drivers/nvme/host/tcp.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * NVMe over Fabrics TCP host.
   4  * Copyright (c) 2018 Lightbits Labs. All rights reserved.
   5  */
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7 #include <linux/module.h>
   8 #include <linux/init.h>
   9 #include <linux/slab.h>
  10 #include <linux/err.h>
  11 #include <linux/nvme-tcp.h>
  12 #include <net/sock.h>
  13 #include <net/tcp.h>
  14 #include <linux/blk-mq.h>
  15 #include <crypto/hash.h>
  16 #include <net/busy_poll.h>
  17
  18 #include "nvme.h"
  19 #include "fabrics.h"
  20
  21 struct nvme_tcp_queue;
  22
  23 enum nvme_tcp_send_state {
  24         NVME_TCP_SEND_CMD_PDU = 0,
  25         NVME_TCP_SEND_H2C_PDU,
  26         NVME_TCP_SEND_DATA,
  27         NVME_TCP_SEND_DDGST,
  28 };
  29
  30 struct nvme_tcp_request {
  31         struct nvme_request     req;
  32         void                    *pdu;
  33         struct nvme_tcp_queue   *queue;
  34         u32                     data_len;
  35         u32                     pdu_len;
  36         u32                     pdu_sent;
  37         u16                     ttag;
  38         struct list_head        entry;
  39         __le32                  ddgst;
  40
  41         struct bio              *curr_bio;
  42         struct iov_iter         iter;
  43
  44         /* send state */
  45         size_t                  offset;
  46         size_t                  data_sent;
  47         enum nvme_tcp_send_state state;
  48 };
  49
  50 enum nvme_tcp_queue_flags {
  51         NVME_TCP_Q_ALLOCATED    = 0,
  52         NVME_TCP_Q_LIVE         = 1,
  53 };
  54
  55 enum nvme_tcp_recv_state {
  56         NVME_TCP_RECV_PDU = 0,
  57         NVME_TCP_RECV_DATA,
  58         NVME_TCP_RECV_DDGST,
  59 };
  60
  61 struct nvme_tcp_ctrl;
  62 struct nvme_tcp_queue {
  63         struct socket           *sock;
  64         struct work_struct      io_work;
  65         int                     io_cpu;
  66
  67         spinlock_t              lock;
  68         struct list_head        send_list;
  69
  70         /* recv state */
  71         void                    *pdu;
  72         int                     pdu_remaining;
  73         int                     pdu_offset;
  74         size_t                  data_remaining;
  75         size_t                  ddgst_remaining;
  76         unsigned int            nr_cqe;
  77
  78         /* send state */
  79         struct nvme_tcp_request *request;
  80
  81         int                     queue_size;
  82         size_t                  cmnd_capsule_len;
  83         struct nvme_tcp_ctrl    *ctrl;
  84         unsigned long           flags;
  85         bool                    rd_enabled;
  86
  87         bool                    hdr_digest;
  88         bool                    data_digest;
  89         struct ahash_request    *rcv_hash;
  90         struct ahash_request    *snd_hash;
  91         __le32                  exp_ddgst;
  92         __le32                  recv_ddgst;
  93
  94         struct page_frag_cache  pf_cache;
  95
  96         void (*state_change)(struct sock *);
  97         void (*data_ready)(struct sock *);
  98         void (*write_space)(struct sock *);
  99 };
 100
 101 struct nvme_tcp_ctrl {
 102         /* read only in the hot path */
 103         struct nvme_tcp_queue   *queues;
 104         struct blk_mq_tag_set   tag_set;
 105
 106         /* other member variables */
 107         struct list_head        list;
 108         struct blk_mq_tag_set   admin_tag_set;
 109         struct sockaddr_storage addr;
 110         struct sockaddr_storage src_addr;
 111         struct nvme_ctrl        ctrl;
 112
 113         struct work_struct      err_work;
 114         struct delayed_work     connect_work;
 115         struct nvme_tcp_request async_req;
 116         u32                     io_queues[HCTX_MAX_TYPES];
 117 };
 118
 119 static LIST_HEAD(nvme_tcp_ctrl_list);
 120 static DEFINE_MUTEX(nvme_tcp_ctrl_mutex);
 121 static struct workqueue_struct *nvme_tcp_wq;
 122 static struct blk_mq_ops nvme_tcp_mq_ops;
 123 static struct blk_mq_ops nvme_tcp_admin_mq_ops;
 124
 125 static inline struct nvme_tcp_ctrl *to_tcp_ctrl(struct nvme_ctrl *ctrl)
 126 {
 127         return container_of(ctrl, struct nvme_tcp_ctrl, ctrl);
 128 }
 129
 130 static inline int nvme_tcp_queue_id(struct nvme_tcp_queue *queue)
 131 {
 132         return queue - queue->ctrl->queues;
 133 }
 134
 135 static inline struct blk_mq_tags *nvme_tcp_tagset(struct nvme_tcp_queue *queue)
 136 {
 137         u32 queue_idx = nvme_tcp_queue_id(queue);
 138
 139         if (queue_idx == 0)
 140                 return queue->ctrl->admin_tag_set.tags[queue_idx];
 141         return queue->ctrl->tag_set.tags[queue_idx - 1];
 142 }
 143
 144 static inline u8 nvme_tcp_hdgst_len(struct nvme_tcp_queue *queue)
 145 {
 146         return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 147 }
 148
 149 static inline u8 nvme_tcp_ddgst_len(struct nvme_tcp_queue *queue)
 150 {
 151         return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 152 }
 153
 154 static inline size_t nvme_tcp_inline_data_size(struct nvme_tcp_queue *queue)
 155 {
 156         return queue->cmnd_capsule_len - sizeof(struct nvme_command);
 157 }
 158
 159 static inline bool nvme_tcp_async_req(struct nvme_tcp_request *req)
 160 {
 161         return req == &req->queue->ctrl->async_req;
 162 }
 163
 164 static inline bool nvme_tcp_has_inline_data(struct nvme_tcp_request *req)
 165 {
 166         struct request *rq;
 167         unsigned int bytes;
 168
 169         if (unlikely(nvme_tcp_async_req(req)))
 170                 return false; /* async events don't have a request */
 171
 172         rq = blk_mq_rq_from_pdu(req);
 173         bytes = blk_rq_payload_bytes(rq);
 174
 175         return rq_data_dir(rq) == WRITE && bytes &&
 176                 bytes <= nvme_tcp_inline_data_size(req->queue);
 177 }
 178
 179 static inline struct page *nvme_tcp_req_cur_page(struct nvme_tcp_request *req)
 180 {
 181         return req->iter.bvec->bv_page;
 182 }
 183
 184 static inline size_t nvme_tcp_req_cur_offset(struct nvme_tcp_request *req)
 185 {
 186         return req->iter.bvec->bv_offset + req->iter.iov_offset;
 187 }
 188
 189 static inline size_t nvme_tcp_req_cur_length(struct nvme_tcp_request *req)
 190 {
 191         return min_t(size_t, req->iter.bvec->bv_len - req->iter.iov_offset,
 192                         req->pdu_len - req->pdu_sent);
 193 }
 194
 195 static inline size_t nvme_tcp_req_offset(struct nvme_tcp_request *req)
 196 {
 197         return req->iter.iov_offset;
 198 }
 199
 200 static inline size_t nvme_tcp_pdu_data_left(struct nvme_tcp_request *req)
 201 {
 202         return rq_data_dir(blk_mq_rq_from_pdu(req)) == WRITE ?
 203                         req->pdu_len - req->pdu_sent : 0;
 204 }
 205
 206 static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req,
 207                 int len)
 208 {
 209         return nvme_tcp_pdu_data_left(req) <= len;
 210 }
 211
 212 static void nvme_tcp_init_iter(struct nvme_tcp_request *req,
 213                 unsigned int dir)
 214 {
 215         struct request *rq = blk_mq_rq_from_pdu(req);
 216         struct bio_vec *vec;
 217         unsigned int size;
 218         int nsegs;
 219         size_t offset;
 220
 221         if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) {
 222                 vec = &rq->special_vec;
 223                 nsegs = 1;
 224                 size = blk_rq_payload_bytes(rq);
 225                 offset = 0;
 226         } else {
 227                 struct bio *bio = req->curr_bio;
 228
 229                 vec = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
 230                 nsegs = bio_segments(bio);
 231                 size = bio->bi_iter.bi_size;
 232                 offset = bio->bi_iter.bi_bvec_done;
 233         }
 234
 235         iov_iter_bvec(&req->iter, dir, vec, nsegs, size);
 236         req->iter.iov_offset = offset;
 237 }
 238
 239 static inline void nvme_tcp_advance_req(struct nvme_tcp_request *req,
 240                 int len)
 241 {
 242         req->data_sent += len;
 243         req->pdu_sent += len;
 244         iov_iter_advance(&req->iter, len);
 245         if (!iov_iter_count(&req->iter) &&
 246             req->data_sent < req->data_len) {
 247                 req->curr_bio = req->curr_bio->bi_next;
 248                 nvme_tcp_init_iter(req, WRITE);
 249         }
 250 }
 251
 252 static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req)
 253 {
 254         struct nvme_tcp_queue *queue = req->queue;
 255
 256         spin_lock(&queue->lock);
 257         list_add_tail(&req->entry, &queue->send_list);
 258         spin_unlock(&queue->lock);
 259
 260         queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 261 }
 262
 263 static inline struct nvme_tcp_request *
 264 nvme_tcp_fetch_request(struct nvme_tcp_queue *queue)
 265 {
 266         struct nvme_tcp_request *req;
 267
 268         spin_lock(&queue->lock);
 269         req = list_first_entry_or_null(&queue->send_list,
 270                         struct nvme_tcp_request, entry);
 271         if (req)
 272                 list_del(&req->entry);
 273         spin_unlock(&queue->lock);
 274
 275         return req;
 276 }
 277
 278 static inline void nvme_tcp_ddgst_final(struct ahash_request *hash,
 279                 __le32 *dgst)
 280 {
 281         ahash_request_set_crypt(hash, NULL, (u8 *)dgst, 0);
 282         crypto_ahash_final(hash);
 283 }
 284
 285 static inline void nvme_tcp_ddgst_update(struct ahash_request *hash,
 286                 struct page *page, off_t off, size_t len)
 287 {
 288         struct scatterlist sg;
 289
 290         sg_init_marker(&sg, 1);
 291         sg_set_page(&sg, page, len, off);
 292         ahash_request_set_crypt(hash, &sg, NULL, len);
 293         crypto_ahash_update(hash);
 294 }
 295
 296 static inline void nvme_tcp_hdgst(struct ahash_request *hash,
 297                 void *pdu, size_t len)
 298 {
 299         struct scatterlist sg;
 300
 301         sg_init_one(&sg, pdu, len);
 302         ahash_request_set_crypt(hash, &sg, pdu + len, len);
 303         crypto_ahash_digest(hash);
 304 }
 305
 306 static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue,
 307                 void *pdu, size_t pdu_len)
 308 {
 309         struct nvme_tcp_hdr *hdr = pdu;
 310         __le32 recv_digest;
 311         __le32 exp_digest;
 312
 313         if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
 314                 dev_err(queue->ctrl->ctrl.device,
 315                         "queue %d: header digest flag is cleared\n",
 316                         nvme_tcp_queue_id(queue));
 317                 return -EPROTO;
 318         }
 319
 320         recv_digest = *(__le32 *)(pdu + hdr->hlen);
 321         nvme_tcp_hdgst(queue->rcv_hash, pdu, pdu_len);
 322         exp_digest = *(__le32 *)(pdu + hdr->hlen);
 323         if (recv_digest != exp_digest) {
 324                 dev_err(queue->ctrl->ctrl.device,
 325                         "header digest error: recv %#x expected %#x\n",
 326                         le32_to_cpu(recv_digest), le32_to_cpu(exp_digest));
 327                 return -EIO;
 328         }
 329
 330         return 0;
 331 }
 332
 333 static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu)
 334 {
 335         struct nvme_tcp_hdr *hdr = pdu;
 336         u8 digest_len = nvme_tcp_hdgst_len(queue);
 337         u32 len;
 338
 339         len = le32_to_cpu(hdr->plen) - hdr->hlen -
 340                 ((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0);
 341
 342         if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
 343                 dev_err(queue->ctrl->ctrl.device,
 344                         "queue %d: data digest flag is cleared\n",
 345                 nvme_tcp_queue_id(queue));
 346                 return -EPROTO;
 347         }
 348         crypto_ahash_init(queue->rcv_hash);
 349
 350         return 0;
 351 }
 352
 353 static void nvme_tcp_exit_request(struct blk_mq_tag_set *set,
 354                 struct request *rq, unsigned int hctx_idx)
 355 {
 356         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 357
 358         page_frag_free(req->pdu);
 359 }
 360
 361 static int nvme_tcp_init_request(struct blk_mq_tag_set *set,
 362                 struct request *rq, unsigned int hctx_idx,
 363                 unsigned int numa_node)
 364 {
 365         struct nvme_tcp_ctrl *ctrl = set->driver_data;
 366         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 367         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 368         struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx];
 369         u8 hdgst = nvme_tcp_hdgst_len(queue);
 370
 371         req->pdu = page_frag_alloc(&queue->pf_cache,
 372                 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
 373                 GFP_KERNEL | __GFP_ZERO);
 374         if (!req->pdu)
 375                 return -ENOMEM;
 376
 377         req->queue = queue;
 378         nvme_req(rq)->ctrl = &ctrl->ctrl;
 379
 380         return 0;
 381 }
 382
 383 static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 384                 unsigned int hctx_idx)
 385 {
 386         struct nvme_tcp_ctrl *ctrl = data;
 387         struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1];
 388
 389         hctx->driver_data = queue;
 390         return 0;
 391 }
 392
 393 static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 394                 unsigned int hctx_idx)
 395 {
 396         struct nvme_tcp_ctrl *ctrl = data;
 397         struct nvme_tcp_queue *queue = &ctrl->queues[0];
 398
 399         hctx->driver_data = queue;
 400         return 0;
 401 }
 402
 403 static enum nvme_tcp_recv_state
 404 nvme_tcp_recv_state(struct nvme_tcp_queue *queue)
 405 {
 406         return  (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
 407                 (queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST :
 408                 NVME_TCP_RECV_DATA;
 409 }
 410
 411 static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue)
 412 {
 413         queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) +
 414                                 nvme_tcp_hdgst_len(queue);
 415         queue->pdu_offset = 0;
 416         queue->data_remaining = -1;
 417         queue->ddgst_remaining = 0;
 418 }
 419
 420 static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl)
 421 {
 422         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
 423                 return;
 424
 425         queue_work(nvme_reset_wq, &to_tcp_ctrl(ctrl)->err_work);
 426 }
 427
 428 static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
 429                 struct nvme_completion *cqe)
 430 {
 431         struct request *rq;
 432
 433         rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), cqe->command_id);
 434         if (!rq) {
 435                 dev_err(queue->ctrl->ctrl.device,
 436                         "queue %d tag 0x%x not found\n",
 437                         nvme_tcp_queue_id(queue), cqe->command_id);
 438                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 439                 return -EINVAL;
 440         }
 441
 442         nvme_end_request(rq, cqe->status, cqe->result);
 443         queue->nr_cqe++;
 444
 445         return 0;
 446 }
 447
 448 static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue,
 449                 struct nvme_tcp_data_pdu *pdu)
 450 {
 451         struct request *rq;
 452
 453         rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
 454         if (!rq) {
 455                 dev_err(queue->ctrl->ctrl.device,
 456                         "queue %d tag %#x not found\n",
 457                         nvme_tcp_queue_id(queue), pdu->command_id);
 458                 return -ENOENT;
 459         }
 460
 461         if (!blk_rq_payload_bytes(rq)) {
 462                 dev_err(queue->ctrl->ctrl.device,
 463                         "queue %d tag %#x unexpected data\n",
 464                         nvme_tcp_queue_id(queue), rq->tag);
 465                 return -EIO;
 466         }
 467
 468         queue->data_remaining = le32_to_cpu(pdu->data_length);
 469
 470         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS &&
 471             unlikely(!(pdu->hdr.flags & NVME_TCP_F_DATA_LAST))) {
 472                 dev_err(queue->ctrl->ctrl.device,
 473                         "queue %d tag %#x SUCCESS set but not last PDU\n",
 474                         nvme_tcp_queue_id(queue), rq->tag);
 475                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 476                 return -EPROTO;
 477         }
 478
 479         return 0;
 480 }
 481
 482 static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue,
 483                 struct nvme_tcp_rsp_pdu *pdu)
 484 {
 485         struct nvme_completion *cqe = &pdu->cqe;
 486         int ret = 0;
 487
 488         /*
 489          * AEN requests are special as they don't time out and can
 490          * survive any kind of queue freeze and often don't respond to
 491          * aborts.  We don't even bother to allocate a struct request
 492          * for them but rather special case them here.
 493          */
 494         if (unlikely(nvme_is_aen_req(nvme_tcp_queue_id(queue),
 495                                      cqe->command_id)))
 496                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
 497                                 &cqe->result);
 498         else
 499                 ret = nvme_tcp_process_nvme_cqe(queue, cqe);
 500
 501         return ret;
 502 }
 503
 504 static int nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req,
 505                 struct nvme_tcp_r2t_pdu *pdu)
 506 {
 507         struct nvme_tcp_data_pdu *data = req->pdu;
 508         struct nvme_tcp_queue *queue = req->queue;
 509         struct request *rq = blk_mq_rq_from_pdu(req);
 510         u8 hdgst = nvme_tcp_hdgst_len(queue);
 511         u8 ddgst = nvme_tcp_ddgst_len(queue);
 512
 513         req->pdu_len = le32_to_cpu(pdu->r2t_length);
 514         req->pdu_sent = 0;
 515
 516         if (unlikely(req->data_sent + req->pdu_len > req->data_len)) {
 517                 dev_err(queue->ctrl->ctrl.device,
 518                         "req %d r2t len %u exceeded data len %u (%zu sent)\n",
 519                         rq->tag, req->pdu_len, req->data_len,
 520                         req->data_sent);
 521                 return -EPROTO;
 522         }
 523
 524         if (unlikely(le32_to_cpu(pdu->r2t_offset) < req->data_sent)) {
 525                 dev_err(queue->ctrl->ctrl.device,
 526                         "req %d unexpected r2t offset %u (expected %zu)\n",
 527                         rq->tag, le32_to_cpu(pdu->r2t_offset),
 528                         req->data_sent);
 529                 return -EPROTO;
 530         }
 531
 532         memset(data, 0, sizeof(*data));
 533         data->hdr.type = nvme_tcp_h2c_data;
 534         data->hdr.flags = NVME_TCP_F_DATA_LAST;
 535         if (queue->hdr_digest)
 536                 data->hdr.flags |= NVME_TCP_F_HDGST;
 537         if (queue->data_digest)
 538                 data->hdr.flags |= NVME_TCP_F_DDGST;
 539         data->hdr.hlen = sizeof(*data);
 540         data->hdr.pdo = data->hdr.hlen + hdgst;
 541         data->hdr.plen =
 542                 cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst);
 543         data->ttag = pdu->ttag;
 544         data->command_id = rq->tag;
 545         data->data_offset = cpu_to_le32(req->data_sent);
 546         data->data_length = cpu_to_le32(req->pdu_len);
 547         return 0;
 548 }
 549
 550 static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue,
 551                 struct nvme_tcp_r2t_pdu *pdu)
 552 {
 553         struct nvme_tcp_request *req;
 554         struct request *rq;
 555         int ret;
 556
 557         rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
 558         if (!rq) {
 559                 dev_err(queue->ctrl->ctrl.device,
 560                         "queue %d tag %#x not found\n",
 561                         nvme_tcp_queue_id(queue), pdu->command_id);
 562                 return -ENOENT;
 563         }
 564         req = blk_mq_rq_to_pdu(rq);
 565
 566         ret = nvme_tcp_setup_h2c_data_pdu(req, pdu);
 567         if (unlikely(ret))
 568                 return ret;
 569
 570         req->state = NVME_TCP_SEND_H2C_PDU;
 571         req->offset = 0;
 572
 573         nvme_tcp_queue_request(req);
 574
 575         return 0;
 576 }
 577
 578 static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 579                 unsigned int *offset, size_t *len)
 580 {
 581         struct nvme_tcp_hdr *hdr;
 582         char *pdu = queue->pdu;
 583         size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
 584         int ret;
 585
 586         ret = skb_copy_bits(skb, *offset,
 587                 &pdu[queue->pdu_offset], rcv_len);
 588         if (unlikely(ret))
 589                 return ret;
 590
 591         queue->pdu_remaining -= rcv_len;
 592         queue->pdu_offset += rcv_len;
 593         *offset += rcv_len;
 594         *len -= rcv_len;
 595         if (queue->pdu_remaining)
 596                 return 0;
 597
 598         hdr = queue->pdu;
 599         if (queue->hdr_digest) {
 600                 ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen);
 601                 if (unlikely(ret))
 602                         return ret;
 603         }
 604
 605
 606         if (queue->data_digest) {
 607                 ret = nvme_tcp_check_ddgst(queue, queue->pdu);
 608                 if (unlikely(ret))
 609                         return ret;
 610         }
 611
 612         switch (hdr->type) {
 613         case nvme_tcp_c2h_data:
 614                 return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
 615         case nvme_tcp_rsp:
 616                 nvme_tcp_init_recv_ctx(queue);
 617                 return nvme_tcp_handle_comp(queue, (void *)queue->pdu);
 618         case nvme_tcp_r2t:
 619                 nvme_tcp_init_recv_ctx(queue);
 620                 return nvme_tcp_handle_r2t(queue, (void *)queue->pdu);
 621         default:
 622                 dev_err(queue->ctrl->ctrl.device,
 623                         "unsupported pdu type (%d)\n", hdr->type);
 624                 return -EINVAL;
 625         }
 626 }
 627
 628 static inline void nvme_tcp_end_request(struct request *rq, u16 status)
 629 {
 630         union nvme_result res = {};
 631
 632         nvme_end_request(rq, cpu_to_le16(status << 1), res);
 633 }
 634
 635 static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 636                               unsigned int *offset, size_t *len)
 637 {
 638         struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
 639         struct nvme_tcp_request *req;
 640         struct request *rq;
 641
 642         rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
 643         if (!rq) {
 644                 dev_err(queue->ctrl->ctrl.device,
 645                         "queue %d tag %#x not found\n",
 646                         nvme_tcp_queue_id(queue), pdu->command_id);
 647                 return -ENOENT;
 648         }
 649         req = blk_mq_rq_to_pdu(rq);
 650
 651         while (true) {
 652                 int recv_len, ret;
 653
 654                 recv_len = min_t(size_t, *len, queue->data_remaining);
 655                 if (!recv_len)
 656                         break;
 657
 658                 if (!iov_iter_count(&req->iter)) {
 659                         req->curr_bio = req->curr_bio->bi_next;
 660
 661                         /*
 662                          * If we don`t have any bios it means that controller
 663                          * sent more data than we requested, hence error
 664                          */
 665                         if (!req->curr_bio) {
 666                                 dev_err(queue->ctrl->ctrl.device,
 667                                         "queue %d no space in request %#x",
 668                                         nvme_tcp_queue_id(queue), rq->tag);
 669                                 nvme_tcp_init_recv_ctx(queue);
 670                                 return -EIO;
 671                         }
 672                         nvme_tcp_init_iter(req, READ);
 673                 }
 674
 675                 /* we can read only from what is left in this bio */
 676                 recv_len = min_t(size_t, recv_len,
 677                                 iov_iter_count(&req->iter));
 678
 679                 if (queue->data_digest)
 680                         ret = skb_copy_and_hash_datagram_iter(skb, *offset,
 681                                 &req->iter, recv_len, queue->rcv_hash);
 682                 else
 683                         ret = skb_copy_datagram_iter(skb, *offset,
 684                                         &req->iter, recv_len);
 685                 if (ret) {
 686                         dev_err(queue->ctrl->ctrl.device,
 687                                 "queue %d failed to copy request %#x data",
 688                                 nvme_tcp_queue_id(queue), rq->tag);
 689                         return ret;
 690                 }
 691
 692                 *len -= recv_len;
 693                 *offset += recv_len;
 694                 queue->data_remaining -= recv_len;
 695         }
 696
 697         if (!queue->data_remaining) {
 698                 if (queue->data_digest) {
 699                         nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst);
 700                         queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
 701                 } else {
 702                         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
 703                                 nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
 704                                 queue->nr_cqe++;
 705                         }
 706                         nvme_tcp_init_recv_ctx(queue);
 707                 }
 708         }
 709
 710         return 0;
 711 }
 712
 713 static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
 714                 struct sk_buff *skb, unsigned int *offset, size_t *len)
 715 {
 716         struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
 717         char *ddgst = (char *)&queue->recv_ddgst;
 718         size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
 719         off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
 720         int ret;
 721
 722         ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
 723         if (unlikely(ret))
 724                 return ret;
 725
 726         queue->ddgst_remaining -= recv_len;
 727         *offset += recv_len;
 728         *len -= recv_len;
 729         if (queue->ddgst_remaining)
 730                 return 0;
 731
 732         if (queue->recv_ddgst != queue->exp_ddgst) {
 733                 dev_err(queue->ctrl->ctrl.device,
 734                         "data digest error: recv %#x expected %#x\n",
 735                         le32_to_cpu(queue->recv_ddgst),
 736                         le32_to_cpu(queue->exp_ddgst));
 737                 return -EIO;
 738         }
 739
 740         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
 741                 struct request *rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue),
 742                                                 pdu->command_id);
 743
 744                 nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
 745                 queue->nr_cqe++;
 746         }
 747
 748         nvme_tcp_init_recv_ctx(queue);
 749         return 0;
 750 }
 751
 752 static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
 753                              unsigned int offset, size_t len)
 754 {
 755         struct nvme_tcp_queue *queue = desc->arg.data;
 756         size_t consumed = len;
 757         int result;
 758
 759         while (len) {
 760                 switch (nvme_tcp_recv_state(queue)) {
 761                 case NVME_TCP_RECV_PDU:
 762                         result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
 763                         break;
 764                 case NVME_TCP_RECV_DATA:
 765                         result = nvme_tcp_recv_data(queue, skb, &offset, &len);
 766                         break;
 767                 case NVME_TCP_RECV_DDGST:
 768                         result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
 769                         break;
 770                 default:
 771                         result = -EFAULT;
 772                 }
 773                 if (result) {
 774                         dev_err(queue->ctrl->ctrl.device,
 775                                 "receive failed:  %d\n", result);
 776                         queue->rd_enabled = false;
 777                         nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 778                         return result;
 779                 }
 780         }
 781
 782         return consumed;
 783 }
 784
 785 static void nvme_tcp_data_ready(struct sock *sk)
 786 {
 787         struct nvme_tcp_queue *queue;
 788
 789         read_lock(&sk->sk_callback_lock);
 790         queue = sk->sk_user_data;
 791         if (likely(queue && queue->rd_enabled))
 792                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 793         read_unlock(&sk->sk_callback_lock);
 794 }
 795
 796 static void nvme_tcp_write_space(struct sock *sk)
 797 {
 798         struct nvme_tcp_queue *queue;
 799
 800         read_lock_bh(&sk->sk_callback_lock);
 801         queue = sk->sk_user_data;
 802         if (likely(queue && sk_stream_is_writeable(sk))) {
 803                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 804                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 805         }
 806         read_unlock_bh(&sk->sk_callback_lock);
 807 }
 808
 809 static void nvme_tcp_state_change(struct sock *sk)
 810 {
 811         struct nvme_tcp_queue *queue;
 812
 813         read_lock(&sk->sk_callback_lock);
 814         queue = sk->sk_user_data;
 815         if (!queue)
 816                 goto done;
 817
 818         switch (sk->sk_state) {
 819         case TCP_CLOSE:
 820         case TCP_CLOSE_WAIT:
 821         case TCP_LAST_ACK:
 822         case TCP_FIN_WAIT1:
 823         case TCP_FIN_WAIT2:
 824                 /* fallthrough */
 825                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 826                 break;
 827         default:
 828                 dev_info(queue->ctrl->ctrl.device,
 829                         "queue %d socket state %d\n",
 830                         nvme_tcp_queue_id(queue), sk->sk_state);
 831         }
 832
 833         queue->state_change(sk);
 834 done:
 835         read_unlock(&sk->sk_callback_lock);
 836 }
 837
 838 static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
 839 {
 840         queue->request = NULL;
 841 }
 842
 843 static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
 844 {
 845         nvme_tcp_end_request(blk_mq_rq_from_pdu(req), NVME_SC_HOST_PATH_ERROR);
 846 }
 847
 848 static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
 849 {
 850         struct nvme_tcp_queue *queue = req->queue;
 851
 852         while (true) {
 853                 struct page *page = nvme_tcp_req_cur_page(req);
 854                 size_t offset = nvme_tcp_req_cur_offset(req);
 855                 size_t len = nvme_tcp_req_cur_length(req);
 856                 bool last = nvme_tcp_pdu_last_send(req, len);
 857                 int ret, flags = MSG_DONTWAIT;
 858
 859                 if (last && !queue->data_digest)
 860                         flags |= MSG_EOR;
 861                 else
 862                         flags |= MSG_MORE;
 863
 864                 /* can't zcopy slab pages */
 865                 if (unlikely(PageSlab(page))) {
 866                         ret = sock_no_sendpage(queue->sock, page, offset, len,
 867                                         flags);
 868                 } else {
 869                         ret = kernel_sendpage(queue->sock, page, offset, len,
 870                                         flags);
 871                 }
 872                 if (ret <= 0)
 873                         return ret;
 874
 875                 nvme_tcp_advance_req(req, ret);
 876                 if (queue->data_digest)
 877                         nvme_tcp_ddgst_update(queue->snd_hash, page,
 878                                         offset, ret);
 879
 880                 /* fully successful last write*/
 881                 if (last && ret == len) {
 882                         if (queue->data_digest) {
 883                                 nvme_tcp_ddgst_final(queue->snd_hash,
 884                                         &req->ddgst);
 885                                 req->state = NVME_TCP_SEND_DDGST;
 886                                 req->offset = 0;
 887                         } else {
 888                                 nvme_tcp_done_send_req(queue);
 889                         }
 890                         return 1;
 891                 }
 892         }
 893         return -EAGAIN;
 894 }
 895
 896 static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
 897 {
 898         struct nvme_tcp_queue *queue = req->queue;
 899         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
 900         bool inline_data = nvme_tcp_has_inline_data(req);
 901         int flags = MSG_DONTWAIT | (inline_data ? MSG_MORE : MSG_EOR);
 902         u8 hdgst = nvme_tcp_hdgst_len(queue);
 903         int len = sizeof(*pdu) + hdgst - req->offset;
 904         int ret;
 905
 906         if (queue->hdr_digest && !req->offset)
 907                 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
 908
 909         ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
 910                         offset_in_page(pdu) + req->offset, len,  flags);
 911         if (unlikely(ret <= 0))
 912                 return ret;
 913
 914         len -= ret;
 915         if (!len) {
 916                 if (inline_data) {
 917                         req->state = NVME_TCP_SEND_DATA;
 918                         if (queue->data_digest)
 919                                 crypto_ahash_init(queue->snd_hash);
 920                         nvme_tcp_init_iter(req, WRITE);
 921                 } else {
 922                         nvme_tcp_done_send_req(queue);
 923                 }
 924                 return 1;
 925         }
 926         req->offset += ret;
 927
 928         return -EAGAIN;
 929 }
 930
 931 static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
 932 {
 933         struct nvme_tcp_queue *queue = req->queue;
 934         struct nvme_tcp_data_pdu *pdu = req->pdu;
 935         u8 hdgst = nvme_tcp_hdgst_len(queue);
 936         int len = sizeof(*pdu) - req->offset + hdgst;
 937         int ret;
 938
 939         if (queue->hdr_digest && !req->offset)
 940                 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
 941
 942         ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
 943                         offset_in_page(pdu) + req->offset, len,
 944                         MSG_DONTWAIT | MSG_MORE);
 945         if (unlikely(ret <= 0))
 946                 return ret;
 947
 948         len -= ret;
 949         if (!len) {
 950                 req->state = NVME_TCP_SEND_DATA;
 951                 if (queue->data_digest)
 952                         crypto_ahash_init(queue->snd_hash);
 953                 if (!req->data_sent)
 954                         nvme_tcp_init_iter(req, WRITE);
 955                 return 1;
 956         }
 957         req->offset += ret;
 958
 959         return -EAGAIN;
 960 }
 961
 962 static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
 963 {
 964         struct nvme_tcp_queue *queue = req->queue;
 965         int ret;
 966         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR };
 967         struct kvec iov = {
 968                 .iov_base = &req->ddgst + req->offset,
 969                 .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
 970         };
 971
 972         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
 973         if (unlikely(ret <= 0))
 974                 return ret;
 975
 976         if (req->offset + ret == NVME_TCP_DIGEST_LENGTH) {
 977                 nvme_tcp_done_send_req(queue);
 978                 return 1;
 979         }
 980
 981         req->offset += ret;
 982         return -EAGAIN;
 983 }
 984
 985 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
 986 {
 987         struct nvme_tcp_request *req;
 988         int ret = 1;
 989
 990         if (!queue->request) {
 991                 queue->request = nvme_tcp_fetch_request(queue);
 992                 if (!queue->request)
 993                         return 0;
 994         }
 995         req = queue->request;
 996
 997         if (req->state == NVME_TCP_SEND_CMD_PDU) {
 998                 ret = nvme_tcp_try_send_cmd_pdu(req);
 999                 if (ret <= 0)
1000                         goto done;
1001                 if (!nvme_tcp_has_inline_data(req))
1002                         return ret;
1003         }
1004
1005         if (req->state == NVME_TCP_SEND_H2C_PDU) {
1006                 ret = nvme_tcp_try_send_data_pdu(req);
1007                 if (ret <= 0)
1008                         goto done;
1009         }
1010
1011         if (req->state == NVME_TCP_SEND_DATA) {
1012                 ret = nvme_tcp_try_send_data(req);
1013                 if (ret <= 0)
1014                         goto done;
1015         }
1016
1017         if (req->state == NVME_TCP_SEND_DDGST)
1018                 ret = nvme_tcp_try_send_ddgst(req);
1019 done:
1020         if (ret == -EAGAIN)
1021                 ret = 0;
1022         return ret;
1023 }
1024
1025 static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1026 {
1027         struct socket *sock = queue->sock;
1028         struct sock *sk = sock->sk;
1029         read_descriptor_t rd_desc;
1030         int consumed;
1031
1032         rd_desc.arg.data = queue;
1033         rd_desc.count = 1;
1034         lock_sock(sk);
1035         queue->nr_cqe = 0;
1036         consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1037         release_sock(sk);
1038         return consumed;
1039 }
1040
1041 static void nvme_tcp_io_work(struct work_struct *w)
1042 {
1043         struct nvme_tcp_queue *queue =
1044                 container_of(w, struct nvme_tcp_queue, io_work);
1045         unsigned long deadline = jiffies + msecs_to_jiffies(1);
1046
1047         do {
1048                 bool pending = false;
1049                 int result;
1050
1051                 result = nvme_tcp_try_send(queue);
1052                 if (result > 0) {
1053                         pending = true;
1054                 } else if (unlikely(result < 0)) {
1055                         dev_err(queue->ctrl->ctrl.device,
1056                                 "failed to send request %d\n", result);
1057
1058                         /*
1059                          * Fail the request unless peer closed the connection,
1060                          * in which case error recovery flow will complete all.
1061                          */
1062                         if ((result != -EPIPE) && (result != -ECONNRESET))
1063                                 nvme_tcp_fail_request(queue->request);
1064                         nvme_tcp_done_send_req(queue);
1065                         return;
1066                 }
1067
1068                 result = nvme_tcp_try_recv(queue);
1069                 if (result > 0)
1070                         pending = true;
1071
1072                 if (!pending)
1073                         return;
1074
1075         } while (!time_after(jiffies, deadline)); /* quota is exhausted */
1076
1077         queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1078 }
1079
1080 static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue)
1081 {
1082         struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
1083
1084         ahash_request_free(queue->rcv_hash);
1085         ahash_request_free(queue->snd_hash);
1086         crypto_free_ahash(tfm);
1087 }
1088
1089 static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue)
1090 {
1091         struct crypto_ahash *tfm;
1092
1093         tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
1094         if (IS_ERR(tfm))
1095                 return PTR_ERR(tfm);
1096
1097         queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1098         if (!queue->snd_hash)
1099                 goto free_tfm;
1100         ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
1101
1102         queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1103         if (!queue->rcv_hash)
1104                 goto free_snd_hash;
1105         ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
1106
1107         return 0;
1108 free_snd_hash:
1109         ahash_request_free(queue->snd_hash);
1110 free_tfm:
1111         crypto_free_ahash(tfm);
1112         return -ENOMEM;
1113 }
1114
1115 static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1116 {
1117         struct nvme_tcp_request *async = &ctrl->async_req;
1118
1119         page_frag_free(async->pdu);
1120 }
1121
1122 static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1123 {
1124         struct nvme_tcp_queue *queue = &ctrl->queues[0];
1125         struct nvme_tcp_request *async = &ctrl->async_req;
1126         u8 hdgst = nvme_tcp_hdgst_len(queue);
1127
1128         async->pdu = page_frag_alloc(&queue->pf_cache,
1129                 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1130                 GFP_KERNEL | __GFP_ZERO);
1131         if (!async->pdu)
1132                 return -ENOMEM;
1133
1134         async->queue = &ctrl->queues[0];
1135         return 0;
1136 }
1137
1138 static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1139 {
1140         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1141         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1142
1143         if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1144                 return;
1145
1146         if (queue->hdr_digest || queue->data_digest)
1147                 nvme_tcp_free_crypto(queue);
1148
1149         sock_release(queue->sock);
1150         kfree(queue->pdu);
1151 }
1152
1153 static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1154 {
1155         struct nvme_tcp_icreq_pdu *icreq;
1156         struct nvme_tcp_icresp_pdu *icresp;
1157         struct msghdr msg = {};
1158         struct kvec iov;
1159         bool ctrl_hdgst, ctrl_ddgst;
1160         int ret;
1161
1162         icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1163         if (!icreq)
1164                 return -ENOMEM;
1165
1166         icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1167         if (!icresp) {
1168                 ret = -ENOMEM;
1169                 goto free_icreq;
1170         }
1171
1172         icreq->hdr.type = nvme_tcp_icreq;
1173         icreq->hdr.hlen = sizeof(*icreq);
1174         icreq->hdr.pdo = 0;
1175         icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1176         icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1177         icreq->maxr2t = 0; /* single inflight r2t supported */
1178         icreq->hpda = 0; /* no alignment constraint */
1179         if (queue->hdr_digest)
1180                 icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1181         if (queue->data_digest)
1182                 icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1183
1184         iov.iov_base = icreq;
1185         iov.iov_len = sizeof(*icreq);
1186         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1187         if (ret < 0)
1188                 goto free_icresp;
1189
1190         memset(&msg, 0, sizeof(msg));
1191         iov.iov_base = icresp;
1192         iov.iov_len = sizeof(*icresp);
1193         ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1194                         iov.iov_len, msg.msg_flags);
1195         if (ret < 0)
1196                 goto free_icresp;
1197
1198         ret = -EINVAL;
1199         if (icresp->hdr.type != nvme_tcp_icresp) {
1200                 pr_err("queue %d: bad type returned %d\n",
1201                         nvme_tcp_queue_id(queue), icresp->hdr.type);
1202                 goto free_icresp;
1203         }
1204
1205         if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1206                 pr_err("queue %d: bad pdu length returned %d\n",
1207                         nvme_tcp_queue_id(queue), icresp->hdr.plen);
1208                 goto free_icresp;
1209         }
1210
1211         if (icresp->pfv != NVME_TCP_PFV_1_0) {
1212                 pr_err("queue %d: bad pfv returned %d\n",
1213                         nvme_tcp_queue_id(queue), icresp->pfv);
1214                 goto free_icresp;
1215         }
1216
1217         ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1218         if ((queue->data_digest && !ctrl_ddgst) ||
1219             (!queue->data_digest && ctrl_ddgst)) {
1220                 pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1221                         nvme_tcp_queue_id(queue),
1222                         queue->data_digest ? "enabled" : "disabled",
1223                         ctrl_ddgst ? "enabled" : "disabled");
1224                 goto free_icresp;
1225         }
1226
1227         ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1228         if ((queue->hdr_digest && !ctrl_hdgst) ||
1229             (!queue->hdr_digest && ctrl_hdgst)) {
1230                 pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1231                         nvme_tcp_queue_id(queue),
1232                         queue->hdr_digest ? "enabled" : "disabled",
1233                         ctrl_hdgst ? "enabled" : "disabled");
1234                 goto free_icresp;
1235         }
1236
1237         if (icresp->cpda != 0) {
1238                 pr_err("queue %d: unsupported cpda returned %d\n",
1239                         nvme_tcp_queue_id(queue), icresp->cpda);
1240                 goto free_icresp;
1241         }
1242
1243         ret = 0;
1244 free_icresp:
1245         kfree(icresp);
1246 free_icreq:
1247         kfree(icreq);
1248         return ret;
1249 }
1250
1251 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl,
1252                 int qid, size_t queue_size)
1253 {
1254         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1255         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1256         struct linger sol = { .l_onoff = 1, .l_linger = 0 };
1257         int ret, opt, rcv_pdu_size, n;
1258
1259         queue->ctrl = ctrl;
1260         INIT_LIST_HEAD(&queue->send_list);
1261         spin_lock_init(&queue->lock);
1262         INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1263         queue->queue_size = queue_size;
1264
1265         if (qid > 0)
1266                 queue->cmnd_capsule_len = nctrl->ioccsz * 16;
1267         else
1268                 queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1269                                                 NVME_TCP_ADMIN_CCSZ;
1270
1271         ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM,
1272                         IPPROTO_TCP, &queue->sock);
1273         if (ret) {
1274                 dev_err(nctrl->device,
1275                         "failed to create socket: %d\n", ret);
1276                 return ret;
1277         }
1278
1279         /* Single syn retry */
1280         opt = 1;
1281         ret = kernel_setsockopt(queue->sock, IPPROTO_TCP, TCP_SYNCNT,
1282                         (char *)&opt, sizeof(opt));
1283         if (ret) {
1284                 dev_err(nctrl->device,
1285                         "failed to set TCP_SYNCNT sock opt %d\n", ret);
1286                 goto err_sock;
1287         }
1288
1289         /* Set TCP no delay */
1290         opt = 1;
1291         ret = kernel_setsockopt(queue->sock, IPPROTO_TCP,
1292                         TCP_NODELAY, (char *)&opt, sizeof(opt));
1293         if (ret) {
1294                 dev_err(nctrl->device,
1295                         "failed to set TCP_NODELAY sock opt %d\n", ret);
1296                 goto err_sock;
1297         }
1298
1299         /*
1300          * Cleanup whatever is sitting in the TCP transmit queue on socket
1301          * close. This is done to prevent stale data from being sent should
1302          * the network connection be restored before TCP times out.
1303          */
1304         ret = kernel_setsockopt(queue->sock, SOL_SOCKET, SO_LINGER,
1305                         (char *)&sol, sizeof(sol));
1306         if (ret) {
1307                 dev_err(nctrl->device,
1308                         "failed to set SO_LINGER sock opt %d\n", ret);
1309                 goto err_sock;
1310         }
1311
1312         /* Set socket type of service */
1313         if (nctrl->opts->tos >= 0) {
1314                 opt = nctrl->opts->tos;
1315                 ret = kernel_setsockopt(queue->sock, SOL_IP, IP_TOS,
1316                                 (char *)&opt, sizeof(opt));
1317                 if (ret) {
1318                         dev_err(nctrl->device,
1319                                 "failed to set IP_TOS sock opt %d\n", ret);
1320                         goto err_sock;
1321                 }
1322         }
1323
1324         queue->sock->sk->sk_allocation = GFP_ATOMIC;
1325         if (!qid)
1326                 n = 0;
1327         else
1328                 n = (qid - 1) % num_online_cpus();
1329         queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
1330         queue->request = NULL;
1331         queue->data_remaining = 0;
1332         queue->ddgst_remaining = 0;
1333         queue->pdu_remaining = 0;
1334         queue->pdu_offset = 0;
1335         sk_set_memalloc(queue->sock->sk);
1336
1337         if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
1338                 ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1339                         sizeof(ctrl->src_addr));
1340                 if (ret) {
1341                         dev_err(nctrl->device,
1342                                 "failed to bind queue %d socket %d\n",
1343                                 qid, ret);
1344                         goto err_sock;
1345                 }
1346         }
1347
1348         queue->hdr_digest = nctrl->opts->hdr_digest;
1349         queue->data_digest = nctrl->opts->data_digest;
1350         if (queue->hdr_digest || queue->data_digest) {
1351                 ret = nvme_tcp_alloc_crypto(queue);
1352                 if (ret) {
1353                         dev_err(nctrl->device,
1354                                 "failed to allocate queue %d crypto\n", qid);
1355                         goto err_sock;
1356                 }
1357         }
1358
1359         rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1360                         nvme_tcp_hdgst_len(queue);
1361         queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1362         if (!queue->pdu) {
1363                 ret = -ENOMEM;
1364                 goto err_crypto;
1365         }
1366
1367         dev_dbg(nctrl->device, "connecting queue %d\n",
1368                         nvme_tcp_queue_id(queue));
1369
1370         ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1371                 sizeof(ctrl->addr), 0);
1372         if (ret) {
1373                 dev_err(nctrl->device,
1374                         "failed to connect socket: %d\n", ret);
1375                 goto err_rcv_pdu;
1376         }
1377
1378         ret = nvme_tcp_init_connection(queue);
1379         if (ret)
1380                 goto err_init_connect;
1381
1382         queue->rd_enabled = true;
1383         set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1384         nvme_tcp_init_recv_ctx(queue);
1385
1386         write_lock_bh(&queue->sock->sk->sk_callback_lock);
1387         queue->sock->sk->sk_user_data = queue;
1388         queue->state_change = queue->sock->sk->sk_state_change;
1389         queue->data_ready = queue->sock->sk->sk_data_ready;
1390         queue->write_space = queue->sock->sk->sk_write_space;
1391         queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1392         queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1393         queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1394 #ifdef CONFIG_NET_RX_BUSY_POLL
1395         queue->sock->sk->sk_ll_usec = 1;
1396 #endif
1397         write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1398
1399         return 0;
1400
1401 err_init_connect:
1402         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1403 err_rcv_pdu:
1404         kfree(queue->pdu);
1405 err_crypto:
1406         if (queue->hdr_digest || queue->data_digest)
1407                 nvme_tcp_free_crypto(queue);
1408 err_sock:
1409         sock_release(queue->sock);
1410         queue->sock = NULL;
1411         return ret;
1412 }
1413
1414 static void nvme_tcp_restore_sock_calls(struct nvme_tcp_queue *queue)
1415 {
1416         struct socket *sock = queue->sock;
1417
1418         write_lock_bh(&sock->sk->sk_callback_lock);
1419         sock->sk->sk_user_data  = NULL;
1420         sock->sk->sk_data_ready = queue->data_ready;
1421         sock->sk->sk_state_change = queue->state_change;
1422         sock->sk->sk_write_space  = queue->write_space;
1423         write_unlock_bh(&sock->sk->sk_callback_lock);
1424 }
1425
1426 static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1427 {
1428         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1429         nvme_tcp_restore_sock_calls(queue);
1430         cancel_work_sync(&queue->io_work);
1431 }
1432
1433 static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1434 {
1435         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1436         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1437
1438         if (!test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1439                 return;
1440
1441         __nvme_tcp_stop_queue(queue);
1442 }
1443
1444 static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1445 {
1446         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1447         int ret;
1448
1449         if (idx)
1450                 ret = nvmf_connect_io_queue(nctrl, idx, false);
1451         else
1452                 ret = nvmf_connect_admin_queue(nctrl);
1453
1454         if (!ret) {
1455                 set_bit(NVME_TCP_Q_LIVE, &ctrl->queues[idx].flags);
1456         } else {
1457                 if (test_bit(NVME_TCP_Q_ALLOCATED, &ctrl->queues[idx].flags))
1458                         __nvme_tcp_stop_queue(&ctrl->queues[idx]);
1459                 dev_err(nctrl->device,
1460                         "failed to connect queue: %d ret=%d\n", idx, ret);
1461         }
1462         return ret;
1463 }
1464
1465 static struct blk_mq_tag_set *nvme_tcp_alloc_tagset(struct nvme_ctrl *nctrl,
1466                 bool admin)
1467 {
1468         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1469         struct blk_mq_tag_set *set;
1470         int ret;
1471
1472         if (admin) {
1473                 set = &ctrl->admin_tag_set;
1474                 memset(set, 0, sizeof(*set));
1475                 set->ops = &nvme_tcp_admin_mq_ops;
1476                 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1477                 set->reserved_tags = 2; /* connect + keep-alive */
1478                 set->numa_node = NUMA_NO_NODE;
1479                 set->cmd_size = sizeof(struct nvme_tcp_request);
1480                 set->driver_data = ctrl;
1481                 set->nr_hw_queues = 1;
1482                 set->timeout = ADMIN_TIMEOUT;
1483         } else {
1484                 set = &ctrl->tag_set;
1485                 memset(set, 0, sizeof(*set));
1486                 set->ops = &nvme_tcp_mq_ops;
1487                 set->queue_depth = nctrl->sqsize + 1;
1488                 set->reserved_tags = 1; /* fabric connect */
1489                 set->numa_node = NUMA_NO_NODE;
1490                 set->flags = BLK_MQ_F_SHOULD_MERGE;
1491                 set->cmd_size = sizeof(struct nvme_tcp_request);
1492                 set->driver_data = ctrl;
1493                 set->nr_hw_queues = nctrl->queue_count - 1;
1494                 set->timeout = NVME_IO_TIMEOUT;
1495                 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
1496         }
1497
1498         ret = blk_mq_alloc_tag_set(set);
1499         if (ret)
1500                 return ERR_PTR(ret);
1501
1502         return set;
1503 }
1504
1505 static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
1506 {
1507         if (to_tcp_ctrl(ctrl)->async_req.pdu) {
1508                 nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
1509                 to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
1510         }
1511
1512         nvme_tcp_free_queue(ctrl, 0);
1513 }
1514
1515 static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
1516 {
1517         int i;
1518
1519         for (i = 1; i < ctrl->queue_count; i++)
1520                 nvme_tcp_free_queue(ctrl, i);
1521 }
1522
1523 static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
1524 {
1525         int i;
1526
1527         for (i = 1; i < ctrl->queue_count; i++)
1528                 nvme_tcp_stop_queue(ctrl, i);
1529 }
1530
1531 static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl)
1532 {
1533         int i, ret = 0;
1534
1535         for (i = 1; i < ctrl->queue_count; i++) {
1536                 ret = nvme_tcp_start_queue(ctrl, i);
1537                 if (ret)
1538                         goto out_stop_queues;
1539         }
1540
1541         return 0;
1542
1543 out_stop_queues:
1544         for (i--; i >= 1; i--)
1545                 nvme_tcp_stop_queue(ctrl, i);
1546         return ret;
1547 }
1548
1549 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
1550 {
1551         int ret;
1552
1553         ret = nvme_tcp_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
1554         if (ret)
1555                 return ret;
1556
1557         ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
1558         if (ret)
1559                 goto out_free_queue;
1560
1561         return 0;
1562
1563 out_free_queue:
1564         nvme_tcp_free_queue(ctrl, 0);
1565         return ret;
1566 }
1567
1568 static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1569 {
1570         int i, ret;
1571
1572         for (i = 1; i < ctrl->queue_count; i++) {
1573                 ret = nvme_tcp_alloc_queue(ctrl, i,
1574                                 ctrl->sqsize + 1);
1575                 if (ret)
1576                         goto out_free_queues;
1577         }
1578
1579         return 0;
1580
1581 out_free_queues:
1582         for (i--; i >= 1; i--)
1583                 nvme_tcp_free_queue(ctrl, i);
1584
1585         return ret;
1586 }
1587
1588 static unsigned int nvme_tcp_nr_io_queues(struct nvme_ctrl *ctrl)
1589 {
1590         unsigned int nr_io_queues;
1591
1592         nr_io_queues = min(ctrl->opts->nr_io_queues, num_online_cpus());
1593         nr_io_queues += min(ctrl->opts->nr_write_queues, num_online_cpus());
1594         nr_io_queues += min(ctrl->opts->nr_poll_queues, num_online_cpus());
1595
1596         return nr_io_queues;
1597 }
1598
1599 static void nvme_tcp_set_io_queues(struct nvme_ctrl *nctrl,
1600                 unsigned int nr_io_queues)
1601 {
1602         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1603         struct nvmf_ctrl_options *opts = nctrl->opts;
1604
1605         if (opts->nr_write_queues && opts->nr_io_queues < nr_io_queues) {
1606                 /*
1607                  * separate read/write queues
1608                  * hand out dedicated default queues only after we have
1609                  * sufficient read queues.
1610                  */
1611                 ctrl->io_queues[HCTX_TYPE_READ] = opts->nr_io_queues;
1612                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
1613                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1614                         min(opts->nr_write_queues, nr_io_queues);
1615                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1616         } else {
1617                 /*
1618                  * shared read/write queues
1619                  * either no write queues were requested, or we don't have
1620                  * sufficient queue count to have dedicated default queues.
1621                  */
1622                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1623                         min(opts->nr_io_queues, nr_io_queues);
1624                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1625         }
1626
1627         if (opts->nr_poll_queues && nr_io_queues) {
1628                 /* map dedicated poll queues only if we have queues left */
1629                 ctrl->io_queues[HCTX_TYPE_POLL] =
1630                         min(opts->nr_poll_queues, nr_io_queues);
1631         }
1632 }
1633
1634 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1635 {
1636         unsigned int nr_io_queues;
1637         int ret;
1638
1639         nr_io_queues = nvme_tcp_nr_io_queues(ctrl);
1640         ret = nvme_set_queue_count(ctrl, &nr_io_queues);
1641         if (ret)
1642                 return ret;
1643
1644         ctrl->queue_count = nr_io_queues + 1;
1645         if (ctrl->queue_count < 2)
1646                 return 0;
1647
1648         dev_info(ctrl->device,
1649                 "creating %d I/O queues.\n", nr_io_queues);
1650
1651         nvme_tcp_set_io_queues(ctrl, nr_io_queues);
1652
1653         return __nvme_tcp_alloc_io_queues(ctrl);
1654 }
1655
1656 static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove)
1657 {
1658         nvme_tcp_stop_io_queues(ctrl);
1659         if (remove) {
1660                 blk_cleanup_queue(ctrl->connect_q);
1661                 blk_mq_free_tag_set(ctrl->tagset);
1662         }
1663         nvme_tcp_free_io_queues(ctrl);
1664 }
1665
1666 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
1667 {
1668         int ret;
1669
1670         ret = nvme_tcp_alloc_io_queues(ctrl);
1671         if (ret)
1672                 return ret;
1673
1674         if (new) {
1675                 ctrl->tagset = nvme_tcp_alloc_tagset(ctrl, false);
1676                 if (IS_ERR(ctrl->tagset)) {
1677                         ret = PTR_ERR(ctrl->tagset);
1678                         goto out_free_io_queues;
1679                 }
1680
1681                 ctrl->connect_q = blk_mq_init_queue(ctrl->tagset);
1682                 if (IS_ERR(ctrl->connect_q)) {
1683                         ret = PTR_ERR(ctrl->connect_q);
1684                         goto out_free_tag_set;
1685                 }
1686         } else {
1687                 blk_mq_update_nr_hw_queues(ctrl->tagset,
1688                         ctrl->queue_count - 1);
1689         }
1690
1691         ret = nvme_tcp_start_io_queues(ctrl);
1692         if (ret)
1693                 goto out_cleanup_connect_q;
1694
1695         return 0;
1696
1697 out_cleanup_connect_q:
1698         if (new)
1699                 blk_cleanup_queue(ctrl->connect_q);
1700 out_free_tag_set:
1701         if (new)
1702                 blk_mq_free_tag_set(ctrl->tagset);
1703 out_free_io_queues:
1704         nvme_tcp_free_io_queues(ctrl);
1705         return ret;
1706 }
1707
1708 static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove)
1709 {
1710         nvme_tcp_stop_queue(ctrl, 0);
1711         if (remove) {
1712                 blk_cleanup_queue(ctrl->admin_q);
1713                 blk_cleanup_queue(ctrl->fabrics_q);
1714                 blk_mq_free_tag_set(ctrl->admin_tagset);
1715         }
1716         nvme_tcp_free_admin_queue(ctrl);
1717 }
1718
1719 static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
1720 {
1721         int error;
1722
1723         error = nvme_tcp_alloc_admin_queue(ctrl);
1724         if (error)
1725                 return error;
1726
1727         if (new) {
1728                 ctrl->admin_tagset = nvme_tcp_alloc_tagset(ctrl, true);
1729                 if (IS_ERR(ctrl->admin_tagset)) {
1730                         error = PTR_ERR(ctrl->admin_tagset);
1731                         goto out_free_queue;
1732                 }
1733
1734                 ctrl->fabrics_q = blk_mq_init_queue(ctrl->admin_tagset);
1735                 if (IS_ERR(ctrl->fabrics_q)) {
1736                         error = PTR_ERR(ctrl->fabrics_q);
1737                         goto out_free_tagset;
1738                 }
1739
1740                 ctrl->admin_q = blk_mq_init_queue(ctrl->admin_tagset);
1741                 if (IS_ERR(ctrl->admin_q)) {
1742                         error = PTR_ERR(ctrl->admin_q);
1743                         goto out_cleanup_fabrics_q;
1744                 }
1745         }
1746
1747         error = nvme_tcp_start_queue(ctrl, 0);
1748         if (error)
1749                 goto out_cleanup_queue;
1750
1751         error = nvme_enable_ctrl(ctrl);
1752         if (error)
1753                 goto out_stop_queue;
1754
1755         blk_mq_unquiesce_queue(ctrl->admin_q);
1756
1757         error = nvme_init_identify(ctrl);
1758         if (error)
1759                 goto out_stop_queue;
1760
1761         return 0;
1762
1763 out_stop_queue:
1764         nvme_tcp_stop_queue(ctrl, 0);
1765 out_cleanup_queue:
1766         if (new)
1767                 blk_cleanup_queue(ctrl->admin_q);
1768 out_cleanup_fabrics_q:
1769         if (new)
1770                 blk_cleanup_queue(ctrl->fabrics_q);
1771 out_free_tagset:
1772         if (new)
1773                 blk_mq_free_tag_set(ctrl->admin_tagset);
1774 out_free_queue:
1775         nvme_tcp_free_admin_queue(ctrl);
1776         return error;
1777 }
1778
1779 static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
1780                 bool remove)
1781 {
1782         blk_mq_quiesce_queue(ctrl->admin_q);
1783         nvme_tcp_stop_queue(ctrl, 0);
1784         if (ctrl->admin_tagset) {
1785                 blk_mq_tagset_busy_iter(ctrl->admin_tagset,
1786                         nvme_cancel_request, ctrl);
1787                 blk_mq_tagset_wait_completed_request(ctrl->admin_tagset);
1788         }
1789         if (remove)
1790                 blk_mq_unquiesce_queue(ctrl->admin_q);
1791         nvme_tcp_destroy_admin_queue(ctrl, remove);
1792 }
1793
1794 static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
1795                 bool remove)
1796 {
1797         if (ctrl->queue_count <= 1)
1798                 return;
1799         nvme_stop_queues(ctrl);
1800         nvme_tcp_stop_io_queues(ctrl);
1801         if (ctrl->tagset) {
1802                 blk_mq_tagset_busy_iter(ctrl->tagset,
1803                         nvme_cancel_request, ctrl);
1804                 blk_mq_tagset_wait_completed_request(ctrl->tagset);
1805         }
1806         if (remove)
1807                 nvme_start_queues(ctrl);
1808         nvme_tcp_destroy_io_queues(ctrl, remove);
1809 }
1810
1811 static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
1812 {
1813         /* If we are resetting/deleting then do nothing */
1814         if (ctrl->state != NVME_CTRL_CONNECTING) {
1815                 WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW ||
1816                         ctrl->state == NVME_CTRL_LIVE);
1817                 return;
1818         }
1819
1820         if (nvmf_should_reconnect(ctrl)) {
1821                 dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
1822                         ctrl->opts->reconnect_delay);
1823                 queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
1824                                 ctrl->opts->reconnect_delay * HZ);
1825         } else {
1826                 dev_info(ctrl->device, "Removing controller...\n");
1827                 nvme_delete_ctrl(ctrl);
1828         }
1829 }
1830
1831 static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
1832 {
1833         struct nvmf_ctrl_options *opts = ctrl->opts;
1834         int ret;
1835
1836         ret = nvme_tcp_configure_admin_queue(ctrl, new);
1837         if (ret)
1838                 return ret;
1839
1840         if (ctrl->icdoff) {
1841                 dev_err(ctrl->device, "icdoff is not supported!\n");
1842                 goto destroy_admin;
1843         }
1844
1845         if (opts->queue_size > ctrl->sqsize + 1)
1846                 dev_warn(ctrl->device,
1847                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
1848                         opts->queue_size, ctrl->sqsize + 1);
1849
1850         if (ctrl->sqsize + 1 > ctrl->maxcmd) {
1851                 dev_warn(ctrl->device,
1852                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
1853                         ctrl->sqsize + 1, ctrl->maxcmd);
1854                 ctrl->sqsize = ctrl->maxcmd - 1;
1855         }
1856
1857         if (ctrl->queue_count > 1) {
1858                 ret = nvme_tcp_configure_io_queues(ctrl, new);
1859                 if (ret)
1860                         goto destroy_admin;
1861         }
1862
1863         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
1864                 /* state change failure is ok if we're in DELETING state */
1865                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
1866                 ret = -EINVAL;
1867                 goto destroy_io;
1868         }
1869
1870         nvme_start_ctrl(ctrl);
1871         return 0;
1872
1873 destroy_io:
1874         if (ctrl->queue_count > 1)
1875                 nvme_tcp_destroy_io_queues(ctrl, new);
1876 destroy_admin:
1877         nvme_tcp_stop_queue(ctrl, 0);
1878         nvme_tcp_destroy_admin_queue(ctrl, new);
1879         return ret;
1880 }
1881
1882 static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
1883 {
1884         struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
1885                         struct nvme_tcp_ctrl, connect_work);
1886         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
1887
1888         ++ctrl->nr_reconnects;
1889
1890         if (nvme_tcp_setup_ctrl(ctrl, false))
1891                 goto requeue;
1892
1893         dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n",
1894                         ctrl->nr_reconnects);
1895
1896         ctrl->nr_reconnects = 0;
1897
1898         return;
1899
1900 requeue:
1901         dev_info(ctrl->device, "Failed reconnect attempt %d\n",
1902                         ctrl->nr_reconnects);
1903         nvme_tcp_reconnect_or_remove(ctrl);
1904 }
1905
1906 static void nvme_tcp_error_recovery_work(struct work_struct *work)
1907 {
1908         struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
1909                                 struct nvme_tcp_ctrl, err_work);
1910         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
1911
1912         nvme_stop_keep_alive(ctrl);
1913         nvme_tcp_teardown_io_queues(ctrl, false);
1914         /* unquiesce to fail fast pending requests */
1915         nvme_start_queues(ctrl);
1916         nvme_tcp_teardown_admin_queue(ctrl, false);
1917         blk_mq_unquiesce_queue(ctrl->admin_q);
1918
1919         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
1920                 /* state change failure is ok if we're in DELETING state */
1921                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
1922                 return;
1923         }
1924
1925         nvme_tcp_reconnect_or_remove(ctrl);
1926 }
1927
1928 static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
1929 {
1930         cancel_work_sync(&to_tcp_ctrl(ctrl)->err_work);
1931         cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
1932
1933         nvme_tcp_teardown_io_queues(ctrl, shutdown);
1934         blk_mq_quiesce_queue(ctrl->admin_q);
1935         if (shutdown)
1936                 nvme_shutdown_ctrl(ctrl);
1937         else
1938                 nvme_disable_ctrl(ctrl);
1939         nvme_tcp_teardown_admin_queue(ctrl, shutdown);
1940 }
1941
1942 static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
1943 {
1944         nvme_tcp_teardown_ctrl(ctrl, true);
1945 }
1946
1947 static void nvme_reset_ctrl_work(struct work_struct *work)
1948 {
1949         struct nvme_ctrl *ctrl =
1950                 container_of(work, struct nvme_ctrl, reset_work);
1951
1952         nvme_stop_ctrl(ctrl);
1953         nvme_tcp_teardown_ctrl(ctrl, false);
1954
1955         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
1956                 /* state change failure is ok if we're in DELETING state */
1957                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
1958                 return;
1959         }
1960
1961         if (nvme_tcp_setup_ctrl(ctrl, false))
1962                 goto out_fail;
1963
1964         return;
1965
1966 out_fail:
1967         ++ctrl->nr_reconnects;
1968         nvme_tcp_reconnect_or_remove(ctrl);
1969 }
1970
1971 static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
1972 {
1973         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1974
1975         if (list_empty(&ctrl->list))
1976                 goto free_ctrl;
1977
1978         mutex_lock(&nvme_tcp_ctrl_mutex);
1979         list_del(&ctrl->list);
1980         mutex_unlock(&nvme_tcp_ctrl_mutex);
1981
1982         nvmf_free_options(nctrl->opts);
1983 free_ctrl:
1984         kfree(ctrl->queues);
1985         kfree(ctrl);
1986 }
1987
1988 static void nvme_tcp_set_sg_null(struct nvme_command *c)
1989 {
1990         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1991
1992         sg->addr = 0;
1993         sg->length = 0;
1994         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
1995                         NVME_SGL_FMT_TRANSPORT_A;
1996 }
1997
1998 static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
1999                 struct nvme_command *c, u32 data_len)
2000 {
2001         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2002
2003         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
2004         sg->length = cpu_to_le32(data_len);
2005         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
2006 }
2007
2008 static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
2009                 u32 data_len)
2010 {
2011         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2012
2013         sg->addr = 0;
2014         sg->length = cpu_to_le32(data_len);
2015         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2016                         NVME_SGL_FMT_TRANSPORT_A;
2017 }
2018
2019 static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
2020 {
2021         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2022         struct nvme_tcp_queue *queue = &ctrl->queues[0];
2023         struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2024         struct nvme_command *cmd = &pdu->cmd;
2025         u8 hdgst = nvme_tcp_hdgst_len(queue);
2026
2027         memset(pdu, 0, sizeof(*pdu));
2028         pdu->hdr.type = nvme_tcp_cmd;
2029         if (queue->hdr_digest)
2030                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2031         pdu->hdr.hlen = sizeof(*pdu);
2032         pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2033
2034         cmd->common.opcode = nvme_admin_async_event;
2035         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2036         cmd->common.flags |= NVME_CMD_SGL_METABUF;
2037         nvme_tcp_set_sg_null(cmd);
2038
2039         ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2040         ctrl->async_req.offset = 0;
2041         ctrl->async_req.curr_bio = NULL;
2042         ctrl->async_req.data_len = 0;
2043
2044         nvme_tcp_queue_request(&ctrl->async_req);
2045 }
2046
2047 static enum blk_eh_timer_return
2048 nvme_tcp_timeout(struct request *rq, bool reserved)
2049 {
2050         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2051         struct nvme_tcp_ctrl *ctrl = req->queue->ctrl;
2052         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2053
2054         /*
2055          * Restart the timer if a controller reset is already scheduled. Any
2056          * timed out commands would be handled before entering the connecting
2057          * state.
2058          */
2059         if (ctrl->ctrl.state == NVME_CTRL_RESETTING)
2060                 return BLK_EH_RESET_TIMER;
2061
2062         dev_warn(ctrl->ctrl.device,
2063                 "queue %d: timeout request %#x type %d\n",
2064                 nvme_tcp_queue_id(req->queue), rq->tag, pdu->hdr.type);
2065
2066         if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
2067                 /*
2068                  * Teardown immediately if controller times out while starting
2069                  * or we are already started error recovery. all outstanding
2070                  * requests are completed on shutdown, so we return BLK_EH_DONE.
2071                  */
2072                 flush_work(&ctrl->err_work);
2073                 nvme_tcp_teardown_io_queues(&ctrl->ctrl, false);
2074                 nvme_tcp_teardown_admin_queue(&ctrl->ctrl, false);
2075                 return BLK_EH_DONE;
2076         }
2077
2078         dev_warn(ctrl->ctrl.device, "starting error recovery\n");
2079         nvme_tcp_error_recovery(&ctrl->ctrl);
2080
2081         return BLK_EH_RESET_TIMER;
2082 }
2083
2084 static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2085                         struct request *rq)
2086 {
2087         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2088         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2089         struct nvme_command *c = &pdu->cmd;
2090
2091         c->common.flags |= NVME_CMD_SGL_METABUF;
2092
2093         if (rq_data_dir(rq) == WRITE && req->data_len &&
2094             req->data_len <= nvme_tcp_inline_data_size(queue))
2095                 nvme_tcp_set_sg_inline(queue, c, req->data_len);
2096         else
2097                 nvme_tcp_set_sg_host_data(c, req->data_len);
2098
2099         return 0;
2100 }
2101
2102 static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2103                 struct request *rq)
2104 {
2105         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2106         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2107         struct nvme_tcp_queue *queue = req->queue;
2108         u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2109         blk_status_t ret;
2110
2111         ret = nvme_setup_cmd(ns, rq, &pdu->cmd);
2112         if (ret)
2113                 return ret;
2114
2115         req->state = NVME_TCP_SEND_CMD_PDU;
2116         req->offset = 0;
2117         req->data_sent = 0;
2118         req->pdu_len = 0;
2119         req->pdu_sent = 0;
2120         req->data_len = blk_rq_payload_bytes(rq);
2121         req->curr_bio = rq->bio;
2122
2123         if (rq_data_dir(rq) == WRITE &&
2124             req->data_len <= nvme_tcp_inline_data_size(queue))
2125                 req->pdu_len = req->data_len;
2126         else if (req->curr_bio)
2127                 nvme_tcp_init_iter(req, READ);
2128
2129         pdu->hdr.type = nvme_tcp_cmd;
2130         pdu->hdr.flags = 0;
2131         if (queue->hdr_digest)
2132                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2133         if (queue->data_digest && req->pdu_len) {
2134                 pdu->hdr.flags |= NVME_TCP_F_DDGST;
2135                 ddgst = nvme_tcp_ddgst_len(queue);
2136         }
2137         pdu->hdr.hlen = sizeof(*pdu);
2138         pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2139         pdu->hdr.plen =
2140                 cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2141
2142         ret = nvme_tcp_map_data(queue, rq);
2143         if (unlikely(ret)) {
2144                 nvme_cleanup_cmd(rq);
2145                 dev_err(queue->ctrl->ctrl.device,
2146                         "Failed to map data (%d)\n", ret);
2147                 return ret;
2148         }
2149
2150         return 0;
2151 }
2152
2153 static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2154                 const struct blk_mq_queue_data *bd)
2155 {
2156         struct nvme_ns *ns = hctx->queue->queuedata;
2157         struct nvme_tcp_queue *queue = hctx->driver_data;
2158         struct request *rq = bd->rq;
2159         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2160         bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2161         blk_status_t ret;
2162
2163         if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2164                 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
2165
2166         ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2167         if (unlikely(ret))
2168                 return ret;
2169
2170         blk_mq_start_request(rq);
2171
2172         nvme_tcp_queue_request(req);
2173
2174         return BLK_STS_OK;
2175 }
2176
2177 static int nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2178 {
2179         struct nvme_tcp_ctrl *ctrl = set->driver_data;
2180         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2181
2182         if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2183                 /* separate read/write queues */
2184                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2185                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2186                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2187                 set->map[HCTX_TYPE_READ].nr_queues =
2188                         ctrl->io_queues[HCTX_TYPE_READ];
2189                 set->map[HCTX_TYPE_READ].queue_offset =
2190                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2191         } else {
2192                 /* shared read/write queues */
2193                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2194                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2195                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2196                 set->map[HCTX_TYPE_READ].nr_queues =
2197                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2198                 set->map[HCTX_TYPE_READ].queue_offset = 0;
2199         }
2200         blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2201         blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
2202
2203         if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2204                 /* map dedicated poll queues only if we have queues left */
2205                 set->map[HCTX_TYPE_POLL].nr_queues =
2206                                 ctrl->io_queues[HCTX_TYPE_POLL];
2207                 set->map[HCTX_TYPE_POLL].queue_offset =
2208                         ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2209                         ctrl->io_queues[HCTX_TYPE_READ];
2210                 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2211         }
2212
2213         dev_info(ctrl->ctrl.device,
2214                 "mapped %d/%d/%d default/read/poll queues.\n",
2215                 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2216                 ctrl->io_queues[HCTX_TYPE_READ],
2217                 ctrl->io_queues[HCTX_TYPE_POLL]);
2218
2219         return 0;
2220 }
2221
2222 static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx)
2223 {
2224         struct nvme_tcp_queue *queue = hctx->driver_data;
2225         struct sock *sk = queue->sock->sk;
2226
2227         if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue))
2228                 sk_busy_loop(sk, true);
2229         nvme_tcp_try_recv(queue);
2230         return queue->nr_cqe;
2231 }
2232
2233 static struct blk_mq_ops nvme_tcp_mq_ops = {
2234         .queue_rq       = nvme_tcp_queue_rq,
2235         .complete       = nvme_complete_rq,
2236         .init_request   = nvme_tcp_init_request,
2237         .exit_request   = nvme_tcp_exit_request,
2238         .init_hctx      = nvme_tcp_init_hctx,
2239         .timeout        = nvme_tcp_timeout,
2240         .map_queues     = nvme_tcp_map_queues,
2241         .poll           = nvme_tcp_poll,
2242 };
2243
2244 static struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2245         .queue_rq       = nvme_tcp_queue_rq,
2246         .complete       = nvme_complete_rq,
2247         .init_request   = nvme_tcp_init_request,
2248         .exit_request   = nvme_tcp_exit_request,
2249         .init_hctx      = nvme_tcp_init_admin_hctx,
2250         .timeout        = nvme_tcp_timeout,
2251 };
2252
2253 static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2254         .name                   = "tcp",
2255         .module                 = THIS_MODULE,
2256         .flags                  = NVME_F_FABRICS,
2257         .reg_read32             = nvmf_reg_read32,
2258         .reg_read64             = nvmf_reg_read64,
2259         .reg_write32            = nvmf_reg_write32,
2260         .free_ctrl              = nvme_tcp_free_ctrl,
2261         .submit_async_event     = nvme_tcp_submit_async_event,
2262         .delete_ctrl            = nvme_tcp_delete_ctrl,
2263         .get_address            = nvmf_get_address,
2264 };
2265
2266 static bool
2267 nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2268 {
2269         struct nvme_tcp_ctrl *ctrl;
2270         bool found = false;
2271
2272         mutex_lock(&nvme_tcp_ctrl_mutex);
2273         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2274                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2275                 if (found)
2276                         break;
2277         }
2278         mutex_unlock(&nvme_tcp_ctrl_mutex);
2279
2280         return found;
2281 }
2282
2283 static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2284                 struct nvmf_ctrl_options *opts)
2285 {
2286         struct nvme_tcp_ctrl *ctrl;
2287         int ret;
2288
2289         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2290         if (!ctrl)
2291                 return ERR_PTR(-ENOMEM);
2292
2293         INIT_LIST_HEAD(&ctrl->list);
2294         ctrl->ctrl.opts = opts;
2295         ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2296                                 opts->nr_poll_queues + 1;
2297         ctrl->ctrl.sqsize = opts->queue_size - 1;
2298         ctrl->ctrl.kato = opts->kato;
2299
2300         INIT_DELAYED_WORK(&ctrl->connect_work,
2301                         nvme_tcp_reconnect_ctrl_work);
2302         INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2303         INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2304
2305         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2306                 opts->trsvcid =
2307                         kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2308                 if (!opts->trsvcid) {
2309                         ret = -ENOMEM;
2310                         goto out_free_ctrl;
2311                 }
2312                 opts->mask |= NVMF_OPT_TRSVCID;
2313         }
2314
2315         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2316                         opts->traddr, opts->trsvcid, &ctrl->addr);
2317         if (ret) {
2318                 pr_err("malformed address passed: %s:%s\n",
2319                         opts->traddr, opts->trsvcid);
2320                 goto out_free_ctrl;
2321         }
2322
2323         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2324                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2325                         opts->host_traddr, NULL, &ctrl->src_addr);
2326                 if (ret) {
2327                         pr_err("malformed src address passed: %s\n",
2328                                opts->host_traddr);
2329                         goto out_free_ctrl;
2330                 }
2331         }
2332
2333         if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2334                 ret = -EALREADY;
2335                 goto out_free_ctrl;
2336         }
2337
2338         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2339                                 GFP_KERNEL);
2340         if (!ctrl->queues) {
2341                 ret = -ENOMEM;
2342                 goto out_free_ctrl;
2343         }
2344
2345         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2346         if (ret)
2347                 goto out_kfree_queues;
2348
2349         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2350                 WARN_ON_ONCE(1);
2351                 ret = -EINTR;
2352                 goto out_uninit_ctrl;
2353         }
2354
2355         ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2356         if (ret)
2357                 goto out_uninit_ctrl;
2358
2359         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
2360                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2361
2362         nvme_get_ctrl(&ctrl->ctrl);
2363
2364         mutex_lock(&nvme_tcp_ctrl_mutex);
2365         list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2366         mutex_unlock(&nvme_tcp_ctrl_mutex);
2367
2368         return &ctrl->ctrl;
2369
2370 out_uninit_ctrl:
2371         nvme_uninit_ctrl(&ctrl->ctrl);
2372         nvme_put_ctrl(&ctrl->ctrl);
2373         if (ret > 0)
2374                 ret = -EIO;
2375         return ERR_PTR(ret);
2376 out_kfree_queues:
2377         kfree(ctrl->queues);
2378 out_free_ctrl:
2379         kfree(ctrl);
2380         return ERR_PTR(ret);
2381 }
2382
2383 static struct nvmf_transport_ops nvme_tcp_transport = {
2384         .name           = "tcp",
2385         .module         = THIS_MODULE,
2386         .required_opts  = NVMF_OPT_TRADDR,
2387         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2388                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2389                           NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
2390                           NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2391                           NVMF_OPT_TOS,
2392         .create_ctrl    = nvme_tcp_create_ctrl,
2393 };
2394
2395 static int __init nvme_tcp_init_module(void)
2396 {
2397         nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
2398                         WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2399         if (!nvme_tcp_wq)
2400                 return -ENOMEM;
2401
2402         nvmf_register_transport(&nvme_tcp_transport);
2403         return 0;
2404 }
2405
2406 static void __exit nvme_tcp_cleanup_module(void)
2407 {
2408         struct nvme_tcp_ctrl *ctrl;
2409
2410         nvmf_unregister_transport(&nvme_tcp_transport);
2411
2412         mutex_lock(&nvme_tcp_ctrl_mutex);
2413         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
2414                 nvme_delete_ctrl(&ctrl->ctrl);
2415         mutex_unlock(&nvme_tcp_ctrl_mutex);
2416         flush_workqueue(nvme_delete_wq);
2417
2418         destroy_workqueue(nvme_tcp_wq);
2419 }
2420
2421 module_init(nvme_tcp_init_module);
2422 module_exit(nvme_tcp_cleanup_module);
2423
2424 MODULE_LICENSE("GPL v2");