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 /* Define the socket priority to use for connections were it is desirable
  24  * that the NIC consider performing optimized packet processing or filtering.
  25  * A non-zero value being sufficient to indicate general consideration of any
  26  * possible optimization.  Making it a module param allows for alternative
  27  * values that may be unique for some NIC implementations.
  28  */
  29 static int so_priority;
  30 module_param(so_priority, int, 0644);
  31 MODULE_PARM_DESC(so_priority, "nvme tcp socket optimize priority");
  32
  33 enum nvme_tcp_send_state {
  34         NVME_TCP_SEND_CMD_PDU = 0,
  35         NVME_TCP_SEND_H2C_PDU,
  36         NVME_TCP_SEND_DATA,
  37         NVME_TCP_SEND_DDGST,
  38 };
  39
  40 struct nvme_tcp_request {
  41         struct nvme_request     req;
  42         void                    *pdu;
  43         struct nvme_tcp_queue   *queue;
  44         u32                     data_len;
  45         u32                     pdu_len;
  46         u32                     pdu_sent;
  47         u16                     ttag;
  48         struct list_head        entry;
  49         __le32                  ddgst;
  50
  51         struct bio              *curr_bio;
  52         struct iov_iter         iter;
  53
  54         /* send state */
  55         size_t                  offset;
  56         size_t                  data_sent;
  57         enum nvme_tcp_send_state state;
  58 };
  59
  60 enum nvme_tcp_queue_flags {
  61         NVME_TCP_Q_ALLOCATED    = 0,
  62         NVME_TCP_Q_LIVE         = 1,
  63         NVME_TCP_Q_POLLING      = 2,
  64 };
  65
  66 enum nvme_tcp_recv_state {
  67         NVME_TCP_RECV_PDU = 0,
  68         NVME_TCP_RECV_DATA,
  69         NVME_TCP_RECV_DDGST,
  70 };
  71
  72 struct nvme_tcp_ctrl;
  73 struct nvme_tcp_queue {
  74         struct socket           *sock;
  75         struct work_struct      io_work;
  76         int                     io_cpu;
  77
  78         spinlock_t              lock;
  79         struct mutex            send_mutex;
  80         struct list_head        send_list;
  81
  82         /* recv state */
  83         void                    *pdu;
  84         int                     pdu_remaining;
  85         int                     pdu_offset;
  86         size_t                  data_remaining;
  87         size_t                  ddgst_remaining;
  88         unsigned int            nr_cqe;
  89
  90         /* send state */
  91         struct nvme_tcp_request *request;
  92
  93         int                     queue_size;
  94         size_t                  cmnd_capsule_len;
  95         struct nvme_tcp_ctrl    *ctrl;
  96         unsigned long           flags;
  97         bool                    rd_enabled;
  98
  99         bool                    hdr_digest;
 100         bool                    data_digest;
 101         struct ahash_request    *rcv_hash;
 102         struct ahash_request    *snd_hash;
 103         __le32                  exp_ddgst;
 104         __le32                  recv_ddgst;
 105
 106         struct page_frag_cache  pf_cache;
 107
 108         void (*state_change)(struct sock *);
 109         void (*data_ready)(struct sock *);
 110         void (*write_space)(struct sock *);
 111 };
 112
 113 struct nvme_tcp_ctrl {
 114         /* read only in the hot path */
 115         struct nvme_tcp_queue   *queues;
 116         struct blk_mq_tag_set   tag_set;
 117
 118         /* other member variables */
 119         struct list_head        list;
 120         struct blk_mq_tag_set   admin_tag_set;
 121         struct sockaddr_storage addr;
 122         struct sockaddr_storage src_addr;
 123         struct nvme_ctrl        ctrl;
 124
 125         struct work_struct      err_work;
 126         struct delayed_work     connect_work;
 127         struct nvme_tcp_request async_req;
 128         u32                     io_queues[HCTX_MAX_TYPES];
 129 };
 130
 131 static LIST_HEAD(nvme_tcp_ctrl_list);
 132 static DEFINE_MUTEX(nvme_tcp_ctrl_mutex);
 133 static struct workqueue_struct *nvme_tcp_wq;
 134 static struct blk_mq_ops nvme_tcp_mq_ops;
 135 static struct blk_mq_ops nvme_tcp_admin_mq_ops;
 136 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue);
 137
 138 static inline struct nvme_tcp_ctrl *to_tcp_ctrl(struct nvme_ctrl *ctrl)
 139 {
 140         return container_of(ctrl, struct nvme_tcp_ctrl, ctrl);
 141 }
 142
 143 static inline int nvme_tcp_queue_id(struct nvme_tcp_queue *queue)
 144 {
 145         return queue - queue->ctrl->queues;
 146 }
 147
 148 static inline struct blk_mq_tags *nvme_tcp_tagset(struct nvme_tcp_queue *queue)
 149 {
 150         u32 queue_idx = nvme_tcp_queue_id(queue);
 151
 152         if (queue_idx == 0)
 153                 return queue->ctrl->admin_tag_set.tags[queue_idx];
 154         return queue->ctrl->tag_set.tags[queue_idx - 1];
 155 }
 156
 157 static inline u8 nvme_tcp_hdgst_len(struct nvme_tcp_queue *queue)
 158 {
 159         return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 160 }
 161
 162 static inline u8 nvme_tcp_ddgst_len(struct nvme_tcp_queue *queue)
 163 {
 164         return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 165 }
 166
 167 static inline size_t nvme_tcp_inline_data_size(struct nvme_tcp_queue *queue)
 168 {
 169         return queue->cmnd_capsule_len - sizeof(struct nvme_command);
 170 }
 171
 172 static inline bool nvme_tcp_async_req(struct nvme_tcp_request *req)
 173 {
 174         return req == &req->queue->ctrl->async_req;
 175 }
 176
 177 static inline bool nvme_tcp_has_inline_data(struct nvme_tcp_request *req)
 178 {
 179         struct request *rq;
 180
 181         if (unlikely(nvme_tcp_async_req(req)))
 182                 return false; /* async events don't have a request */
 183
 184         rq = blk_mq_rq_from_pdu(req);
 185
 186         return rq_data_dir(rq) == WRITE && req->data_len &&
 187                 req->data_len <= nvme_tcp_inline_data_size(req->queue);
 188 }
 189
 190 static inline struct page *nvme_tcp_req_cur_page(struct nvme_tcp_request *req)
 191 {
 192         return req->iter.bvec->bv_page;
 193 }
 194
 195 static inline size_t nvme_tcp_req_cur_offset(struct nvme_tcp_request *req)
 196 {
 197         return req->iter.bvec->bv_offset + req->iter.iov_offset;
 198 }
 199
 200 static inline size_t nvme_tcp_req_cur_length(struct nvme_tcp_request *req)
 201 {
 202         return min_t(size_t, req->iter.bvec->bv_len - req->iter.iov_offset,
 203                         req->pdu_len - req->pdu_sent);
 204 }
 205
 206 static inline size_t nvme_tcp_req_offset(struct nvme_tcp_request *req)
 207 {
 208         return req->iter.iov_offset;
 209 }
 210
 211 static inline size_t nvme_tcp_pdu_data_left(struct nvme_tcp_request *req)
 212 {
 213         return rq_data_dir(blk_mq_rq_from_pdu(req)) == WRITE ?
 214                         req->pdu_len - req->pdu_sent : 0;
 215 }
 216
 217 static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req,
 218                 int len)
 219 {
 220         return nvme_tcp_pdu_data_left(req) <= len;
 221 }
 222
 223 static void nvme_tcp_init_iter(struct nvme_tcp_request *req,
 224                 unsigned int dir)
 225 {
 226         struct request *rq = blk_mq_rq_from_pdu(req);
 227         struct bio_vec *vec;
 228         unsigned int size;
 229         int nsegs;
 230         size_t offset;
 231
 232         if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) {
 233                 vec = &rq->special_vec;
 234                 nsegs = 1;
 235                 size = blk_rq_payload_bytes(rq);
 236                 offset = 0;
 237         } else {
 238                 struct bio *bio = req->curr_bio;
 239
 240                 vec = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
 241                 nsegs = bio_segments(bio);
 242                 size = bio->bi_iter.bi_size;
 243                 offset = bio->bi_iter.bi_bvec_done;
 244         }
 245
 246         iov_iter_bvec(&req->iter, dir, vec, nsegs, size);
 247         req->iter.iov_offset = offset;
 248 }
 249
 250 static inline void nvme_tcp_advance_req(struct nvme_tcp_request *req,
 251                 int len)
 252 {
 253         req->data_sent += len;
 254         req->pdu_sent += len;
 255         iov_iter_advance(&req->iter, len);
 256         if (!iov_iter_count(&req->iter) &&
 257             req->data_sent < req->data_len) {
 258                 req->curr_bio = req->curr_bio->bi_next;
 259                 nvme_tcp_init_iter(req, WRITE);
 260         }
 261 }
 262
 263 static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req,
 264                 bool sync)
 265 {
 266         struct nvme_tcp_queue *queue = req->queue;
 267         bool empty;
 268
 269         spin_lock(&queue->lock);
 270         empty = list_empty(&queue->send_list) && !queue->request;
 271         list_add_tail(&req->entry, &queue->send_list);
 272         spin_unlock(&queue->lock);
 273
 274         /*
 275          * if we're the first on the send_list and we can try to send
 276          * directly, otherwise queue io_work. Also, only do that if we
 277          * are on the same cpu, so we don't introduce contention.
 278          */
 279         if (queue->io_cpu == smp_processor_id() &&
 280             sync && empty && mutex_trylock(&queue->send_mutex)) {
 281                 nvme_tcp_try_send(queue);
 282                 mutex_unlock(&queue->send_mutex);
 283         } else {
 284                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 285         }
 286 }
 287
 288 static inline struct nvme_tcp_request *
 289 nvme_tcp_fetch_request(struct nvme_tcp_queue *queue)
 290 {
 291         struct nvme_tcp_request *req;
 292
 293         spin_lock(&queue->lock);
 294         req = list_first_entry_or_null(&queue->send_list,
 295                         struct nvme_tcp_request, entry);
 296         if (req)
 297                 list_del(&req->entry);
 298         spin_unlock(&queue->lock);
 299
 300         return req;
 301 }
 302
 303 static inline void nvme_tcp_ddgst_final(struct ahash_request *hash,
 304                 __le32 *dgst)
 305 {
 306         ahash_request_set_crypt(hash, NULL, (u8 *)dgst, 0);
 307         crypto_ahash_final(hash);
 308 }
 309
 310 static inline void nvme_tcp_ddgst_update(struct ahash_request *hash,
 311                 struct page *page, off_t off, size_t len)
 312 {
 313         struct scatterlist sg;
 314
 315         sg_init_marker(&sg, 1);
 316         sg_set_page(&sg, page, len, off);
 317         ahash_request_set_crypt(hash, &sg, NULL, len);
 318         crypto_ahash_update(hash);
 319 }
 320
 321 static inline void nvme_tcp_hdgst(struct ahash_request *hash,
 322                 void *pdu, size_t len)
 323 {
 324         struct scatterlist sg;
 325
 326         sg_init_one(&sg, pdu, len);
 327         ahash_request_set_crypt(hash, &sg, pdu + len, len);
 328         crypto_ahash_digest(hash);
 329 }
 330
 331 static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue,
 332                 void *pdu, size_t pdu_len)
 333 {
 334         struct nvme_tcp_hdr *hdr = pdu;
 335         __le32 recv_digest;
 336         __le32 exp_digest;
 337
 338         if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
 339                 dev_err(queue->ctrl->ctrl.device,
 340                         "queue %d: header digest flag is cleared\n",
 341                         nvme_tcp_queue_id(queue));
 342                 return -EPROTO;
 343         }
 344
 345         recv_digest = *(__le32 *)(pdu + hdr->hlen);
 346         nvme_tcp_hdgst(queue->rcv_hash, pdu, pdu_len);
 347         exp_digest = *(__le32 *)(pdu + hdr->hlen);
 348         if (recv_digest != exp_digest) {
 349                 dev_err(queue->ctrl->ctrl.device,
 350                         "header digest error: recv %#x expected %#x\n",
 351                         le32_to_cpu(recv_digest), le32_to_cpu(exp_digest));
 352                 return -EIO;
 353         }
 354
 355         return 0;
 356 }
 357
 358 static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu)
 359 {
 360         struct nvme_tcp_hdr *hdr = pdu;
 361         u8 digest_len = nvme_tcp_hdgst_len(queue);
 362         u32 len;
 363
 364         len = le32_to_cpu(hdr->plen) - hdr->hlen -
 365                 ((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0);
 366
 367         if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
 368                 dev_err(queue->ctrl->ctrl.device,
 369                         "queue %d: data digest flag is cleared\n",
 370                 nvme_tcp_queue_id(queue));
 371                 return -EPROTO;
 372         }
 373         crypto_ahash_init(queue->rcv_hash);
 374
 375         return 0;
 376 }
 377
 378 static void nvme_tcp_exit_request(struct blk_mq_tag_set *set,
 379                 struct request *rq, unsigned int hctx_idx)
 380 {
 381         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 382
 383         page_frag_free(req->pdu);
 384 }
 385
 386 static int nvme_tcp_init_request(struct blk_mq_tag_set *set,
 387                 struct request *rq, unsigned int hctx_idx,
 388                 unsigned int numa_node)
 389 {
 390         struct nvme_tcp_ctrl *ctrl = set->driver_data;
 391         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 392         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 393         struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx];
 394         u8 hdgst = nvme_tcp_hdgst_len(queue);
 395
 396         req->pdu = page_frag_alloc(&queue->pf_cache,
 397                 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
 398                 GFP_KERNEL | __GFP_ZERO);
 399         if (!req->pdu)
 400                 return -ENOMEM;
 401
 402         req->queue = queue;
 403         nvme_req(rq)->ctrl = &ctrl->ctrl;
 404
 405         return 0;
 406 }
 407
 408 static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 409                 unsigned int hctx_idx)
 410 {
 411         struct nvme_tcp_ctrl *ctrl = data;
 412         struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1];
 413
 414         hctx->driver_data = queue;
 415         return 0;
 416 }
 417
 418 static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 419                 unsigned int hctx_idx)
 420 {
 421         struct nvme_tcp_ctrl *ctrl = data;
 422         struct nvme_tcp_queue *queue = &ctrl->queues[0];
 423
 424         hctx->driver_data = queue;
 425         return 0;
 426 }
 427
 428 static enum nvme_tcp_recv_state
 429 nvme_tcp_recv_state(struct nvme_tcp_queue *queue)
 430 {
 431         return  (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
 432                 (queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST :
 433                 NVME_TCP_RECV_DATA;
 434 }
 435
 436 static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue)
 437 {
 438         queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) +
 439                                 nvme_tcp_hdgst_len(queue);
 440         queue->pdu_offset = 0;
 441         queue->data_remaining = -1;
 442         queue->ddgst_remaining = 0;
 443 }
 444
 445 static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl)
 446 {
 447         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
 448                 return;
 449
 450         queue_work(nvme_reset_wq, &to_tcp_ctrl(ctrl)->err_work);
 451 }
 452
 453 static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
 454                 struct nvme_completion *cqe)
 455 {
 456         struct request *rq;
 457
 458         rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), cqe->command_id);
 459         if (!rq) {
 460                 dev_err(queue->ctrl->ctrl.device,
 461                         "queue %d tag 0x%x not found\n",
 462                         nvme_tcp_queue_id(queue), cqe->command_id);
 463                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 464                 return -EINVAL;
 465         }
 466
 467         nvme_end_request(rq, cqe->status, cqe->result);
 468         queue->nr_cqe++;
 469
 470         return 0;
 471 }
 472
 473 static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue,
 474                 struct nvme_tcp_data_pdu *pdu)
 475 {
 476         struct request *rq;
 477
 478         rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
 479         if (!rq) {
 480                 dev_err(queue->ctrl->ctrl.device,
 481                         "queue %d tag %#x not found\n",
 482                         nvme_tcp_queue_id(queue), pdu->command_id);
 483                 return -ENOENT;
 484         }
 485
 486         if (!blk_rq_payload_bytes(rq)) {
 487                 dev_err(queue->ctrl->ctrl.device,
 488                         "queue %d tag %#x unexpected data\n",
 489                         nvme_tcp_queue_id(queue), rq->tag);
 490                 return -EIO;
 491         }
 492
 493         queue->data_remaining = le32_to_cpu(pdu->data_length);
 494
 495         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS &&
 496             unlikely(!(pdu->hdr.flags & NVME_TCP_F_DATA_LAST))) {
 497                 dev_err(queue->ctrl->ctrl.device,
 498                         "queue %d tag %#x SUCCESS set but not last PDU\n",
 499                         nvme_tcp_queue_id(queue), rq->tag);
 500                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 501                 return -EPROTO;
 502         }
 503
 504         return 0;
 505 }
 506
 507 static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue,
 508                 struct nvme_tcp_rsp_pdu *pdu)
 509 {
 510         struct nvme_completion *cqe = &pdu->cqe;
 511         int ret = 0;
 512
 513         /*
 514          * AEN requests are special as they don't time out and can
 515          * survive any kind of queue freeze and often don't respond to
 516          * aborts.  We don't even bother to allocate a struct request
 517          * for them but rather special case them here.
 518          */
 519         if (unlikely(nvme_is_aen_req(nvme_tcp_queue_id(queue),
 520                                      cqe->command_id)))
 521                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
 522                                 &cqe->result);
 523         else
 524                 ret = nvme_tcp_process_nvme_cqe(queue, cqe);
 525
 526         return ret;
 527 }
 528
 529 static int nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req,
 530                 struct nvme_tcp_r2t_pdu *pdu)
 531 {
 532         struct nvme_tcp_data_pdu *data = req->pdu;
 533         struct nvme_tcp_queue *queue = req->queue;
 534         struct request *rq = blk_mq_rq_from_pdu(req);
 535         u8 hdgst = nvme_tcp_hdgst_len(queue);
 536         u8 ddgst = nvme_tcp_ddgst_len(queue);
 537
 538         req->pdu_len = le32_to_cpu(pdu->r2t_length);
 539         req->pdu_sent = 0;
 540
 541         if (unlikely(req->data_sent + req->pdu_len > req->data_len)) {
 542                 dev_err(queue->ctrl->ctrl.device,
 543                         "req %d r2t len %u exceeded data len %u (%zu sent)\n",
 544                         rq->tag, req->pdu_len, req->data_len,
 545                         req->data_sent);
 546                 return -EPROTO;
 547         }
 548
 549         if (unlikely(le32_to_cpu(pdu->r2t_offset) < req->data_sent)) {
 550                 dev_err(queue->ctrl->ctrl.device,
 551                         "req %d unexpected r2t offset %u (expected %zu)\n",
 552                         rq->tag, le32_to_cpu(pdu->r2t_offset),
 553                         req->data_sent);
 554                 return -EPROTO;
 555         }
 556
 557         memset(data, 0, sizeof(*data));
 558         data->hdr.type = nvme_tcp_h2c_data;
 559         data->hdr.flags = NVME_TCP_F_DATA_LAST;
 560         if (queue->hdr_digest)
 561                 data->hdr.flags |= NVME_TCP_F_HDGST;
 562         if (queue->data_digest)
 563                 data->hdr.flags |= NVME_TCP_F_DDGST;
 564         data->hdr.hlen = sizeof(*data);
 565         data->hdr.pdo = data->hdr.hlen + hdgst;
 566         data->hdr.plen =
 567                 cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst);
 568         data->ttag = pdu->ttag;
 569         data->command_id = rq->tag;
 570         data->data_offset = cpu_to_le32(req->data_sent);
 571         data->data_length = cpu_to_le32(req->pdu_len);
 572         return 0;
 573 }
 574
 575 static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue,
 576                 struct nvme_tcp_r2t_pdu *pdu)
 577 {
 578         struct nvme_tcp_request *req;
 579         struct request *rq;
 580         int ret;
 581
 582         rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
 583         if (!rq) {
 584                 dev_err(queue->ctrl->ctrl.device,
 585                         "queue %d tag %#x not found\n",
 586                         nvme_tcp_queue_id(queue), pdu->command_id);
 587                 return -ENOENT;
 588         }
 589         req = blk_mq_rq_to_pdu(rq);
 590
 591         ret = nvme_tcp_setup_h2c_data_pdu(req, pdu);
 592         if (unlikely(ret))
 593                 return ret;
 594
 595         req->state = NVME_TCP_SEND_H2C_PDU;
 596         req->offset = 0;
 597
 598         nvme_tcp_queue_request(req, false);
 599
 600         return 0;
 601 }
 602
 603 static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 604                 unsigned int *offset, size_t *len)
 605 {
 606         struct nvme_tcp_hdr *hdr;
 607         char *pdu = queue->pdu;
 608         size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
 609         int ret;
 610
 611         ret = skb_copy_bits(skb, *offset,
 612                 &pdu[queue->pdu_offset], rcv_len);
 613         if (unlikely(ret))
 614                 return ret;
 615
 616         queue->pdu_remaining -= rcv_len;
 617         queue->pdu_offset += rcv_len;
 618         *offset += rcv_len;
 619         *len -= rcv_len;
 620         if (queue->pdu_remaining)
 621                 return 0;
 622
 623         hdr = queue->pdu;
 624         if (queue->hdr_digest) {
 625                 ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen);
 626                 if (unlikely(ret))
 627                         return ret;
 628         }
 629
 630
 631         if (queue->data_digest) {
 632                 ret = nvme_tcp_check_ddgst(queue, queue->pdu);
 633                 if (unlikely(ret))
 634                         return ret;
 635         }
 636
 637         switch (hdr->type) {
 638         case nvme_tcp_c2h_data:
 639                 return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
 640         case nvme_tcp_rsp:
 641                 nvme_tcp_init_recv_ctx(queue);
 642                 return nvme_tcp_handle_comp(queue, (void *)queue->pdu);
 643         case nvme_tcp_r2t:
 644                 nvme_tcp_init_recv_ctx(queue);
 645                 return nvme_tcp_handle_r2t(queue, (void *)queue->pdu);
 646         default:
 647                 dev_err(queue->ctrl->ctrl.device,
 648                         "unsupported pdu type (%d)\n", hdr->type);
 649                 return -EINVAL;
 650         }
 651 }
 652
 653 static inline void nvme_tcp_end_request(struct request *rq, u16 status)
 654 {
 655         union nvme_result res = {};
 656
 657         nvme_end_request(rq, cpu_to_le16(status << 1), res);
 658 }
 659
 660 static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 661                               unsigned int *offset, size_t *len)
 662 {
 663         struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
 664         struct nvme_tcp_request *req;
 665         struct request *rq;
 666
 667         rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
 668         if (!rq) {
 669                 dev_err(queue->ctrl->ctrl.device,
 670                         "queue %d tag %#x not found\n",
 671                         nvme_tcp_queue_id(queue), pdu->command_id);
 672                 return -ENOENT;
 673         }
 674         req = blk_mq_rq_to_pdu(rq);
 675
 676         while (true) {
 677                 int recv_len, ret;
 678
 679                 recv_len = min_t(size_t, *len, queue->data_remaining);
 680                 if (!recv_len)
 681                         break;
 682
 683                 if (!iov_iter_count(&req->iter)) {
 684                         req->curr_bio = req->curr_bio->bi_next;
 685
 686                         /*
 687                          * If we don`t have any bios it means that controller
 688                          * sent more data than we requested, hence error
 689                          */
 690                         if (!req->curr_bio) {
 691                                 dev_err(queue->ctrl->ctrl.device,
 692                                         "queue %d no space in request %#x",
 693                                         nvme_tcp_queue_id(queue), rq->tag);
 694                                 nvme_tcp_init_recv_ctx(queue);
 695                                 return -EIO;
 696                         }
 697                         nvme_tcp_init_iter(req, READ);
 698                 }
 699
 700                 /* we can read only from what is left in this bio */
 701                 recv_len = min_t(size_t, recv_len,
 702                                 iov_iter_count(&req->iter));
 703
 704                 if (queue->data_digest)
 705                         ret = skb_copy_and_hash_datagram_iter(skb, *offset,
 706                                 &req->iter, recv_len, queue->rcv_hash);
 707                 else
 708                         ret = skb_copy_datagram_iter(skb, *offset,
 709                                         &req->iter, recv_len);
 710                 if (ret) {
 711                         dev_err(queue->ctrl->ctrl.device,
 712                                 "queue %d failed to copy request %#x data",
 713                                 nvme_tcp_queue_id(queue), rq->tag);
 714                         return ret;
 715                 }
 716
 717                 *len -= recv_len;
 718                 *offset += recv_len;
 719                 queue->data_remaining -= recv_len;
 720         }
 721
 722         if (!queue->data_remaining) {
 723                 if (queue->data_digest) {
 724                         nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst);
 725                         queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
 726                 } else {
 727                         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
 728                                 nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
 729                                 queue->nr_cqe++;
 730                         }
 731                         nvme_tcp_init_recv_ctx(queue);
 732                 }
 733         }
 734
 735         return 0;
 736 }
 737
 738 static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
 739                 struct sk_buff *skb, unsigned int *offset, size_t *len)
 740 {
 741         struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
 742         char *ddgst = (char *)&queue->recv_ddgst;
 743         size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
 744         off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
 745         int ret;
 746
 747         ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
 748         if (unlikely(ret))
 749                 return ret;
 750
 751         queue->ddgst_remaining -= recv_len;
 752         *offset += recv_len;
 753         *len -= recv_len;
 754         if (queue->ddgst_remaining)
 755                 return 0;
 756
 757         if (queue->recv_ddgst != queue->exp_ddgst) {
 758                 dev_err(queue->ctrl->ctrl.device,
 759                         "data digest error: recv %#x expected %#x\n",
 760                         le32_to_cpu(queue->recv_ddgst),
 761                         le32_to_cpu(queue->exp_ddgst));
 762                 return -EIO;
 763         }
 764
 765         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
 766                 struct request *rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue),
 767                                                 pdu->command_id);
 768
 769                 nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
 770                 queue->nr_cqe++;
 771         }
 772
 773         nvme_tcp_init_recv_ctx(queue);
 774         return 0;
 775 }
 776
 777 static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
 778                              unsigned int offset, size_t len)
 779 {
 780         struct nvme_tcp_queue *queue = desc->arg.data;
 781         size_t consumed = len;
 782         int result;
 783
 784         while (len) {
 785                 switch (nvme_tcp_recv_state(queue)) {
 786                 case NVME_TCP_RECV_PDU:
 787                         result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
 788                         break;
 789                 case NVME_TCP_RECV_DATA:
 790                         result = nvme_tcp_recv_data(queue, skb, &offset, &len);
 791                         break;
 792                 case NVME_TCP_RECV_DDGST:
 793                         result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
 794                         break;
 795                 default:
 796                         result = -EFAULT;
 797                 }
 798                 if (result) {
 799                         dev_err(queue->ctrl->ctrl.device,
 800                                 "receive failed:  %d\n", result);
 801                         queue->rd_enabled = false;
 802                         nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 803                         return result;
 804                 }
 805         }
 806
 807         return consumed;
 808 }
 809
 810 static void nvme_tcp_data_ready(struct sock *sk)
 811 {
 812         struct nvme_tcp_queue *queue;
 813
 814         read_lock_bh(&sk->sk_callback_lock);
 815         queue = sk->sk_user_data;
 816         if (likely(queue && queue->rd_enabled) &&
 817             !test_bit(NVME_TCP_Q_POLLING, &queue->flags))
 818                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 819         read_unlock_bh(&sk->sk_callback_lock);
 820 }
 821
 822 static void nvme_tcp_write_space(struct sock *sk)
 823 {
 824         struct nvme_tcp_queue *queue;
 825
 826         read_lock_bh(&sk->sk_callback_lock);
 827         queue = sk->sk_user_data;
 828         if (likely(queue && sk_stream_is_writeable(sk))) {
 829                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 830                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 831         }
 832         read_unlock_bh(&sk->sk_callback_lock);
 833 }
 834
 835 static void nvme_tcp_state_change(struct sock *sk)
 836 {
 837         struct nvme_tcp_queue *queue;
 838
 839         read_lock(&sk->sk_callback_lock);
 840         queue = sk->sk_user_data;
 841         if (!queue)
 842                 goto done;
 843
 844         switch (sk->sk_state) {
 845         case TCP_CLOSE:
 846         case TCP_CLOSE_WAIT:
 847         case TCP_LAST_ACK:
 848         case TCP_FIN_WAIT1:
 849         case TCP_FIN_WAIT2:
 850                 /* fallthrough */
 851                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 852                 break;
 853         default:
 854                 dev_info(queue->ctrl->ctrl.device,
 855                         "queue %d socket state %d\n",
 856                         nvme_tcp_queue_id(queue), sk->sk_state);
 857         }
 858
 859         queue->state_change(sk);
 860 done:
 861         read_unlock(&sk->sk_callback_lock);
 862 }
 863
 864 static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
 865 {
 866         queue->request = NULL;
 867 }
 868
 869 static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
 870 {
 871         nvme_tcp_end_request(blk_mq_rq_from_pdu(req), NVME_SC_HOST_PATH_ERROR);
 872 }
 873
 874 static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
 875 {
 876         struct nvme_tcp_queue *queue = req->queue;
 877
 878         while (true) {
 879                 struct page *page = nvme_tcp_req_cur_page(req);
 880                 size_t offset = nvme_tcp_req_cur_offset(req);
 881                 size_t len = nvme_tcp_req_cur_length(req);
 882                 bool last = nvme_tcp_pdu_last_send(req, len);
 883                 int ret, flags = MSG_DONTWAIT;
 884
 885                 if (last && !queue->data_digest)
 886                         flags |= MSG_EOR;
 887                 else
 888                         flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
 889
 890                 /* can't zcopy slab pages */
 891                 if (unlikely(PageSlab(page))) {
 892                         ret = sock_no_sendpage(queue->sock, page, offset, len,
 893                                         flags);
 894                 } else {
 895                         ret = kernel_sendpage(queue->sock, page, offset, len,
 896                                         flags);
 897                 }
 898                 if (ret <= 0)
 899                         return ret;
 900
 901                 nvme_tcp_advance_req(req, ret);
 902                 if (queue->data_digest)
 903                         nvme_tcp_ddgst_update(queue->snd_hash, page,
 904                                         offset, ret);
 905
 906                 /* fully successful last write*/
 907                 if (last && ret == len) {
 908                         if (queue->data_digest) {
 909                                 nvme_tcp_ddgst_final(queue->snd_hash,
 910                                         &req->ddgst);
 911                                 req->state = NVME_TCP_SEND_DDGST;
 912                                 req->offset = 0;
 913                         } else {
 914                                 nvme_tcp_done_send_req(queue);
 915                         }
 916                         return 1;
 917                 }
 918         }
 919         return -EAGAIN;
 920 }
 921
 922 static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
 923 {
 924         struct nvme_tcp_queue *queue = req->queue;
 925         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
 926         bool inline_data = nvme_tcp_has_inline_data(req);
 927         u8 hdgst = nvme_tcp_hdgst_len(queue);
 928         int len = sizeof(*pdu) + hdgst - req->offset;
 929         int flags = MSG_DONTWAIT;
 930         int ret;
 931
 932         if (inline_data)
 933                 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
 934         else
 935                 flags |= MSG_EOR;
 936
 937         if (queue->hdr_digest && !req->offset)
 938                 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
 939
 940         ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
 941                         offset_in_page(pdu) + req->offset, len,  flags);
 942         if (unlikely(ret <= 0))
 943                 return ret;
 944
 945         len -= ret;
 946         if (!len) {
 947                 if (inline_data) {
 948                         req->state = NVME_TCP_SEND_DATA;
 949                         if (queue->data_digest)
 950                                 crypto_ahash_init(queue->snd_hash);
 951                         nvme_tcp_init_iter(req, WRITE);
 952                 } else {
 953                         nvme_tcp_done_send_req(queue);
 954                 }
 955                 return 1;
 956         }
 957         req->offset += ret;
 958
 959         return -EAGAIN;
 960 }
 961
 962 static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
 963 {
 964         struct nvme_tcp_queue *queue = req->queue;
 965         struct nvme_tcp_data_pdu *pdu = req->pdu;
 966         u8 hdgst = nvme_tcp_hdgst_len(queue);
 967         int len = sizeof(*pdu) - req->offset + hdgst;
 968         int ret;
 969
 970         if (queue->hdr_digest && !req->offset)
 971                 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
 972
 973         ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
 974                         offset_in_page(pdu) + req->offset, len,
 975                         MSG_DONTWAIT | MSG_MORE | MSG_SENDPAGE_NOTLAST);
 976         if (unlikely(ret <= 0))
 977                 return ret;
 978
 979         len -= ret;
 980         if (!len) {
 981                 req->state = NVME_TCP_SEND_DATA;
 982                 if (queue->data_digest)
 983                         crypto_ahash_init(queue->snd_hash);
 984                 if (!req->data_sent)
 985                         nvme_tcp_init_iter(req, WRITE);
 986                 return 1;
 987         }
 988         req->offset += ret;
 989
 990         return -EAGAIN;
 991 }
 992
 993 static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
 994 {
 995         struct nvme_tcp_queue *queue = req->queue;
 996         int ret;
 997         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR };
 998         struct kvec iov = {
 999                 .iov_base = &req->ddgst + req->offset,
1000                 .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
1001         };
1002
1003         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1004         if (unlikely(ret <= 0))
1005                 return ret;
1006
1007         if (req->offset + ret == NVME_TCP_DIGEST_LENGTH) {
1008                 nvme_tcp_done_send_req(queue);
1009                 return 1;
1010         }
1011
1012         req->offset += ret;
1013         return -EAGAIN;
1014 }
1015
1016 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
1017 {
1018         struct nvme_tcp_request *req;
1019         int ret = 1;
1020
1021         if (!queue->request) {
1022                 queue->request = nvme_tcp_fetch_request(queue);
1023                 if (!queue->request)
1024                         return 0;
1025         }
1026         req = queue->request;
1027
1028         if (req->state == NVME_TCP_SEND_CMD_PDU) {
1029                 ret = nvme_tcp_try_send_cmd_pdu(req);
1030                 if (ret <= 0)
1031                         goto done;
1032                 if (!nvme_tcp_has_inline_data(req))
1033                         return ret;
1034         }
1035
1036         if (req->state == NVME_TCP_SEND_H2C_PDU) {
1037                 ret = nvme_tcp_try_send_data_pdu(req);
1038                 if (ret <= 0)
1039                         goto done;
1040         }
1041
1042         if (req->state == NVME_TCP_SEND_DATA) {
1043                 ret = nvme_tcp_try_send_data(req);
1044                 if (ret <= 0)
1045                         goto done;
1046         }
1047
1048         if (req->state == NVME_TCP_SEND_DDGST)
1049                 ret = nvme_tcp_try_send_ddgst(req);
1050 done:
1051         if (ret == -EAGAIN) {
1052                 ret = 0;
1053         } else if (ret < 0) {
1054                 dev_err(queue->ctrl->ctrl.device,
1055                         "failed to send request %d\n", ret);
1056                 if (ret != -EPIPE && ret != -ECONNRESET)
1057                         nvme_tcp_fail_request(queue->request);
1058                 nvme_tcp_done_send_req(queue);
1059         }
1060         return ret;
1061 }
1062
1063 static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1064 {
1065         struct socket *sock = queue->sock;
1066         struct sock *sk = sock->sk;
1067         read_descriptor_t rd_desc;
1068         int consumed;
1069
1070         rd_desc.arg.data = queue;
1071         rd_desc.count = 1;
1072         lock_sock(sk);
1073         queue->nr_cqe = 0;
1074         consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1075         release_sock(sk);
1076         return consumed;
1077 }
1078
1079 static void nvme_tcp_io_work(struct work_struct *w)
1080 {
1081         struct nvme_tcp_queue *queue =
1082                 container_of(w, struct nvme_tcp_queue, io_work);
1083         unsigned long deadline = jiffies + msecs_to_jiffies(1);
1084
1085         do {
1086                 bool pending = false;
1087                 int result;
1088
1089                 if (mutex_trylock(&queue->send_mutex)) {
1090                         result = nvme_tcp_try_send(queue);
1091                         mutex_unlock(&queue->send_mutex);
1092                         if (result > 0)
1093                                 pending = true;
1094                         else if (unlikely(result < 0))
1095                                 break;
1096                 }
1097
1098                 result = nvme_tcp_try_recv(queue);
1099                 if (result > 0)
1100                         pending = true;
1101                 else if (unlikely(result < 0))
1102                         return;
1103
1104                 if (!pending)
1105                         return;
1106
1107         } while (!time_after(jiffies, deadline)); /* quota is exhausted */
1108
1109         queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1110 }
1111
1112 static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue)
1113 {
1114         struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
1115
1116         ahash_request_free(queue->rcv_hash);
1117         ahash_request_free(queue->snd_hash);
1118         crypto_free_ahash(tfm);
1119 }
1120
1121 static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue)
1122 {
1123         struct crypto_ahash *tfm;
1124
1125         tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
1126         if (IS_ERR(tfm))
1127                 return PTR_ERR(tfm);
1128
1129         queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1130         if (!queue->snd_hash)
1131                 goto free_tfm;
1132         ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
1133
1134         queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1135         if (!queue->rcv_hash)
1136                 goto free_snd_hash;
1137         ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
1138
1139         return 0;
1140 free_snd_hash:
1141         ahash_request_free(queue->snd_hash);
1142 free_tfm:
1143         crypto_free_ahash(tfm);
1144         return -ENOMEM;
1145 }
1146
1147 static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1148 {
1149         struct nvme_tcp_request *async = &ctrl->async_req;
1150
1151         page_frag_free(async->pdu);
1152 }
1153
1154 static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1155 {
1156         struct nvme_tcp_queue *queue = &ctrl->queues[0];
1157         struct nvme_tcp_request *async = &ctrl->async_req;
1158         u8 hdgst = nvme_tcp_hdgst_len(queue);
1159
1160         async->pdu = page_frag_alloc(&queue->pf_cache,
1161                 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1162                 GFP_KERNEL | __GFP_ZERO);
1163         if (!async->pdu)
1164                 return -ENOMEM;
1165
1166         async->queue = &ctrl->queues[0];
1167         return 0;
1168 }
1169
1170 static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1171 {
1172         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1173         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1174
1175         if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1176                 return;
1177
1178         if (queue->hdr_digest || queue->data_digest)
1179                 nvme_tcp_free_crypto(queue);
1180
1181         sock_release(queue->sock);
1182         kfree(queue->pdu);
1183 }
1184
1185 static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1186 {
1187         struct nvme_tcp_icreq_pdu *icreq;
1188         struct nvme_tcp_icresp_pdu *icresp;
1189         struct msghdr msg = {};
1190         struct kvec iov;
1191         bool ctrl_hdgst, ctrl_ddgst;
1192         int ret;
1193
1194         icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1195         if (!icreq)
1196                 return -ENOMEM;
1197
1198         icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1199         if (!icresp) {
1200                 ret = -ENOMEM;
1201                 goto free_icreq;
1202         }
1203
1204         icreq->hdr.type = nvme_tcp_icreq;
1205         icreq->hdr.hlen = sizeof(*icreq);
1206         icreq->hdr.pdo = 0;
1207         icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1208         icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1209         icreq->maxr2t = 0; /* single inflight r2t supported */
1210         icreq->hpda = 0; /* no alignment constraint */
1211         if (queue->hdr_digest)
1212                 icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1213         if (queue->data_digest)
1214                 icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1215
1216         iov.iov_base = icreq;
1217         iov.iov_len = sizeof(*icreq);
1218         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1219         if (ret < 0)
1220                 goto free_icresp;
1221
1222         memset(&msg, 0, sizeof(msg));
1223         iov.iov_base = icresp;
1224         iov.iov_len = sizeof(*icresp);
1225         ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1226                         iov.iov_len, msg.msg_flags);
1227         if (ret < 0)
1228                 goto free_icresp;
1229
1230         ret = -EINVAL;
1231         if (icresp->hdr.type != nvme_tcp_icresp) {
1232                 pr_err("queue %d: bad type returned %d\n",
1233                         nvme_tcp_queue_id(queue), icresp->hdr.type);
1234                 goto free_icresp;
1235         }
1236
1237         if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1238                 pr_err("queue %d: bad pdu length returned %d\n",
1239                         nvme_tcp_queue_id(queue), icresp->hdr.plen);
1240                 goto free_icresp;
1241         }
1242
1243         if (icresp->pfv != NVME_TCP_PFV_1_0) {
1244                 pr_err("queue %d: bad pfv returned %d\n",
1245                         nvme_tcp_queue_id(queue), icresp->pfv);
1246                 goto free_icresp;
1247         }
1248
1249         ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1250         if ((queue->data_digest && !ctrl_ddgst) ||
1251             (!queue->data_digest && ctrl_ddgst)) {
1252                 pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1253                         nvme_tcp_queue_id(queue),
1254                         queue->data_digest ? "enabled" : "disabled",
1255                         ctrl_ddgst ? "enabled" : "disabled");
1256                 goto free_icresp;
1257         }
1258
1259         ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1260         if ((queue->hdr_digest && !ctrl_hdgst) ||
1261             (!queue->hdr_digest && ctrl_hdgst)) {
1262                 pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1263                         nvme_tcp_queue_id(queue),
1264                         queue->hdr_digest ? "enabled" : "disabled",
1265                         ctrl_hdgst ? "enabled" : "disabled");
1266                 goto free_icresp;
1267         }
1268
1269         if (icresp->cpda != 0) {
1270                 pr_err("queue %d: unsupported cpda returned %d\n",
1271                         nvme_tcp_queue_id(queue), icresp->cpda);
1272                 goto free_icresp;
1273         }
1274
1275         ret = 0;
1276 free_icresp:
1277         kfree(icresp);
1278 free_icreq:
1279         kfree(icreq);
1280         return ret;
1281 }
1282
1283 static bool nvme_tcp_admin_queue(struct nvme_tcp_queue *queue)
1284 {
1285         return nvme_tcp_queue_id(queue) == 0;
1286 }
1287
1288 static bool nvme_tcp_default_queue(struct nvme_tcp_queue *queue)
1289 {
1290         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1291         int qid = nvme_tcp_queue_id(queue);
1292
1293         return !nvme_tcp_admin_queue(queue) &&
1294                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT];
1295 }
1296
1297 static bool nvme_tcp_read_queue(struct nvme_tcp_queue *queue)
1298 {
1299         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1300         int qid = nvme_tcp_queue_id(queue);
1301
1302         return !nvme_tcp_admin_queue(queue) &&
1303                 !nvme_tcp_default_queue(queue) &&
1304                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1305                           ctrl->io_queues[HCTX_TYPE_READ];
1306 }
1307
1308 static bool nvme_tcp_poll_queue(struct nvme_tcp_queue *queue)
1309 {
1310         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1311         int qid = nvme_tcp_queue_id(queue);
1312
1313         return !nvme_tcp_admin_queue(queue) &&
1314                 !nvme_tcp_default_queue(queue) &&
1315                 !nvme_tcp_read_queue(queue) &&
1316                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1317                           ctrl->io_queues[HCTX_TYPE_READ] +
1318                           ctrl->io_queues[HCTX_TYPE_POLL];
1319 }
1320
1321 static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
1322 {
1323         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1324         int qid = nvme_tcp_queue_id(queue);
1325         int n = 0;
1326
1327         if (nvme_tcp_default_queue(queue))
1328                 n = qid - 1;
1329         else if (nvme_tcp_read_queue(queue))
1330                 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] - 1;
1331         else if (nvme_tcp_poll_queue(queue))
1332                 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] -
1333                                 ctrl->io_queues[HCTX_TYPE_READ] - 1;
1334         queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
1335 }
1336
1337 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl,
1338                 int qid, size_t queue_size)
1339 {
1340         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1341         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1342         int ret, rcv_pdu_size;
1343
1344         queue->ctrl = ctrl;
1345         INIT_LIST_HEAD(&queue->send_list);
1346         spin_lock_init(&queue->lock);
1347         mutex_init(&queue->send_mutex);
1348         INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1349         queue->queue_size = queue_size;
1350
1351         if (qid > 0)
1352                 queue->cmnd_capsule_len = nctrl->ioccsz * 16;
1353         else
1354                 queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1355                                                 NVME_TCP_ADMIN_CCSZ;
1356
1357         ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM,
1358                         IPPROTO_TCP, &queue->sock);
1359         if (ret) {
1360                 dev_err(nctrl->device,
1361                         "failed to create socket: %d\n", ret);
1362                 return ret;
1363         }
1364
1365         /* Single syn retry */
1366         tcp_sock_set_syncnt(queue->sock->sk, 1);
1367
1368         /* Set TCP no delay */
1369         tcp_sock_set_nodelay(queue->sock->sk);
1370
1371         /*
1372          * Cleanup whatever is sitting in the TCP transmit queue on socket
1373          * close. This is done to prevent stale data from being sent should
1374          * the network connection be restored before TCP times out.
1375          */
1376         sock_no_linger(queue->sock->sk);
1377
1378         if (so_priority > 0)
1379                 sock_set_priority(queue->sock->sk, so_priority);
1380
1381         /* Set socket type of service */
1382         if (nctrl->opts->tos >= 0)
1383                 ip_sock_set_tos(queue->sock->sk, nctrl->opts->tos);
1384
1385         queue->sock->sk->sk_allocation = GFP_ATOMIC;
1386         nvme_tcp_set_queue_io_cpu(queue);
1387         queue->request = NULL;
1388         queue->data_remaining = 0;
1389         queue->ddgst_remaining = 0;
1390         queue->pdu_remaining = 0;
1391         queue->pdu_offset = 0;
1392         sk_set_memalloc(queue->sock->sk);
1393
1394         if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
1395                 ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1396                         sizeof(ctrl->src_addr));
1397                 if (ret) {
1398                         dev_err(nctrl->device,
1399                                 "failed to bind queue %d socket %d\n",
1400                                 qid, ret);
1401                         goto err_sock;
1402                 }
1403         }
1404
1405         queue->hdr_digest = nctrl->opts->hdr_digest;
1406         queue->data_digest = nctrl->opts->data_digest;
1407         if (queue->hdr_digest || queue->data_digest) {
1408                 ret = nvme_tcp_alloc_crypto(queue);
1409                 if (ret) {
1410                         dev_err(nctrl->device,
1411                                 "failed to allocate queue %d crypto\n", qid);
1412                         goto err_sock;
1413                 }
1414         }
1415
1416         rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1417                         nvme_tcp_hdgst_len(queue);
1418         queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1419         if (!queue->pdu) {
1420                 ret = -ENOMEM;
1421                 goto err_crypto;
1422         }
1423
1424         dev_dbg(nctrl->device, "connecting queue %d\n",
1425                         nvme_tcp_queue_id(queue));
1426
1427         ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1428                 sizeof(ctrl->addr), 0);
1429         if (ret) {
1430                 dev_err(nctrl->device,
1431                         "failed to connect socket: %d\n", ret);
1432                 goto err_rcv_pdu;
1433         }
1434
1435         ret = nvme_tcp_init_connection(queue);
1436         if (ret)
1437                 goto err_init_connect;
1438
1439         queue->rd_enabled = true;
1440         set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1441         nvme_tcp_init_recv_ctx(queue);
1442
1443         write_lock_bh(&queue->sock->sk->sk_callback_lock);
1444         queue->sock->sk->sk_user_data = queue;
1445         queue->state_change = queue->sock->sk->sk_state_change;
1446         queue->data_ready = queue->sock->sk->sk_data_ready;
1447         queue->write_space = queue->sock->sk->sk_write_space;
1448         queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1449         queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1450         queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1451 #ifdef CONFIG_NET_RX_BUSY_POLL
1452         queue->sock->sk->sk_ll_usec = 1;
1453 #endif
1454         write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1455
1456         return 0;
1457
1458 err_init_connect:
1459         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1460 err_rcv_pdu:
1461         kfree(queue->pdu);
1462 err_crypto:
1463         if (queue->hdr_digest || queue->data_digest)
1464                 nvme_tcp_free_crypto(queue);
1465 err_sock:
1466         sock_release(queue->sock);
1467         queue->sock = NULL;
1468         return ret;
1469 }
1470
1471 static void nvme_tcp_restore_sock_calls(struct nvme_tcp_queue *queue)
1472 {
1473         struct socket *sock = queue->sock;
1474
1475         write_lock_bh(&sock->sk->sk_callback_lock);
1476         sock->sk->sk_user_data  = NULL;
1477         sock->sk->sk_data_ready = queue->data_ready;
1478         sock->sk->sk_state_change = queue->state_change;
1479         sock->sk->sk_write_space  = queue->write_space;
1480         write_unlock_bh(&sock->sk->sk_callback_lock);
1481 }
1482
1483 static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1484 {
1485         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1486         nvme_tcp_restore_sock_calls(queue);
1487         cancel_work_sync(&queue->io_work);
1488 }
1489
1490 static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1491 {
1492         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1493         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1494
1495         if (!test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1496                 return;
1497
1498         __nvme_tcp_stop_queue(queue);
1499 }
1500
1501 static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1502 {
1503         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1504         int ret;
1505
1506         if (idx)
1507                 ret = nvmf_connect_io_queue(nctrl, idx, false);
1508         else
1509                 ret = nvmf_connect_admin_queue(nctrl);
1510
1511         if (!ret) {
1512                 set_bit(NVME_TCP_Q_LIVE, &ctrl->queues[idx].flags);
1513         } else {
1514                 if (test_bit(NVME_TCP_Q_ALLOCATED, &ctrl->queues[idx].flags))
1515                         __nvme_tcp_stop_queue(&ctrl->queues[idx]);
1516                 dev_err(nctrl->device,
1517                         "failed to connect queue: %d ret=%d\n", idx, ret);
1518         }
1519         return ret;
1520 }
1521
1522 static struct blk_mq_tag_set *nvme_tcp_alloc_tagset(struct nvme_ctrl *nctrl,
1523                 bool admin)
1524 {
1525         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1526         struct blk_mq_tag_set *set;
1527         int ret;
1528
1529         if (admin) {
1530                 set = &ctrl->admin_tag_set;
1531                 memset(set, 0, sizeof(*set));
1532                 set->ops = &nvme_tcp_admin_mq_ops;
1533                 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1534                 set->reserved_tags = 2; /* connect + keep-alive */
1535                 set->numa_node = NUMA_NO_NODE;
1536                 set->flags = BLK_MQ_F_BLOCKING;
1537                 set->cmd_size = sizeof(struct nvme_tcp_request);
1538                 set->driver_data = ctrl;
1539                 set->nr_hw_queues = 1;
1540                 set->timeout = ADMIN_TIMEOUT;
1541         } else {
1542                 set = &ctrl->tag_set;
1543                 memset(set, 0, sizeof(*set));
1544                 set->ops = &nvme_tcp_mq_ops;
1545                 set->queue_depth = nctrl->sqsize + 1;
1546                 set->reserved_tags = 1; /* fabric connect */
1547                 set->numa_node = NUMA_NO_NODE;
1548                 set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_BLOCKING;
1549                 set->cmd_size = sizeof(struct nvme_tcp_request);
1550                 set->driver_data = ctrl;
1551                 set->nr_hw_queues = nctrl->queue_count - 1;
1552                 set->timeout = NVME_IO_TIMEOUT;
1553                 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
1554         }
1555
1556         ret = blk_mq_alloc_tag_set(set);
1557         if (ret)
1558                 return ERR_PTR(ret);
1559
1560         return set;
1561 }
1562
1563 static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
1564 {
1565         if (to_tcp_ctrl(ctrl)->async_req.pdu) {
1566                 nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
1567                 to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
1568         }
1569
1570         nvme_tcp_free_queue(ctrl, 0);
1571 }
1572
1573 static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
1574 {
1575         int i;
1576
1577         for (i = 1; i < ctrl->queue_count; i++)
1578                 nvme_tcp_free_queue(ctrl, i);
1579 }
1580
1581 static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
1582 {
1583         int i;
1584
1585         for (i = 1; i < ctrl->queue_count; i++)
1586                 nvme_tcp_stop_queue(ctrl, i);
1587 }
1588
1589 static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl)
1590 {
1591         int i, ret = 0;
1592
1593         for (i = 1; i < ctrl->queue_count; i++) {
1594                 ret = nvme_tcp_start_queue(ctrl, i);
1595                 if (ret)
1596                         goto out_stop_queues;
1597         }
1598
1599         return 0;
1600
1601 out_stop_queues:
1602         for (i--; i >= 1; i--)
1603                 nvme_tcp_stop_queue(ctrl, i);
1604         return ret;
1605 }
1606
1607 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
1608 {
1609         int ret;
1610
1611         ret = nvme_tcp_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
1612         if (ret)
1613                 return ret;
1614
1615         ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
1616         if (ret)
1617                 goto out_free_queue;
1618
1619         return 0;
1620
1621 out_free_queue:
1622         nvme_tcp_free_queue(ctrl, 0);
1623         return ret;
1624 }
1625
1626 static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1627 {
1628         int i, ret;
1629
1630         for (i = 1; i < ctrl->queue_count; i++) {
1631                 ret = nvme_tcp_alloc_queue(ctrl, i,
1632                                 ctrl->sqsize + 1);
1633                 if (ret)
1634                         goto out_free_queues;
1635         }
1636
1637         return 0;
1638
1639 out_free_queues:
1640         for (i--; i >= 1; i--)
1641                 nvme_tcp_free_queue(ctrl, i);
1642
1643         return ret;
1644 }
1645
1646 static unsigned int nvme_tcp_nr_io_queues(struct nvme_ctrl *ctrl)
1647 {
1648         unsigned int nr_io_queues;
1649
1650         nr_io_queues = min(ctrl->opts->nr_io_queues, num_online_cpus());
1651         nr_io_queues += min(ctrl->opts->nr_write_queues, num_online_cpus());
1652         nr_io_queues += min(ctrl->opts->nr_poll_queues, num_online_cpus());
1653
1654         return nr_io_queues;
1655 }
1656
1657 static void nvme_tcp_set_io_queues(struct nvme_ctrl *nctrl,
1658                 unsigned int nr_io_queues)
1659 {
1660         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1661         struct nvmf_ctrl_options *opts = nctrl->opts;
1662
1663         if (opts->nr_write_queues && opts->nr_io_queues < nr_io_queues) {
1664                 /*
1665                  * separate read/write queues
1666                  * hand out dedicated default queues only after we have
1667                  * sufficient read queues.
1668                  */
1669                 ctrl->io_queues[HCTX_TYPE_READ] = opts->nr_io_queues;
1670                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
1671                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1672                         min(opts->nr_write_queues, nr_io_queues);
1673                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1674         } else {
1675                 /*
1676                  * shared read/write queues
1677                  * either no write queues were requested, or we don't have
1678                  * sufficient queue count to have dedicated default queues.
1679                  */
1680                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1681                         min(opts->nr_io_queues, nr_io_queues);
1682                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1683         }
1684
1685         if (opts->nr_poll_queues && nr_io_queues) {
1686                 /* map dedicated poll queues only if we have queues left */
1687                 ctrl->io_queues[HCTX_TYPE_POLL] =
1688                         min(opts->nr_poll_queues, nr_io_queues);
1689         }
1690 }
1691
1692 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1693 {
1694         unsigned int nr_io_queues;
1695         int ret;
1696
1697         nr_io_queues = nvme_tcp_nr_io_queues(ctrl);
1698         ret = nvme_set_queue_count(ctrl, &nr_io_queues);
1699         if (ret)
1700                 return ret;
1701
1702         ctrl->queue_count = nr_io_queues + 1;
1703         if (ctrl->queue_count < 2)
1704                 return 0;
1705
1706         dev_info(ctrl->device,
1707                 "creating %d I/O queues.\n", nr_io_queues);
1708
1709         nvme_tcp_set_io_queues(ctrl, nr_io_queues);
1710
1711         return __nvme_tcp_alloc_io_queues(ctrl);
1712 }
1713
1714 static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove)
1715 {
1716         nvme_tcp_stop_io_queues(ctrl);
1717         if (remove) {
1718                 blk_cleanup_queue(ctrl->connect_q);
1719                 blk_mq_free_tag_set(ctrl->tagset);
1720         }
1721         nvme_tcp_free_io_queues(ctrl);
1722 }
1723
1724 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
1725 {
1726         int ret;
1727
1728         ret = nvme_tcp_alloc_io_queues(ctrl);
1729         if (ret)
1730                 return ret;
1731
1732         if (new) {
1733                 ctrl->tagset = nvme_tcp_alloc_tagset(ctrl, false);
1734                 if (IS_ERR(ctrl->tagset)) {
1735                         ret = PTR_ERR(ctrl->tagset);
1736                         goto out_free_io_queues;
1737                 }
1738
1739                 ctrl->connect_q = blk_mq_init_queue(ctrl->tagset);
1740                 if (IS_ERR(ctrl->connect_q)) {
1741                         ret = PTR_ERR(ctrl->connect_q);
1742                         goto out_free_tag_set;
1743                 }
1744         } else {
1745                 blk_mq_update_nr_hw_queues(ctrl->tagset,
1746                         ctrl->queue_count - 1);
1747         }
1748
1749         ret = nvme_tcp_start_io_queues(ctrl);
1750         if (ret)
1751                 goto out_cleanup_connect_q;
1752
1753         return 0;
1754
1755 out_cleanup_connect_q:
1756         if (new)
1757                 blk_cleanup_queue(ctrl->connect_q);
1758 out_free_tag_set:
1759         if (new)
1760                 blk_mq_free_tag_set(ctrl->tagset);
1761 out_free_io_queues:
1762         nvme_tcp_free_io_queues(ctrl);
1763         return ret;
1764 }
1765
1766 static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove)
1767 {
1768         nvme_tcp_stop_queue(ctrl, 0);
1769         if (remove) {
1770                 blk_cleanup_queue(ctrl->admin_q);
1771                 blk_cleanup_queue(ctrl->fabrics_q);
1772                 blk_mq_free_tag_set(ctrl->admin_tagset);
1773         }
1774         nvme_tcp_free_admin_queue(ctrl);
1775 }
1776
1777 static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
1778 {
1779         int error;
1780
1781         error = nvme_tcp_alloc_admin_queue(ctrl);
1782         if (error)
1783                 return error;
1784
1785         if (new) {
1786                 ctrl->admin_tagset = nvme_tcp_alloc_tagset(ctrl, true);
1787                 if (IS_ERR(ctrl->admin_tagset)) {
1788                         error = PTR_ERR(ctrl->admin_tagset);
1789                         goto out_free_queue;
1790                 }
1791
1792                 ctrl->fabrics_q = blk_mq_init_queue(ctrl->admin_tagset);
1793                 if (IS_ERR(ctrl->fabrics_q)) {
1794                         error = PTR_ERR(ctrl->fabrics_q);
1795                         goto out_free_tagset;
1796                 }
1797
1798                 ctrl->admin_q = blk_mq_init_queue(ctrl->admin_tagset);
1799                 if (IS_ERR(ctrl->admin_q)) {
1800                         error = PTR_ERR(ctrl->admin_q);
1801                         goto out_cleanup_fabrics_q;
1802                 }
1803         }
1804
1805         error = nvme_tcp_start_queue(ctrl, 0);
1806         if (error)
1807                 goto out_cleanup_queue;
1808
1809         error = nvme_enable_ctrl(ctrl);
1810         if (error)
1811                 goto out_stop_queue;
1812
1813         blk_mq_unquiesce_queue(ctrl->admin_q);
1814
1815         error = nvme_init_identify(ctrl);
1816         if (error)
1817                 goto out_stop_queue;
1818
1819         return 0;
1820
1821 out_stop_queue:
1822         nvme_tcp_stop_queue(ctrl, 0);
1823 out_cleanup_queue:
1824         if (new)
1825                 blk_cleanup_queue(ctrl->admin_q);
1826 out_cleanup_fabrics_q:
1827         if (new)
1828                 blk_cleanup_queue(ctrl->fabrics_q);
1829 out_free_tagset:
1830         if (new)
1831                 blk_mq_free_tag_set(ctrl->admin_tagset);
1832 out_free_queue:
1833         nvme_tcp_free_admin_queue(ctrl);
1834         return error;
1835 }
1836
1837 static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
1838                 bool remove)
1839 {
1840         blk_mq_quiesce_queue(ctrl->admin_q);
1841         nvme_tcp_stop_queue(ctrl, 0);
1842         if (ctrl->admin_tagset) {
1843                 blk_mq_tagset_busy_iter(ctrl->admin_tagset,
1844                         nvme_cancel_request, ctrl);
1845                 blk_mq_tagset_wait_completed_request(ctrl->admin_tagset);
1846         }
1847         if (remove)
1848                 blk_mq_unquiesce_queue(ctrl->admin_q);
1849         nvme_tcp_destroy_admin_queue(ctrl, remove);
1850 }
1851
1852 static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
1853                 bool remove)
1854 {
1855         if (ctrl->queue_count <= 1)
1856                 return;
1857         nvme_stop_queues(ctrl);
1858         nvme_tcp_stop_io_queues(ctrl);
1859         if (ctrl->tagset) {
1860                 blk_mq_tagset_busy_iter(ctrl->tagset,
1861                         nvme_cancel_request, ctrl);
1862                 blk_mq_tagset_wait_completed_request(ctrl->tagset);
1863         }
1864         if (remove)
1865                 nvme_start_queues(ctrl);
1866         nvme_tcp_destroy_io_queues(ctrl, remove);
1867 }
1868
1869 static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
1870 {
1871         /* If we are resetting/deleting then do nothing */
1872         if (ctrl->state != NVME_CTRL_CONNECTING) {
1873                 WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW ||
1874                         ctrl->state == NVME_CTRL_LIVE);
1875                 return;
1876         }
1877
1878         if (nvmf_should_reconnect(ctrl)) {
1879                 dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
1880                         ctrl->opts->reconnect_delay);
1881                 queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
1882                                 ctrl->opts->reconnect_delay * HZ);
1883         } else {
1884                 dev_info(ctrl->device, "Removing controller...\n");
1885                 nvme_delete_ctrl(ctrl);
1886         }
1887 }
1888
1889 static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
1890 {
1891         struct nvmf_ctrl_options *opts = ctrl->opts;
1892         int ret;
1893
1894         ret = nvme_tcp_configure_admin_queue(ctrl, new);
1895         if (ret)
1896                 return ret;
1897
1898         if (ctrl->icdoff) {
1899                 dev_err(ctrl->device, "icdoff is not supported!\n");
1900                 goto destroy_admin;
1901         }
1902
1903         if (opts->queue_size > ctrl->sqsize + 1)
1904                 dev_warn(ctrl->device,
1905                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
1906                         opts->queue_size, ctrl->sqsize + 1);
1907
1908         if (ctrl->sqsize + 1 > ctrl->maxcmd) {
1909                 dev_warn(ctrl->device,
1910                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
1911                         ctrl->sqsize + 1, ctrl->maxcmd);
1912                 ctrl->sqsize = ctrl->maxcmd - 1;
1913         }
1914
1915         if (ctrl->queue_count > 1) {
1916                 ret = nvme_tcp_configure_io_queues(ctrl, new);
1917                 if (ret)
1918                         goto destroy_admin;
1919         }
1920
1921         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
1922                 /*
1923                  * state change failure is ok if we're in DELETING state,
1924                  * unless we're during creation of a new controller to
1925                  * avoid races with teardown flow.
1926                  */
1927                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
1928                 WARN_ON_ONCE(new);
1929                 ret = -EINVAL;
1930                 goto destroy_io;
1931         }
1932
1933         nvme_start_ctrl(ctrl);
1934         return 0;
1935
1936 destroy_io:
1937         if (ctrl->queue_count > 1)
1938                 nvme_tcp_destroy_io_queues(ctrl, new);
1939 destroy_admin:
1940         nvme_tcp_stop_queue(ctrl, 0);
1941         nvme_tcp_destroy_admin_queue(ctrl, new);
1942         return ret;
1943 }
1944
1945 static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
1946 {
1947         struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
1948                         struct nvme_tcp_ctrl, connect_work);
1949         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
1950
1951         ++ctrl->nr_reconnects;
1952
1953         if (nvme_tcp_setup_ctrl(ctrl, false))
1954                 goto requeue;
1955
1956         dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n",
1957                         ctrl->nr_reconnects);
1958
1959         ctrl->nr_reconnects = 0;
1960
1961         return;
1962
1963 requeue:
1964         dev_info(ctrl->device, "Failed reconnect attempt %d\n",
1965                         ctrl->nr_reconnects);
1966         nvme_tcp_reconnect_or_remove(ctrl);
1967 }
1968
1969 static void nvme_tcp_error_recovery_work(struct work_struct *work)
1970 {
1971         struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
1972                                 struct nvme_tcp_ctrl, err_work);
1973         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
1974
1975         nvme_stop_keep_alive(ctrl);
1976         nvme_tcp_teardown_io_queues(ctrl, false);
1977         /* unquiesce to fail fast pending requests */
1978         nvme_start_queues(ctrl);
1979         nvme_tcp_teardown_admin_queue(ctrl, false);
1980         blk_mq_unquiesce_queue(ctrl->admin_q);
1981
1982         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
1983                 /* state change failure is ok if we're in DELETING state */
1984                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
1985                 return;
1986         }
1987
1988         nvme_tcp_reconnect_or_remove(ctrl);
1989 }
1990
1991 static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
1992 {
1993         cancel_work_sync(&to_tcp_ctrl(ctrl)->err_work);
1994         cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
1995
1996         nvme_tcp_teardown_io_queues(ctrl, shutdown);
1997         blk_mq_quiesce_queue(ctrl->admin_q);
1998         if (shutdown)
1999                 nvme_shutdown_ctrl(ctrl);
2000         else
2001                 nvme_disable_ctrl(ctrl);
2002         nvme_tcp_teardown_admin_queue(ctrl, shutdown);
2003 }
2004
2005 static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
2006 {
2007         nvme_tcp_teardown_ctrl(ctrl, true);
2008 }
2009
2010 static void nvme_reset_ctrl_work(struct work_struct *work)
2011 {
2012         struct nvme_ctrl *ctrl =
2013                 container_of(work, struct nvme_ctrl, reset_work);
2014
2015         nvme_stop_ctrl(ctrl);
2016         nvme_tcp_teardown_ctrl(ctrl, false);
2017
2018         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2019                 /* state change failure is ok if we're in DELETING state */
2020                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
2021                 return;
2022         }
2023
2024         if (nvme_tcp_setup_ctrl(ctrl, false))
2025                 goto out_fail;
2026
2027         return;
2028
2029 out_fail:
2030         ++ctrl->nr_reconnects;
2031         nvme_tcp_reconnect_or_remove(ctrl);
2032 }
2033
2034 static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
2035 {
2036         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
2037
2038         if (list_empty(&ctrl->list))
2039                 goto free_ctrl;
2040
2041         mutex_lock(&nvme_tcp_ctrl_mutex);
2042         list_del(&ctrl->list);
2043         mutex_unlock(&nvme_tcp_ctrl_mutex);
2044
2045         nvmf_free_options(nctrl->opts);
2046 free_ctrl:
2047         kfree(ctrl->queues);
2048         kfree(ctrl);
2049 }
2050
2051 static void nvme_tcp_set_sg_null(struct nvme_command *c)
2052 {
2053         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2054
2055         sg->addr = 0;
2056         sg->length = 0;
2057         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2058                         NVME_SGL_FMT_TRANSPORT_A;
2059 }
2060
2061 static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
2062                 struct nvme_command *c, u32 data_len)
2063 {
2064         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2065
2066         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
2067         sg->length = cpu_to_le32(data_len);
2068         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
2069 }
2070
2071 static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
2072                 u32 data_len)
2073 {
2074         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2075
2076         sg->addr = 0;
2077         sg->length = cpu_to_le32(data_len);
2078         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2079                         NVME_SGL_FMT_TRANSPORT_A;
2080 }
2081
2082 static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
2083 {
2084         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2085         struct nvme_tcp_queue *queue = &ctrl->queues[0];
2086         struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2087         struct nvme_command *cmd = &pdu->cmd;
2088         u8 hdgst = nvme_tcp_hdgst_len(queue);
2089
2090         memset(pdu, 0, sizeof(*pdu));
2091         pdu->hdr.type = nvme_tcp_cmd;
2092         if (queue->hdr_digest)
2093                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2094         pdu->hdr.hlen = sizeof(*pdu);
2095         pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2096
2097         cmd->common.opcode = nvme_admin_async_event;
2098         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2099         cmd->common.flags |= NVME_CMD_SGL_METABUF;
2100         nvme_tcp_set_sg_null(cmd);
2101
2102         ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2103         ctrl->async_req.offset = 0;
2104         ctrl->async_req.curr_bio = NULL;
2105         ctrl->async_req.data_len = 0;
2106
2107         nvme_tcp_queue_request(&ctrl->async_req, true);
2108 }
2109
2110 static enum blk_eh_timer_return
2111 nvme_tcp_timeout(struct request *rq, bool reserved)
2112 {
2113         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2114         struct nvme_tcp_ctrl *ctrl = req->queue->ctrl;
2115         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2116
2117         /*
2118          * Restart the timer if a controller reset is already scheduled. Any
2119          * timed out commands would be handled before entering the connecting
2120          * state.
2121          */
2122         if (ctrl->ctrl.state == NVME_CTRL_RESETTING)
2123                 return BLK_EH_RESET_TIMER;
2124
2125         dev_warn(ctrl->ctrl.device,
2126                 "queue %d: timeout request %#x type %d\n",
2127                 nvme_tcp_queue_id(req->queue), rq->tag, pdu->hdr.type);
2128
2129         if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
2130                 /*
2131                  * Teardown immediately if controller times out while starting
2132                  * or we are already started error recovery. all outstanding
2133                  * requests are completed on shutdown, so we return BLK_EH_DONE.
2134                  */
2135                 flush_work(&ctrl->err_work);
2136                 nvme_tcp_teardown_io_queues(&ctrl->ctrl, false);
2137                 nvme_tcp_teardown_admin_queue(&ctrl->ctrl, false);
2138                 return BLK_EH_DONE;
2139         }
2140
2141         dev_warn(ctrl->ctrl.device, "starting error recovery\n");
2142         nvme_tcp_error_recovery(&ctrl->ctrl);
2143
2144         return BLK_EH_RESET_TIMER;
2145 }
2146
2147 static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2148                         struct request *rq)
2149 {
2150         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2151         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2152         struct nvme_command *c = &pdu->cmd;
2153
2154         c->common.flags |= NVME_CMD_SGL_METABUF;
2155
2156         if (!blk_rq_nr_phys_segments(rq))
2157                 nvme_tcp_set_sg_null(c);
2158         else if (rq_data_dir(rq) == WRITE &&
2159             req->data_len <= nvme_tcp_inline_data_size(queue))
2160                 nvme_tcp_set_sg_inline(queue, c, req->data_len);
2161         else
2162                 nvme_tcp_set_sg_host_data(c, req->data_len);
2163
2164         return 0;
2165 }
2166
2167 static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2168                 struct request *rq)
2169 {
2170         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2171         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2172         struct nvme_tcp_queue *queue = req->queue;
2173         u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2174         blk_status_t ret;
2175
2176         ret = nvme_setup_cmd(ns, rq, &pdu->cmd);
2177         if (ret)
2178                 return ret;
2179
2180         req->state = NVME_TCP_SEND_CMD_PDU;
2181         req->offset = 0;
2182         req->data_sent = 0;
2183         req->pdu_len = 0;
2184         req->pdu_sent = 0;
2185         req->data_len = blk_rq_nr_phys_segments(rq) ?
2186                                 blk_rq_payload_bytes(rq) : 0;
2187         req->curr_bio = rq->bio;
2188
2189         if (rq_data_dir(rq) == WRITE &&
2190             req->data_len <= nvme_tcp_inline_data_size(queue))
2191                 req->pdu_len = req->data_len;
2192         else if (req->curr_bio)
2193                 nvme_tcp_init_iter(req, READ);
2194
2195         pdu->hdr.type = nvme_tcp_cmd;
2196         pdu->hdr.flags = 0;
2197         if (queue->hdr_digest)
2198                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2199         if (queue->data_digest && req->pdu_len) {
2200                 pdu->hdr.flags |= NVME_TCP_F_DDGST;
2201                 ddgst = nvme_tcp_ddgst_len(queue);
2202         }
2203         pdu->hdr.hlen = sizeof(*pdu);
2204         pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2205         pdu->hdr.plen =
2206                 cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2207
2208         ret = nvme_tcp_map_data(queue, rq);
2209         if (unlikely(ret)) {
2210                 nvme_cleanup_cmd(rq);
2211                 dev_err(queue->ctrl->ctrl.device,
2212                         "Failed to map data (%d)\n", ret);
2213                 return ret;
2214         }
2215
2216         return 0;
2217 }
2218
2219 static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2220                 const struct blk_mq_queue_data *bd)
2221 {
2222         struct nvme_ns *ns = hctx->queue->queuedata;
2223         struct nvme_tcp_queue *queue = hctx->driver_data;
2224         struct request *rq = bd->rq;
2225         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2226         bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2227         blk_status_t ret;
2228
2229         if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2230                 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
2231
2232         ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2233         if (unlikely(ret))
2234                 return ret;
2235
2236         blk_mq_start_request(rq);
2237
2238         nvme_tcp_queue_request(req, true);
2239
2240         return BLK_STS_OK;
2241 }
2242
2243 static int nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2244 {
2245         struct nvme_tcp_ctrl *ctrl = set->driver_data;
2246         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2247
2248         if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2249                 /* separate read/write queues */
2250                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2251                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2252                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2253                 set->map[HCTX_TYPE_READ].nr_queues =
2254                         ctrl->io_queues[HCTX_TYPE_READ];
2255                 set->map[HCTX_TYPE_READ].queue_offset =
2256                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2257         } else {
2258                 /* shared read/write queues */
2259                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2260                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2261                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2262                 set->map[HCTX_TYPE_READ].nr_queues =
2263                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2264                 set->map[HCTX_TYPE_READ].queue_offset = 0;
2265         }
2266         blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2267         blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
2268
2269         if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2270                 /* map dedicated poll queues only if we have queues left */
2271                 set->map[HCTX_TYPE_POLL].nr_queues =
2272                                 ctrl->io_queues[HCTX_TYPE_POLL];
2273                 set->map[HCTX_TYPE_POLL].queue_offset =
2274                         ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2275                         ctrl->io_queues[HCTX_TYPE_READ];
2276                 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2277         }
2278
2279         dev_info(ctrl->ctrl.device,
2280                 "mapped %d/%d/%d default/read/poll queues.\n",
2281                 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2282                 ctrl->io_queues[HCTX_TYPE_READ],
2283                 ctrl->io_queues[HCTX_TYPE_POLL]);
2284
2285         return 0;
2286 }
2287
2288 static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx)
2289 {
2290         struct nvme_tcp_queue *queue = hctx->driver_data;
2291         struct sock *sk = queue->sock->sk;
2292
2293         if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2294                 return 0;
2295
2296         set_bit(NVME_TCP_Q_POLLING, &queue->flags);
2297         if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue))
2298                 sk_busy_loop(sk, true);
2299         nvme_tcp_try_recv(queue);
2300         clear_bit(NVME_TCP_Q_POLLING, &queue->flags);
2301         return queue->nr_cqe;
2302 }
2303
2304 static struct blk_mq_ops nvme_tcp_mq_ops = {
2305         .queue_rq       = nvme_tcp_queue_rq,
2306         .complete       = nvme_complete_rq,
2307         .init_request   = nvme_tcp_init_request,
2308         .exit_request   = nvme_tcp_exit_request,
2309         .init_hctx      = nvme_tcp_init_hctx,
2310         .timeout        = nvme_tcp_timeout,
2311         .map_queues     = nvme_tcp_map_queues,
2312         .poll           = nvme_tcp_poll,
2313 };
2314
2315 static struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2316         .queue_rq       = nvme_tcp_queue_rq,
2317         .complete       = nvme_complete_rq,
2318         .init_request   = nvme_tcp_init_request,
2319         .exit_request   = nvme_tcp_exit_request,
2320         .init_hctx      = nvme_tcp_init_admin_hctx,
2321         .timeout        = nvme_tcp_timeout,
2322 };
2323
2324 static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2325         .name                   = "tcp",
2326         .module                 = THIS_MODULE,
2327         .flags                  = NVME_F_FABRICS,
2328         .reg_read32             = nvmf_reg_read32,
2329         .reg_read64             = nvmf_reg_read64,
2330         .reg_write32            = nvmf_reg_write32,
2331         .free_ctrl              = nvme_tcp_free_ctrl,
2332         .submit_async_event     = nvme_tcp_submit_async_event,
2333         .delete_ctrl            = nvme_tcp_delete_ctrl,
2334         .get_address            = nvmf_get_address,
2335 };
2336
2337 static bool
2338 nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2339 {
2340         struct nvme_tcp_ctrl *ctrl;
2341         bool found = false;
2342
2343         mutex_lock(&nvme_tcp_ctrl_mutex);
2344         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2345                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2346                 if (found)
2347                         break;
2348         }
2349         mutex_unlock(&nvme_tcp_ctrl_mutex);
2350
2351         return found;
2352 }
2353
2354 static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2355                 struct nvmf_ctrl_options *opts)
2356 {
2357         struct nvme_tcp_ctrl *ctrl;
2358         int ret;
2359
2360         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2361         if (!ctrl)
2362                 return ERR_PTR(-ENOMEM);
2363
2364         INIT_LIST_HEAD(&ctrl->list);
2365         ctrl->ctrl.opts = opts;
2366         ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2367                                 opts->nr_poll_queues + 1;
2368         ctrl->ctrl.sqsize = opts->queue_size - 1;
2369         ctrl->ctrl.kato = opts->kato;
2370
2371         INIT_DELAYED_WORK(&ctrl->connect_work,
2372                         nvme_tcp_reconnect_ctrl_work);
2373         INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2374         INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2375
2376         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2377                 opts->trsvcid =
2378                         kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2379                 if (!opts->trsvcid) {
2380                         ret = -ENOMEM;
2381                         goto out_free_ctrl;
2382                 }
2383                 opts->mask |= NVMF_OPT_TRSVCID;
2384         }
2385
2386         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2387                         opts->traddr, opts->trsvcid, &ctrl->addr);
2388         if (ret) {
2389                 pr_err("malformed address passed: %s:%s\n",
2390                         opts->traddr, opts->trsvcid);
2391                 goto out_free_ctrl;
2392         }
2393
2394         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2395                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2396                         opts->host_traddr, NULL, &ctrl->src_addr);
2397                 if (ret) {
2398                         pr_err("malformed src address passed: %s\n",
2399                                opts->host_traddr);
2400                         goto out_free_ctrl;
2401                 }
2402         }
2403
2404         if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2405                 ret = -EALREADY;
2406                 goto out_free_ctrl;
2407         }
2408
2409         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2410                                 GFP_KERNEL);
2411         if (!ctrl->queues) {
2412                 ret = -ENOMEM;
2413                 goto out_free_ctrl;
2414         }
2415
2416         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2417         if (ret)
2418                 goto out_kfree_queues;
2419
2420         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2421                 WARN_ON_ONCE(1);
2422                 ret = -EINTR;
2423                 goto out_uninit_ctrl;
2424         }
2425
2426         ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2427         if (ret)
2428                 goto out_uninit_ctrl;
2429
2430         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
2431                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2432
2433         mutex_lock(&nvme_tcp_ctrl_mutex);
2434         list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2435         mutex_unlock(&nvme_tcp_ctrl_mutex);
2436
2437         return &ctrl->ctrl;
2438
2439 out_uninit_ctrl:
2440         nvme_uninit_ctrl(&ctrl->ctrl);
2441         nvme_put_ctrl(&ctrl->ctrl);
2442         if (ret > 0)
2443                 ret = -EIO;
2444         return ERR_PTR(ret);
2445 out_kfree_queues:
2446         kfree(ctrl->queues);
2447 out_free_ctrl:
2448         kfree(ctrl);
2449         return ERR_PTR(ret);
2450 }
2451
2452 static struct nvmf_transport_ops nvme_tcp_transport = {
2453         .name           = "tcp",
2454         .module         = THIS_MODULE,
2455         .required_opts  = NVMF_OPT_TRADDR,
2456         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2457                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2458                           NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
2459                           NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2460                           NVMF_OPT_TOS,
2461         .create_ctrl    = nvme_tcp_create_ctrl,
2462 };
2463
2464 static int __init nvme_tcp_init_module(void)
2465 {
2466         nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
2467                         WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2468         if (!nvme_tcp_wq)
2469                 return -ENOMEM;
2470
2471         nvmf_register_transport(&nvme_tcp_transport);
2472         return 0;
2473 }
2474
2475 static void __exit nvme_tcp_cleanup_module(void)
2476 {
2477         struct nvme_tcp_ctrl *ctrl;
2478
2479         nvmf_unregister_transport(&nvme_tcp_transport);
2480
2481         mutex_lock(&nvme_tcp_ctrl_mutex);
2482         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
2483                 nvme_delete_ctrl(&ctrl->ctrl);
2484         mutex_unlock(&nvme_tcp_ctrl_mutex);
2485         flush_workqueue(nvme_delete_wq);
2486
2487         destroy_workqueue(nvme_tcp_wq);
2488 }
2489
2490 module_init(nvme_tcp_init_module);
2491 module_exit(nvme_tcp_cleanup_module);
2492
2493 MODULE_LICENSE("GPL v2");