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         struct llist_node       lentry;
  50         __le32                  ddgst;
  51
  52         struct bio              *curr_bio;
  53         struct iov_iter         iter;
  54
  55         /* send state */
  56         size_t                  offset;
  57         size_t                  data_sent;
  58         enum nvme_tcp_send_state state;
  59 };
  60
  61 enum nvme_tcp_queue_flags {
  62         NVME_TCP_Q_ALLOCATED    = 0,
  63         NVME_TCP_Q_LIVE         = 1,
  64         NVME_TCP_Q_POLLING      = 2,
  65 };
  66
  67 enum nvme_tcp_recv_state {
  68         NVME_TCP_RECV_PDU = 0,
  69         NVME_TCP_RECV_DATA,
  70         NVME_TCP_RECV_DDGST,
  71 };
  72
  73 struct nvme_tcp_ctrl;
  74 struct nvme_tcp_queue {
  75         struct socket           *sock;
  76         struct work_struct      io_work;
  77         int                     io_cpu;
  78
  79         struct mutex            send_mutex;
  80         struct llist_head       req_list;
  81         struct list_head        send_list;
  82
  83         /* recv state */
  84         void                    *pdu;
  85         int                     pdu_remaining;
  86         int                     pdu_offset;
  87         size_t                  data_remaining;
  88         size_t                  ddgst_remaining;
  89         unsigned int            nr_cqe;
  90
  91         /* send state */
  92         struct nvme_tcp_request *request;
  93
  94         int                     queue_size;
  95         size_t                  cmnd_capsule_len;
  96         struct nvme_tcp_ctrl    *ctrl;
  97         unsigned long           flags;
  98         bool                    rd_enabled;
  99
 100         bool                    hdr_digest;
 101         bool                    data_digest;
 102         struct ahash_request    *rcv_hash;
 103         struct ahash_request    *snd_hash;
 104         __le32                  exp_ddgst;
 105         __le32                  recv_ddgst;
 106
 107         struct page_frag_cache  pf_cache;
 108
 109         void (*state_change)(struct sock *);
 110         void (*data_ready)(struct sock *);
 111         void (*write_space)(struct sock *);
 112 };
 113
 114 struct nvme_tcp_ctrl {
 115         /* read only in the hot path */
 116         struct nvme_tcp_queue   *queues;
 117         struct blk_mq_tag_set   tag_set;
 118
 119         /* other member variables */
 120         struct list_head        list;
 121         struct blk_mq_tag_set   admin_tag_set;
 122         struct sockaddr_storage addr;
 123         struct sockaddr_storage src_addr;
 124         struct nvme_ctrl        ctrl;
 125
 126         struct work_struct      err_work;
 127         struct delayed_work     connect_work;
 128         struct nvme_tcp_request async_req;
 129         u32                     io_queues[HCTX_MAX_TYPES];
 130 };
 131
 132 static LIST_HEAD(nvme_tcp_ctrl_list);
 133 static DEFINE_MUTEX(nvme_tcp_ctrl_mutex);
 134 static struct workqueue_struct *nvme_tcp_wq;
 135 static const struct blk_mq_ops nvme_tcp_mq_ops;
 136 static const struct blk_mq_ops nvme_tcp_admin_mq_ops;
 137 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue);
 138
 139 static inline struct nvme_tcp_ctrl *to_tcp_ctrl(struct nvme_ctrl *ctrl)
 140 {
 141         return container_of(ctrl, struct nvme_tcp_ctrl, ctrl);
 142 }
 143
 144 static inline int nvme_tcp_queue_id(struct nvme_tcp_queue *queue)
 145 {
 146         return queue - queue->ctrl->queues;
 147 }
 148
 149 static inline struct blk_mq_tags *nvme_tcp_tagset(struct nvme_tcp_queue *queue)
 150 {
 151         u32 queue_idx = nvme_tcp_queue_id(queue);
 152
 153         if (queue_idx == 0)
 154                 return queue->ctrl->admin_tag_set.tags[queue_idx];
 155         return queue->ctrl->tag_set.tags[queue_idx - 1];
 156 }
 157
 158 static inline u8 nvme_tcp_hdgst_len(struct nvme_tcp_queue *queue)
 159 {
 160         return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 161 }
 162
 163 static inline u8 nvme_tcp_ddgst_len(struct nvme_tcp_queue *queue)
 164 {
 165         return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 166 }
 167
 168 static inline size_t nvme_tcp_inline_data_size(struct nvme_tcp_queue *queue)
 169 {
 170         return queue->cmnd_capsule_len - sizeof(struct nvme_command);
 171 }
 172
 173 static inline bool nvme_tcp_async_req(struct nvme_tcp_request *req)
 174 {
 175         return req == &req->queue->ctrl->async_req;
 176 }
 177
 178 static inline bool nvme_tcp_has_inline_data(struct nvme_tcp_request *req)
 179 {
 180         struct request *rq;
 181
 182         if (unlikely(nvme_tcp_async_req(req)))
 183                 return false; /* async events don't have a request */
 184
 185         rq = blk_mq_rq_from_pdu(req);
 186
 187         return rq_data_dir(rq) == WRITE && req->data_len &&
 188                 req->data_len <= nvme_tcp_inline_data_size(req->queue);
 189 }
 190
 191 static inline struct page *nvme_tcp_req_cur_page(struct nvme_tcp_request *req)
 192 {
 193         return req->iter.bvec->bv_page;
 194 }
 195
 196 static inline size_t nvme_tcp_req_cur_offset(struct nvme_tcp_request *req)
 197 {
 198         return req->iter.bvec->bv_offset + req->iter.iov_offset;
 199 }
 200
 201 static inline size_t nvme_tcp_req_cur_length(struct nvme_tcp_request *req)
 202 {
 203         return min_t(size_t, req->iter.bvec->bv_len - req->iter.iov_offset,
 204                         req->pdu_len - req->pdu_sent);
 205 }
 206
 207 static inline size_t nvme_tcp_req_offset(struct nvme_tcp_request *req)
 208 {
 209         return req->iter.iov_offset;
 210 }
 211
 212 static inline size_t nvme_tcp_pdu_data_left(struct nvme_tcp_request *req)
 213 {
 214         return rq_data_dir(blk_mq_rq_from_pdu(req)) == WRITE ?
 215                         req->pdu_len - req->pdu_sent : 0;
 216 }
 217
 218 static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req,
 219                 int len)
 220 {
 221         return nvme_tcp_pdu_data_left(req) <= len;
 222 }
 223
 224 static void nvme_tcp_init_iter(struct nvme_tcp_request *req,
 225                 unsigned int dir)
 226 {
 227         struct request *rq = blk_mq_rq_from_pdu(req);
 228         struct bio_vec *vec;
 229         unsigned int size;
 230         int nsegs;
 231         size_t offset;
 232
 233         if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) {
 234                 vec = &rq->special_vec;
 235                 nsegs = 1;
 236                 size = blk_rq_payload_bytes(rq);
 237                 offset = 0;
 238         } else {
 239                 struct bio *bio = req->curr_bio;
 240
 241                 vec = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
 242                 nsegs = bio_segments(bio);
 243                 size = bio->bi_iter.bi_size;
 244                 offset = bio->bi_iter.bi_bvec_done;
 245         }
 246
 247         iov_iter_bvec(&req->iter, dir, vec, nsegs, size);
 248         req->iter.iov_offset = offset;
 249 }
 250
 251 static inline void nvme_tcp_advance_req(struct nvme_tcp_request *req,
 252                 int len)
 253 {
 254         req->data_sent += len;
 255         req->pdu_sent += len;
 256         iov_iter_advance(&req->iter, len);
 257         if (!iov_iter_count(&req->iter) &&
 258             req->data_sent < req->data_len) {
 259                 req->curr_bio = req->curr_bio->bi_next;
 260                 nvme_tcp_init_iter(req, WRITE);
 261         }
 262 }
 263
 264 static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req,
 265                 bool sync)
 266 {
 267         struct nvme_tcp_queue *queue = req->queue;
 268         bool empty;
 269
 270         empty = llist_add(&req->lentry, &queue->req_list) &&
 271                 list_empty(&queue->send_list) && !queue->request;
 272
 273         /*
 274          * if we're the first on the send_list and we can try to send
 275          * directly, otherwise queue io_work. Also, only do that if we
 276          * are on the same cpu, so we don't introduce contention.
 277          */
 278         if (queue->io_cpu == smp_processor_id() &&
 279             sync && empty && mutex_trylock(&queue->send_mutex)) {
 280                 nvme_tcp_try_send(queue);
 281                 mutex_unlock(&queue->send_mutex);
 282         } else {
 283                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 284         }
 285 }
 286
 287 static void nvme_tcp_process_req_list(struct nvme_tcp_queue *queue)
 288 {
 289         struct nvme_tcp_request *req;
 290         struct llist_node *node;
 291
 292         for (node = llist_del_all(&queue->req_list); node; node = node->next) {
 293                 req = llist_entry(node, struct nvme_tcp_request, lentry);
 294                 list_add(&req->entry, &queue->send_list);
 295         }
 296 }
 297
 298 static inline struct nvme_tcp_request *
 299 nvme_tcp_fetch_request(struct nvme_tcp_queue *queue)
 300 {
 301         struct nvme_tcp_request *req;
 302
 303         req = list_first_entry_or_null(&queue->send_list,
 304                         struct nvme_tcp_request, entry);
 305         if (!req) {
 306                 nvme_tcp_process_req_list(queue);
 307                 req = list_first_entry_or_null(&queue->send_list,
 308                                 struct nvme_tcp_request, entry);
 309                 if (unlikely(!req))
 310                         return NULL;
 311         }
 312
 313         list_del(&req->entry);
 314         return req;
 315 }
 316
 317 static inline void nvme_tcp_ddgst_final(struct ahash_request *hash,
 318                 __le32 *dgst)
 319 {
 320         ahash_request_set_crypt(hash, NULL, (u8 *)dgst, 0);
 321         crypto_ahash_final(hash);
 322 }
 323
 324 static inline void nvme_tcp_ddgst_update(struct ahash_request *hash,
 325                 struct page *page, off_t off, size_t len)
 326 {
 327         struct scatterlist sg;
 328
 329         sg_init_marker(&sg, 1);
 330         sg_set_page(&sg, page, len, off);
 331         ahash_request_set_crypt(hash, &sg, NULL, len);
 332         crypto_ahash_update(hash);
 333 }
 334
 335 static inline void nvme_tcp_hdgst(struct ahash_request *hash,
 336                 void *pdu, size_t len)
 337 {
 338         struct scatterlist sg;
 339
 340         sg_init_one(&sg, pdu, len);
 341         ahash_request_set_crypt(hash, &sg, pdu + len, len);
 342         crypto_ahash_digest(hash);
 343 }
 344
 345 static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue,
 346                 void *pdu, size_t pdu_len)
 347 {
 348         struct nvme_tcp_hdr *hdr = pdu;
 349         __le32 recv_digest;
 350         __le32 exp_digest;
 351
 352         if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
 353                 dev_err(queue->ctrl->ctrl.device,
 354                         "queue %d: header digest flag is cleared\n",
 355                         nvme_tcp_queue_id(queue));
 356                 return -EPROTO;
 357         }
 358
 359         recv_digest = *(__le32 *)(pdu + hdr->hlen);
 360         nvme_tcp_hdgst(queue->rcv_hash, pdu, pdu_len);
 361         exp_digest = *(__le32 *)(pdu + hdr->hlen);
 362         if (recv_digest != exp_digest) {
 363                 dev_err(queue->ctrl->ctrl.device,
 364                         "header digest error: recv %#x expected %#x\n",
 365                         le32_to_cpu(recv_digest), le32_to_cpu(exp_digest));
 366                 return -EIO;
 367         }
 368
 369         return 0;
 370 }
 371
 372 static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu)
 373 {
 374         struct nvme_tcp_hdr *hdr = pdu;
 375         u8 digest_len = nvme_tcp_hdgst_len(queue);
 376         u32 len;
 377
 378         len = le32_to_cpu(hdr->plen) - hdr->hlen -
 379                 ((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0);
 380
 381         if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
 382                 dev_err(queue->ctrl->ctrl.device,
 383                         "queue %d: data digest flag is cleared\n",
 384                 nvme_tcp_queue_id(queue));
 385                 return -EPROTO;
 386         }
 387         crypto_ahash_init(queue->rcv_hash);
 388
 389         return 0;
 390 }
 391
 392 static void nvme_tcp_exit_request(struct blk_mq_tag_set *set,
 393                 struct request *rq, unsigned int hctx_idx)
 394 {
 395         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 396
 397         page_frag_free(req->pdu);
 398 }
 399
 400 static int nvme_tcp_init_request(struct blk_mq_tag_set *set,
 401                 struct request *rq, unsigned int hctx_idx,
 402                 unsigned int numa_node)
 403 {
 404         struct nvme_tcp_ctrl *ctrl = set->driver_data;
 405         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 406         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 407         struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx];
 408         u8 hdgst = nvme_tcp_hdgst_len(queue);
 409
 410         req->pdu = page_frag_alloc(&queue->pf_cache,
 411                 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
 412                 GFP_KERNEL | __GFP_ZERO);
 413         if (!req->pdu)
 414                 return -ENOMEM;
 415
 416         req->queue = queue;
 417         nvme_req(rq)->ctrl = &ctrl->ctrl;
 418
 419         return 0;
 420 }
 421
 422 static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 423                 unsigned int hctx_idx)
 424 {
 425         struct nvme_tcp_ctrl *ctrl = data;
 426         struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1];
 427
 428         hctx->driver_data = queue;
 429         return 0;
 430 }
 431
 432 static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 433                 unsigned int hctx_idx)
 434 {
 435         struct nvme_tcp_ctrl *ctrl = data;
 436         struct nvme_tcp_queue *queue = &ctrl->queues[0];
 437
 438         hctx->driver_data = queue;
 439         return 0;
 440 }
 441
 442 static enum nvme_tcp_recv_state
 443 nvme_tcp_recv_state(struct nvme_tcp_queue *queue)
 444 {
 445         return  (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
 446                 (queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST :
 447                 NVME_TCP_RECV_DATA;
 448 }
 449
 450 static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue)
 451 {
 452         queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) +
 453                                 nvme_tcp_hdgst_len(queue);
 454         queue->pdu_offset = 0;
 455         queue->data_remaining = -1;
 456         queue->ddgst_remaining = 0;
 457 }
 458
 459 static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl)
 460 {
 461         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
 462                 return;
 463
 464         queue_work(nvme_reset_wq, &to_tcp_ctrl(ctrl)->err_work);
 465 }
 466
 467 static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
 468                 struct nvme_completion *cqe)
 469 {
 470         struct request *rq;
 471
 472         rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), cqe->command_id);
 473         if (!rq) {
 474                 dev_err(queue->ctrl->ctrl.device,
 475                         "queue %d tag 0x%x not found\n",
 476                         nvme_tcp_queue_id(queue), cqe->command_id);
 477                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 478                 return -EINVAL;
 479         }
 480
 481         if (!nvme_end_request(rq, cqe->status, cqe->result))
 482                 nvme_complete_rq(rq);
 483         queue->nr_cqe++;
 484
 485         return 0;
 486 }
 487
 488 static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue,
 489                 struct nvme_tcp_data_pdu *pdu)
 490 {
 491         struct request *rq;
 492
 493         rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
 494         if (!rq) {
 495                 dev_err(queue->ctrl->ctrl.device,
 496                         "queue %d tag %#x not found\n",
 497                         nvme_tcp_queue_id(queue), pdu->command_id);
 498                 return -ENOENT;
 499         }
 500
 501         if (!blk_rq_payload_bytes(rq)) {
 502                 dev_err(queue->ctrl->ctrl.device,
 503                         "queue %d tag %#x unexpected data\n",
 504                         nvme_tcp_queue_id(queue), rq->tag);
 505                 return -EIO;
 506         }
 507
 508         queue->data_remaining = le32_to_cpu(pdu->data_length);
 509
 510         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS &&
 511             unlikely(!(pdu->hdr.flags & NVME_TCP_F_DATA_LAST))) {
 512                 dev_err(queue->ctrl->ctrl.device,
 513                         "queue %d tag %#x SUCCESS set but not last PDU\n",
 514                         nvme_tcp_queue_id(queue), rq->tag);
 515                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 516                 return -EPROTO;
 517         }
 518
 519         return 0;
 520 }
 521
 522 static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue,
 523                 struct nvme_tcp_rsp_pdu *pdu)
 524 {
 525         struct nvme_completion *cqe = &pdu->cqe;
 526         int ret = 0;
 527
 528         /*
 529          * AEN requests are special as they don't time out and can
 530          * survive any kind of queue freeze and often don't respond to
 531          * aborts.  We don't even bother to allocate a struct request
 532          * for them but rather special case them here.
 533          */
 534         if (unlikely(nvme_is_aen_req(nvme_tcp_queue_id(queue),
 535                                      cqe->command_id)))
 536                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
 537                                 &cqe->result);
 538         else
 539                 ret = nvme_tcp_process_nvme_cqe(queue, cqe);
 540
 541         return ret;
 542 }
 543
 544 static int nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req,
 545                 struct nvme_tcp_r2t_pdu *pdu)
 546 {
 547         struct nvme_tcp_data_pdu *data = req->pdu;
 548         struct nvme_tcp_queue *queue = req->queue;
 549         struct request *rq = blk_mq_rq_from_pdu(req);
 550         u8 hdgst = nvme_tcp_hdgst_len(queue);
 551         u8 ddgst = nvme_tcp_ddgst_len(queue);
 552
 553         req->pdu_len = le32_to_cpu(pdu->r2t_length);
 554         req->pdu_sent = 0;
 555
 556         if (unlikely(req->data_sent + req->pdu_len > req->data_len)) {
 557                 dev_err(queue->ctrl->ctrl.device,
 558                         "req %d r2t len %u exceeded data len %u (%zu sent)\n",
 559                         rq->tag, req->pdu_len, req->data_len,
 560                         req->data_sent);
 561                 return -EPROTO;
 562         }
 563
 564         if (unlikely(le32_to_cpu(pdu->r2t_offset) < req->data_sent)) {
 565                 dev_err(queue->ctrl->ctrl.device,
 566                         "req %d unexpected r2t offset %u (expected %zu)\n",
 567                         rq->tag, le32_to_cpu(pdu->r2t_offset),
 568                         req->data_sent);
 569                 return -EPROTO;
 570         }
 571
 572         memset(data, 0, sizeof(*data));
 573         data->hdr.type = nvme_tcp_h2c_data;
 574         data->hdr.flags = NVME_TCP_F_DATA_LAST;
 575         if (queue->hdr_digest)
 576                 data->hdr.flags |= NVME_TCP_F_HDGST;
 577         if (queue->data_digest)
 578                 data->hdr.flags |= NVME_TCP_F_DDGST;
 579         data->hdr.hlen = sizeof(*data);
 580         data->hdr.pdo = data->hdr.hlen + hdgst;
 581         data->hdr.plen =
 582                 cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst);
 583         data->ttag = pdu->ttag;
 584         data->command_id = rq->tag;
 585         data->data_offset = cpu_to_le32(req->data_sent);
 586         data->data_length = cpu_to_le32(req->pdu_len);
 587         return 0;
 588 }
 589
 590 static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue,
 591                 struct nvme_tcp_r2t_pdu *pdu)
 592 {
 593         struct nvme_tcp_request *req;
 594         struct request *rq;
 595         int ret;
 596
 597         rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
 598         if (!rq) {
 599                 dev_err(queue->ctrl->ctrl.device,
 600                         "queue %d tag %#x not found\n",
 601                         nvme_tcp_queue_id(queue), pdu->command_id);
 602                 return -ENOENT;
 603         }
 604         req = blk_mq_rq_to_pdu(rq);
 605
 606         ret = nvme_tcp_setup_h2c_data_pdu(req, pdu);
 607         if (unlikely(ret))
 608                 return ret;
 609
 610         req->state = NVME_TCP_SEND_H2C_PDU;
 611         req->offset = 0;
 612
 613         nvme_tcp_queue_request(req, false);
 614
 615         return 0;
 616 }
 617
 618 static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 619                 unsigned int *offset, size_t *len)
 620 {
 621         struct nvme_tcp_hdr *hdr;
 622         char *pdu = queue->pdu;
 623         size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
 624         int ret;
 625
 626         ret = skb_copy_bits(skb, *offset,
 627                 &pdu[queue->pdu_offset], rcv_len);
 628         if (unlikely(ret))
 629                 return ret;
 630
 631         queue->pdu_remaining -= rcv_len;
 632         queue->pdu_offset += rcv_len;
 633         *offset += rcv_len;
 634         *len -= rcv_len;
 635         if (queue->pdu_remaining)
 636                 return 0;
 637
 638         hdr = queue->pdu;
 639         if (queue->hdr_digest) {
 640                 ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen);
 641                 if (unlikely(ret))
 642                         return ret;
 643         }
 644
 645
 646         if (queue->data_digest) {
 647                 ret = nvme_tcp_check_ddgst(queue, queue->pdu);
 648                 if (unlikely(ret))
 649                         return ret;
 650         }
 651
 652         switch (hdr->type) {
 653         case nvme_tcp_c2h_data:
 654                 return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
 655         case nvme_tcp_rsp:
 656                 nvme_tcp_init_recv_ctx(queue);
 657                 return nvme_tcp_handle_comp(queue, (void *)queue->pdu);
 658         case nvme_tcp_r2t:
 659                 nvme_tcp_init_recv_ctx(queue);
 660                 return nvme_tcp_handle_r2t(queue, (void *)queue->pdu);
 661         default:
 662                 dev_err(queue->ctrl->ctrl.device,
 663                         "unsupported pdu type (%d)\n", hdr->type);
 664                 return -EINVAL;
 665         }
 666 }
 667
 668 static inline void nvme_tcp_end_request(struct request *rq, u16 status)
 669 {
 670         union nvme_result res = {};
 671
 672         if (!nvme_end_request(rq, cpu_to_le16(status << 1), res))
 673                 nvme_complete_rq(rq);
 674 }
 675
 676 static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 677                               unsigned int *offset, size_t *len)
 678 {
 679         struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
 680         struct nvme_tcp_request *req;
 681         struct request *rq;
 682
 683         rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
 684         if (!rq) {
 685                 dev_err(queue->ctrl->ctrl.device,
 686                         "queue %d tag %#x not found\n",
 687                         nvme_tcp_queue_id(queue), pdu->command_id);
 688                 return -ENOENT;
 689         }
 690         req = blk_mq_rq_to_pdu(rq);
 691
 692         while (true) {
 693                 int recv_len, ret;
 694
 695                 recv_len = min_t(size_t, *len, queue->data_remaining);
 696                 if (!recv_len)
 697                         break;
 698
 699                 if (!iov_iter_count(&req->iter)) {
 700                         req->curr_bio = req->curr_bio->bi_next;
 701
 702                         /*
 703                          * If we don`t have any bios it means that controller
 704                          * sent more data than we requested, hence error
 705                          */
 706                         if (!req->curr_bio) {
 707                                 dev_err(queue->ctrl->ctrl.device,
 708                                         "queue %d no space in request %#x",
 709                                         nvme_tcp_queue_id(queue), rq->tag);
 710                                 nvme_tcp_init_recv_ctx(queue);
 711                                 return -EIO;
 712                         }
 713                         nvme_tcp_init_iter(req, READ);
 714                 }
 715
 716                 /* we can read only from what is left in this bio */
 717                 recv_len = min_t(size_t, recv_len,
 718                                 iov_iter_count(&req->iter));
 719
 720                 if (queue->data_digest)
 721                         ret = skb_copy_and_hash_datagram_iter(skb, *offset,
 722                                 &req->iter, recv_len, queue->rcv_hash);
 723                 else
 724                         ret = skb_copy_datagram_iter(skb, *offset,
 725                                         &req->iter, recv_len);
 726                 if (ret) {
 727                         dev_err(queue->ctrl->ctrl.device,
 728                                 "queue %d failed to copy request %#x data",
 729                                 nvme_tcp_queue_id(queue), rq->tag);
 730                         return ret;
 731                 }
 732
 733                 *len -= recv_len;
 734                 *offset += recv_len;
 735                 queue->data_remaining -= recv_len;
 736         }
 737
 738         if (!queue->data_remaining) {
 739                 if (queue->data_digest) {
 740                         nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst);
 741                         queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
 742                 } else {
 743                         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
 744                                 nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
 745                                 queue->nr_cqe++;
 746                         }
 747                         nvme_tcp_init_recv_ctx(queue);
 748                 }
 749         }
 750
 751         return 0;
 752 }
 753
 754 static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
 755                 struct sk_buff *skb, unsigned int *offset, size_t *len)
 756 {
 757         struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
 758         char *ddgst = (char *)&queue->recv_ddgst;
 759         size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
 760         off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
 761         int ret;
 762
 763         ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
 764         if (unlikely(ret))
 765                 return ret;
 766
 767         queue->ddgst_remaining -= recv_len;
 768         *offset += recv_len;
 769         *len -= recv_len;
 770         if (queue->ddgst_remaining)
 771                 return 0;
 772
 773         if (queue->recv_ddgst != queue->exp_ddgst) {
 774                 dev_err(queue->ctrl->ctrl.device,
 775                         "data digest error: recv %#x expected %#x\n",
 776                         le32_to_cpu(queue->recv_ddgst),
 777                         le32_to_cpu(queue->exp_ddgst));
 778                 return -EIO;
 779         }
 780
 781         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
 782                 struct request *rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue),
 783                                                 pdu->command_id);
 784
 785                 nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
 786                 queue->nr_cqe++;
 787         }
 788
 789         nvme_tcp_init_recv_ctx(queue);
 790         return 0;
 791 }
 792
 793 static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
 794                              unsigned int offset, size_t len)
 795 {
 796         struct nvme_tcp_queue *queue = desc->arg.data;
 797         size_t consumed = len;
 798         int result;
 799
 800         while (len) {
 801                 switch (nvme_tcp_recv_state(queue)) {
 802                 case NVME_TCP_RECV_PDU:
 803                         result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
 804                         break;
 805                 case NVME_TCP_RECV_DATA:
 806                         result = nvme_tcp_recv_data(queue, skb, &offset, &len);
 807                         break;
 808                 case NVME_TCP_RECV_DDGST:
 809                         result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
 810                         break;
 811                 default:
 812                         result = -EFAULT;
 813                 }
 814                 if (result) {
 815                         dev_err(queue->ctrl->ctrl.device,
 816                                 "receive failed:  %d\n", result);
 817                         queue->rd_enabled = false;
 818                         nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 819                         return result;
 820                 }
 821         }
 822
 823         return consumed;
 824 }
 825
 826 static void nvme_tcp_data_ready(struct sock *sk)
 827 {
 828         struct nvme_tcp_queue *queue;
 829
 830         read_lock_bh(&sk->sk_callback_lock);
 831         queue = sk->sk_user_data;
 832         if (likely(queue && queue->rd_enabled) &&
 833             !test_bit(NVME_TCP_Q_POLLING, &queue->flags))
 834                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 835         read_unlock_bh(&sk->sk_callback_lock);
 836 }
 837
 838 static void nvme_tcp_write_space(struct sock *sk)
 839 {
 840         struct nvme_tcp_queue *queue;
 841
 842         read_lock_bh(&sk->sk_callback_lock);
 843         queue = sk->sk_user_data;
 844         if (likely(queue && sk_stream_is_writeable(sk))) {
 845                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 846                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 847         }
 848         read_unlock_bh(&sk->sk_callback_lock);
 849 }
 850
 851 static void nvme_tcp_state_change(struct sock *sk)
 852 {
 853         struct nvme_tcp_queue *queue;
 854
 855         read_lock(&sk->sk_callback_lock);
 856         queue = sk->sk_user_data;
 857         if (!queue)
 858                 goto done;
 859
 860         switch (sk->sk_state) {
 861         case TCP_CLOSE:
 862         case TCP_CLOSE_WAIT:
 863         case TCP_LAST_ACK:
 864         case TCP_FIN_WAIT1:
 865         case TCP_FIN_WAIT2:
 866                 /* fallthrough */
 867                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 868                 break;
 869         default:
 870                 dev_info(queue->ctrl->ctrl.device,
 871                         "queue %d socket state %d\n",
 872                         nvme_tcp_queue_id(queue), sk->sk_state);
 873         }
 874
 875         queue->state_change(sk);
 876 done:
 877         read_unlock(&sk->sk_callback_lock);
 878 }
 879
 880 static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
 881 {
 882         queue->request = NULL;
 883 }
 884
 885 static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
 886 {
 887         nvme_tcp_end_request(blk_mq_rq_from_pdu(req), NVME_SC_HOST_PATH_ERROR);
 888 }
 889
 890 static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
 891 {
 892         struct nvme_tcp_queue *queue = req->queue;
 893
 894         while (true) {
 895                 struct page *page = nvme_tcp_req_cur_page(req);
 896                 size_t offset = nvme_tcp_req_cur_offset(req);
 897                 size_t len = nvme_tcp_req_cur_length(req);
 898                 bool last = nvme_tcp_pdu_last_send(req, len);
 899                 int ret, flags = MSG_DONTWAIT;
 900
 901                 if (last && !queue->data_digest)
 902                         flags |= MSG_EOR;
 903                 else
 904                         flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
 905
 906                 /* can't zcopy slab pages */
 907                 if (unlikely(PageSlab(page))) {
 908                         ret = sock_no_sendpage(queue->sock, page, offset, len,
 909                                         flags);
 910                 } else {
 911                         ret = kernel_sendpage(queue->sock, page, offset, len,
 912                                         flags);
 913                 }
 914                 if (ret <= 0)
 915                         return ret;
 916
 917                 nvme_tcp_advance_req(req, ret);
 918                 if (queue->data_digest)
 919                         nvme_tcp_ddgst_update(queue->snd_hash, page,
 920                                         offset, ret);
 921
 922                 /* fully successful last write*/
 923                 if (last && ret == len) {
 924                         if (queue->data_digest) {
 925                                 nvme_tcp_ddgst_final(queue->snd_hash,
 926                                         &req->ddgst);
 927                                 req->state = NVME_TCP_SEND_DDGST;
 928                                 req->offset = 0;
 929                         } else {
 930                                 nvme_tcp_done_send_req(queue);
 931                         }
 932                         return 1;
 933                 }
 934         }
 935         return -EAGAIN;
 936 }
 937
 938 static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
 939 {
 940         struct nvme_tcp_queue *queue = req->queue;
 941         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
 942         bool inline_data = nvme_tcp_has_inline_data(req);
 943         u8 hdgst = nvme_tcp_hdgst_len(queue);
 944         int len = sizeof(*pdu) + hdgst - req->offset;
 945         int flags = MSG_DONTWAIT;
 946         int ret;
 947
 948         if (inline_data)
 949                 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
 950         else
 951                 flags |= MSG_EOR;
 952
 953         if (queue->hdr_digest && !req->offset)
 954                 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
 955
 956         ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
 957                         offset_in_page(pdu) + req->offset, len,  flags);
 958         if (unlikely(ret <= 0))
 959                 return ret;
 960
 961         len -= ret;
 962         if (!len) {
 963                 if (inline_data) {
 964                         req->state = NVME_TCP_SEND_DATA;
 965                         if (queue->data_digest)
 966                                 crypto_ahash_init(queue->snd_hash);
 967                         nvme_tcp_init_iter(req, WRITE);
 968                 } else {
 969                         nvme_tcp_done_send_req(queue);
 970                 }
 971                 return 1;
 972         }
 973         req->offset += ret;
 974
 975         return -EAGAIN;
 976 }
 977
 978 static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
 979 {
 980         struct nvme_tcp_queue *queue = req->queue;
 981         struct nvme_tcp_data_pdu *pdu = req->pdu;
 982         u8 hdgst = nvme_tcp_hdgst_len(queue);
 983         int len = sizeof(*pdu) - req->offset + hdgst;
 984         int ret;
 985
 986         if (queue->hdr_digest && !req->offset)
 987                 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
 988
 989         ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
 990                         offset_in_page(pdu) + req->offset, len,
 991                         MSG_DONTWAIT | MSG_MORE | MSG_SENDPAGE_NOTLAST);
 992         if (unlikely(ret <= 0))
 993                 return ret;
 994
 995         len -= ret;
 996         if (!len) {
 997                 req->state = NVME_TCP_SEND_DATA;
 998                 if (queue->data_digest)
 999                         crypto_ahash_init(queue->snd_hash);
1000                 if (!req->data_sent)
1001                         nvme_tcp_init_iter(req, WRITE);
1002                 return 1;
1003         }
1004         req->offset += ret;
1005
1006         return -EAGAIN;
1007 }
1008
1009 static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
1010 {
1011         struct nvme_tcp_queue *queue = req->queue;
1012         int ret;
1013         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR };
1014         struct kvec iov = {
1015                 .iov_base = &req->ddgst + req->offset,
1016                 .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
1017         };
1018
1019         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1020         if (unlikely(ret <= 0))
1021                 return ret;
1022
1023         if (req->offset + ret == NVME_TCP_DIGEST_LENGTH) {
1024                 nvme_tcp_done_send_req(queue);
1025                 return 1;
1026         }
1027
1028         req->offset += ret;
1029         return -EAGAIN;
1030 }
1031
1032 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
1033 {
1034         struct nvme_tcp_request *req;
1035         int ret = 1;
1036
1037         if (!queue->request) {
1038                 queue->request = nvme_tcp_fetch_request(queue);
1039                 if (!queue->request)
1040                         return 0;
1041         }
1042         req = queue->request;
1043
1044         if (req->state == NVME_TCP_SEND_CMD_PDU) {
1045                 ret = nvme_tcp_try_send_cmd_pdu(req);
1046                 if (ret <= 0)
1047                         goto done;
1048                 if (!nvme_tcp_has_inline_data(req))
1049                         return ret;
1050         }
1051
1052         if (req->state == NVME_TCP_SEND_H2C_PDU) {
1053                 ret = nvme_tcp_try_send_data_pdu(req);
1054                 if (ret <= 0)
1055                         goto done;
1056         }
1057
1058         if (req->state == NVME_TCP_SEND_DATA) {
1059                 ret = nvme_tcp_try_send_data(req);
1060                 if (ret <= 0)
1061                         goto done;
1062         }
1063
1064         if (req->state == NVME_TCP_SEND_DDGST)
1065                 ret = nvme_tcp_try_send_ddgst(req);
1066 done:
1067         if (ret == -EAGAIN) {
1068                 ret = 0;
1069         } else if (ret < 0) {
1070                 dev_err(queue->ctrl->ctrl.device,
1071                         "failed to send request %d\n", ret);
1072                 if (ret != -EPIPE && ret != -ECONNRESET)
1073                         nvme_tcp_fail_request(queue->request);
1074                 nvme_tcp_done_send_req(queue);
1075         }
1076         return ret;
1077 }
1078
1079 static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1080 {
1081         struct socket *sock = queue->sock;
1082         struct sock *sk = sock->sk;
1083         read_descriptor_t rd_desc;
1084         int consumed;
1085
1086         rd_desc.arg.data = queue;
1087         rd_desc.count = 1;
1088         lock_sock(sk);
1089         queue->nr_cqe = 0;
1090         consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1091         release_sock(sk);
1092         return consumed;
1093 }
1094
1095 static void nvme_tcp_io_work(struct work_struct *w)
1096 {
1097         struct nvme_tcp_queue *queue =
1098                 container_of(w, struct nvme_tcp_queue, io_work);
1099         unsigned long deadline = jiffies + msecs_to_jiffies(1);
1100
1101         do {
1102                 bool pending = false;
1103                 int result;
1104
1105                 if (mutex_trylock(&queue->send_mutex)) {
1106                         result = nvme_tcp_try_send(queue);
1107                         mutex_unlock(&queue->send_mutex);
1108                         if (result > 0)
1109                                 pending = true;
1110                         else if (unlikely(result < 0))
1111                                 break;
1112                 }
1113
1114                 result = nvme_tcp_try_recv(queue);
1115                 if (result > 0)
1116                         pending = true;
1117                 else if (unlikely(result < 0))
1118                         return;
1119
1120                 if (!pending)
1121                         return;
1122
1123         } while (!time_after(jiffies, deadline)); /* quota is exhausted */
1124
1125         queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1126 }
1127
1128 static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue)
1129 {
1130         struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
1131
1132         ahash_request_free(queue->rcv_hash);
1133         ahash_request_free(queue->snd_hash);
1134         crypto_free_ahash(tfm);
1135 }
1136
1137 static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue)
1138 {
1139         struct crypto_ahash *tfm;
1140
1141         tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
1142         if (IS_ERR(tfm))
1143                 return PTR_ERR(tfm);
1144
1145         queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1146         if (!queue->snd_hash)
1147                 goto free_tfm;
1148         ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
1149
1150         queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1151         if (!queue->rcv_hash)
1152                 goto free_snd_hash;
1153         ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
1154
1155         return 0;
1156 free_snd_hash:
1157         ahash_request_free(queue->snd_hash);
1158 free_tfm:
1159         crypto_free_ahash(tfm);
1160         return -ENOMEM;
1161 }
1162
1163 static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1164 {
1165         struct nvme_tcp_request *async = &ctrl->async_req;
1166
1167         page_frag_free(async->pdu);
1168 }
1169
1170 static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1171 {
1172         struct nvme_tcp_queue *queue = &ctrl->queues[0];
1173         struct nvme_tcp_request *async = &ctrl->async_req;
1174         u8 hdgst = nvme_tcp_hdgst_len(queue);
1175
1176         async->pdu = page_frag_alloc(&queue->pf_cache,
1177                 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1178                 GFP_KERNEL | __GFP_ZERO);
1179         if (!async->pdu)
1180                 return -ENOMEM;
1181
1182         async->queue = &ctrl->queues[0];
1183         return 0;
1184 }
1185
1186 static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1187 {
1188         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1189         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1190
1191         if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1192                 return;
1193
1194         if (queue->hdr_digest || queue->data_digest)
1195                 nvme_tcp_free_crypto(queue);
1196
1197         sock_release(queue->sock);
1198         kfree(queue->pdu);
1199 }
1200
1201 static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1202 {
1203         struct nvme_tcp_icreq_pdu *icreq;
1204         struct nvme_tcp_icresp_pdu *icresp;
1205         struct msghdr msg = {};
1206         struct kvec iov;
1207         bool ctrl_hdgst, ctrl_ddgst;
1208         int ret;
1209
1210         icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1211         if (!icreq)
1212                 return -ENOMEM;
1213
1214         icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1215         if (!icresp) {
1216                 ret = -ENOMEM;
1217                 goto free_icreq;
1218         }
1219
1220         icreq->hdr.type = nvme_tcp_icreq;
1221         icreq->hdr.hlen = sizeof(*icreq);
1222         icreq->hdr.pdo = 0;
1223         icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1224         icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1225         icreq->maxr2t = 0; /* single inflight r2t supported */
1226         icreq->hpda = 0; /* no alignment constraint */
1227         if (queue->hdr_digest)
1228                 icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1229         if (queue->data_digest)
1230                 icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1231
1232         iov.iov_base = icreq;
1233         iov.iov_len = sizeof(*icreq);
1234         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1235         if (ret < 0)
1236                 goto free_icresp;
1237
1238         memset(&msg, 0, sizeof(msg));
1239         iov.iov_base = icresp;
1240         iov.iov_len = sizeof(*icresp);
1241         ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1242                         iov.iov_len, msg.msg_flags);
1243         if (ret < 0)
1244                 goto free_icresp;
1245
1246         ret = -EINVAL;
1247         if (icresp->hdr.type != nvme_tcp_icresp) {
1248                 pr_err("queue %d: bad type returned %d\n",
1249                         nvme_tcp_queue_id(queue), icresp->hdr.type);
1250                 goto free_icresp;
1251         }
1252
1253         if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1254                 pr_err("queue %d: bad pdu length returned %d\n",
1255                         nvme_tcp_queue_id(queue), icresp->hdr.plen);
1256                 goto free_icresp;
1257         }
1258
1259         if (icresp->pfv != NVME_TCP_PFV_1_0) {
1260                 pr_err("queue %d: bad pfv returned %d\n",
1261                         nvme_tcp_queue_id(queue), icresp->pfv);
1262                 goto free_icresp;
1263         }
1264
1265         ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1266         if ((queue->data_digest && !ctrl_ddgst) ||
1267             (!queue->data_digest && ctrl_ddgst)) {
1268                 pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1269                         nvme_tcp_queue_id(queue),
1270                         queue->data_digest ? "enabled" : "disabled",
1271                         ctrl_ddgst ? "enabled" : "disabled");
1272                 goto free_icresp;
1273         }
1274
1275         ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1276         if ((queue->hdr_digest && !ctrl_hdgst) ||
1277             (!queue->hdr_digest && ctrl_hdgst)) {
1278                 pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1279                         nvme_tcp_queue_id(queue),
1280                         queue->hdr_digest ? "enabled" : "disabled",
1281                         ctrl_hdgst ? "enabled" : "disabled");
1282                 goto free_icresp;
1283         }
1284
1285         if (icresp->cpda != 0) {
1286                 pr_err("queue %d: unsupported cpda returned %d\n",
1287                         nvme_tcp_queue_id(queue), icresp->cpda);
1288                 goto free_icresp;
1289         }
1290
1291         ret = 0;
1292 free_icresp:
1293         kfree(icresp);
1294 free_icreq:
1295         kfree(icreq);
1296         return ret;
1297 }
1298
1299 static bool nvme_tcp_admin_queue(struct nvme_tcp_queue *queue)
1300 {
1301         return nvme_tcp_queue_id(queue) == 0;
1302 }
1303
1304 static bool nvme_tcp_default_queue(struct nvme_tcp_queue *queue)
1305 {
1306         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1307         int qid = nvme_tcp_queue_id(queue);
1308
1309         return !nvme_tcp_admin_queue(queue) &&
1310                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT];
1311 }
1312
1313 static bool nvme_tcp_read_queue(struct nvme_tcp_queue *queue)
1314 {
1315         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1316         int qid = nvme_tcp_queue_id(queue);
1317
1318         return !nvme_tcp_admin_queue(queue) &&
1319                 !nvme_tcp_default_queue(queue) &&
1320                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1321                           ctrl->io_queues[HCTX_TYPE_READ];
1322 }
1323
1324 static bool nvme_tcp_poll_queue(struct nvme_tcp_queue *queue)
1325 {
1326         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1327         int qid = nvme_tcp_queue_id(queue);
1328
1329         return !nvme_tcp_admin_queue(queue) &&
1330                 !nvme_tcp_default_queue(queue) &&
1331                 !nvme_tcp_read_queue(queue) &&
1332                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1333                           ctrl->io_queues[HCTX_TYPE_READ] +
1334                           ctrl->io_queues[HCTX_TYPE_POLL];
1335 }
1336
1337 static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
1338 {
1339         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1340         int qid = nvme_tcp_queue_id(queue);
1341         int n = 0;
1342
1343         if (nvme_tcp_default_queue(queue))
1344                 n = qid - 1;
1345         else if (nvme_tcp_read_queue(queue))
1346                 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] - 1;
1347         else if (nvme_tcp_poll_queue(queue))
1348                 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] -
1349                                 ctrl->io_queues[HCTX_TYPE_READ] - 1;
1350         queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
1351 }
1352
1353 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl,
1354                 int qid, size_t queue_size)
1355 {
1356         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1357         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1358         int ret, rcv_pdu_size;
1359
1360         queue->ctrl = ctrl;
1361         init_llist_head(&queue->req_list);
1362         INIT_LIST_HEAD(&queue->send_list);
1363         mutex_init(&queue->send_mutex);
1364         INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1365         queue->queue_size = queue_size;
1366
1367         if (qid > 0)
1368                 queue->cmnd_capsule_len = nctrl->ioccsz * 16;
1369         else
1370                 queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1371                                                 NVME_TCP_ADMIN_CCSZ;
1372
1373         ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM,
1374                         IPPROTO_TCP, &queue->sock);
1375         if (ret) {
1376                 dev_err(nctrl->device,
1377                         "failed to create socket: %d\n", ret);
1378                 return ret;
1379         }
1380
1381         /* Single syn retry */
1382         tcp_sock_set_syncnt(queue->sock->sk, 1);
1383
1384         /* Set TCP no delay */
1385         tcp_sock_set_nodelay(queue->sock->sk);
1386
1387         /*
1388          * Cleanup whatever is sitting in the TCP transmit queue on socket
1389          * close. This is done to prevent stale data from being sent should
1390          * the network connection be restored before TCP times out.
1391          */
1392         sock_no_linger(queue->sock->sk);
1393
1394         if (so_priority > 0)
1395                 sock_set_priority(queue->sock->sk, so_priority);
1396
1397         /* Set socket type of service */
1398         if (nctrl->opts->tos >= 0)
1399                 ip_sock_set_tos(queue->sock->sk, nctrl->opts->tos);
1400
1401         queue->sock->sk->sk_allocation = GFP_ATOMIC;
1402         nvme_tcp_set_queue_io_cpu(queue);
1403         queue->request = NULL;
1404         queue->data_remaining = 0;
1405         queue->ddgst_remaining = 0;
1406         queue->pdu_remaining = 0;
1407         queue->pdu_offset = 0;
1408         sk_set_memalloc(queue->sock->sk);
1409
1410         if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
1411                 ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1412                         sizeof(ctrl->src_addr));
1413                 if (ret) {
1414                         dev_err(nctrl->device,
1415                                 "failed to bind queue %d socket %d\n",
1416                                 qid, ret);
1417                         goto err_sock;
1418                 }
1419         }
1420
1421         queue->hdr_digest = nctrl->opts->hdr_digest;
1422         queue->data_digest = nctrl->opts->data_digest;
1423         if (queue->hdr_digest || queue->data_digest) {
1424                 ret = nvme_tcp_alloc_crypto(queue);
1425                 if (ret) {
1426                         dev_err(nctrl->device,
1427                                 "failed to allocate queue %d crypto\n", qid);
1428                         goto err_sock;
1429                 }
1430         }
1431
1432         rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1433                         nvme_tcp_hdgst_len(queue);
1434         queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1435         if (!queue->pdu) {
1436                 ret = -ENOMEM;
1437                 goto err_crypto;
1438         }
1439
1440         dev_dbg(nctrl->device, "connecting queue %d\n",
1441                         nvme_tcp_queue_id(queue));
1442
1443         ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1444                 sizeof(ctrl->addr), 0);
1445         if (ret) {
1446                 dev_err(nctrl->device,
1447                         "failed to connect socket: %d\n", ret);
1448                 goto err_rcv_pdu;
1449         }
1450
1451         ret = nvme_tcp_init_connection(queue);
1452         if (ret)
1453                 goto err_init_connect;
1454
1455         queue->rd_enabled = true;
1456         set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1457         nvme_tcp_init_recv_ctx(queue);
1458
1459         write_lock_bh(&queue->sock->sk->sk_callback_lock);
1460         queue->sock->sk->sk_user_data = queue;
1461         queue->state_change = queue->sock->sk->sk_state_change;
1462         queue->data_ready = queue->sock->sk->sk_data_ready;
1463         queue->write_space = queue->sock->sk->sk_write_space;
1464         queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1465         queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1466         queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1467 #ifdef CONFIG_NET_RX_BUSY_POLL
1468         queue->sock->sk->sk_ll_usec = 1;
1469 #endif
1470         write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1471
1472         return 0;
1473
1474 err_init_connect:
1475         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1476 err_rcv_pdu:
1477         kfree(queue->pdu);
1478 err_crypto:
1479         if (queue->hdr_digest || queue->data_digest)
1480                 nvme_tcp_free_crypto(queue);
1481 err_sock:
1482         sock_release(queue->sock);
1483         queue->sock = NULL;
1484         return ret;
1485 }
1486
1487 static void nvme_tcp_restore_sock_calls(struct nvme_tcp_queue *queue)
1488 {
1489         struct socket *sock = queue->sock;
1490
1491         write_lock_bh(&sock->sk->sk_callback_lock);
1492         sock->sk->sk_user_data  = NULL;
1493         sock->sk->sk_data_ready = queue->data_ready;
1494         sock->sk->sk_state_change = queue->state_change;
1495         sock->sk->sk_write_space  = queue->write_space;
1496         write_unlock_bh(&sock->sk->sk_callback_lock);
1497 }
1498
1499 static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1500 {
1501         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1502         nvme_tcp_restore_sock_calls(queue);
1503         cancel_work_sync(&queue->io_work);
1504 }
1505
1506 static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1507 {
1508         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1509         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1510
1511         if (!test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1512                 return;
1513
1514         __nvme_tcp_stop_queue(queue);
1515 }
1516
1517 static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1518 {
1519         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1520         int ret;
1521
1522         if (idx)
1523                 ret = nvmf_connect_io_queue(nctrl, idx, false);
1524         else
1525                 ret = nvmf_connect_admin_queue(nctrl);
1526
1527         if (!ret) {
1528                 set_bit(NVME_TCP_Q_LIVE, &ctrl->queues[idx].flags);
1529         } else {
1530                 if (test_bit(NVME_TCP_Q_ALLOCATED, &ctrl->queues[idx].flags))
1531                         __nvme_tcp_stop_queue(&ctrl->queues[idx]);
1532                 dev_err(nctrl->device,
1533                         "failed to connect queue: %d ret=%d\n", idx, ret);
1534         }
1535         return ret;
1536 }
1537
1538 static struct blk_mq_tag_set *nvme_tcp_alloc_tagset(struct nvme_ctrl *nctrl,
1539                 bool admin)
1540 {
1541         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1542         struct blk_mq_tag_set *set;
1543         int ret;
1544
1545         if (admin) {
1546                 set = &ctrl->admin_tag_set;
1547                 memset(set, 0, sizeof(*set));
1548                 set->ops = &nvme_tcp_admin_mq_ops;
1549                 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1550                 set->reserved_tags = 2; /* connect + keep-alive */
1551                 set->numa_node = nctrl->numa_node;
1552                 set->flags = BLK_MQ_F_BLOCKING;
1553                 set->cmd_size = sizeof(struct nvme_tcp_request);
1554                 set->driver_data = ctrl;
1555                 set->nr_hw_queues = 1;
1556                 set->timeout = ADMIN_TIMEOUT;
1557         } else {
1558                 set = &ctrl->tag_set;
1559                 memset(set, 0, sizeof(*set));
1560                 set->ops = &nvme_tcp_mq_ops;
1561                 set->queue_depth = nctrl->sqsize + 1;
1562                 set->reserved_tags = 1; /* fabric connect */
1563                 set->numa_node = nctrl->numa_node;
1564                 set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_BLOCKING;
1565                 set->cmd_size = sizeof(struct nvme_tcp_request);
1566                 set->driver_data = ctrl;
1567                 set->nr_hw_queues = nctrl->queue_count - 1;
1568                 set->timeout = NVME_IO_TIMEOUT;
1569                 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
1570         }
1571
1572         ret = blk_mq_alloc_tag_set(set);
1573         if (ret)
1574                 return ERR_PTR(ret);
1575
1576         return set;
1577 }
1578
1579 static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
1580 {
1581         if (to_tcp_ctrl(ctrl)->async_req.pdu) {
1582                 nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
1583                 to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
1584         }
1585
1586         nvme_tcp_free_queue(ctrl, 0);
1587 }
1588
1589 static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
1590 {
1591         int i;
1592
1593         for (i = 1; i < ctrl->queue_count; i++)
1594                 nvme_tcp_free_queue(ctrl, i);
1595 }
1596
1597 static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
1598 {
1599         int i;
1600
1601         for (i = 1; i < ctrl->queue_count; i++)
1602                 nvme_tcp_stop_queue(ctrl, i);
1603 }
1604
1605 static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl)
1606 {
1607         int i, ret = 0;
1608
1609         for (i = 1; i < ctrl->queue_count; i++) {
1610                 ret = nvme_tcp_start_queue(ctrl, i);
1611                 if (ret)
1612                         goto out_stop_queues;
1613         }
1614
1615         return 0;
1616
1617 out_stop_queues:
1618         for (i--; i >= 1; i--)
1619                 nvme_tcp_stop_queue(ctrl, i);
1620         return ret;
1621 }
1622
1623 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
1624 {
1625         int ret;
1626
1627         ret = nvme_tcp_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
1628         if (ret)
1629                 return ret;
1630
1631         ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
1632         if (ret)
1633                 goto out_free_queue;
1634
1635         return 0;
1636
1637 out_free_queue:
1638         nvme_tcp_free_queue(ctrl, 0);
1639         return ret;
1640 }
1641
1642 static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1643 {
1644         int i, ret;
1645
1646         for (i = 1; i < ctrl->queue_count; i++) {
1647                 ret = nvme_tcp_alloc_queue(ctrl, i,
1648                                 ctrl->sqsize + 1);
1649                 if (ret)
1650                         goto out_free_queues;
1651         }
1652
1653         return 0;
1654
1655 out_free_queues:
1656         for (i--; i >= 1; i--)
1657                 nvme_tcp_free_queue(ctrl, i);
1658
1659         return ret;
1660 }
1661
1662 static unsigned int nvme_tcp_nr_io_queues(struct nvme_ctrl *ctrl)
1663 {
1664         unsigned int nr_io_queues;
1665
1666         nr_io_queues = min(ctrl->opts->nr_io_queues, num_online_cpus());
1667         nr_io_queues += min(ctrl->opts->nr_write_queues, num_online_cpus());
1668         nr_io_queues += min(ctrl->opts->nr_poll_queues, num_online_cpus());
1669
1670         return nr_io_queues;
1671 }
1672
1673 static void nvme_tcp_set_io_queues(struct nvme_ctrl *nctrl,
1674                 unsigned int nr_io_queues)
1675 {
1676         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1677         struct nvmf_ctrl_options *opts = nctrl->opts;
1678
1679         if (opts->nr_write_queues && opts->nr_io_queues < nr_io_queues) {
1680                 /*
1681                  * separate read/write queues
1682                  * hand out dedicated default queues only after we have
1683                  * sufficient read queues.
1684                  */
1685                 ctrl->io_queues[HCTX_TYPE_READ] = opts->nr_io_queues;
1686                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
1687                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1688                         min(opts->nr_write_queues, nr_io_queues);
1689                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1690         } else {
1691                 /*
1692                  * shared read/write queues
1693                  * either no write queues were requested, or we don't have
1694                  * sufficient queue count to have dedicated default queues.
1695                  */
1696                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1697                         min(opts->nr_io_queues, nr_io_queues);
1698                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1699         }
1700
1701         if (opts->nr_poll_queues && nr_io_queues) {
1702                 /* map dedicated poll queues only if we have queues left */
1703                 ctrl->io_queues[HCTX_TYPE_POLL] =
1704                         min(opts->nr_poll_queues, nr_io_queues);
1705         }
1706 }
1707
1708 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1709 {
1710         unsigned int nr_io_queues;
1711         int ret;
1712
1713         nr_io_queues = nvme_tcp_nr_io_queues(ctrl);
1714         ret = nvme_set_queue_count(ctrl, &nr_io_queues);
1715         if (ret)
1716                 return ret;
1717
1718         ctrl->queue_count = nr_io_queues + 1;
1719         if (ctrl->queue_count < 2)
1720                 return 0;
1721
1722         dev_info(ctrl->device,
1723                 "creating %d I/O queues.\n", nr_io_queues);
1724
1725         nvme_tcp_set_io_queues(ctrl, nr_io_queues);
1726
1727         return __nvme_tcp_alloc_io_queues(ctrl);
1728 }
1729
1730 static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove)
1731 {
1732         nvme_tcp_stop_io_queues(ctrl);
1733         if (remove) {
1734                 blk_cleanup_queue(ctrl->connect_q);
1735                 blk_mq_free_tag_set(ctrl->tagset);
1736         }
1737         nvme_tcp_free_io_queues(ctrl);
1738 }
1739
1740 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
1741 {
1742         int ret;
1743
1744         ret = nvme_tcp_alloc_io_queues(ctrl);
1745         if (ret)
1746                 return ret;
1747
1748         if (new) {
1749                 ctrl->tagset = nvme_tcp_alloc_tagset(ctrl, false);
1750                 if (IS_ERR(ctrl->tagset)) {
1751                         ret = PTR_ERR(ctrl->tagset);
1752                         goto out_free_io_queues;
1753                 }
1754
1755                 ctrl->connect_q = blk_mq_init_queue(ctrl->tagset);
1756                 if (IS_ERR(ctrl->connect_q)) {
1757                         ret = PTR_ERR(ctrl->connect_q);
1758                         goto out_free_tag_set;
1759                 }
1760         } else {
1761                 blk_mq_update_nr_hw_queues(ctrl->tagset,
1762                         ctrl->queue_count - 1);
1763         }
1764
1765         ret = nvme_tcp_start_io_queues(ctrl);
1766         if (ret)
1767                 goto out_cleanup_connect_q;
1768
1769         return 0;
1770
1771 out_cleanup_connect_q:
1772         if (new)
1773                 blk_cleanup_queue(ctrl->connect_q);
1774 out_free_tag_set:
1775         if (new)
1776                 blk_mq_free_tag_set(ctrl->tagset);
1777 out_free_io_queues:
1778         nvme_tcp_free_io_queues(ctrl);
1779         return ret;
1780 }
1781
1782 static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove)
1783 {
1784         nvme_tcp_stop_queue(ctrl, 0);
1785         if (remove) {
1786                 blk_cleanup_queue(ctrl->admin_q);
1787                 blk_cleanup_queue(ctrl->fabrics_q);
1788                 blk_mq_free_tag_set(ctrl->admin_tagset);
1789         }
1790         nvme_tcp_free_admin_queue(ctrl);
1791 }
1792
1793 static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
1794 {
1795         int error;
1796
1797         error = nvme_tcp_alloc_admin_queue(ctrl);
1798         if (error)
1799                 return error;
1800
1801         if (new) {
1802                 ctrl->admin_tagset = nvme_tcp_alloc_tagset(ctrl, true);
1803                 if (IS_ERR(ctrl->admin_tagset)) {
1804                         error = PTR_ERR(ctrl->admin_tagset);
1805                         goto out_free_queue;
1806                 }
1807
1808                 ctrl->fabrics_q = blk_mq_init_queue(ctrl->admin_tagset);
1809                 if (IS_ERR(ctrl->fabrics_q)) {
1810                         error = PTR_ERR(ctrl->fabrics_q);
1811                         goto out_free_tagset;
1812                 }
1813
1814                 ctrl->admin_q = blk_mq_init_queue(ctrl->admin_tagset);
1815                 if (IS_ERR(ctrl->admin_q)) {
1816                         error = PTR_ERR(ctrl->admin_q);
1817                         goto out_cleanup_fabrics_q;
1818                 }
1819         }
1820
1821         error = nvme_tcp_start_queue(ctrl, 0);
1822         if (error)
1823                 goto out_cleanup_queue;
1824
1825         error = nvme_enable_ctrl(ctrl);
1826         if (error)
1827                 goto out_stop_queue;
1828
1829         blk_mq_unquiesce_queue(ctrl->admin_q);
1830
1831         error = nvme_init_identify(ctrl);
1832         if (error)
1833                 goto out_stop_queue;
1834
1835         return 0;
1836
1837 out_stop_queue:
1838         nvme_tcp_stop_queue(ctrl, 0);
1839 out_cleanup_queue:
1840         if (new)
1841                 blk_cleanup_queue(ctrl->admin_q);
1842 out_cleanup_fabrics_q:
1843         if (new)
1844                 blk_cleanup_queue(ctrl->fabrics_q);
1845 out_free_tagset:
1846         if (new)
1847                 blk_mq_free_tag_set(ctrl->admin_tagset);
1848 out_free_queue:
1849         nvme_tcp_free_admin_queue(ctrl);
1850         return error;
1851 }
1852
1853 static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
1854                 bool remove)
1855 {
1856         blk_mq_quiesce_queue(ctrl->admin_q);
1857         nvme_tcp_stop_queue(ctrl, 0);
1858         if (ctrl->admin_tagset) {
1859                 blk_mq_tagset_busy_iter(ctrl->admin_tagset,
1860                         nvme_cancel_request, ctrl);
1861                 blk_mq_tagset_wait_completed_request(ctrl->admin_tagset);
1862         }
1863         if (remove)
1864                 blk_mq_unquiesce_queue(ctrl->admin_q);
1865         nvme_tcp_destroy_admin_queue(ctrl, remove);
1866 }
1867
1868 static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
1869                 bool remove)
1870 {
1871         if (ctrl->queue_count <= 1)
1872                 return;
1873         nvme_stop_queues(ctrl);
1874         nvme_tcp_stop_io_queues(ctrl);
1875         if (ctrl->tagset) {
1876                 blk_mq_tagset_busy_iter(ctrl->tagset,
1877                         nvme_cancel_request, ctrl);
1878                 blk_mq_tagset_wait_completed_request(ctrl->tagset);
1879         }
1880         if (remove)
1881                 nvme_start_queues(ctrl);
1882         nvme_tcp_destroy_io_queues(ctrl, remove);
1883 }
1884
1885 static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
1886 {
1887         /* If we are resetting/deleting then do nothing */
1888         if (ctrl->state != NVME_CTRL_CONNECTING) {
1889                 WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW ||
1890                         ctrl->state == NVME_CTRL_LIVE);
1891                 return;
1892         }
1893
1894         if (nvmf_should_reconnect(ctrl)) {
1895                 dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
1896                         ctrl->opts->reconnect_delay);
1897                 queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
1898                                 ctrl->opts->reconnect_delay * HZ);
1899         } else {
1900                 dev_info(ctrl->device, "Removing controller...\n");
1901                 nvme_delete_ctrl(ctrl);
1902         }
1903 }
1904
1905 static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
1906 {
1907         struct nvmf_ctrl_options *opts = ctrl->opts;
1908         int ret;
1909
1910         ret = nvme_tcp_configure_admin_queue(ctrl, new);
1911         if (ret)
1912                 return ret;
1913
1914         if (ctrl->icdoff) {
1915                 dev_err(ctrl->device, "icdoff is not supported!\n");
1916                 goto destroy_admin;
1917         }
1918
1919         if (opts->queue_size > ctrl->sqsize + 1)
1920                 dev_warn(ctrl->device,
1921                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
1922                         opts->queue_size, ctrl->sqsize + 1);
1923
1924         if (ctrl->sqsize + 1 > ctrl->maxcmd) {
1925                 dev_warn(ctrl->device,
1926                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
1927                         ctrl->sqsize + 1, ctrl->maxcmd);
1928                 ctrl->sqsize = ctrl->maxcmd - 1;
1929         }
1930
1931         if (ctrl->queue_count > 1) {
1932                 ret = nvme_tcp_configure_io_queues(ctrl, new);
1933                 if (ret)
1934                         goto destroy_admin;
1935         }
1936
1937         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
1938                 /*
1939                  * state change failure is ok if we're in DELETING state,
1940                  * unless we're during creation of a new controller to
1941                  * avoid races with teardown flow.
1942                  */
1943                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
1944                 WARN_ON_ONCE(new);
1945                 ret = -EINVAL;
1946                 goto destroy_io;
1947         }
1948
1949         nvme_start_ctrl(ctrl);
1950         return 0;
1951
1952 destroy_io:
1953         if (ctrl->queue_count > 1)
1954                 nvme_tcp_destroy_io_queues(ctrl, new);
1955 destroy_admin:
1956         nvme_tcp_stop_queue(ctrl, 0);
1957         nvme_tcp_destroy_admin_queue(ctrl, new);
1958         return ret;
1959 }
1960
1961 static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
1962 {
1963         struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
1964                         struct nvme_tcp_ctrl, connect_work);
1965         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
1966
1967         ++ctrl->nr_reconnects;
1968
1969         if (nvme_tcp_setup_ctrl(ctrl, false))
1970                 goto requeue;
1971
1972         dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n",
1973                         ctrl->nr_reconnects);
1974
1975         ctrl->nr_reconnects = 0;
1976
1977         return;
1978
1979 requeue:
1980         dev_info(ctrl->device, "Failed reconnect attempt %d\n",
1981                         ctrl->nr_reconnects);
1982         nvme_tcp_reconnect_or_remove(ctrl);
1983 }
1984
1985 static void nvme_tcp_error_recovery_work(struct work_struct *work)
1986 {
1987         struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
1988                                 struct nvme_tcp_ctrl, err_work);
1989         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
1990
1991         nvme_stop_keep_alive(ctrl);
1992         nvme_tcp_teardown_io_queues(ctrl, false);
1993         /* unquiesce to fail fast pending requests */
1994         nvme_start_queues(ctrl);
1995         nvme_tcp_teardown_admin_queue(ctrl, false);
1996         blk_mq_unquiesce_queue(ctrl->admin_q);
1997
1998         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
1999                 /* state change failure is ok if we're in DELETING state */
2000                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
2001                 return;
2002         }
2003
2004         nvme_tcp_reconnect_or_remove(ctrl);
2005 }
2006
2007 static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
2008 {
2009         cancel_work_sync(&to_tcp_ctrl(ctrl)->err_work);
2010         cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
2011
2012         nvme_tcp_teardown_io_queues(ctrl, shutdown);
2013         blk_mq_quiesce_queue(ctrl->admin_q);
2014         if (shutdown)
2015                 nvme_shutdown_ctrl(ctrl);
2016         else
2017                 nvme_disable_ctrl(ctrl);
2018         nvme_tcp_teardown_admin_queue(ctrl, shutdown);
2019 }
2020
2021 static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
2022 {
2023         nvme_tcp_teardown_ctrl(ctrl, true);
2024 }
2025
2026 static void nvme_reset_ctrl_work(struct work_struct *work)
2027 {
2028         struct nvme_ctrl *ctrl =
2029                 container_of(work, struct nvme_ctrl, reset_work);
2030
2031         nvme_stop_ctrl(ctrl);
2032         nvme_tcp_teardown_ctrl(ctrl, false);
2033
2034         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2035                 /* state change failure is ok if we're in DELETING state */
2036                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
2037                 return;
2038         }
2039
2040         if (nvme_tcp_setup_ctrl(ctrl, false))
2041                 goto out_fail;
2042
2043         return;
2044
2045 out_fail:
2046         ++ctrl->nr_reconnects;
2047         nvme_tcp_reconnect_or_remove(ctrl);
2048 }
2049
2050 static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
2051 {
2052         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
2053
2054         if (list_empty(&ctrl->list))
2055                 goto free_ctrl;
2056
2057         mutex_lock(&nvme_tcp_ctrl_mutex);
2058         list_del(&ctrl->list);
2059         mutex_unlock(&nvme_tcp_ctrl_mutex);
2060
2061         nvmf_free_options(nctrl->opts);
2062 free_ctrl:
2063         kfree(ctrl->queues);
2064         kfree(ctrl);
2065 }
2066
2067 static void nvme_tcp_set_sg_null(struct nvme_command *c)
2068 {
2069         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2070
2071         sg->addr = 0;
2072         sg->length = 0;
2073         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2074                         NVME_SGL_FMT_TRANSPORT_A;
2075 }
2076
2077 static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
2078                 struct nvme_command *c, u32 data_len)
2079 {
2080         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2081
2082         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
2083         sg->length = cpu_to_le32(data_len);
2084         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
2085 }
2086
2087 static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
2088                 u32 data_len)
2089 {
2090         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2091
2092         sg->addr = 0;
2093         sg->length = cpu_to_le32(data_len);
2094         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2095                         NVME_SGL_FMT_TRANSPORT_A;
2096 }
2097
2098 static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
2099 {
2100         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2101         struct nvme_tcp_queue *queue = &ctrl->queues[0];
2102         struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2103         struct nvme_command *cmd = &pdu->cmd;
2104         u8 hdgst = nvme_tcp_hdgst_len(queue);
2105
2106         memset(pdu, 0, sizeof(*pdu));
2107         pdu->hdr.type = nvme_tcp_cmd;
2108         if (queue->hdr_digest)
2109                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2110         pdu->hdr.hlen = sizeof(*pdu);
2111         pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2112
2113         cmd->common.opcode = nvme_admin_async_event;
2114         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2115         cmd->common.flags |= NVME_CMD_SGL_METABUF;
2116         nvme_tcp_set_sg_null(cmd);
2117
2118         ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2119         ctrl->async_req.offset = 0;
2120         ctrl->async_req.curr_bio = NULL;
2121         ctrl->async_req.data_len = 0;
2122
2123         nvme_tcp_queue_request(&ctrl->async_req, true);
2124 }
2125
2126 static enum blk_eh_timer_return
2127 nvme_tcp_timeout(struct request *rq, bool reserved)
2128 {
2129         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2130         struct nvme_tcp_ctrl *ctrl = req->queue->ctrl;
2131         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2132
2133         /*
2134          * Restart the timer if a controller reset is already scheduled. Any
2135          * timed out commands would be handled before entering the connecting
2136          * state.
2137          */
2138         if (ctrl->ctrl.state == NVME_CTRL_RESETTING)
2139                 return BLK_EH_RESET_TIMER;
2140
2141         dev_warn(ctrl->ctrl.device,
2142                 "queue %d: timeout request %#x type %d\n",
2143                 nvme_tcp_queue_id(req->queue), rq->tag, pdu->hdr.type);
2144
2145         if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
2146                 /*
2147                  * Teardown immediately if controller times out while starting
2148                  * or we are already started error recovery. all outstanding
2149                  * requests are completed on shutdown, so we return BLK_EH_DONE.
2150                  */
2151                 flush_work(&ctrl->err_work);
2152                 nvme_tcp_teardown_io_queues(&ctrl->ctrl, false);
2153                 nvme_tcp_teardown_admin_queue(&ctrl->ctrl, false);
2154                 return BLK_EH_DONE;
2155         }
2156
2157         dev_warn(ctrl->ctrl.device, "starting error recovery\n");
2158         nvme_tcp_error_recovery(&ctrl->ctrl);
2159
2160         return BLK_EH_RESET_TIMER;
2161 }
2162
2163 static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2164                         struct request *rq)
2165 {
2166         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2167         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2168         struct nvme_command *c = &pdu->cmd;
2169
2170         c->common.flags |= NVME_CMD_SGL_METABUF;
2171
2172         if (!blk_rq_nr_phys_segments(rq))
2173                 nvme_tcp_set_sg_null(c);
2174         else if (rq_data_dir(rq) == WRITE &&
2175             req->data_len <= nvme_tcp_inline_data_size(queue))
2176                 nvme_tcp_set_sg_inline(queue, c, req->data_len);
2177         else
2178                 nvme_tcp_set_sg_host_data(c, req->data_len);
2179
2180         return 0;
2181 }
2182
2183 static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2184                 struct request *rq)
2185 {
2186         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2187         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2188         struct nvme_tcp_queue *queue = req->queue;
2189         u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2190         blk_status_t ret;
2191
2192         ret = nvme_setup_cmd(ns, rq, &pdu->cmd);
2193         if (ret)
2194                 return ret;
2195
2196         req->state = NVME_TCP_SEND_CMD_PDU;
2197         req->offset = 0;
2198         req->data_sent = 0;
2199         req->pdu_len = 0;
2200         req->pdu_sent = 0;
2201         req->data_len = blk_rq_nr_phys_segments(rq) ?
2202                                 blk_rq_payload_bytes(rq) : 0;
2203         req->curr_bio = rq->bio;
2204
2205         if (rq_data_dir(rq) == WRITE &&
2206             req->data_len <= nvme_tcp_inline_data_size(queue))
2207                 req->pdu_len = req->data_len;
2208         else if (req->curr_bio)
2209                 nvme_tcp_init_iter(req, READ);
2210
2211         pdu->hdr.type = nvme_tcp_cmd;
2212         pdu->hdr.flags = 0;
2213         if (queue->hdr_digest)
2214                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2215         if (queue->data_digest && req->pdu_len) {
2216                 pdu->hdr.flags |= NVME_TCP_F_DDGST;
2217                 ddgst = nvme_tcp_ddgst_len(queue);
2218         }
2219         pdu->hdr.hlen = sizeof(*pdu);
2220         pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2221         pdu->hdr.plen =
2222                 cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2223
2224         ret = nvme_tcp_map_data(queue, rq);
2225         if (unlikely(ret)) {
2226                 nvme_cleanup_cmd(rq);
2227                 dev_err(queue->ctrl->ctrl.device,
2228                         "Failed to map data (%d)\n", ret);
2229                 return ret;
2230         }
2231
2232         return 0;
2233 }
2234
2235 static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2236                 const struct blk_mq_queue_data *bd)
2237 {
2238         struct nvme_ns *ns = hctx->queue->queuedata;
2239         struct nvme_tcp_queue *queue = hctx->driver_data;
2240         struct request *rq = bd->rq;
2241         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2242         bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2243         blk_status_t ret;
2244
2245         if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2246                 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
2247
2248         ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2249         if (unlikely(ret))
2250                 return ret;
2251
2252         blk_mq_start_request(rq);
2253
2254         nvme_tcp_queue_request(req, true);
2255
2256         return BLK_STS_OK;
2257 }
2258
2259 static int nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2260 {
2261         struct nvme_tcp_ctrl *ctrl = set->driver_data;
2262         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2263
2264         if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2265                 /* separate read/write queues */
2266                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2267                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2268                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2269                 set->map[HCTX_TYPE_READ].nr_queues =
2270                         ctrl->io_queues[HCTX_TYPE_READ];
2271                 set->map[HCTX_TYPE_READ].queue_offset =
2272                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2273         } else {
2274                 /* shared read/write queues */
2275                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2276                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2277                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2278                 set->map[HCTX_TYPE_READ].nr_queues =
2279                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2280                 set->map[HCTX_TYPE_READ].queue_offset = 0;
2281         }
2282         blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2283         blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
2284
2285         if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2286                 /* map dedicated poll queues only if we have queues left */
2287                 set->map[HCTX_TYPE_POLL].nr_queues =
2288                                 ctrl->io_queues[HCTX_TYPE_POLL];
2289                 set->map[HCTX_TYPE_POLL].queue_offset =
2290                         ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2291                         ctrl->io_queues[HCTX_TYPE_READ];
2292                 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2293         }
2294
2295         dev_info(ctrl->ctrl.device,
2296                 "mapped %d/%d/%d default/read/poll queues.\n",
2297                 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2298                 ctrl->io_queues[HCTX_TYPE_READ],
2299                 ctrl->io_queues[HCTX_TYPE_POLL]);
2300
2301         return 0;
2302 }
2303
2304 static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx)
2305 {
2306         struct nvme_tcp_queue *queue = hctx->driver_data;
2307         struct sock *sk = queue->sock->sk;
2308
2309         if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2310                 return 0;
2311
2312         set_bit(NVME_TCP_Q_POLLING, &queue->flags);
2313         if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue))
2314                 sk_busy_loop(sk, true);
2315         nvme_tcp_try_recv(queue);
2316         clear_bit(NVME_TCP_Q_POLLING, &queue->flags);
2317         return queue->nr_cqe;
2318 }
2319
2320 static const struct blk_mq_ops nvme_tcp_mq_ops = {
2321         .queue_rq       = nvme_tcp_queue_rq,
2322         .complete       = nvme_complete_rq,
2323         .init_request   = nvme_tcp_init_request,
2324         .exit_request   = nvme_tcp_exit_request,
2325         .init_hctx      = nvme_tcp_init_hctx,
2326         .timeout        = nvme_tcp_timeout,
2327         .map_queues     = nvme_tcp_map_queues,
2328         .poll           = nvme_tcp_poll,
2329 };
2330
2331 static const struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2332         .queue_rq       = nvme_tcp_queue_rq,
2333         .complete       = nvme_complete_rq,
2334         .init_request   = nvme_tcp_init_request,
2335         .exit_request   = nvme_tcp_exit_request,
2336         .init_hctx      = nvme_tcp_init_admin_hctx,
2337         .timeout        = nvme_tcp_timeout,
2338 };
2339
2340 static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2341         .name                   = "tcp",
2342         .module                 = THIS_MODULE,
2343         .flags                  = NVME_F_FABRICS,
2344         .reg_read32             = nvmf_reg_read32,
2345         .reg_read64             = nvmf_reg_read64,
2346         .reg_write32            = nvmf_reg_write32,
2347         .free_ctrl              = nvme_tcp_free_ctrl,
2348         .submit_async_event     = nvme_tcp_submit_async_event,
2349         .delete_ctrl            = nvme_tcp_delete_ctrl,
2350         .get_address            = nvmf_get_address,
2351 };
2352
2353 static bool
2354 nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2355 {
2356         struct nvme_tcp_ctrl *ctrl;
2357         bool found = false;
2358
2359         mutex_lock(&nvme_tcp_ctrl_mutex);
2360         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2361                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2362                 if (found)
2363                         break;
2364         }
2365         mutex_unlock(&nvme_tcp_ctrl_mutex);
2366
2367         return found;
2368 }
2369
2370 static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2371                 struct nvmf_ctrl_options *opts)
2372 {
2373         struct nvme_tcp_ctrl *ctrl;
2374         int ret;
2375
2376         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2377         if (!ctrl)
2378                 return ERR_PTR(-ENOMEM);
2379
2380         INIT_LIST_HEAD(&ctrl->list);
2381         ctrl->ctrl.opts = opts;
2382         ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2383                                 opts->nr_poll_queues + 1;
2384         ctrl->ctrl.sqsize = opts->queue_size - 1;
2385         ctrl->ctrl.kato = opts->kato;
2386
2387         INIT_DELAYED_WORK(&ctrl->connect_work,
2388                         nvme_tcp_reconnect_ctrl_work);
2389         INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2390         INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2391
2392         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2393                 opts->trsvcid =
2394                         kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2395                 if (!opts->trsvcid) {
2396                         ret = -ENOMEM;
2397                         goto out_free_ctrl;
2398                 }
2399                 opts->mask |= NVMF_OPT_TRSVCID;
2400         }
2401
2402         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2403                         opts->traddr, opts->trsvcid, &ctrl->addr);
2404         if (ret) {
2405                 pr_err("malformed address passed: %s:%s\n",
2406                         opts->traddr, opts->trsvcid);
2407                 goto out_free_ctrl;
2408         }
2409
2410         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2411                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2412                         opts->host_traddr, NULL, &ctrl->src_addr);
2413                 if (ret) {
2414                         pr_err("malformed src address passed: %s\n",
2415                                opts->host_traddr);
2416                         goto out_free_ctrl;
2417                 }
2418         }
2419
2420         if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2421                 ret = -EALREADY;
2422                 goto out_free_ctrl;
2423         }
2424
2425         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2426                                 GFP_KERNEL);
2427         if (!ctrl->queues) {
2428                 ret = -ENOMEM;
2429                 goto out_free_ctrl;
2430         }
2431
2432         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2433         if (ret)
2434                 goto out_kfree_queues;
2435
2436         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2437                 WARN_ON_ONCE(1);
2438                 ret = -EINTR;
2439                 goto out_uninit_ctrl;
2440         }
2441
2442         ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2443         if (ret)
2444                 goto out_uninit_ctrl;
2445
2446         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
2447                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2448
2449         mutex_lock(&nvme_tcp_ctrl_mutex);
2450         list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2451         mutex_unlock(&nvme_tcp_ctrl_mutex);
2452
2453         return &ctrl->ctrl;
2454
2455 out_uninit_ctrl:
2456         nvme_uninit_ctrl(&ctrl->ctrl);
2457         nvme_put_ctrl(&ctrl->ctrl);
2458         if (ret > 0)
2459                 ret = -EIO;
2460         return ERR_PTR(ret);
2461 out_kfree_queues:
2462         kfree(ctrl->queues);
2463 out_free_ctrl:
2464         kfree(ctrl);
2465         return ERR_PTR(ret);
2466 }
2467
2468 static struct nvmf_transport_ops nvme_tcp_transport = {
2469         .name           = "tcp",
2470         .module         = THIS_MODULE,
2471         .required_opts  = NVMF_OPT_TRADDR,
2472         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2473                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2474                           NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
2475                           NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2476                           NVMF_OPT_TOS,
2477         .create_ctrl    = nvme_tcp_create_ctrl,
2478 };
2479
2480 static int __init nvme_tcp_init_module(void)
2481 {
2482         nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
2483                         WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2484         if (!nvme_tcp_wq)
2485                 return -ENOMEM;
2486
2487         nvmf_register_transport(&nvme_tcp_transport);
2488         return 0;
2489 }
2490
2491 static void __exit nvme_tcp_cleanup_module(void)
2492 {
2493         struct nvme_tcp_ctrl *ctrl;
2494
2495         nvmf_unregister_transport(&nvme_tcp_transport);
2496
2497         mutex_lock(&nvme_tcp_ctrl_mutex);
2498         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
2499                 nvme_delete_ctrl(&ctrl->ctrl);
2500         mutex_unlock(&nvme_tcp_ctrl_mutex);
2501         flush_workqueue(nvme_delete_wq);
2502
2503         destroy_workqueue(nvme_tcp_wq);
2504 }
2505
2506 module_init(nvme_tcp_init_module);
2507 module_exit(nvme_tcp_cleanup_module);
2508
2509 MODULE_LICENSE("GPL v2");