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
  95         nvme_cleanup_cmd(req);
  96         sg_free_table_chained(&iod->sg_table, true);
  97         nvme_complete_rq(req);
  98 }
  99
 100 static struct blk_mq_tags *nvme_loop_tagset(struct nvme_loop_queue *queue)
 101 {
 102         u32 queue_idx = nvme_loop_queue_idx(queue);
 103
 104         if (queue_idx == 0)
 105                 return queue->ctrl->admin_tag_set.tags[queue_idx];
 106         return queue->ctrl->tag_set.tags[queue_idx - 1];
 107 }
 108
 109 static void nvme_loop_queue_response(struct nvmet_req *req)
 110 {
 111         struct nvme_loop_queue *queue =
 112                 container_of(req->sq, struct nvme_loop_queue, nvme_sq);
 113         struct nvme_completion *cqe = req->rsp;
 114
 115         /*
 116          * AEN requests are special as they don't time out and can
 117          * survive any kind of queue freeze and often don't respond to
 118          * aborts.  We don't even bother to allocate a struct request
 119          * for them but rather special case them here.
 120          */
 121         if (unlikely(nvme_loop_queue_idx(queue) == 0 &&
 122                         cqe->command_id >= NVME_LOOP_AQ_BLKMQ_DEPTH)) {
 123                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
 124                                 &cqe->result);
 125         } else {
 126                 struct request *rq;
 127
 128                 rq = blk_mq_tag_to_rq(nvme_loop_tagset(queue), cqe->command_id);
 129                 if (!rq) {
 130                         dev_err(queue->ctrl->ctrl.device,
 131                                 "tag 0x%x on queue %d not found\n",
 132                                 cqe->command_id, nvme_loop_queue_idx(queue));
 133                         return;
 134                 }
 135
 136                 nvme_end_request(rq, cqe->status, cqe->result);
 137         }
 138 }
 139
 140 static void nvme_loop_execute_work(struct work_struct *work)
 141 {
 142         struct nvme_loop_iod *iod =
 143                 container_of(work, struct nvme_loop_iod, work);
 144
 145         iod->req.execute(&iod->req);
 146 }
 147
 148 static enum blk_eh_timer_return
 149 nvme_loop_timeout(struct request *rq, bool reserved)
 150 {
 151         struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(rq);
 152
 153         /* queue error recovery */
 154         schedule_work(&iod->queue->ctrl->reset_work);
 155
 156         /* fail with DNR on admin cmd timeout */
 157         nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
 158
 159         return BLK_EH_HANDLED;
 160 }
 161
 162 static int nvme_loop_queue_rq(struct blk_mq_hw_ctx *hctx,
 163                 const struct blk_mq_queue_data *bd)
 164 {
 165         struct nvme_ns *ns = hctx->queue->queuedata;
 166         struct nvme_loop_queue *queue = hctx->driver_data;
 167         struct request *req = bd->rq;
 168         struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
 169         int ret;
 170
 171         ret = nvme_setup_cmd(ns, req, &iod->cmd);
 172         if (ret != BLK_MQ_RQ_QUEUE_OK)
 173                 return ret;
 174
 175         iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
 176         iod->req.port = nvmet_loop_port;
 177         if (!nvmet_req_init(&iod->req, &queue->nvme_cq,
 178                         &queue->nvme_sq, &nvme_loop_ops)) {
 179                 nvme_cleanup_cmd(req);
 180                 blk_mq_start_request(req);
 181                 nvme_loop_queue_response(&iod->req);
 182                 return BLK_MQ_RQ_QUEUE_OK;
 183         }
 184
 185         if (blk_rq_bytes(req)) {
 186                 iod->sg_table.sgl = iod->first_sgl;
 187                 ret = sg_alloc_table_chained(&iod->sg_table,
 188                                 blk_rq_nr_phys_segments(req),
 189                                 iod->sg_table.sgl);
 190                 if (ret)
 191                         return BLK_MQ_RQ_QUEUE_BUSY;
 192
 193                 iod->req.sg = iod->sg_table.sgl;
 194                 iod->req.sg_cnt = blk_rq_map_sg(req->q, req, iod->sg_table.sgl);
 195         }
 196
 197         blk_mq_start_request(req);
 198
 199         schedule_work(&iod->work);
 200         return BLK_MQ_RQ_QUEUE_OK;
 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 const 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 const 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_connect_io_queues(struct nvme_loop_ctrl *ctrl)
 353 {
 354         int i, ret;
 355
 356         for (i = 1; i < ctrl->queue_count; i++) {
 357                 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
 358                 if (ret)
 359                         return ret;
 360         }
 361
 362         return 0;
 363 }
 364
 365 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
 366 {
 367         int error;
 368
 369         memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
 370         ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
 371         ctrl->admin_tag_set.queue_depth = NVME_LOOP_AQ_BLKMQ_DEPTH;
 372         ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
 373         ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
 374         ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
 375                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
 376         ctrl->admin_tag_set.driver_data = ctrl;
 377         ctrl->admin_tag_set.nr_hw_queues = 1;
 378         ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
 379
 380         ctrl->queues[0].ctrl = ctrl;
 381         error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
 382         if (error)
 383                 return error;
 384         ctrl->queue_count = 1;
 385
 386         error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
 387         if (error)
 388                 goto out_free_sq;
 389
 390         ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 391         if (IS_ERR(ctrl->ctrl.admin_q)) {
 392                 error = PTR_ERR(ctrl->ctrl.admin_q);
 393                 goto out_free_tagset;
 394         }
 395
 396         error = nvmf_connect_admin_queue(&ctrl->ctrl);
 397         if (error)
 398                 goto out_cleanup_queue;
 399
 400         error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->cap);
 401         if (error) {
 402                 dev_err(ctrl->ctrl.device,
 403                         "prop_get NVME_REG_CAP failed\n");
 404                 goto out_cleanup_queue;
 405         }
 406
 407         ctrl->ctrl.sqsize =
 408                 min_t(int, NVME_CAP_MQES(ctrl->cap), ctrl->ctrl.sqsize);
 409
 410         error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
 411         if (error)
 412                 goto out_cleanup_queue;
 413
 414         ctrl->ctrl.max_hw_sectors =
 415                 (NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
 416
 417         error = nvme_init_identify(&ctrl->ctrl);
 418         if (error)
 419                 goto out_cleanup_queue;
 420
 421         nvme_start_keep_alive(&ctrl->ctrl);
 422
 423         return 0;
 424
 425 out_cleanup_queue:
 426         blk_cleanup_queue(ctrl->ctrl.admin_q);
 427 out_free_tagset:
 428         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 429 out_free_sq:
 430         nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
 431         return error;
 432 }
 433
 434 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
 435 {
 436         nvme_stop_keep_alive(&ctrl->ctrl);
 437
 438         if (ctrl->queue_count > 1) {
 439                 nvme_stop_queues(&ctrl->ctrl);
 440                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
 441                                         nvme_cancel_request, &ctrl->ctrl);
 442                 nvme_loop_destroy_io_queues(ctrl);
 443         }
 444
 445         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
 446                 nvme_shutdown_ctrl(&ctrl->ctrl);
 447
 448         blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
 449         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
 450                                 nvme_cancel_request, &ctrl->ctrl);
 451         nvme_loop_destroy_admin_queue(ctrl);
 452 }
 453
 454 static void nvme_loop_del_ctrl_work(struct work_struct *work)
 455 {
 456         struct nvme_loop_ctrl *ctrl = container_of(work,
 457                                 struct nvme_loop_ctrl, delete_work);
 458
 459         nvme_uninit_ctrl(&ctrl->ctrl);
 460         nvme_loop_shutdown_ctrl(ctrl);
 461         nvme_put_ctrl(&ctrl->ctrl);
 462 }
 463
 464 static int __nvme_loop_del_ctrl(struct nvme_loop_ctrl *ctrl)
 465 {
 466         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
 467                 return -EBUSY;
 468
 469         if (!schedule_work(&ctrl->delete_work))
 470                 return -EBUSY;
 471
 472         return 0;
 473 }
 474
 475 static int nvme_loop_del_ctrl(struct nvme_ctrl *nctrl)
 476 {
 477         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
 478         int ret;
 479
 480         ret = __nvme_loop_del_ctrl(ctrl);
 481         if (ret)
 482                 return ret;
 483
 484         flush_work(&ctrl->delete_work);
 485
 486         return 0;
 487 }
 488
 489 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
 490 {
 491         struct nvme_loop_ctrl *ctrl;
 492
 493         mutex_lock(&nvme_loop_ctrl_mutex);
 494         list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
 495                 if (ctrl->ctrl.cntlid == nctrl->cntlid)
 496                         __nvme_loop_del_ctrl(ctrl);
 497         }
 498         mutex_unlock(&nvme_loop_ctrl_mutex);
 499 }
 500
 501 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
 502 {
 503         struct nvme_loop_ctrl *ctrl = container_of(work,
 504                                         struct nvme_loop_ctrl, reset_work);
 505         bool changed;
 506         int ret;
 507
 508         nvme_loop_shutdown_ctrl(ctrl);
 509
 510         ret = nvme_loop_configure_admin_queue(ctrl);
 511         if (ret)
 512                 goto out_disable;
 513
 514         ret = nvme_loop_init_io_queues(ctrl);
 515         if (ret)
 516                 goto out_destroy_admin;
 517
 518         ret = nvme_loop_connect_io_queues(ctrl);
 519         if (ret)
 520                 goto out_destroy_io;
 521
 522         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 523         WARN_ON_ONCE(!changed);
 524
 525         nvme_queue_scan(&ctrl->ctrl);
 526         nvme_queue_async_events(&ctrl->ctrl);
 527
 528         nvme_start_queues(&ctrl->ctrl);
 529
 530         return;
 531
 532 out_destroy_io:
 533         nvme_loop_destroy_io_queues(ctrl);
 534 out_destroy_admin:
 535         nvme_loop_destroy_admin_queue(ctrl);
 536 out_disable:
 537         dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
 538         nvme_uninit_ctrl(&ctrl->ctrl);
 539         nvme_put_ctrl(&ctrl->ctrl);
 540 }
 541
 542 static int nvme_loop_reset_ctrl(struct nvme_ctrl *nctrl)
 543 {
 544         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
 545
 546         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
 547                 return -EBUSY;
 548
 549         if (!schedule_work(&ctrl->reset_work))
 550                 return -EBUSY;
 551
 552         flush_work(&ctrl->reset_work);
 553
 554         return 0;
 555 }
 556
 557 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
 558         .name                   = "loop",
 559         .module                 = THIS_MODULE,
 560         .is_fabrics             = true,
 561         .reg_read32             = nvmf_reg_read32,
 562         .reg_read64             = nvmf_reg_read64,
 563         .reg_write32            = nvmf_reg_write32,
 564         .reset_ctrl             = nvme_loop_reset_ctrl,
 565         .free_ctrl              = nvme_loop_free_ctrl,
 566         .submit_async_event     = nvme_loop_submit_async_event,
 567         .delete_ctrl            = nvme_loop_del_ctrl,
 568         .get_subsysnqn          = nvmf_get_subsysnqn,
 569 };
 570
 571 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
 572 {
 573         int ret;
 574
 575         ret = nvme_loop_init_io_queues(ctrl);
 576         if (ret)
 577                 return ret;
 578
 579         memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
 580         ctrl->tag_set.ops = &nvme_loop_mq_ops;
 581         ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
 582         ctrl->tag_set.reserved_tags = 1; /* fabric connect */
 583         ctrl->tag_set.numa_node = NUMA_NO_NODE;
 584         ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
 585         ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
 586                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
 587         ctrl->tag_set.driver_data = ctrl;
 588         ctrl->tag_set.nr_hw_queues = ctrl->queue_count - 1;
 589         ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
 590         ctrl->ctrl.tagset = &ctrl->tag_set;
 591
 592         ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
 593         if (ret)
 594                 goto out_destroy_queues;
 595
 596         ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 597         if (IS_ERR(ctrl->ctrl.connect_q)) {
 598                 ret = PTR_ERR(ctrl->ctrl.connect_q);
 599                 goto out_free_tagset;
 600         }
 601
 602         ret = nvme_loop_connect_io_queues(ctrl);
 603         if (ret)
 604                 goto out_cleanup_connect_q;
 605
 606         return 0;
 607
 608 out_cleanup_connect_q:
 609         blk_cleanup_queue(ctrl->ctrl.connect_q);
 610 out_free_tagset:
 611         blk_mq_free_tag_set(&ctrl->tag_set);
 612 out_destroy_queues:
 613         nvme_loop_destroy_io_queues(ctrl);
 614         return ret;
 615 }
 616
 617 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
 618                 struct nvmf_ctrl_options *opts)
 619 {
 620         struct nvme_loop_ctrl *ctrl;
 621         bool changed;
 622         int ret;
 623
 624         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
 625         if (!ctrl)
 626                 return ERR_PTR(-ENOMEM);
 627         ctrl->ctrl.opts = opts;
 628         INIT_LIST_HEAD(&ctrl->list);
 629
 630         INIT_WORK(&ctrl->delete_work, nvme_loop_del_ctrl_work);
 631         INIT_WORK(&ctrl->reset_work, nvme_loop_reset_ctrl_work);
 632
 633         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
 634                                 0 /* no quirks, we're perfect! */);
 635         if (ret)
 636                 goto out_put_ctrl;
 637
 638         spin_lock_init(&ctrl->lock);
 639
 640         ret = -ENOMEM;
 641
 642         ctrl->ctrl.sqsize = opts->queue_size - 1;
 643         ctrl->ctrl.kato = opts->kato;
 644
 645         ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
 646                         GFP_KERNEL);
 647         if (!ctrl->queues)
 648                 goto out_uninit_ctrl;
 649
 650         ret = nvme_loop_configure_admin_queue(ctrl);
 651         if (ret)
 652                 goto out_free_queues;
 653
 654         if (opts->queue_size > ctrl->ctrl.maxcmd) {
 655                 /* warn if maxcmd is lower than queue_size */
 656                 dev_warn(ctrl->ctrl.device,
 657                         "queue_size %zu > ctrl maxcmd %u, clamping down\n",
 658                         opts->queue_size, ctrl->ctrl.maxcmd);
 659                 opts->queue_size = ctrl->ctrl.maxcmd;
 660         }
 661
 662         if (opts->nr_io_queues) {
 663                 ret = nvme_loop_create_io_queues(ctrl);
 664                 if (ret)
 665                         goto out_remove_admin_queue;
 666         }
 667
 668         nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
 669
 670         dev_info(ctrl->ctrl.device,
 671                  "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
 672
 673         kref_get(&ctrl->ctrl.kref);
 674
 675         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 676         WARN_ON_ONCE(!changed);
 677
 678         mutex_lock(&nvme_loop_ctrl_mutex);
 679         list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
 680         mutex_unlock(&nvme_loop_ctrl_mutex);
 681
 682         if (opts->nr_io_queues) {
 683                 nvme_queue_scan(&ctrl->ctrl);
 684                 nvme_queue_async_events(&ctrl->ctrl);
 685         }
 686
 687         return &ctrl->ctrl;
 688
 689 out_remove_admin_queue:
 690         nvme_loop_destroy_admin_queue(ctrl);
 691 out_free_queues:
 692         kfree(ctrl->queues);
 693 out_uninit_ctrl:
 694         nvme_uninit_ctrl(&ctrl->ctrl);
 695 out_put_ctrl:
 696         nvme_put_ctrl(&ctrl->ctrl);
 697         if (ret > 0)
 698                 ret = -EIO;
 699         return ERR_PTR(ret);
 700 }
 701
 702 static int nvme_loop_add_port(struct nvmet_port *port)
 703 {
 704         /*
 705          * XXX: disalow adding more than one port so
 706          * there is no connection rejections when a
 707          * a subsystem is assigned to a port for which
 708          * loop doesn't have a pointer.
 709          * This scenario would be possible if we allowed
 710          * more than one port to be added and a subsystem
 711          * was assigned to a port other than nvmet_loop_port.
 712          */
 713
 714         if (nvmet_loop_port)
 715                 return -EPERM;
 716
 717         nvmet_loop_port = port;
 718         return 0;
 719 }
 720
 721 static void nvme_loop_remove_port(struct nvmet_port *port)
 722 {
 723         if (port == nvmet_loop_port)
 724                 nvmet_loop_port = NULL;
 725 }
 726
 727 static struct nvmet_fabrics_ops nvme_loop_ops = {
 728         .owner          = THIS_MODULE,
 729         .type           = NVMF_TRTYPE_LOOP,
 730         .add_port       = nvme_loop_add_port,
 731         .remove_port    = nvme_loop_remove_port,
 732         .queue_response = nvme_loop_queue_response,
 733         .delete_ctrl    = nvme_loop_delete_ctrl,
 734 };
 735
 736 static struct nvmf_transport_ops nvme_loop_transport = {
 737         .name           = "loop",
 738         .create_ctrl    = nvme_loop_create_ctrl,
 739 };
 740
 741 static int __init nvme_loop_init_module(void)
 742 {
 743         int ret;
 744
 745         ret = nvmet_register_transport(&nvme_loop_ops);
 746         if (ret)
 747                 return ret;
 748
 749         ret = nvmf_register_transport(&nvme_loop_transport);
 750         if (ret)
 751                 nvmet_unregister_transport(&nvme_loop_ops);
 752
 753         return ret;
 754 }
 755
 756 static void __exit nvme_loop_cleanup_module(void)
 757 {
 758         struct nvme_loop_ctrl *ctrl, *next;
 759
 760         nvmf_unregister_transport(&nvme_loop_transport);
 761         nvmet_unregister_transport(&nvme_loop_ops);
 762
 763         mutex_lock(&nvme_loop_ctrl_mutex);
 764         list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
 765                 __nvme_loop_del_ctrl(ctrl);
 766         mutex_unlock(&nvme_loop_ctrl_mutex);
 767
 768         flush_scheduled_work();
 769 }
 770
 771 module_init(nvme_loop_init_module);
 772 module_exit(nvme_loop_cleanup_module);
 773
 774 MODULE_LICENSE("GPL v2");
 775 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */