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_end_request(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 (!nvmf_check_ready(&queue->ctrl->ctrl, req, queue_ready))
 142                 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, req);
 143
 144         ret = nvme_setup_cmd(ns, req, &iod->cmd);
 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
 209         nvme_req(req)->ctrl = &ctrl->ctrl;
 210         return nvme_loop_init_iod(ctrl, blk_mq_rq_to_pdu(req),
 211                         (set == &ctrl->tag_set) ? hctx_idx + 1 : 0);
 212 }
 213
 214 static int nvme_loop_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 215                 unsigned int hctx_idx)
 216 {
 217         struct nvme_loop_ctrl *ctrl = data;
 218         struct nvme_loop_queue *queue = &ctrl->queues[hctx_idx + 1];
 219
 220         BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
 221
 222         hctx->driver_data = queue;
 223         return 0;
 224 }
 225
 226 static int nvme_loop_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 227                 unsigned int hctx_idx)
 228 {
 229         struct nvme_loop_ctrl *ctrl = data;
 230         struct nvme_loop_queue *queue = &ctrl->queues[0];
 231
 232         BUG_ON(hctx_idx != 0);
 233
 234         hctx->driver_data = queue;
 235         return 0;
 236 }
 237
 238 static const struct blk_mq_ops nvme_loop_mq_ops = {
 239         .queue_rq       = nvme_loop_queue_rq,
 240         .complete       = nvme_loop_complete_rq,
 241         .init_request   = nvme_loop_init_request,
 242         .init_hctx      = nvme_loop_init_hctx,
 243 };
 244
 245 static const struct blk_mq_ops nvme_loop_admin_mq_ops = {
 246         .queue_rq       = nvme_loop_queue_rq,
 247         .complete       = nvme_loop_complete_rq,
 248         .init_request   = nvme_loop_init_request,
 249         .init_hctx      = nvme_loop_init_admin_hctx,
 250 };
 251
 252 static void nvme_loop_destroy_admin_queue(struct nvme_loop_ctrl *ctrl)
 253 {
 254         clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
 255         nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
 256         blk_cleanup_queue(ctrl->ctrl.admin_q);
 257         blk_cleanup_queue(ctrl->ctrl.fabrics_q);
 258         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 259 }
 260
 261 static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl)
 262 {
 263         struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
 264
 265         if (list_empty(&ctrl->list))
 266                 goto free_ctrl;
 267
 268         mutex_lock(&nvme_loop_ctrl_mutex);
 269         list_del(&ctrl->list);
 270         mutex_unlock(&nvme_loop_ctrl_mutex);
 271
 272         if (nctrl->tagset) {
 273                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 274                 blk_mq_free_tag_set(&ctrl->tag_set);
 275         }
 276         kfree(ctrl->queues);
 277         nvmf_free_options(nctrl->opts);
 278 free_ctrl:
 279         kfree(ctrl);
 280 }
 281
 282 static void nvme_loop_destroy_io_queues(struct nvme_loop_ctrl *ctrl)
 283 {
 284         int i;
 285
 286         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 287                 clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
 288                 nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
 289         }
 290 }
 291
 292 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl)
 293 {
 294         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 295         unsigned int nr_io_queues;
 296         int ret, i;
 297
 298         nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
 299         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 300         if (ret || !nr_io_queues)
 301                 return ret;
 302
 303         dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues);
 304
 305         for (i = 1; i <= nr_io_queues; i++) {
 306                 ctrl->queues[i].ctrl = ctrl;
 307                 ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
 308                 if (ret)
 309                         goto out_destroy_queues;
 310
 311                 ctrl->ctrl.queue_count++;
 312         }
 313
 314         return 0;
 315
 316 out_destroy_queues:
 317         nvme_loop_destroy_io_queues(ctrl);
 318         return ret;
 319 }
 320
 321 static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl)
 322 {
 323         int i, ret;
 324
 325         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 326                 ret = nvmf_connect_io_queue(&ctrl->ctrl, i, false);
 327                 if (ret)
 328                         return ret;
 329                 set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
 330         }
 331
 332         return 0;
 333 }
 334
 335 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
 336 {
 337         int error;
 338
 339         memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
 340         ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
 341         ctrl->admin_tag_set.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
 342         ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
 343         ctrl->admin_tag_set.numa_node = ctrl->ctrl.numa_node;
 344         ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
 345                 NVME_INLINE_SG_CNT * sizeof(struct scatterlist);
 346         ctrl->admin_tag_set.driver_data = ctrl;
 347         ctrl->admin_tag_set.nr_hw_queues = 1;
 348         ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
 349         ctrl->admin_tag_set.flags = BLK_MQ_F_NO_SCHED;
 350
 351         ctrl->queues[0].ctrl = ctrl;
 352         error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
 353         if (error)
 354                 return error;
 355         ctrl->ctrl.queue_count = 1;
 356
 357         error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
 358         if (error)
 359                 goto out_free_sq;
 360         ctrl->ctrl.admin_tagset = &ctrl->admin_tag_set;
 361
 362         ctrl->ctrl.fabrics_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 363         if (IS_ERR(ctrl->ctrl.fabrics_q)) {
 364                 error = PTR_ERR(ctrl->ctrl.fabrics_q);
 365                 goto out_free_tagset;
 366         }
 367
 368         ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 369         if (IS_ERR(ctrl->ctrl.admin_q)) {
 370                 error = PTR_ERR(ctrl->ctrl.admin_q);
 371                 goto out_cleanup_fabrics_q;
 372         }
 373
 374         error = nvmf_connect_admin_queue(&ctrl->ctrl);
 375         if (error)
 376                 goto out_cleanup_queue;
 377
 378         set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
 379
 380         error = nvme_enable_ctrl(&ctrl->ctrl);
 381         if (error)
 382                 goto out_cleanup_queue;
 383
 384         ctrl->ctrl.max_hw_sectors =
 385                 (NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
 386
 387         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
 388
 389         error = nvme_init_identify(&ctrl->ctrl);
 390         if (error)
 391                 goto out_cleanup_queue;
 392
 393         return 0;
 394
 395 out_cleanup_queue:
 396         blk_cleanup_queue(ctrl->ctrl.admin_q);
 397 out_cleanup_fabrics_q:
 398         blk_cleanup_queue(ctrl->ctrl.fabrics_q);
 399 out_free_tagset:
 400         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 401 out_free_sq:
 402         nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
 403         return error;
 404 }
 405
 406 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
 407 {
 408         if (ctrl->ctrl.queue_count > 1) {
 409                 nvme_stop_queues(&ctrl->ctrl);
 410                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
 411                                         nvme_cancel_request, &ctrl->ctrl);
 412                 blk_mq_tagset_wait_completed_request(&ctrl->tag_set);
 413                 nvme_loop_destroy_io_queues(ctrl);
 414         }
 415
 416         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
 417         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
 418                 nvme_shutdown_ctrl(&ctrl->ctrl);
 419
 420         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
 421                                 nvme_cancel_request, &ctrl->ctrl);
 422         blk_mq_tagset_wait_completed_request(&ctrl->admin_tag_set);
 423         nvme_loop_destroy_admin_queue(ctrl);
 424 }
 425
 426 static void nvme_loop_delete_ctrl_host(struct nvme_ctrl *ctrl)
 427 {
 428         nvme_loop_shutdown_ctrl(to_loop_ctrl(ctrl));
 429 }
 430
 431 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
 432 {
 433         struct nvme_loop_ctrl *ctrl;
 434
 435         mutex_lock(&nvme_loop_ctrl_mutex);
 436         list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
 437                 if (ctrl->ctrl.cntlid == nctrl->cntlid)
 438                         nvme_delete_ctrl(&ctrl->ctrl);
 439         }
 440         mutex_unlock(&nvme_loop_ctrl_mutex);
 441 }
 442
 443 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
 444 {
 445         struct nvme_loop_ctrl *ctrl =
 446                 container_of(work, struct nvme_loop_ctrl, ctrl.reset_work);
 447         int ret;
 448
 449         nvme_stop_ctrl(&ctrl->ctrl);
 450         nvme_loop_shutdown_ctrl(ctrl);
 451
 452         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
 453                 /* state change failure should never happen */
 454                 WARN_ON_ONCE(1);
 455                 return;
 456         }
 457
 458         ret = nvme_loop_configure_admin_queue(ctrl);
 459         if (ret)
 460                 goto out_disable;
 461
 462         ret = nvme_loop_init_io_queues(ctrl);
 463         if (ret)
 464                 goto out_destroy_admin;
 465
 466         ret = nvme_loop_connect_io_queues(ctrl);
 467         if (ret)
 468                 goto out_destroy_io;
 469
 470         blk_mq_update_nr_hw_queues(&ctrl->tag_set,
 471                         ctrl->ctrl.queue_count - 1);
 472
 473         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE))
 474                 WARN_ON_ONCE(1);
 475
 476         nvme_start_ctrl(&ctrl->ctrl);
 477
 478         return;
 479
 480 out_destroy_io:
 481         nvme_loop_destroy_io_queues(ctrl);
 482 out_destroy_admin:
 483         nvme_loop_destroy_admin_queue(ctrl);
 484 out_disable:
 485         dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
 486         nvme_uninit_ctrl(&ctrl->ctrl);
 487 }
 488
 489 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
 490         .name                   = "loop",
 491         .module                 = THIS_MODULE,
 492         .flags                  = NVME_F_FABRICS,
 493         .reg_read32             = nvmf_reg_read32,
 494         .reg_read64             = nvmf_reg_read64,
 495         .reg_write32            = nvmf_reg_write32,
 496         .free_ctrl              = nvme_loop_free_ctrl,
 497         .submit_async_event     = nvme_loop_submit_async_event,
 498         .delete_ctrl            = nvme_loop_delete_ctrl_host,
 499         .get_address            = nvmf_get_address,
 500 };
 501
 502 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
 503 {
 504         int ret;
 505
 506         ret = nvme_loop_init_io_queues(ctrl);
 507         if (ret)
 508                 return ret;
 509
 510         memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
 511         ctrl->tag_set.ops = &nvme_loop_mq_ops;
 512         ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
 513         ctrl->tag_set.reserved_tags = 1; /* fabric connect */
 514         ctrl->tag_set.numa_node = ctrl->ctrl.numa_node;
 515         ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
 516         ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
 517                 NVME_INLINE_SG_CNT * sizeof(struct scatterlist);
 518         ctrl->tag_set.driver_data = ctrl;
 519         ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
 520         ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
 521         ctrl->ctrl.tagset = &ctrl->tag_set;
 522
 523         ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
 524         if (ret)
 525                 goto out_destroy_queues;
 526
 527         ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 528         if (IS_ERR(ctrl->ctrl.connect_q)) {
 529                 ret = PTR_ERR(ctrl->ctrl.connect_q);
 530                 goto out_free_tagset;
 531         }
 532
 533         ret = nvme_loop_connect_io_queues(ctrl);
 534         if (ret)
 535                 goto out_cleanup_connect_q;
 536
 537         return 0;
 538
 539 out_cleanup_connect_q:
 540         blk_cleanup_queue(ctrl->ctrl.connect_q);
 541 out_free_tagset:
 542         blk_mq_free_tag_set(&ctrl->tag_set);
 543 out_destroy_queues:
 544         nvme_loop_destroy_io_queues(ctrl);
 545         return ret;
 546 }
 547
 548 static struct nvmet_port *nvme_loop_find_port(struct nvme_ctrl *ctrl)
 549 {
 550         struct nvmet_port *p, *found = NULL;
 551
 552         mutex_lock(&nvme_loop_ports_mutex);
 553         list_for_each_entry(p, &nvme_loop_ports, entry) {
 554                 /* if no transport address is specified use the first port */
 555                 if ((ctrl->opts->mask & NVMF_OPT_TRADDR) &&
 556                     strcmp(ctrl->opts->traddr, p->disc_addr.traddr))
 557                         continue;
 558                 found = p;
 559                 break;
 560         }
 561         mutex_unlock(&nvme_loop_ports_mutex);
 562         return found;
 563 }
 564
 565 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
 566                 struct nvmf_ctrl_options *opts)
 567 {
 568         struct nvme_loop_ctrl *ctrl;
 569         int ret;
 570
 571         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
 572         if (!ctrl)
 573                 return ERR_PTR(-ENOMEM);
 574         ctrl->ctrl.opts = opts;
 575         INIT_LIST_HEAD(&ctrl->list);
 576
 577         INIT_WORK(&ctrl->ctrl.reset_work, nvme_loop_reset_ctrl_work);
 578
 579         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
 580                                 0 /* no quirks, we're perfect! */);
 581         if (ret)
 582                 goto out_put_ctrl;
 583
 584         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING))
 585                 WARN_ON_ONCE(1);
 586
 587         ret = -ENOMEM;
 588
 589         ctrl->ctrl.sqsize = opts->queue_size - 1;
 590         ctrl->ctrl.kato = opts->kato;
 591         ctrl->port = nvme_loop_find_port(&ctrl->ctrl);
 592
 593         ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
 594                         GFP_KERNEL);
 595         if (!ctrl->queues)
 596                 goto out_uninit_ctrl;
 597
 598         ret = nvme_loop_configure_admin_queue(ctrl);
 599         if (ret)
 600                 goto out_free_queues;
 601
 602         if (opts->queue_size > ctrl->ctrl.maxcmd) {
 603                 /* warn if maxcmd is lower than queue_size */
 604                 dev_warn(ctrl->ctrl.device,
 605                         "queue_size %zu > ctrl maxcmd %u, clamping down\n",
 606                         opts->queue_size, ctrl->ctrl.maxcmd);
 607                 opts->queue_size = ctrl->ctrl.maxcmd;
 608         }
 609
 610         if (opts->nr_io_queues) {
 611                 ret = nvme_loop_create_io_queues(ctrl);
 612                 if (ret)
 613                         goto out_remove_admin_queue;
 614         }
 615
 616         nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
 617
 618         dev_info(ctrl->ctrl.device,
 619                  "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
 620
 621         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE))
 622                 WARN_ON_ONCE(1);
 623
 624         mutex_lock(&nvme_loop_ctrl_mutex);
 625         list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
 626         mutex_unlock(&nvme_loop_ctrl_mutex);
 627
 628         nvme_start_ctrl(&ctrl->ctrl);
 629
 630         return &ctrl->ctrl;
 631
 632 out_remove_admin_queue:
 633         nvme_loop_destroy_admin_queue(ctrl);
 634 out_free_queues:
 635         kfree(ctrl->queues);
 636 out_uninit_ctrl:
 637         nvme_uninit_ctrl(&ctrl->ctrl);
 638 out_put_ctrl:
 639         nvme_put_ctrl(&ctrl->ctrl);
 640         if (ret > 0)
 641                 ret = -EIO;
 642         return ERR_PTR(ret);
 643 }
 644
 645 static int nvme_loop_add_port(struct nvmet_port *port)
 646 {
 647         mutex_lock(&nvme_loop_ports_mutex);
 648         list_add_tail(&port->entry, &nvme_loop_ports);
 649         mutex_unlock(&nvme_loop_ports_mutex);
 650         return 0;
 651 }
 652
 653 static void nvme_loop_remove_port(struct nvmet_port *port)
 654 {
 655         mutex_lock(&nvme_loop_ports_mutex);
 656         list_del_init(&port->entry);
 657         mutex_unlock(&nvme_loop_ports_mutex);
 658
 659         /*
 660          * Ensure any ctrls that are in the process of being
 661          * deleted are in fact deleted before we return
 662          * and free the port. This is to prevent active
 663          * ctrls from using a port after it's freed.
 664          */
 665         flush_workqueue(nvme_delete_wq);
 666 }
 667
 668 static const struct nvmet_fabrics_ops nvme_loop_ops = {
 669         .owner          = THIS_MODULE,
 670         .type           = NVMF_TRTYPE_LOOP,
 671         .add_port       = nvme_loop_add_port,
 672         .remove_port    = nvme_loop_remove_port,
 673         .queue_response = nvme_loop_queue_response,
 674         .delete_ctrl    = nvme_loop_delete_ctrl,
 675 };
 676
 677 static struct nvmf_transport_ops nvme_loop_transport = {
 678         .name           = "loop",
 679         .module         = THIS_MODULE,
 680         .create_ctrl    = nvme_loop_create_ctrl,
 681         .allowed_opts   = NVMF_OPT_TRADDR,
 682 };
 683
 684 static int __init nvme_loop_init_module(void)
 685 {
 686         int ret;
 687
 688         ret = nvmet_register_transport(&nvme_loop_ops);
 689         if (ret)
 690                 return ret;
 691
 692         ret = nvmf_register_transport(&nvme_loop_transport);
 693         if (ret)
 694                 nvmet_unregister_transport(&nvme_loop_ops);
 695
 696         return ret;
 697 }
 698
 699 static void __exit nvme_loop_cleanup_module(void)
 700 {
 701         struct nvme_loop_ctrl *ctrl, *next;
 702
 703         nvmf_unregister_transport(&nvme_loop_transport);
 704         nvmet_unregister_transport(&nvme_loop_ops);
 705
 706         mutex_lock(&nvme_loop_ctrl_mutex);
 707         list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
 708                 nvme_delete_ctrl(&ctrl->ctrl);
 709         mutex_unlock(&nvme_loop_ctrl_mutex);
 710
 711         flush_workqueue(nvme_delete_wq);
 712 }
 713
 714 module_init(nvme_loop_init_module);
 715 module_exit(nvme_loop_cleanup_module);
 716
 717 MODULE_LICENSE("GPL v2");
 718 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */