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(struct blk_mq_tag_set *set,
 234                 struct request *req, unsigned int hctx_idx,
 235                 unsigned int numa_node)
 236 {
 237         return nvme_loop_init_iod(set->driver_data, blk_mq_rq_to_pdu(req),
 238                         hctx_idx + 1);
 239 }
 240
 241 static int nvme_loop_init_admin_request(struct blk_mq_tag_set *set,
 242                 struct request *req, unsigned int hctx_idx,
 243                 unsigned int numa_node)
 244 {
 245         return nvme_loop_init_iod(set->driver_data, blk_mq_rq_to_pdu(req), 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->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_admin_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         nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
 291         blk_cleanup_queue(ctrl->ctrl.admin_q);
 292         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 293 }
 294
 295 static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl)
 296 {
 297         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
 298
 299         if (list_empty(&ctrl->list))
 300                 goto free_ctrl;
 301
 302         mutex_lock(&nvme_loop_ctrl_mutex);
 303         list_del(&ctrl->list);
 304         mutex_unlock(&nvme_loop_ctrl_mutex);
 305
 306         if (nctrl->tagset) {
 307                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 308                 blk_mq_free_tag_set(&ctrl->tag_set);
 309         }
 310         kfree(ctrl->queues);
 311         nvmf_free_options(nctrl->opts);
 312 free_ctrl:
 313         kfree(ctrl);
 314 }
 315
 316 static void nvme_loop_destroy_io_queues(struct nvme_loop_ctrl *ctrl)
 317 {
 318         int i;
 319
 320         for (i = 1; i < ctrl->queue_count; i++)
 321                 nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
 322 }
 323
 324 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl)
 325 {
 326         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 327         unsigned int nr_io_queues;
 328         int ret, i;
 329
 330         nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
 331         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 332         if (ret || !nr_io_queues)
 333                 return ret;
 334
 335         dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues);
 336
 337         for (i = 1; i <= nr_io_queues; i++) {
 338                 ctrl->queues[i].ctrl = ctrl;
 339                 ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
 340                 if (ret)
 341                         goto out_destroy_queues;
 342
 343                 ctrl->queue_count++;
 344         }
 345
 346         return 0;
 347
 348 out_destroy_queues:
 349         nvme_loop_destroy_io_queues(ctrl);
 350         return ret;
 351 }
 352
 353 static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl)
 354 {
 355         int i, ret;
 356
 357         for (i = 1; i < ctrl->queue_count; i++) {
 358                 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
 359                 if (ret)
 360                         return ret;
 361         }
 362
 363         return 0;
 364 }
 365
 366 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
 367 {
 368         int error;
 369
 370         memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
 371         ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
 372         ctrl->admin_tag_set.queue_depth = NVME_LOOP_AQ_BLKMQ_DEPTH;
 373         ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
 374         ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
 375         ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
 376                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
 377         ctrl->admin_tag_set.driver_data = ctrl;
 378         ctrl->admin_tag_set.nr_hw_queues = 1;
 379         ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
 380
 381         ctrl->queues[0].ctrl = ctrl;
 382         error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
 383         if (error)
 384                 return error;
 385         ctrl->queue_count = 1;
 386
 387         error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
 388         if (error)
 389                 goto out_free_sq;
 390
 391         ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 392         if (IS_ERR(ctrl->ctrl.admin_q)) {
 393                 error = PTR_ERR(ctrl->ctrl.admin_q);
 394                 goto out_free_tagset;
 395         }
 396
 397         error = nvmf_connect_admin_queue(&ctrl->ctrl);
 398         if (error)
 399                 goto out_cleanup_queue;
 400
 401         error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->cap);
 402         if (error) {
 403                 dev_err(ctrl->ctrl.device,
 404                         "prop_get NVME_REG_CAP failed\n");
 405                 goto out_cleanup_queue;
 406         }
 407
 408         ctrl->ctrl.sqsize =
 409                 min_t(int, NVME_CAP_MQES(ctrl->cap), ctrl->ctrl.sqsize);
 410
 411         error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
 412         if (error)
 413                 goto out_cleanup_queue;
 414
 415         ctrl->ctrl.max_hw_sectors =
 416                 (NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
 417
 418         error = nvme_init_identify(&ctrl->ctrl);
 419         if (error)
 420                 goto out_cleanup_queue;
 421
 422         nvme_start_keep_alive(&ctrl->ctrl);
 423
 424         return 0;
 425
 426 out_cleanup_queue:
 427         blk_cleanup_queue(ctrl->ctrl.admin_q);
 428 out_free_tagset:
 429         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 430 out_free_sq:
 431         nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
 432         return error;
 433 }
 434
 435 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
 436 {
 437         nvme_stop_keep_alive(&ctrl->ctrl);
 438
 439         if (ctrl->queue_count > 1) {
 440                 nvme_stop_queues(&ctrl->ctrl);
 441                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
 442                                         nvme_cancel_request, &ctrl->ctrl);
 443                 nvme_loop_destroy_io_queues(ctrl);
 444         }
 445
 446         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
 447                 nvme_shutdown_ctrl(&ctrl->ctrl);
 448
 449         blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
 450         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
 451                                 nvme_cancel_request, &ctrl->ctrl);
 452         nvme_loop_destroy_admin_queue(ctrl);
 453 }
 454
 455 static void nvme_loop_del_ctrl_work(struct work_struct *work)
 456 {
 457         struct nvme_loop_ctrl *ctrl = container_of(work,
 458                                 struct nvme_loop_ctrl, delete_work);
 459
 460         nvme_uninit_ctrl(&ctrl->ctrl);
 461         nvme_loop_shutdown_ctrl(ctrl);
 462         nvme_put_ctrl(&ctrl->ctrl);
 463 }
 464
 465 static int __nvme_loop_del_ctrl(struct nvme_loop_ctrl *ctrl)
 466 {
 467         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
 468                 return -EBUSY;
 469
 470         if (!schedule_work(&ctrl->delete_work))
 471                 return -EBUSY;
 472
 473         return 0;
 474 }
 475
 476 static int nvme_loop_del_ctrl(struct nvme_ctrl *nctrl)
 477 {
 478         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
 479         int ret;
 480
 481         ret = __nvme_loop_del_ctrl(ctrl);
 482         if (ret)
 483                 return ret;
 484
 485         flush_work(&ctrl->delete_work);
 486
 487         return 0;
 488 }
 489
 490 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
 491 {
 492         struct nvme_loop_ctrl *ctrl;
 493
 494         mutex_lock(&nvme_loop_ctrl_mutex);
 495         list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
 496                 if (ctrl->ctrl.cntlid == nctrl->cntlid)
 497                         __nvme_loop_del_ctrl(ctrl);
 498         }
 499         mutex_unlock(&nvme_loop_ctrl_mutex);
 500 }
 501
 502 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
 503 {
 504         struct nvme_loop_ctrl *ctrl = container_of(work,
 505                                         struct nvme_loop_ctrl, reset_work);
 506         bool changed;
 507         int ret;
 508
 509         nvme_loop_shutdown_ctrl(ctrl);
 510
 511         ret = nvme_loop_configure_admin_queue(ctrl);
 512         if (ret)
 513                 goto out_disable;
 514
 515         ret = nvme_loop_init_io_queues(ctrl);
 516         if (ret)
 517                 goto out_destroy_admin;
 518
 519         ret = nvme_loop_connect_io_queues(ctrl);
 520         if (ret)
 521                 goto out_destroy_io;
 522
 523         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 524         WARN_ON_ONCE(!changed);
 525
 526         nvme_queue_scan(&ctrl->ctrl);
 527         nvme_queue_async_events(&ctrl->ctrl);
 528
 529         nvme_start_queues(&ctrl->ctrl);
 530
 531         return;
 532
 533 out_destroy_io:
 534         nvme_loop_destroy_io_queues(ctrl);
 535 out_destroy_admin:
 536         nvme_loop_destroy_admin_queue(ctrl);
 537 out_disable:
 538         dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
 539         nvme_uninit_ctrl(&ctrl->ctrl);
 540         nvme_put_ctrl(&ctrl->ctrl);
 541 }
 542
 543 static int nvme_loop_reset_ctrl(struct nvme_ctrl *nctrl)
 544 {
 545         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
 546
 547         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
 548                 return -EBUSY;
 549
 550         if (!schedule_work(&ctrl->reset_work))
 551                 return -EBUSY;
 552
 553         flush_work(&ctrl->reset_work);
 554
 555         return 0;
 556 }
 557
 558 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
 559         .name                   = "loop",
 560         .module                 = THIS_MODULE,
 561         .flags                  = NVME_F_FABRICS,
 562         .reg_read32             = nvmf_reg_read32,
 563         .reg_read64             = nvmf_reg_read64,
 564         .reg_write32            = nvmf_reg_write32,
 565         .reset_ctrl             = nvme_loop_reset_ctrl,
 566         .free_ctrl              = nvme_loop_free_ctrl,
 567         .submit_async_event     = nvme_loop_submit_async_event,
 568         .delete_ctrl            = nvme_loop_del_ctrl,
 569         .get_subsysnqn          = nvmf_get_subsysnqn,
 570 };
 571
 572 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
 573 {
 574         int ret;
 575
 576         ret = nvme_loop_init_io_queues(ctrl);
 577         if (ret)
 578                 return ret;
 579
 580         memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
 581         ctrl->tag_set.ops = &nvme_loop_mq_ops;
 582         ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
 583         ctrl->tag_set.reserved_tags = 1; /* fabric connect */
 584         ctrl->tag_set.numa_node = NUMA_NO_NODE;
 585         ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
 586         ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
 587                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
 588         ctrl->tag_set.driver_data = ctrl;
 589         ctrl->tag_set.nr_hw_queues = ctrl->queue_count - 1;
 590         ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
 591         ctrl->ctrl.tagset = &ctrl->tag_set;
 592
 593         ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
 594         if (ret)
 595                 goto out_destroy_queues;
 596
 597         ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 598         if (IS_ERR(ctrl->ctrl.connect_q)) {
 599                 ret = PTR_ERR(ctrl->ctrl.connect_q);
 600                 goto out_free_tagset;
 601         }
 602
 603         ret = nvme_loop_connect_io_queues(ctrl);
 604         if (ret)
 605                 goto out_cleanup_connect_q;
 606
 607         return 0;
 608
 609 out_cleanup_connect_q:
 610         blk_cleanup_queue(ctrl->ctrl.connect_q);
 611 out_free_tagset:
 612         blk_mq_free_tag_set(&ctrl->tag_set);
 613 out_destroy_queues:
 614         nvme_loop_destroy_io_queues(ctrl);
 615         return ret;
 616 }
 617
 618 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
 619                 struct nvmf_ctrl_options *opts)
 620 {
 621         struct nvme_loop_ctrl *ctrl;
 622         bool changed;
 623         int ret;
 624
 625         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
 626         if (!ctrl)
 627                 return ERR_PTR(-ENOMEM);
 628         ctrl->ctrl.opts = opts;
 629         INIT_LIST_HEAD(&ctrl->list);
 630
 631         INIT_WORK(&ctrl->delete_work, nvme_loop_del_ctrl_work);
 632         INIT_WORK(&ctrl->reset_work, nvme_loop_reset_ctrl_work);
 633
 634         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
 635                                 0 /* no quirks, we're perfect! */);
 636         if (ret)
 637                 goto out_put_ctrl;
 638
 639         spin_lock_init(&ctrl->lock);
 640
 641         ret = -ENOMEM;
 642
 643         ctrl->ctrl.sqsize = opts->queue_size - 1;
 644         ctrl->ctrl.kato = opts->kato;
 645
 646         ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
 647                         GFP_KERNEL);
 648         if (!ctrl->queues)
 649                 goto out_uninit_ctrl;
 650
 651         ret = nvme_loop_configure_admin_queue(ctrl);
 652         if (ret)
 653                 goto out_free_queues;
 654
 655         if (opts->queue_size > ctrl->ctrl.maxcmd) {
 656                 /* warn if maxcmd is lower than queue_size */
 657                 dev_warn(ctrl->ctrl.device,
 658                         "queue_size %zu > ctrl maxcmd %u, clamping down\n",
 659                         opts->queue_size, ctrl->ctrl.maxcmd);
 660                 opts->queue_size = ctrl->ctrl.maxcmd;
 661         }
 662
 663         if (opts->nr_io_queues) {
 664                 ret = nvme_loop_create_io_queues(ctrl);
 665                 if (ret)
 666                         goto out_remove_admin_queue;
 667         }
 668
 669         nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
 670
 671         dev_info(ctrl->ctrl.device,
 672                  "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
 673
 674         kref_get(&ctrl->ctrl.kref);
 675
 676         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 677         WARN_ON_ONCE(!changed);
 678
 679         mutex_lock(&nvme_loop_ctrl_mutex);
 680         list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
 681         mutex_unlock(&nvme_loop_ctrl_mutex);
 682
 683         if (opts->nr_io_queues) {
 684                 nvme_queue_scan(&ctrl->ctrl);
 685                 nvme_queue_async_events(&ctrl->ctrl);
 686         }
 687
 688         return &ctrl->ctrl;
 689
 690 out_remove_admin_queue:
 691         nvme_loop_destroy_admin_queue(ctrl);
 692 out_free_queues:
 693         kfree(ctrl->queues);
 694 out_uninit_ctrl:
 695         nvme_uninit_ctrl(&ctrl->ctrl);
 696 out_put_ctrl:
 697         nvme_put_ctrl(&ctrl->ctrl);
 698         if (ret > 0)
 699                 ret = -EIO;
 700         return ERR_PTR(ret);
 701 }
 702
 703 static int nvme_loop_add_port(struct nvmet_port *port)
 704 {
 705         /*
 706          * XXX: disalow adding more than one port so
 707          * there is no connection rejections when a
 708          * a subsystem is assigned to a port for which
 709          * loop doesn't have a pointer.
 710          * This scenario would be possible if we allowed
 711          * more than one port to be added and a subsystem
 712          * was assigned to a port other than nvmet_loop_port.
 713          */
 714
 715         if (nvmet_loop_port)
 716                 return -EPERM;
 717
 718         nvmet_loop_port = port;
 719         return 0;
 720 }
 721
 722 static void nvme_loop_remove_port(struct nvmet_port *port)
 723 {
 724         if (port == nvmet_loop_port)
 725                 nvmet_loop_port = NULL;
 726 }
 727
 728 static struct nvmet_fabrics_ops nvme_loop_ops = {
 729         .owner          = THIS_MODULE,
 730         .type           = NVMF_TRTYPE_LOOP,
 731         .add_port       = nvme_loop_add_port,
 732         .remove_port    = nvme_loop_remove_port,
 733         .queue_response = nvme_loop_queue_response,
 734         .delete_ctrl    = nvme_loop_delete_ctrl,
 735 };
 736
 737 static struct nvmf_transport_ops nvme_loop_transport = {
 738         .name           = "loop",
 739         .create_ctrl    = nvme_loop_create_ctrl,
 740 };
 741
 742 static int __init nvme_loop_init_module(void)
 743 {
 744         int ret;
 745
 746         ret = nvmet_register_transport(&nvme_loop_ops);
 747         if (ret)
 748                 return ret;
 749
 750         ret = nvmf_register_transport(&nvme_loop_transport);
 751         if (ret)
 752                 nvmet_unregister_transport(&nvme_loop_ops);
 753
 754         return ret;
 755 }
 756
 757 static void __exit nvme_loop_cleanup_module(void)
 758 {
 759         struct nvme_loop_ctrl *ctrl, *next;
 760
 761         nvmf_unregister_transport(&nvme_loop_transport);
 762         nvmet_unregister_transport(&nvme_loop_ops);
 763
 764         mutex_lock(&nvme_loop_ctrl_mutex);
 765         list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
 766                 __nvme_loop_del_ctrl(ctrl);
 767         mutex_unlock(&nvme_loop_ctrl_mutex);
 768
 769         flush_scheduled_work();
 770 }
 771
 772 module_init(nvme_loop_init_module);
 773 module_exit(nvme_loop_cleanup_module);
 774
 775 MODULE_LICENSE("GPL v2");
 776 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */