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_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 enum nvme_loop_queue_flags {
  65         NVME_LOOP_Q_LIVE        = 0,
  66 };
  67
  68 struct nvme_loop_queue {
  69         struct nvmet_cq         nvme_cq;
  70         struct nvmet_sq         nvme_sq;
  71         struct nvme_loop_ctrl   *ctrl;
  72         unsigned long           flags;
  73 };
  74
  75 static struct nvmet_port *nvmet_loop_port;
  76
  77 static LIST_HEAD(nvme_loop_ctrl_list);
  78 static DEFINE_MUTEX(nvme_loop_ctrl_mutex);
  79
  80 static void nvme_loop_queue_response(struct nvmet_req *nvme_req);
  81 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *ctrl);
  82
  83 static struct nvmet_fabrics_ops nvme_loop_ops;
  84
  85 static inline int nvme_loop_queue_idx(struct nvme_loop_queue *queue)
  86 {
  87         return queue - queue->ctrl->queues;
  88 }
  89
  90 static void nvme_loop_complete_rq(struct request *req)
  91 {
  92         struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
  93
  94         nvme_cleanup_cmd(req);
  95         sg_free_table_chained(&iod->sg_table, true);
  96         nvme_complete_rq(req);
  97 }
  98
  99 static struct blk_mq_tags *nvme_loop_tagset(struct nvme_loop_queue *queue)
 100 {
 101         u32 queue_idx = nvme_loop_queue_idx(queue);
 102
 103         if (queue_idx == 0)
 104                 return queue->ctrl->admin_tag_set.tags[queue_idx];
 105         return queue->ctrl->tag_set.tags[queue_idx - 1];
 106 }
 107
 108 static void nvme_loop_queue_response(struct nvmet_req *req)
 109 {
 110         struct nvme_loop_queue *queue =
 111                 container_of(req->sq, struct nvme_loop_queue, nvme_sq);
 112         struct nvme_completion *cqe = req->rsp;
 113
 114         /*
 115          * AEN requests are special as they don't time out and can
 116          * survive any kind of queue freeze and often don't respond to
 117          * aborts.  We don't even bother to allocate a struct request
 118          * for them but rather special case them here.
 119          */
 120         if (unlikely(nvme_loop_queue_idx(queue) == 0 &&
 121                         cqe->command_id >= NVME_LOOP_AQ_BLKMQ_DEPTH)) {
 122                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
 123                                 &cqe->result);
 124         } else {
 125                 struct request *rq;
 126
 127                 rq = blk_mq_tag_to_rq(nvme_loop_tagset(queue), cqe->command_id);
 128                 if (!rq) {
 129                         dev_err(queue->ctrl->ctrl.device,
 130                                 "tag 0x%x on queue %d not found\n",
 131                                 cqe->command_id, nvme_loop_queue_idx(queue));
 132                         return;
 133                 }
 134
 135                 nvme_end_request(rq, cqe->status, cqe->result);
 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         nvme_reset_ctrl(&iod->queue->ctrl->ctrl);
 154
 155         /* fail with DNR on admin cmd timeout */
 156         nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
 157
 158         return BLK_EH_HANDLED;
 159 }
 160
 161 static inline blk_status_t nvme_loop_is_ready(struct nvme_loop_queue *queue,
 162                 struct request *rq)
 163 {
 164         if (unlikely(!test_bit(NVME_LOOP_Q_LIVE, &queue->flags)))
 165                 return nvmf_check_init_req(&queue->ctrl->ctrl, rq);
 166         return BLK_STS_OK;
 167 }
 168
 169 static blk_status_t nvme_loop_queue_rq(struct blk_mq_hw_ctx *hctx,
 170                 const struct blk_mq_queue_data *bd)
 171 {
 172         struct nvme_ns *ns = hctx->queue->queuedata;
 173         struct nvme_loop_queue *queue = hctx->driver_data;
 174         struct request *req = bd->rq;
 175         struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
 176         blk_status_t ret;
 177
 178         ret = nvme_loop_is_ready(queue, req);
 179         if (unlikely(ret))
 180                 return ret;
 181
 182         ret = nvme_setup_cmd(ns, req, &iod->cmd);
 183         if (ret)
 184                 return ret;
 185
 186         blk_mq_start_request(req);
 187         iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
 188         iod->req.port = nvmet_loop_port;
 189         if (!nvmet_req_init(&iod->req, &queue->nvme_cq,
 190                         &queue->nvme_sq, &nvme_loop_ops))
 191                 return BLK_STS_OK;
 192
 193         if (blk_rq_bytes(req)) {
 194                 iod->sg_table.sgl = iod->first_sgl;
 195                 if (sg_alloc_table_chained(&iod->sg_table,
 196                                 blk_rq_nr_phys_segments(req),
 197                                 iod->sg_table.sgl))
 198                         return BLK_STS_RESOURCE;
 199
 200                 iod->req.sg = iod->sg_table.sgl;
 201                 iod->req.sg_cnt = blk_rq_map_sg(req->q, req, iod->sg_table.sgl);
 202         }
 203
 204         schedule_work(&iod->work);
 205         return BLK_STS_OK;
 206 }
 207
 208 static void nvme_loop_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
 209 {
 210         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(arg);
 211         struct nvme_loop_queue *queue = &ctrl->queues[0];
 212         struct nvme_loop_iod *iod = &ctrl->async_event_iod;
 213
 214         memset(&iod->cmd, 0, sizeof(iod->cmd));
 215         iod->cmd.common.opcode = nvme_admin_async_event;
 216         iod->cmd.common.command_id = NVME_LOOP_AQ_BLKMQ_DEPTH;
 217         iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
 218
 219         if (!nvmet_req_init(&iod->req, &queue->nvme_cq, &queue->nvme_sq,
 220                         &nvme_loop_ops)) {
 221                 dev_err(ctrl->ctrl.device, "failed async event work\n");
 222                 return;
 223         }
 224
 225         schedule_work(&iod->work);
 226 }
 227
 228 static int nvme_loop_init_iod(struct nvme_loop_ctrl *ctrl,
 229                 struct nvme_loop_iod *iod, unsigned int queue_idx)
 230 {
 231         iod->req.cmd = &iod->cmd;
 232         iod->req.rsp = &iod->rsp;
 233         iod->queue = &ctrl->queues[queue_idx];
 234         INIT_WORK(&iod->work, nvme_loop_execute_work);
 235         return 0;
 236 }
 237
 238 static int nvme_loop_init_request(struct blk_mq_tag_set *set,
 239                 struct request *req, unsigned int hctx_idx,
 240                 unsigned int numa_node)
 241 {
 242         struct nvme_loop_ctrl *ctrl = set->driver_data;
 243
 244         return nvme_loop_init_iod(ctrl, blk_mq_rq_to_pdu(req),
 245                         (set == &ctrl->tag_set) ? hctx_idx + 1 : 0);
 246 }
 247
 248 static int nvme_loop_init_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[hctx_idx + 1];
 253
 254         BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
 255
 256         hctx->driver_data = queue;
 257         return 0;
 258 }
 259
 260 static int nvme_loop_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 261                 unsigned int hctx_idx)
 262 {
 263         struct nvme_loop_ctrl *ctrl = data;
 264         struct nvme_loop_queue *queue = &ctrl->queues[0];
 265
 266         BUG_ON(hctx_idx != 0);
 267
 268         hctx->driver_data = queue;
 269         return 0;
 270 }
 271
 272 static const struct blk_mq_ops nvme_loop_mq_ops = {
 273         .queue_rq       = nvme_loop_queue_rq,
 274         .complete       = nvme_loop_complete_rq,
 275         .init_request   = nvme_loop_init_request,
 276         .init_hctx      = nvme_loop_init_hctx,
 277         .timeout        = nvme_loop_timeout,
 278 };
 279
 280 static const struct blk_mq_ops nvme_loop_admin_mq_ops = {
 281         .queue_rq       = nvme_loop_queue_rq,
 282         .complete       = nvme_loop_complete_rq,
 283         .init_request   = nvme_loop_init_request,
 284         .init_hctx      = nvme_loop_init_admin_hctx,
 285         .timeout        = nvme_loop_timeout,
 286 };
 287
 288 static void nvme_loop_destroy_admin_queue(struct nvme_loop_ctrl *ctrl)
 289 {
 290         if (!test_and_clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags))
 291                 return;
 292         nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
 293         blk_cleanup_queue(ctrl->ctrl.admin_q);
 294         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 295 }
 296
 297 static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl)
 298 {
 299         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
 300
 301         if (list_empty(&ctrl->list))
 302                 goto free_ctrl;
 303
 304         mutex_lock(&nvme_loop_ctrl_mutex);
 305         list_del(&ctrl->list);
 306         mutex_unlock(&nvme_loop_ctrl_mutex);
 307
 308         if (nctrl->tagset) {
 309                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 310                 blk_mq_free_tag_set(&ctrl->tag_set);
 311         }
 312         kfree(ctrl->queues);
 313         nvmf_free_options(nctrl->opts);
 314 free_ctrl:
 315         kfree(ctrl);
 316 }
 317
 318 static void nvme_loop_destroy_io_queues(struct nvme_loop_ctrl *ctrl)
 319 {
 320         int i;
 321
 322         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 323                 clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
 324                 nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
 325         }
 326         ctrl->ctrl.queue_count = 1;
 327 }
 328
 329 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl)
 330 {
 331         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 332         unsigned int nr_io_queues;
 333         int ret, i;
 334
 335         nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
 336         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 337         if (ret || !nr_io_queues)
 338                 return ret;
 339
 340         dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues);
 341
 342         for (i = 1; i <= nr_io_queues; i++) {
 343                 ctrl->queues[i].ctrl = ctrl;
 344                 ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
 345                 if (ret)
 346                         goto out_destroy_queues;
 347
 348                 ctrl->ctrl.queue_count++;
 349         }
 350
 351         return 0;
 352
 353 out_destroy_queues:
 354         nvme_loop_destroy_io_queues(ctrl);
 355         return ret;
 356 }
 357
 358 static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl)
 359 {
 360         int i, ret;
 361
 362         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 363                 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
 364                 if (ret)
 365                         return ret;
 366                 set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
 367         }
 368
 369         return 0;
 370 }
 371
 372 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
 373 {
 374         int error;
 375
 376         memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
 377         ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
 378         ctrl->admin_tag_set.queue_depth = NVME_LOOP_AQ_BLKMQ_DEPTH;
 379         ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
 380         ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
 381         ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
 382                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
 383         ctrl->admin_tag_set.driver_data = ctrl;
 384         ctrl->admin_tag_set.nr_hw_queues = 1;
 385         ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
 386
 387         ctrl->queues[0].ctrl = ctrl;
 388         error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
 389         if (error)
 390                 return error;
 391         ctrl->ctrl.queue_count = 1;
 392
 393         error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
 394         if (error)
 395                 goto out_free_sq;
 396         ctrl->ctrl.admin_tagset = &ctrl->admin_tag_set;
 397
 398         ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 399         if (IS_ERR(ctrl->ctrl.admin_q)) {
 400                 error = PTR_ERR(ctrl->ctrl.admin_q);
 401                 goto out_free_tagset;
 402         }
 403
 404         error = nvmf_connect_admin_queue(&ctrl->ctrl);
 405         if (error)
 406                 goto out_cleanup_queue;
 407
 408         set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
 409
 410         error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->ctrl.cap);
 411         if (error) {
 412                 dev_err(ctrl->ctrl.device,
 413                         "prop_get NVME_REG_CAP failed\n");
 414                 goto out_cleanup_queue;
 415         }
 416
 417         ctrl->ctrl.sqsize =
 418                 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
 419
 420         error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
 421         if (error)
 422                 goto out_cleanup_queue;
 423
 424         ctrl->ctrl.max_hw_sectors =
 425                 (NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
 426
 427         error = nvme_init_identify(&ctrl->ctrl);
 428         if (error)
 429                 goto out_cleanup_queue;
 430
 431         return 0;
 432
 433 out_cleanup_queue:
 434         clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
 435         blk_cleanup_queue(ctrl->ctrl.admin_q);
 436 out_free_tagset:
 437         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 438 out_free_sq:
 439         nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
 440         return error;
 441 }
 442
 443 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
 444 {
 445         if (ctrl->ctrl.queue_count > 1) {
 446                 nvme_stop_queues(&ctrl->ctrl);
 447                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
 448                                         nvme_cancel_request, &ctrl->ctrl);
 449                 nvme_loop_destroy_io_queues(ctrl);
 450         }
 451
 452         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
 453                 nvme_shutdown_ctrl(&ctrl->ctrl);
 454
 455         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
 456         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
 457                                 nvme_cancel_request, &ctrl->ctrl);
 458         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
 459         nvme_loop_destroy_admin_queue(ctrl);
 460 }
 461
 462 static void nvme_loop_del_ctrl_work(struct work_struct *work)
 463 {
 464         struct nvme_loop_ctrl *ctrl = container_of(work,
 465                                 struct nvme_loop_ctrl, delete_work);
 466
 467         nvme_stop_ctrl(&ctrl->ctrl);
 468         nvme_remove_namespaces(&ctrl->ctrl);
 469         nvme_loop_shutdown_ctrl(ctrl);
 470         nvme_uninit_ctrl(&ctrl->ctrl);
 471         nvme_put_ctrl(&ctrl->ctrl);
 472 }
 473
 474 static int __nvme_loop_del_ctrl(struct nvme_loop_ctrl *ctrl)
 475 {
 476         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
 477                 return -EBUSY;
 478
 479         if (!queue_work(nvme_wq, &ctrl->delete_work))
 480                 return -EBUSY;
 481
 482         return 0;
 483 }
 484
 485 static int nvme_loop_del_ctrl(struct nvme_ctrl *nctrl)
 486 {
 487         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
 488         int ret;
 489
 490         ret = __nvme_loop_del_ctrl(ctrl);
 491         if (ret)
 492                 return ret;
 493
 494         flush_work(&ctrl->delete_work);
 495
 496         return 0;
 497 }
 498
 499 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
 500 {
 501         struct nvme_loop_ctrl *ctrl;
 502
 503         mutex_lock(&nvme_loop_ctrl_mutex);
 504         list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
 505                 if (ctrl->ctrl.cntlid == nctrl->cntlid)
 506                         __nvme_loop_del_ctrl(ctrl);
 507         }
 508         mutex_unlock(&nvme_loop_ctrl_mutex);
 509 }
 510
 511 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
 512 {
 513         struct nvme_loop_ctrl *ctrl =
 514                 container_of(work, struct nvme_loop_ctrl, ctrl.reset_work);
 515         bool changed;
 516         int ret;
 517
 518         nvme_stop_ctrl(&ctrl->ctrl);
 519         nvme_loop_shutdown_ctrl(ctrl);
 520
 521         ret = nvme_loop_configure_admin_queue(ctrl);
 522         if (ret)
 523                 goto out_disable;
 524
 525         ret = nvme_loop_init_io_queues(ctrl);
 526         if (ret)
 527                 goto out_destroy_admin;
 528
 529         ret = nvme_loop_connect_io_queues(ctrl);
 530         if (ret)
 531                 goto out_destroy_io;
 532
 533         blk_mq_update_nr_hw_queues(&ctrl->tag_set,
 534                         ctrl->ctrl.queue_count - 1);
 535
 536         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 537         WARN_ON_ONCE(!changed);
 538
 539         nvme_start_ctrl(&ctrl->ctrl);
 540
 541         return;
 542
 543 out_destroy_io:
 544         nvme_loop_destroy_io_queues(ctrl);
 545 out_destroy_admin:
 546         nvme_loop_destroy_admin_queue(ctrl);
 547 out_disable:
 548         dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
 549         nvme_uninit_ctrl(&ctrl->ctrl);
 550         nvme_put_ctrl(&ctrl->ctrl);
 551 }
 552
 553 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
 554         .name                   = "loop",
 555         .module                 = THIS_MODULE,
 556         .flags                  = NVME_F_FABRICS,
 557         .reg_read32             = nvmf_reg_read32,
 558         .reg_read64             = nvmf_reg_read64,
 559         .reg_write32            = nvmf_reg_write32,
 560         .free_ctrl              = nvme_loop_free_ctrl,
 561         .submit_async_event     = nvme_loop_submit_async_event,
 562         .delete_ctrl            = nvme_loop_del_ctrl,
 563 };
 564
 565 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
 566 {
 567         int ret;
 568
 569         ret = nvme_loop_init_io_queues(ctrl);
 570         if (ret)
 571                 return ret;
 572
 573         memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
 574         ctrl->tag_set.ops = &nvme_loop_mq_ops;
 575         ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
 576         ctrl->tag_set.reserved_tags = 1; /* fabric connect */
 577         ctrl->tag_set.numa_node = NUMA_NO_NODE;
 578         ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
 579         ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
 580                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
 581         ctrl->tag_set.driver_data = ctrl;
 582         ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
 583         ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
 584         ctrl->ctrl.tagset = &ctrl->tag_set;
 585
 586         ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
 587         if (ret)
 588                 goto out_destroy_queues;
 589
 590         ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 591         if (IS_ERR(ctrl->ctrl.connect_q)) {
 592                 ret = PTR_ERR(ctrl->ctrl.connect_q);
 593                 goto out_free_tagset;
 594         }
 595
 596         ret = nvme_loop_connect_io_queues(ctrl);
 597         if (ret)
 598                 goto out_cleanup_connect_q;
 599
 600         return 0;
 601
 602 out_cleanup_connect_q:
 603         blk_cleanup_queue(ctrl->ctrl.connect_q);
 604 out_free_tagset:
 605         blk_mq_free_tag_set(&ctrl->tag_set);
 606 out_destroy_queues:
 607         nvme_loop_destroy_io_queues(ctrl);
 608         return ret;
 609 }
 610
 611 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
 612                 struct nvmf_ctrl_options *opts)
 613 {
 614         struct nvme_loop_ctrl *ctrl;
 615         bool changed;
 616         int ret;
 617
 618         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
 619         if (!ctrl)
 620                 return ERR_PTR(-ENOMEM);
 621         ctrl->ctrl.opts = opts;
 622         INIT_LIST_HEAD(&ctrl->list);
 623
 624         INIT_WORK(&ctrl->delete_work, nvme_loop_del_ctrl_work);
 625         INIT_WORK(&ctrl->ctrl.reset_work, nvme_loop_reset_ctrl_work);
 626
 627         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
 628                                 0 /* no quirks, we're perfect! */);
 629         if (ret)
 630                 goto out_put_ctrl;
 631
 632         ret = -ENOMEM;
 633
 634         ctrl->ctrl.sqsize = opts->queue_size - 1;
 635         ctrl->ctrl.kato = opts->kato;
 636
 637         ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
 638                         GFP_KERNEL);
 639         if (!ctrl->queues)
 640                 goto out_uninit_ctrl;
 641
 642         ret = nvme_loop_configure_admin_queue(ctrl);
 643         if (ret)
 644                 goto out_free_queues;
 645
 646         if (opts->queue_size > ctrl->ctrl.maxcmd) {
 647                 /* warn if maxcmd is lower than queue_size */
 648                 dev_warn(ctrl->ctrl.device,
 649                         "queue_size %zu > ctrl maxcmd %u, clamping down\n",
 650                         opts->queue_size, ctrl->ctrl.maxcmd);
 651                 opts->queue_size = ctrl->ctrl.maxcmd;
 652         }
 653
 654         if (opts->nr_io_queues) {
 655                 ret = nvme_loop_create_io_queues(ctrl);
 656                 if (ret)
 657                         goto out_remove_admin_queue;
 658         }
 659
 660         nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
 661
 662         dev_info(ctrl->ctrl.device,
 663                  "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
 664
 665         kref_get(&ctrl->ctrl.kref);
 666
 667         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 668         WARN_ON_ONCE(!changed);
 669
 670         mutex_lock(&nvme_loop_ctrl_mutex);
 671         list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
 672         mutex_unlock(&nvme_loop_ctrl_mutex);
 673
 674         nvme_start_ctrl(&ctrl->ctrl);
 675
 676         return &ctrl->ctrl;
 677
 678 out_remove_admin_queue:
 679         nvme_loop_destroy_admin_queue(ctrl);
 680 out_free_queues:
 681         kfree(ctrl->queues);
 682 out_uninit_ctrl:
 683         nvme_uninit_ctrl(&ctrl->ctrl);
 684 out_put_ctrl:
 685         nvme_put_ctrl(&ctrl->ctrl);
 686         if (ret > 0)
 687                 ret = -EIO;
 688         return ERR_PTR(ret);
 689 }
 690
 691 static int nvme_loop_add_port(struct nvmet_port *port)
 692 {
 693         /*
 694          * XXX: disalow adding more than one port so
 695          * there is no connection rejections when a
 696          * a subsystem is assigned to a port for which
 697          * loop doesn't have a pointer.
 698          * This scenario would be possible if we allowed
 699          * more than one port to be added and a subsystem
 700          * was assigned to a port other than nvmet_loop_port.
 701          */
 702
 703         if (nvmet_loop_port)
 704                 return -EPERM;
 705
 706         nvmet_loop_port = port;
 707         return 0;
 708 }
 709
 710 static void nvme_loop_remove_port(struct nvmet_port *port)
 711 {
 712         if (port == nvmet_loop_port)
 713                 nvmet_loop_port = NULL;
 714 }
 715
 716 static struct nvmet_fabrics_ops nvme_loop_ops = {
 717         .owner          = THIS_MODULE,
 718         .type           = NVMF_TRTYPE_LOOP,
 719         .add_port       = nvme_loop_add_port,
 720         .remove_port    = nvme_loop_remove_port,
 721         .queue_response = nvme_loop_queue_response,
 722         .delete_ctrl    = nvme_loop_delete_ctrl,
 723 };
 724
 725 static struct nvmf_transport_ops nvme_loop_transport = {
 726         .name           = "loop",
 727         .create_ctrl    = nvme_loop_create_ctrl,
 728 };
 729
 730 static int __init nvme_loop_init_module(void)
 731 {
 732         int ret;
 733
 734         ret = nvmet_register_transport(&nvme_loop_ops);
 735         if (ret)
 736                 return ret;
 737
 738         ret = nvmf_register_transport(&nvme_loop_transport);
 739         if (ret)
 740                 nvmet_unregister_transport(&nvme_loop_ops);
 741
 742         return ret;
 743 }
 744
 745 static void __exit nvme_loop_cleanup_module(void)
 746 {
 747         struct nvme_loop_ctrl *ctrl, *next;
 748
 749         nvmf_unregister_transport(&nvme_loop_transport);
 750         nvmet_unregister_transport(&nvme_loop_ops);
 751
 752         mutex_lock(&nvme_loop_ctrl_mutex);
 753         list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
 754                 __nvme_loop_del_ctrl(ctrl);
 755         mutex_unlock(&nvme_loop_ctrl_mutex);
 756
 757         flush_workqueue(nvme_wq);
 758 }
 759
 760 module_init(nvme_loop_init_module);
 761 module_exit(nvme_loop_cleanup_module);
 762
 763 MODULE_LICENSE("GPL v2");
 764 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */