420                 dev_err(ctrl->ctrl.device,
 421                         "prop_get NVME_REG_CAP failed\n");
 422                 goto out_cleanup_queue;
 423         }
 424
 425         ctrl->ctrl.sqsize =
 426                 min_t(int, NVME_CAP_MQES(ctrl->cap) + 1, ctrl->ctrl.sqsize);
 427
 428         error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
 429         if (error)
 430                 goto out_cleanup_queue;
 431
 432         ctrl->ctrl.max_hw_sectors =
 433                 (NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
 434
 435         error = nvme_init_identify(&ctrl->ctrl);
 436         if (error)
 437                 goto out_cleanup_queue;
 438
 439         nvme_start_keep_alive(&ctrl->ctrl);
 440
 441         return 0;
 442
 443 out_cleanup_queue:
 444         blk_cleanup_queue(ctrl->ctrl.admin_q);
 445 out_free_tagset:
 446         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 447 out_free_sq:
 448         nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
 449         return error;
 450 }
 451
 452 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
 453 {
 454         nvme_stop_keep_alive(&ctrl->ctrl);
 455
 456         if (ctrl->queue_count > 1) {
 457                 nvme_stop_queues(&ctrl->ctrl);
 458                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
 459                                         nvme_cancel_request, &ctrl->ctrl);
 460                 nvme_loop_destroy_io_queues(ctrl);
 461         }
 462
 463         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
 464                 nvme_shutdown_ctrl(&ctrl->ctrl);
 465
 466         blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
 467         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
 468                                 nvme_cancel_request, &ctrl->ctrl);
 469         nvme_loop_destroy_admin_queue(ctrl);
 470 }
 471
 472 static void nvme_loop_del_ctrl_work(struct work_struct *work)
 473 {
 474         struct nvme_loop_ctrl *ctrl = container_of(work,
 475                                 struct nvme_loop_ctrl, delete_work);
 476
 477         nvme_uninit_ctrl(&ctrl->ctrl);
 478         nvme_loop_shutdown_ctrl(ctrl);
 479         nvme_put_ctrl(&ctrl->ctrl);
 480 }
 481
 482 static int __nvme_loop_del_ctrl(struct nvme_loop_ctrl *ctrl)
 483 {
 484         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
 485                 return -EBUSY;
 486
 487         if (!schedule_work(&ctrl->delete_work))
 488                 return -EBUSY;
 489
 490         return 0;
 491 }
 492
 493 static int nvme_loop_del_ctrl(struct nvme_ctrl *nctrl)
 494 {
 495         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
 496         int ret;
 497
 498         ret = __nvme_loop_del_ctrl(ctrl);
 499         if (ret)
 500                 return ret;
 501
 502         flush_work(&ctrl->delete_work);
 503
 504         return 0;
 505 }
 506
 507 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
 508 {
 509         struct nvme_loop_ctrl *ctrl;
 510
 511         mutex_lock(&nvme_loop_ctrl_mutex);
 512         list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
 513                 if (ctrl->ctrl.cntlid == nctrl->cntlid)
 514                         __nvme_loop_del_ctrl(ctrl);
 515         }
 516         mutex_unlock(&nvme_loop_ctrl_mutex);
 517 }
 518
 519 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
 520 {
 521         struct nvme_loop_ctrl *ctrl = container_of(work,
 522                                         struct nvme_loop_ctrl, reset_work);
 523         bool changed;
 524         int ret;
 525
 526         nvme_loop_shutdown_ctrl(ctrl);
 527
 528         ret = nvme_loop_configure_admin_queue(ctrl);
 529         if (ret)
 530                 goto out_disable;
 531
 532         ret = nvme_loop_init_io_queues(ctrl);
 533         if (ret)
 534                 goto out_destroy_admin;
 535
 536         ret = nvme_loop_connect_io_queues(ctrl);
 537         if (ret)
 538                 goto out_destroy_io;
 539
 540         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 541         WARN_ON_ONCE(!changed);
 542
 543         nvme_queue_scan(&ctrl->ctrl);
 544         nvme_queue_async_events(&ctrl->ctrl);
 545
 546         nvme_start_queues(&ctrl->ctrl);
 547
 548         return;
 549
 550 out_destroy_io:
 551         nvme_loop_destroy_io_queues(ctrl);
 552 out_destroy_admin:
 553         nvme_loop_destroy_admin_queue(ctrl);
 554 out_disable:
 555         dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
 556         nvme_uninit_ctrl(&ctrl->ctrl);
 557         nvme_put_ctrl(&ctrl->ctrl);
 558 }
 559
 560 static int nvme_loop_reset_ctrl(struct nvme_ctrl *nctrl)
 561 {
 562         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
 563
 564         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
 565                 return -EBUSY;
 566
 567         if (!schedule_work(&ctrl->reset_work))
 568                 return -EBUSY;
 569
 570         flush_work(&ctrl->reset_work);
 571
 572         return 0;
 573 }
 574
 575 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
 576         .name                   = "loop",
 577         .module                 = THIS_MODULE,
 578         .is_fabrics             = true,
 579         .reg_read32             = nvmf_reg_read32,
 580         .reg_read64             = nvmf_reg_read64,
 581         .reg_write32            = nvmf_reg_write32,
 582         .reset_ctrl             = nvme_loop_reset_ctrl,
 583         .free_ctrl              = nvme_loop_free_ctrl,
 584         .submit_async_event     = nvme_loop_submit_async_event,
 585         .delete_ctrl            = nvme_loop_del_ctrl,
 586         .get_subsysnqn          = nvmf_get_subsysnqn,
 587 };
 588
 589 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
 590 {
 591         int ret;
 592
 593         ret = nvme_loop_init_io_queues(ctrl);
 594         if (ret)
 595                 return ret;
 596
 597         memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
 598         ctrl->tag_set.ops = &nvme_loop_mq_ops;
 599         ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
 600         ctrl->tag_set.reserved_tags = 1; /* fabric connect */
 601         ctrl->tag_set.numa_node = NUMA_NO_NODE;
 602         ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
 603         ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
 604                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
 605         ctrl->tag_set.driver_data = ctrl;
 606         ctrl->tag_set.nr_hw_queues = ctrl->queue_count - 1;
 607         ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
 608         ctrl->ctrl.tagset = &ctrl->tag_set;
 609
 610         ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
 611         if (ret)
 612                 goto out_destroy_queues;
 613
 614         ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 615         if (IS_ERR(ctrl->ctrl.connect_q)) {
 616                 ret = PTR_ERR(ctrl->ctrl.connect_q);
 617                 goto out_free_tagset;
 618         }
 619
 620         ret = nvme_loop_connect_io_queues(ctrl);
 621         if (ret)
 622                 goto out_cleanup_connect_q;
 623
 624         return 0;
 625
 626 out_cleanup_connect_q:
 627         blk_cleanup_queue(ctrl->ctrl.connect_q);
 628 out_free_tagset:
 629         blk_mq_free_tag_set(&ctrl->tag_set);
 630 out_destroy_queues:
 631         nvme_loop_destroy_io_queues(ctrl);
 632         return ret;
 633 }
 634
 635 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
 636                 struct nvmf_ctrl_options *opts)
 637 {
 638         struct nvme_loop_ctrl *ctrl;
 639         bool changed;
 640         int ret;
 641
 642         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
 643         if (!ctrl)
 644                 return ERR_PTR(-ENOMEM);
 645         ctrl->ctrl.opts = opts;
 646         INIT_LIST_HEAD(&ctrl->list);
 647
 648         INIT_WORK(&ctrl->delete_work, nvme_loop_del_ctrl_work);
 649         INIT_WORK(&ctrl->reset_work, nvme_loop_reset_ctrl_work);
 650
 651         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
 652                                 0 /* no quirks, we're perfect! */);
 653         if (ret)
 654                 goto out_put_ctrl;
 655
 656         spin_lock_init(&ctrl->lock);
 657
 658         ret = -ENOMEM;
 659
 660         ctrl->ctrl.sqsize = opts->queue_size - 1;
 661         ctrl->ctrl.kato = opts->kato;
 662
 663         ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
 664                         GFP_KERNEL);
 665         if (!ctrl->queues)
 666                 goto out_uninit_ctrl;
 667
 668         ret = nvme_loop_configure_admin_queue(ctrl);
 669         if (ret)
 670                 goto out_free_queues;
 671
 672         if (opts->queue_size > ctrl->ctrl.maxcmd) {
 673                 /* warn if maxcmd is lower than queue_size */
 674                 dev_warn(ctrl->ctrl.device,
 675                         "queue_size %zu > ctrl maxcmd %u, clamping down\n",
 676                         opts->queue_size, ctrl->ctrl.maxcmd);
 677                 opts->queue_size = ctrl->ctrl.maxcmd;
 678         }
 679
 680         if (opts->nr_io_queues) {
 681                 ret = nvme_loop_create_io_queues(ctrl);
 682                 if (ret)
 683                         goto out_remove_admin_queue;
 684         }
 685
 686         nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
 687
 688         dev_info(ctrl->ctrl.device,
 689                  "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
 690
 691         kref_get(&ctrl->ctrl.kref);
 692
 693         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 694         WARN_ON_ONCE(!changed);
 695
 696         mutex_lock(&nvme_loop_ctrl_mutex);
 697         list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
 698         mutex_unlock(&nvme_loop_ctrl_mutex);
 699
 700         if (opts->nr_io_queues) {
 701                 nvme_queue_scan(&ctrl->ctrl);
 702                 nvme_queue_async_events(&ctrl->ctrl);
 703         }
 704
 705         return &ctrl->ctrl;
 706
 707 out_remove_admin_queue:
 708         nvme_loop_destroy_admin_queue(ctrl);
 709 out_free_queues:
 710         kfree(ctrl->queues);
 711 out_uninit_ctrl:
 712         nvme_uninit_ctrl(&ctrl->ctrl);
 713 out_put_ctrl:
 714         nvme_put_ctrl(&ctrl->ctrl);
 715         if (ret > 0)
 716                 ret = -EIO;
 717         return ERR_PTR(ret);
 718 }
 719
 720 static int nvme_loop_add_port(struct nvmet_port *port)
 721 {
 722         /*
 723          * XXX: disalow adding more than one port so
 724          * there is no connection rejections when a
 725          * a subsystem is assigned to a port for which
 726          * loop doesn't have a pointer.
 727          * This scenario would be possible if we allowed
 728          * more than one port to be added and a subsystem
 729          * was assigned to a port other than nvmet_loop_port.
 730          */
 731
 732         if (nvmet_loop_port)
 733                 return -EPERM;
 734
 735         nvmet_loop_port = port;
 736         return 0;
 737 }
 738
 739 static void nvme_loop_remove_port(struct nvmet_port *port)
 740 {
 741         if (port == nvmet_loop_port)
 742                 nvmet_loop_port = NULL;
 743 }
 744
 745 static struct nvmet_fabrics_ops nvme_loop_ops = {
 746         .owner          = THIS_MODULE,
 747         .type           = NVMF_TRTYPE_LOOP,
 748         .add_port       = nvme_loop_add_port,
 749         .remove_port    = nvme_loop_remove_port,
 750         .queue_response = nvme_loop_queue_response,
 751         .delete_ctrl    = nvme_loop_delete_ctrl,
 752 };
 753
 754 static struct nvmf_transport_ops nvme_loop_transport = {
 755         .name           = "loop",
 756         .create_ctrl    = nvme_loop_create_ctrl,
 757 };
 758
 759 static int __init nvme_loop_init_module(void)
 760 {
 761         int ret;
 762
 763         ret = nvmet_register_transport(&nvme_loop_ops);
 764         if (ret)
 765                 return ret;
 766
 767         ret = nvmf_register_transport(&nvme_loop_transport);
 768         if (ret)
 769                 nvmet_unregister_transport(&nvme_loop_ops);
 770
 771         return ret;
 772 }
 773
 774 static void __exit nvme_loop_cleanup_module(void)
 775 {
 776         struct nvme_loop_ctrl *ctrl, *next;
 777
 778         nvmf_unregister_transport(&nvme_loop_transport);
 779         nvmet_unregister_transport(&nvme_loop_ops);
 780
 781         mutex_lock(&nvme_loop_ctrl_mutex);
 782         list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
 783                 __nvme_loop_del_ctrl(ctrl);
 784         mutex_unlock(&nvme_loop_ctrl_mutex);
 785
 786         flush_scheduled_work();
 787 }
 788
 789 module_init(nvme_loop_init_module);
 790 module_exit(nvme_loop_cleanup_module);
 791
 792 MODULE_LICENSE("GPL v2");
 793 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */