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