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 }
 303
 304 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl)
 305 {
 306         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 307         unsigned int nr_io_queues;
 308         int ret, i;
 309
 310         nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
 311         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 312         if (ret || !nr_io_queues)
 313                 return ret;
 314
 315         dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues);
 316
 317         for (i = 1; i <= nr_io_queues; i++) {
 318                 ctrl->queues[i].ctrl = ctrl;
 319                 ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
 320                 if (ret)
 321                         goto out_destroy_queues;
 322
 323                 ctrl->ctrl.queue_count++;
 324         }
 325
 326         return 0;
 327
 328 out_destroy_queues:
 329         nvme_loop_destroy_io_queues(ctrl);
 330         return ret;
 331 }
 332
 333 static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl)
 334 {
 335         int i, ret;
 336
 337         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 338                 ret = nvmf_connect_io_queue(&ctrl->ctrl, i, false);
 339                 if (ret)
 340                         return ret;
 341                 set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
 342         }
 343
 344         return 0;
 345 }
 346
 347 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
 348 {
 349         int error;
 350
 351         memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
 352         ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
 353         ctrl->admin_tag_set.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
 354         ctrl->admin_tag_set.reserved_tags = NVMF_RESERVED_TAGS;
 355         ctrl->admin_tag_set.numa_node = ctrl->ctrl.numa_node;
 356         ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
 357                 NVME_INLINE_SG_CNT * sizeof(struct scatterlist);
 358         ctrl->admin_tag_set.driver_data = ctrl;
 359         ctrl->admin_tag_set.nr_hw_queues = 1;
 360         ctrl->admin_tag_set.timeout = NVME_ADMIN_TIMEOUT;
 361         ctrl->admin_tag_set.flags = BLK_MQ_F_NO_SCHED;
 362
 363         ctrl->queues[0].ctrl = ctrl;
 364         error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
 365         if (error)
 366                 return error;
 367         ctrl->ctrl.queue_count = 1;
 368
 369         error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
 370         if (error)
 371                 goto out_free_sq;
 372         ctrl->ctrl.admin_tagset = &ctrl->admin_tag_set;
 373
 374         ctrl->ctrl.fabrics_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 375         if (IS_ERR(ctrl->ctrl.fabrics_q)) {
 376                 error = PTR_ERR(ctrl->ctrl.fabrics_q);
 377                 goto out_free_tagset;
 378         }
 379
 380         ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 381         if (IS_ERR(ctrl->ctrl.admin_q)) {
 382                 error = PTR_ERR(ctrl->ctrl.admin_q);
 383                 goto out_cleanup_fabrics_q;
 384         }
 385
 386         error = nvmf_connect_admin_queue(&ctrl->ctrl);
 387         if (error)
 388                 goto out_cleanup_queue;
 389
 390         set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
 391
 392         error = nvme_enable_ctrl(&ctrl->ctrl);
 393         if (error)
 394                 goto out_cleanup_queue;
 395
 396         ctrl->ctrl.max_hw_sectors =
 397                 (NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
 398
 399         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
 400
 401         error = nvme_init_ctrl_finish(&ctrl->ctrl);
 402         if (error)
 403                 goto out_cleanup_queue;
 404
 405         return 0;
 406
 407 out_cleanup_queue:
 408         blk_cleanup_queue(ctrl->ctrl.admin_q);
 409 out_cleanup_fabrics_q:
 410         blk_cleanup_queue(ctrl->ctrl.fabrics_q);
 411 out_free_tagset:
 412         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 413 out_free_sq:
 414         nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
 415         return error;
 416 }
 417
 418 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
 419 {
 420         if (ctrl->ctrl.queue_count > 1) {
 421                 nvme_stop_queues(&ctrl->ctrl);
 422                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
 423                                         nvme_cancel_request, &ctrl->ctrl);
 424                 blk_mq_tagset_wait_completed_request(&ctrl->tag_set);
 425                 nvme_loop_destroy_io_queues(ctrl);
 426         }
 427
 428         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
 429         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
 430                 nvme_shutdown_ctrl(&ctrl->ctrl);
 431
 432         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
 433                                 nvme_cancel_request, &ctrl->ctrl);
 434         blk_mq_tagset_wait_completed_request(&ctrl->admin_tag_set);
 435         nvme_loop_destroy_admin_queue(ctrl);
 436 }
 437
 438 static void nvme_loop_delete_ctrl_host(struct nvme_ctrl *ctrl)
 439 {
 440         nvme_loop_shutdown_ctrl(to_loop_ctrl(ctrl));
 441 }
 442
 443 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
 444 {
 445         struct nvme_loop_ctrl *ctrl;
 446
 447         mutex_lock(&nvme_loop_ctrl_mutex);
 448         list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
 449                 if (ctrl->ctrl.cntlid == nctrl->cntlid)
 450                         nvme_delete_ctrl(&ctrl->ctrl);
 451         }
 452         mutex_unlock(&nvme_loop_ctrl_mutex);
 453 }
 454
 455 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
 456 {
 457         struct nvme_loop_ctrl *ctrl =
 458                 container_of(work, struct nvme_loop_ctrl, ctrl.reset_work);
 459         int ret;
 460
 461         nvme_stop_ctrl(&ctrl->ctrl);
 462         nvme_loop_shutdown_ctrl(ctrl);
 463
 464         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
 465                 /* state change failure should never happen */
 466                 WARN_ON_ONCE(1);
 467                 return;
 468         }
 469
 470         ret = nvme_loop_configure_admin_queue(ctrl);
 471         if (ret)
 472                 goto out_disable;
 473
 474         ret = nvme_loop_init_io_queues(ctrl);
 475         if (ret)
 476                 goto out_destroy_admin;
 477
 478         ret = nvme_loop_connect_io_queues(ctrl);
 479         if (ret)
 480                 goto out_destroy_io;
 481
 482         blk_mq_update_nr_hw_queues(&ctrl->tag_set,
 483                         ctrl->ctrl.queue_count - 1);
 484
 485         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE))
 486                 WARN_ON_ONCE(1);
 487
 488         nvme_start_ctrl(&ctrl->ctrl);
 489
 490         return;
 491
 492 out_destroy_io:
 493         nvme_loop_destroy_io_queues(ctrl);
 494 out_destroy_admin:
 495         nvme_loop_destroy_admin_queue(ctrl);
 496 out_disable:
 497         dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
 498         nvme_uninit_ctrl(&ctrl->ctrl);
 499 }
 500
 501 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
 502         .name                   = "loop",
 503         .module                 = THIS_MODULE,
 504         .flags                  = NVME_F_FABRICS,
 505         .reg_read32             = nvmf_reg_read32,
 506         .reg_read64             = nvmf_reg_read64,
 507         .reg_write32            = nvmf_reg_write32,
 508         .free_ctrl              = nvme_loop_free_ctrl,
 509         .submit_async_event     = nvme_loop_submit_async_event,
 510         .delete_ctrl            = nvme_loop_delete_ctrl_host,
 511         .get_address            = nvmf_get_address,
 512 };
 513
 514 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
 515 {
 516         int ret;
 517
 518         ret = nvme_loop_init_io_queues(ctrl);
 519         if (ret)
 520                 return ret;
 521
 522         memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
 523         ctrl->tag_set.ops = &nvme_loop_mq_ops;
 524         ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
 525         ctrl->tag_set.reserved_tags = NVMF_RESERVED_TAGS;
 526         ctrl->tag_set.numa_node = ctrl->ctrl.numa_node;
 527         ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
 528         ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
 529                 NVME_INLINE_SG_CNT * sizeof(struct scatterlist);
 530         ctrl->tag_set.driver_data = ctrl;
 531         ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
 532         ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
 533         ctrl->ctrl.tagset = &ctrl->tag_set;
 534
 535         ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
 536         if (ret)
 537                 goto out_destroy_queues;
 538
 539         ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 540         if (IS_ERR(ctrl->ctrl.connect_q)) {
 541                 ret = PTR_ERR(ctrl->ctrl.connect_q);
 542                 goto out_free_tagset;
 543         }
 544
 545         ret = nvme_loop_connect_io_queues(ctrl);
 546         if (ret)
 547                 goto out_cleanup_connect_q;
 548
 549         return 0;
 550
 551 out_cleanup_connect_q:
 552         blk_cleanup_queue(ctrl->ctrl.connect_q);
 553 out_free_tagset:
 554         blk_mq_free_tag_set(&ctrl->tag_set);
 555 out_destroy_queues:
 556         nvme_loop_destroy_io_queues(ctrl);
 557         return ret;
 558 }
 559
 560 static struct nvmet_port *nvme_loop_find_port(struct nvme_ctrl *ctrl)
 561 {
 562         struct nvmet_port *p, *found = NULL;
 563
 564         mutex_lock(&nvme_loop_ports_mutex);
 565         list_for_each_entry(p, &nvme_loop_ports, entry) {
 566                 /* if no transport address is specified use the first port */
 567                 if ((ctrl->opts->mask & NVMF_OPT_TRADDR) &&
 568                     strcmp(ctrl->opts->traddr, p->disc_addr.traddr))
 569                         continue;
 570                 found = p;
 571                 break;
 572         }
 573         mutex_unlock(&nvme_loop_ports_mutex);
 574         return found;
 575 }
 576
 577 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
 578                 struct nvmf_ctrl_options *opts)
 579 {
 580         struct nvme_loop_ctrl *ctrl;
 581         int ret;
 582
 583         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
 584         if (!ctrl)
 585                 return ERR_PTR(-ENOMEM);
 586         ctrl->ctrl.opts = opts;
 587         INIT_LIST_HEAD(&ctrl->list);
 588
 589         INIT_WORK(&ctrl->ctrl.reset_work, nvme_loop_reset_ctrl_work);
 590
 591         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
 592                                 0 /* no quirks, we're perfect! */);
 593         if (ret)
 594                 goto out;
 595
 596         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING))
 597                 WARN_ON_ONCE(1);
 598
 599         ret = -ENOMEM;
 600
 601         ctrl->ctrl.sqsize = opts->queue_size - 1;
 602         ctrl->ctrl.kato = opts->kato;
 603         ctrl->port = nvme_loop_find_port(&ctrl->ctrl);
 604
 605         ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
 606                         GFP_KERNEL);
 607         if (!ctrl->queues)
 608                 goto out_uninit_ctrl;
 609
 610         ret = nvme_loop_configure_admin_queue(ctrl);
 611         if (ret)
 612                 goto out_free_queues;
 613
 614         if (opts->queue_size > ctrl->ctrl.maxcmd) {
 615                 /* warn if maxcmd is lower than queue_size */
 616                 dev_warn(ctrl->ctrl.device,
 617                         "queue_size %zu > ctrl maxcmd %u, clamping down\n",
 618                         opts->queue_size, ctrl->ctrl.maxcmd);
 619                 opts->queue_size = ctrl->ctrl.maxcmd;
 620         }
 621
 622         if (opts->nr_io_queues) {
 623                 ret = nvme_loop_create_io_queues(ctrl);
 624                 if (ret)
 625                         goto out_remove_admin_queue;
 626         }
 627
 628         nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
 629
 630         dev_info(ctrl->ctrl.device,
 631                  "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
 632
 633         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE))
 634                 WARN_ON_ONCE(1);
 635
 636         mutex_lock(&nvme_loop_ctrl_mutex);
 637         list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
 638         mutex_unlock(&nvme_loop_ctrl_mutex);
 639
 640         nvme_start_ctrl(&ctrl->ctrl);
 641
 642         return &ctrl->ctrl;
 643
 644 out_remove_admin_queue:
 645         nvme_loop_destroy_admin_queue(ctrl);
 646 out_free_queues:
 647         kfree(ctrl->queues);
 648 out_uninit_ctrl:
 649         nvme_uninit_ctrl(&ctrl->ctrl);
 650         nvme_put_ctrl(&ctrl->ctrl);
 651 out:
 652         if (ret > 0)
 653                 ret = -EIO;
 654         return ERR_PTR(ret);
 655 }
 656
 657 static int nvme_loop_add_port(struct nvmet_port *port)
 658 {
 659         mutex_lock(&nvme_loop_ports_mutex);
 660         list_add_tail(&port->entry, &nvme_loop_ports);
 661         mutex_unlock(&nvme_loop_ports_mutex);
 662         return 0;
 663 }
 664
 665 static void nvme_loop_remove_port(struct nvmet_port *port)
 666 {
 667         mutex_lock(&nvme_loop_ports_mutex);
 668         list_del_init(&port->entry);
 669         mutex_unlock(&nvme_loop_ports_mutex);
 670
 671         /*
 672          * Ensure any ctrls that are in the process of being
 673          * deleted are in fact deleted before we return
 674          * and free the port. This is to prevent active
 675          * ctrls from using a port after it's freed.
 676          */
 677         flush_workqueue(nvme_delete_wq);
 678 }
 679
 680 static const struct nvmet_fabrics_ops nvme_loop_ops = {
 681         .owner          = THIS_MODULE,
 682         .type           = NVMF_TRTYPE_LOOP,
 683         .add_port       = nvme_loop_add_port,
 684         .remove_port    = nvme_loop_remove_port,
 685         .queue_response = nvme_loop_queue_response,
 686         .delete_ctrl    = nvme_loop_delete_ctrl,
 687 };
 688
 689 static struct nvmf_transport_ops nvme_loop_transport = {
 690         .name           = "loop",
 691         .module         = THIS_MODULE,
 692         .create_ctrl    = nvme_loop_create_ctrl,
 693         .allowed_opts   = NVMF_OPT_TRADDR,
 694 };
 695
 696 static int __init nvme_loop_init_module(void)
 697 {
 698         int ret;
 699
 700         ret = nvmet_register_transport(&nvme_loop_ops);
 701         if (ret)
 702                 return ret;
 703
 704         ret = nvmf_register_transport(&nvme_loop_transport);
 705         if (ret)
 706                 nvmet_unregister_transport(&nvme_loop_ops);
 707
 708         return ret;
 709 }
 710
 711 static void __exit nvme_loop_cleanup_module(void)
 712 {
 713         struct nvme_loop_ctrl *ctrl, *next;
 714
 715         nvmf_unregister_transport(&nvme_loop_transport);
 716         nvmet_unregister_transport(&nvme_loop_ops);
 717
 718         mutex_lock(&nvme_loop_ctrl_mutex);
 719         list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
 720                 nvme_delete_ctrl(&ctrl->ctrl);
 721         mutex_unlock(&nvme_loop_ctrl_mutex);
 722
 723         flush_workqueue(nvme_delete_wq);
 724 }
 725
 726 module_init(nvme_loop_init_module);
 727 module_exit(nvme_loop_cleanup_module);
 728
 729 MODULE_LICENSE("GPL v2");
 730 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */