drivers/nvme/target/loop.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * NVMe over Fabrics loopback device.
   4  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
   5  */
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7 #include <linux/scatterlist.h>
   8 #include <linux/blk-mq.h>
   9 #include <linux/nvme.h>
  10 #include <linux/module.h>
  11 #include <linux/parser.h>
  12 #include "nvmet.h"
  13 #include "../host/nvme.h"
  14 #include "../host/fabrics.h"
  15
  16 #define NVME_LOOP_MAX_SEGMENTS          256
  17
  18 struct nvme_loop_iod {
  19         struct nvme_request     nvme_req;
  20         struct nvme_command     cmd;
  21         struct nvme_completion  cqe;
  22         struct nvmet_req        req;
  23         struct nvme_loop_queue  *queue;
  24         struct work_struct      work;
  25         struct sg_table         sg_table;
  26         struct scatterlist      first_sgl[];
  27 };
  28
  29 struct nvme_loop_ctrl {
  30         struct nvme_loop_queue  *queues;
  31
  32         struct blk_mq_tag_set   admin_tag_set;
  33
  34         struct list_head        list;
  35         struct blk_mq_tag_set   tag_set;
  36         struct nvme_loop_iod    async_event_iod;
  37         struct nvme_ctrl        ctrl;
  38
  39         struct nvmet_port       *port;
  40 };
  41
  42 static inline struct nvme_loop_ctrl *to_loop_ctrl(struct nvme_ctrl *ctrl)
  43 {
  44         return container_of(ctrl, struct nvme_loop_ctrl, ctrl);
  45 }
  46
  47 enum nvme_loop_queue_flags {
  48         NVME_LOOP_Q_LIVE        = 0,
  49 };
  50
  51 struct nvme_loop_queue {
  52         struct nvmet_cq         nvme_cq;
  53         struct nvmet_sq         nvme_sq;
  54         struct nvme_loop_ctrl   *ctrl;
  55         unsigned long           flags;
  56 };
  57
  58 static LIST_HEAD(nvme_loop_ports);
  59 static DEFINE_MUTEX(nvme_loop_ports_mutex);
  60
  61 static LIST_HEAD(nvme_loop_ctrl_list);
  62 static DEFINE_MUTEX(nvme_loop_ctrl_mutex);
  63
  64 static void nvme_loop_queue_response(struct nvmet_req *nvme_req);
  65 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *ctrl);
  66
  67 static const struct nvmet_fabrics_ops nvme_loop_ops;
  68
  69 static inline int nvme_loop_queue_idx(struct nvme_loop_queue *queue)
  70 {
  71         return queue - queue->ctrl->queues;
  72 }
  73
  74 static void nvme_loop_complete_rq(struct request *req)
  75 {
  76         struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
  77
  78         sg_free_table_chained(&iod->sg_table, NVME_INLINE_SG_CNT);
  79         nvme_complete_rq(req);
  80 }
  81
  82 static struct blk_mq_tags *nvme_loop_tagset(struct nvme_loop_queue *queue)
  83 {
  84         u32 queue_idx = nvme_loop_queue_idx(queue);
  85
  86         if (queue_idx == 0)
  87                 return queue->ctrl->admin_tag_set.tags[queue_idx];
  88         return queue->ctrl->tag_set.tags[queue_idx - 1];
  89 }
  90
  91 static void nvme_loop_queue_response(struct nvmet_req *req)
  92 {
  93         struct nvme_loop_queue *queue =
  94                 container_of(req->sq, struct nvme_loop_queue, nvme_sq);
  95         struct nvme_completion *cqe = req->cqe;
  96
  97         /*
  98          * AEN requests are special as they don't time out and can
  99          * survive any kind of queue freeze and often don't respond to
 100          * aborts.  We don't even bother to allocate a struct request
 101          * for them but rather special case them here.
 102          */
 103         if (unlikely(nvme_is_aen_req(nvme_loop_queue_idx(queue),
 104                                      cqe->command_id))) {
 105                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
 106                                 &cqe->result);
 107         } else {
 108                 struct request *rq;
 109
 110                 rq = blk_mq_tag_to_rq(nvme_loop_tagset(queue), cqe->command_id);
 111                 if (!rq) {
 112                         dev_err(queue->ctrl->ctrl.device,
 113                                 "tag 0x%x on queue %d not found\n",
 114                                 cqe->command_id, nvme_loop_queue_idx(queue));
 115                         return;
 116                 }
 117
 118                 if (!nvme_try_complete_req(rq, cqe->status, cqe->result))
 119                         nvme_loop_complete_rq(rq);
 120         }
 121 }
 122
 123 static void nvme_loop_execute_work(struct work_struct *work)
 124 {
 125         struct nvme_loop_iod *iod =
 126                 container_of(work, struct nvme_loop_iod, work);
 127
 128         iod->req.execute(&iod->req);
 129 }
 130
 131 static blk_status_t nvme_loop_queue_rq(struct blk_mq_hw_ctx *hctx,
 132                 const struct blk_mq_queue_data *bd)
 133 {
 134         struct nvme_ns *ns = hctx->queue->queuedata;
 135         struct nvme_loop_queue *queue = hctx->driver_data;
 136         struct request *req = bd->rq;
 137         struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
 138         bool queue_ready = test_bit(NVME_LOOP_Q_LIVE, &queue->flags);
 139         blk_status_t ret;
 140
 141         if (!nvme_check_ready(&queue->ctrl->ctrl, req, queue_ready))
 142                 return nvme_fail_nonready_command(&queue->ctrl->ctrl, req);
 143
 144         ret = nvme_setup_cmd(ns, req);
 145         if (ret)
 146                 return ret;
 147
 148         blk_mq_start_request(req);
 149         iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
 150         iod->req.port = queue->ctrl->port;
 151         if (!nvmet_req_init(&iod->req, &queue->nvme_cq,
 152                         &queue->nvme_sq, &nvme_loop_ops))
 153                 return BLK_STS_OK;
 154
 155         if (blk_rq_nr_phys_segments(req)) {
 156                 iod->sg_table.sgl = iod->first_sgl;
 157                 if (sg_alloc_table_chained(&iod->sg_table,
 158                                 blk_rq_nr_phys_segments(req),
 159                                 iod->sg_table.sgl, NVME_INLINE_SG_CNT)) {
 160                         nvme_cleanup_cmd(req);
 161                         return BLK_STS_RESOURCE;
 162                 }
 163
 164                 iod->req.sg = iod->sg_table.sgl;
 165                 iod->req.sg_cnt = blk_rq_map_sg(req->q, req, iod->sg_table.sgl);
 166                 iod->req.transfer_len = blk_rq_payload_bytes(req);
 167         }
 168
 169         schedule_work(&iod->work);
 170         return BLK_STS_OK;
 171 }
 172
 173 static void nvme_loop_submit_async_event(struct nvme_ctrl *arg)
 174 {
 175         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(arg);
 176         struct nvme_loop_queue *queue = &ctrl->queues[0];
 177         struct nvme_loop_iod *iod = &ctrl->async_event_iod;
 178
 179         memset(&iod->cmd, 0, sizeof(iod->cmd));
 180         iod->cmd.common.opcode = nvme_admin_async_event;
 181         iod->cmd.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
 182         iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
 183
 184         if (!nvmet_req_init(&iod->req, &queue->nvme_cq, &queue->nvme_sq,
 185                         &nvme_loop_ops)) {
 186                 dev_err(ctrl->ctrl.device, "failed async event work\n");
 187                 return;
 188         }
 189
 190         schedule_work(&iod->work);
 191 }
 192
 193 static int nvme_loop_init_iod(struct nvme_loop_ctrl *ctrl,
 194                 struct nvme_loop_iod *iod, unsigned int queue_idx)
 195 {
 196         iod->req.cmd = &iod->cmd;
 197         iod->req.cqe = &iod->cqe;
 198         iod->queue = &ctrl->queues[queue_idx];
 199         INIT_WORK(&iod->work, nvme_loop_execute_work);
 200         return 0;
 201 }
 202
 203 static int nvme_loop_init_request(struct blk_mq_tag_set *set,
 204                 struct request *req, unsigned int hctx_idx,
 205                 unsigned int numa_node)
 206 {
 207         struct nvme_loop_ctrl *ctrl = set->driver_data;
 208         struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
 209
 210         nvme_req(req)->ctrl = &ctrl->ctrl;
 211         nvme_req(req)->cmd = &iod->cmd;
 212         return nvme_loop_init_iod(ctrl, blk_mq_rq_to_pdu(req),
 213                         (set == &ctrl->tag_set) ? hctx_idx + 1 : 0);
 214 }
 215
 216 static struct lock_class_key loop_hctx_fq_lock_key;
 217
 218 static int nvme_loop_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 219                 unsigned int hctx_idx)
 220 {
 221         struct nvme_loop_ctrl *ctrl = data;
 222         struct nvme_loop_queue *queue = &ctrl->queues[hctx_idx + 1];
 223
 224         BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
 225
 226         /*
 227          * flush_end_io() can be called recursively for us, so use our own
 228          * lock class key for avoiding lockdep possible recursive locking,
 229          * then we can remove the dynamically allocated lock class for each
 230          * flush queue, that way may cause horrible boot delay.
 231          */
 232         blk_mq_hctx_set_fq_lock_class(hctx, &loop_hctx_fq_lock_key);
 233
 234         hctx->driver_data = queue;
 235         return 0;
 236 }
 237
 238 static int nvme_loop_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 239                 unsigned int hctx_idx)
 240 {
 241         struct nvme_loop_ctrl *ctrl = data;
 242         struct nvme_loop_queue *queue = &ctrl->queues[0];
 243
 244         BUG_ON(hctx_idx != 0);
 245
 246         hctx->driver_data = queue;
 247         return 0;
 248 }
 249
 250 static const struct blk_mq_ops nvme_loop_mq_ops = {
 251         .queue_rq       = nvme_loop_queue_rq,
 252         .complete       = nvme_loop_complete_rq,
 253         .init_request   = nvme_loop_init_request,
 254         .init_hctx      = nvme_loop_init_hctx,
 255 };
 256
 257 static const struct blk_mq_ops nvme_loop_admin_mq_ops = {
 258         .queue_rq       = nvme_loop_queue_rq,
 259         .complete       = nvme_loop_complete_rq,
 260         .init_request   = nvme_loop_init_request,
 261         .init_hctx      = nvme_loop_init_admin_hctx,
 262 };
 263
 264 static void nvme_loop_destroy_admin_queue(struct nvme_loop_ctrl *ctrl)
 265 {
 266         clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
 267         nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
 268         blk_cleanup_queue(ctrl->ctrl.admin_q);
 269         blk_cleanup_queue(ctrl->ctrl.fabrics_q);
 270         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 271 }
 272
 273 static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl)
 274 {
 275         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
 276
 277         if (list_empty(&ctrl->list))
 278                 goto free_ctrl;
 279
 280         mutex_lock(&nvme_loop_ctrl_mutex);
 281         list_del(&ctrl->list);
 282         mutex_unlock(&nvme_loop_ctrl_mutex);
 283
 284         if (nctrl->tagset) {
 285                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 286                 blk_mq_free_tag_set(&ctrl->tag_set);
 287         }
 288         kfree(ctrl->queues);
 289         nvmf_free_options(nctrl->opts);
 290 free_ctrl:
 291         kfree(ctrl);
 292 }
 293
 294 static void nvme_loop_destroy_io_queues(struct nvme_loop_ctrl *ctrl)
 295 {
 296         int i;
 297
 298         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 299                 clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
 300                 nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
 301         }
 302         ctrl->ctrl.queue_count = 1;
 303 }
 304
 305 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl)
 306 {
 307         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 308         unsigned int nr_io_queues;
 309         int ret, i;
 310
 311         nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
 312         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 313         if (ret || !nr_io_queues)
 314                 return ret;
 315
 316         dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues);
 317
 318         for (i = 1; i <= nr_io_queues; i++) {
 319                 ctrl->queues[i].ctrl = ctrl;
 320                 ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
 321                 if (ret)
 322                         goto out_destroy_queues;
 323
 324                 ctrl->ctrl.queue_count++;
 325         }
 326
 327         return 0;
 328
 329 out_destroy_queues:
 330         nvme_loop_destroy_io_queues(ctrl);
 331         return ret;
 332 }
 333
 334 static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl)
 335 {
 336         int i, ret;
 337
 338         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 339                 ret = nvmf_connect_io_queue(&ctrl->ctrl, i, false);
 340                 if (ret)
 341                         return ret;
 342                 set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
 343         }
 344
 345         return 0;
 346 }
 347
 348 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
 349 {
 350         int error;
 351
 352         memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
 353         ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
 354         ctrl->admin_tag_set.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
 355         ctrl->admin_tag_set.reserved_tags = NVMF_RESERVED_TAGS;
 356         ctrl->admin_tag_set.numa_node = ctrl->ctrl.numa_node;
 357         ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
 358                 NVME_INLINE_SG_CNT * sizeof(struct scatterlist);
 359         ctrl->admin_tag_set.driver_data = ctrl;
 360         ctrl->admin_tag_set.nr_hw_queues = 1;
 361         ctrl->admin_tag_set.timeout = NVME_ADMIN_TIMEOUT;
 362         ctrl->admin_tag_set.flags = BLK_MQ_F_NO_SCHED;
 363
 364         ctrl->queues[0].ctrl = ctrl;
 365         error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
 366         if (error)
 367                 return error;
 368         ctrl->ctrl.queue_count = 1;
 369
 370         error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
 371         if (error)
 372                 goto out_free_sq;
 373         ctrl->ctrl.admin_tagset = &ctrl->admin_tag_set;
 374
 375         ctrl->ctrl.fabrics_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 376         if (IS_ERR(ctrl->ctrl.fabrics_q)) {
 377                 error = PTR_ERR(ctrl->ctrl.fabrics_q);
 378                 goto out_free_tagset;
 379         }
 380
 381         ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 382         if (IS_ERR(ctrl->ctrl.admin_q)) {
 383                 error = PTR_ERR(ctrl->ctrl.admin_q);
 384                 goto out_cleanup_fabrics_q;
 385         }
 386
 387         error = nvmf_connect_admin_queue(&ctrl->ctrl);
 388         if (error)
 389                 goto out_cleanup_queue;
 390
 391         set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
 392
 393         error = nvme_enable_ctrl(&ctrl->ctrl);
 394         if (error)
 395                 goto out_cleanup_queue;
 396
 397         ctrl->ctrl.max_hw_sectors =
 398                 (NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
 399
 400         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
 401
 402         error = nvme_init_ctrl_finish(&ctrl->ctrl);
 403         if (error)
 404                 goto out_cleanup_queue;
 405
 406         return 0;
 407
 408 out_cleanup_queue:
 409         clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
 410         blk_cleanup_queue(ctrl->ctrl.admin_q);
 411 out_cleanup_fabrics_q:
 412         blk_cleanup_queue(ctrl->ctrl.fabrics_q);
 413 out_free_tagset:
 414         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 415 out_free_sq:
 416         nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
 417         return error;
 418 }
 419
 420 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
 421 {
 422         if (ctrl->ctrl.queue_count > 1) {
 423                 nvme_stop_queues(&ctrl->ctrl);
 424                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
 425                                         nvme_cancel_request, &ctrl->ctrl);
 426                 blk_mq_tagset_wait_completed_request(&ctrl->tag_set);
 427                 nvme_loop_destroy_io_queues(ctrl);
 428         }
 429
 430         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
 431         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
 432                 nvme_shutdown_ctrl(&ctrl->ctrl);
 433
 434         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
 435                                 nvme_cancel_request, &ctrl->ctrl);
 436         blk_mq_tagset_wait_completed_request(&ctrl->admin_tag_set);
 437         nvme_loop_destroy_admin_queue(ctrl);
 438 }
 439
 440 static void nvme_loop_delete_ctrl_host(struct nvme_ctrl *ctrl)
 441 {
 442         nvme_loop_shutdown_ctrl(to_loop_ctrl(ctrl));
 443 }
 444
 445 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
 446 {
 447         struct nvme_loop_ctrl *ctrl;
 448
 449         mutex_lock(&nvme_loop_ctrl_mutex);
 450         list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
 451                 if (ctrl->ctrl.cntlid == nctrl->cntlid)
 452                         nvme_delete_ctrl(&ctrl->ctrl);
 453         }
 454         mutex_unlock(&nvme_loop_ctrl_mutex);
 455 }
 456
 457 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
 458 {
 459         struct nvme_loop_ctrl *ctrl =
 460                 container_of(work, struct nvme_loop_ctrl, ctrl.reset_work);
 461         int ret;
 462
 463         nvme_stop_ctrl(&ctrl->ctrl);
 464         nvme_loop_shutdown_ctrl(ctrl);
 465
 466         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
 467                 /* state change failure should never happen */
 468                 WARN_ON_ONCE(1);
 469                 return;
 470         }
 471
 472         ret = nvme_loop_configure_admin_queue(ctrl);
 473         if (ret)
 474                 goto out_disable;
 475
 476         ret = nvme_loop_init_io_queues(ctrl);
 477         if (ret)
 478                 goto out_destroy_admin;
 479
 480         ret = nvme_loop_connect_io_queues(ctrl);
 481         if (ret)
 482                 goto out_destroy_io;
 483
 484         blk_mq_update_nr_hw_queues(&ctrl->tag_set,
 485                         ctrl->ctrl.queue_count - 1);
 486
 487         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE))
 488                 WARN_ON_ONCE(1);
 489
 490         nvme_start_ctrl(&ctrl->ctrl);
 491
 492         return;
 493
 494 out_destroy_io:
 495         nvme_loop_destroy_io_queues(ctrl);
 496 out_destroy_admin:
 497         nvme_loop_destroy_admin_queue(ctrl);
 498 out_disable:
 499         dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
 500         nvme_uninit_ctrl(&ctrl->ctrl);
 501 }
 502
 503 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
 504         .name                   = "loop",
 505         .module                 = THIS_MODULE,
 506         .flags                  = NVME_F_FABRICS,
 507         .reg_read32             = nvmf_reg_read32,
 508         .reg_read64             = nvmf_reg_read64,
 509         .reg_write32            = nvmf_reg_write32,
 510         .free_ctrl              = nvme_loop_free_ctrl,
 511         .submit_async_event     = nvme_loop_submit_async_event,
 512         .delete_ctrl            = nvme_loop_delete_ctrl_host,
 513         .get_address            = nvmf_get_address,
 514 };
 515
 516 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
 517 {
 518         int ret;
 519
 520         ret = nvme_loop_init_io_queues(ctrl);
 521         if (ret)
 522                 return ret;
 523
 524         memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
 525         ctrl->tag_set.ops = &nvme_loop_mq_ops;
 526         ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
 527         ctrl->tag_set.reserved_tags = NVMF_RESERVED_TAGS;
 528         ctrl->tag_set.numa_node = ctrl->ctrl.numa_node;
 529         ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
 530         ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
 531                 NVME_INLINE_SG_CNT * sizeof(struct scatterlist);
 532         ctrl->tag_set.driver_data = ctrl;
 533         ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
 534         ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
 535         ctrl->ctrl.tagset = &ctrl->tag_set;
 536
 537         ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
 538         if (ret)
 539                 goto out_destroy_queues;
 540
 541         ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 542         if (IS_ERR(ctrl->ctrl.connect_q)) {
 543                 ret = PTR_ERR(ctrl->ctrl.connect_q);
 544                 goto out_free_tagset;
 545         }
 546
 547         ret = nvme_loop_connect_io_queues(ctrl);
 548         if (ret)
 549                 goto out_cleanup_connect_q;
 550
 551         return 0;
 552
 553 out_cleanup_connect_q:
 554         blk_cleanup_queue(ctrl->ctrl.connect_q);
 555 out_free_tagset:
 556         blk_mq_free_tag_set(&ctrl->tag_set);
 557 out_destroy_queues:
 558         nvme_loop_destroy_io_queues(ctrl);
 559         return ret;
 560 }
 561
 562 static struct nvmet_port *nvme_loop_find_port(struct nvme_ctrl *ctrl)
 563 {
 564         struct nvmet_port *p, *found = NULL;
 565
 566         mutex_lock(&nvme_loop_ports_mutex);
 567         list_for_each_entry(p, &nvme_loop_ports, entry) {
 568                 /* if no transport address is specified use the first port */
 569                 if ((ctrl->opts->mask & NVMF_OPT_TRADDR) &&
 570                     strcmp(ctrl->opts->traddr, p->disc_addr.traddr))
 571                         continue;
 572                 found = p;
 573                 break;
 574         }
 575         mutex_unlock(&nvme_loop_ports_mutex);
 576         return found;
 577 }
 578
 579 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
 580                 struct nvmf_ctrl_options *opts)
 581 {
 582         struct nvme_loop_ctrl *ctrl;
 583         int ret;
 584
 585         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
 586         if (!ctrl)
 587                 return ERR_PTR(-ENOMEM);
 588         ctrl->ctrl.opts = opts;
 589         INIT_LIST_HEAD(&ctrl->list);
 590
 591         INIT_WORK(&ctrl->ctrl.reset_work, nvme_loop_reset_ctrl_work);
 592
 593         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
 594                                 0 /* no quirks, we're perfect! */);
 595         if (ret) {
 596                 kfree(ctrl);
 597                 goto out;
 598         }
 599
 600         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING))
 601                 WARN_ON_ONCE(1);
 602
 603         ret = -ENOMEM;
 604
 605         ctrl->ctrl.sqsize = opts->queue_size - 1;
 606         ctrl->ctrl.kato = opts->kato;
 607         ctrl->port = nvme_loop_find_port(&ctrl->ctrl);
 608
 609         ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
 610                         GFP_KERNEL);
 611         if (!ctrl->queues)
 612                 goto out_uninit_ctrl;
 613
 614         ret = nvme_loop_configure_admin_queue(ctrl);
 615         if (ret)
 616                 goto out_free_queues;
 617
 618         if (opts->queue_size > ctrl->ctrl.maxcmd) {
 619                 /* warn if maxcmd is lower than queue_size */
 620                 dev_warn(ctrl->ctrl.device,
 621                         "queue_size %zu > ctrl maxcmd %u, clamping down\n",
 622                         opts->queue_size, ctrl->ctrl.maxcmd);
 623                 opts->queue_size = ctrl->ctrl.maxcmd;
 624         }
 625
 626         if (opts->nr_io_queues) {
 627                 ret = nvme_loop_create_io_queues(ctrl);
 628                 if (ret)
 629                         goto out_remove_admin_queue;
 630         }
 631
 632         nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
 633
 634         dev_info(ctrl->ctrl.device,
 635                  "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
 636
 637         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE))
 638                 WARN_ON_ONCE(1);
 639
 640         mutex_lock(&nvme_loop_ctrl_mutex);
 641         list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
 642         mutex_unlock(&nvme_loop_ctrl_mutex);
 643
 644         nvme_start_ctrl(&ctrl->ctrl);
 645
 646         return &ctrl->ctrl;
 647
 648 out_remove_admin_queue:
 649         nvme_loop_destroy_admin_queue(ctrl);
 650 out_free_queues:
 651         kfree(ctrl->queues);
 652 out_uninit_ctrl:
 653         nvme_uninit_ctrl(&ctrl->ctrl);
 654         nvme_put_ctrl(&ctrl->ctrl);
 655 out:
 656         if (ret > 0)
 657                 ret = -EIO;
 658         return ERR_PTR(ret);
 659 }
 660
 661 static int nvme_loop_add_port(struct nvmet_port *port)
 662 {
 663         mutex_lock(&nvme_loop_ports_mutex);
 664         list_add_tail(&port->entry, &nvme_loop_ports);
 665         mutex_unlock(&nvme_loop_ports_mutex);
 666         return 0;
 667 }
 668
 669 static void nvme_loop_remove_port(struct nvmet_port *port)
 670 {
 671         mutex_lock(&nvme_loop_ports_mutex);
 672         list_del_init(&port->entry);
 673         mutex_unlock(&nvme_loop_ports_mutex);
 674
 675         /*
 676          * Ensure any ctrls that are in the process of being
 677          * deleted are in fact deleted before we return
 678          * and free the port. This is to prevent active
 679          * ctrls from using a port after it's freed.
 680          */
 681         flush_workqueue(nvme_delete_wq);
 682 }
 683
 684 static const struct nvmet_fabrics_ops nvme_loop_ops = {
 685         .owner          = THIS_MODULE,
 686         .type           = NVMF_TRTYPE_LOOP,
 687         .add_port       = nvme_loop_add_port,
 688         .remove_port    = nvme_loop_remove_port,
 689         .queue_response = nvme_loop_queue_response,
 690         .delete_ctrl    = nvme_loop_delete_ctrl,
 691 };
 692
 693 static struct nvmf_transport_ops nvme_loop_transport = {
 694         .name           = "loop",
 695         .module         = THIS_MODULE,
 696         .create_ctrl    = nvme_loop_create_ctrl,
 697         .allowed_opts   = NVMF_OPT_TRADDR,
 698 };
 699
 700 static int __init nvme_loop_init_module(void)
 701 {
 702         int ret;
 703
 704         ret = nvmet_register_transport(&nvme_loop_ops);
 705         if (ret)
 706                 return ret;
 707
 708         ret = nvmf_register_transport(&nvme_loop_transport);
 709         if (ret)
 710                 nvmet_unregister_transport(&nvme_loop_ops);
 711
 712         return ret;
 713 }
 714
 715 static void __exit nvme_loop_cleanup_module(void)
 716 {
 717         struct nvme_loop_ctrl *ctrl, *next;
 718
 719         nvmf_unregister_transport(&nvme_loop_transport);
 720         nvmet_unregister_transport(&nvme_loop_ops);
 721
 722         mutex_lock(&nvme_loop_ctrl_mutex);
 723         list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
 724                 nvme_delete_ctrl(&ctrl->ctrl);
 725         mutex_unlock(&nvme_loop_ctrl_mutex);
 726
 727         flush_workqueue(nvme_delete_wq);
 728 }
 729
 730 module_init(nvme_loop_init_module);
 731 module_exit(nvme_loop_cleanup_module);
 732
 733 MODULE_LICENSE("GPL v2");
 734 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */