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 #ifdef CONFIG_DEBUG_LOCK_ALLOC
  34 /* lockdep can detect a circular dependency of the form
  35  *   sk_lock -> mmap_lock (page fault) -> fs locks -> sk_lock
  36  * because dependencies are tracked for both nvme-tcp and user contexts. Using
  37  * a separate class prevents lockdep from conflating nvme-tcp socket use with
  38  * user-space socket API use.
  39  */
  40 static struct lock_class_key nvme_tcp_sk_key[2];
  41 static struct lock_class_key nvme_tcp_slock_key[2];
  42
  43 static void nvme_tcp_reclassify_socket(struct socket *sock)
  44 {
  45         struct sock *sk = sock->sk;
  46
  47         if (WARN_ON_ONCE(!sock_allow_reclassification(sk)))
  48                 return;
  49
  50         switch (sk->sk_family) {
  51         case AF_INET:
  52                 sock_lock_init_class_and_name(sk, "slock-AF_INET-NVME",
  53                                               &nvme_tcp_slock_key[0],
  54                                               "sk_lock-AF_INET-NVME",
  55                                               &nvme_tcp_sk_key[0]);
  56                 break;
  57         case AF_INET6:
  58                 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NVME",
  59                                               &nvme_tcp_slock_key[1],
  60                                               "sk_lock-AF_INET6-NVME",
  61                                               &nvme_tcp_sk_key[1]);
  62                 break;
  63         default:
  64                 WARN_ON_ONCE(1);
  65         }
  66 }
  67 #else
  68 static void nvme_tcp_reclassify_socket(struct socket *sock) { }
  69 #endif
  70
  71 enum nvme_tcp_send_state {
  72         NVME_TCP_SEND_CMD_PDU = 0,
  73         NVME_TCP_SEND_H2C_PDU,
  74         NVME_TCP_SEND_DATA,
  75         NVME_TCP_SEND_DDGST,
  76 };
  77
  78 struct nvme_tcp_request {
  79         struct nvme_request     req;
  80         void                    *pdu;
  81         struct nvme_tcp_queue   *queue;
  82         u32                     data_len;
  83         u32                     pdu_len;
  84         u32                     pdu_sent;
  85         u16                     ttag;
  86         __le16                  status;
  87         struct list_head        entry;
  88         struct llist_node       lentry;
  89         __le32                  ddgst;
  90
  91         struct bio              *curr_bio;
  92         struct iov_iter         iter;
  93
  94         /* send state */
  95         size_t                  offset;
  96         size_t                  data_sent;
  97         enum nvme_tcp_send_state state;
  98 };
  99
 100 enum nvme_tcp_queue_flags {
 101         NVME_TCP_Q_ALLOCATED    = 0,
 102         NVME_TCP_Q_LIVE         = 1,
 103         NVME_TCP_Q_POLLING      = 2,
 104 };
 105
 106 enum nvme_tcp_recv_state {
 107         NVME_TCP_RECV_PDU = 0,
 108         NVME_TCP_RECV_DATA,
 109         NVME_TCP_RECV_DDGST,
 110 };
 111
 112 struct nvme_tcp_ctrl;
 113 struct nvme_tcp_queue {
 114         struct socket           *sock;
 115         struct work_struct      io_work;
 116         int                     io_cpu;
 117
 118         struct mutex            queue_lock;
 119         struct mutex            send_mutex;
 120         struct llist_head       req_list;
 121         struct list_head        send_list;
 122         bool                    more_requests;
 123
 124         /* recv state */
 125         void                    *pdu;
 126         int                     pdu_remaining;
 127         int                     pdu_offset;
 128         size_t                  data_remaining;
 129         size_t                  ddgst_remaining;
 130         unsigned int            nr_cqe;
 131
 132         /* send state */
 133         struct nvme_tcp_request *request;
 134
 135         int                     queue_size;
 136         size_t                  cmnd_capsule_len;
 137         struct nvme_tcp_ctrl    *ctrl;
 138         unsigned long           flags;
 139         bool                    rd_enabled;
 140
 141         bool                    hdr_digest;
 142         bool                    data_digest;
 143         struct ahash_request    *rcv_hash;
 144         struct ahash_request    *snd_hash;
 145         __le32                  exp_ddgst;
 146         __le32                  recv_ddgst;
 147
 148         struct page_frag_cache  pf_cache;
 149
 150         void (*state_change)(struct sock *);
 151         void (*data_ready)(struct sock *);
 152         void (*write_space)(struct sock *);
 153 };
 154
 155 struct nvme_tcp_ctrl {
 156         /* read only in the hot path */
 157         struct nvme_tcp_queue   *queues;
 158         struct blk_mq_tag_set   tag_set;
 159
 160         /* other member variables */
 161         struct list_head        list;
 162         struct blk_mq_tag_set   admin_tag_set;
 163         struct sockaddr_storage addr;
 164         struct sockaddr_storage src_addr;
 165         struct nvme_ctrl        ctrl;
 166
 167         struct work_struct      err_work;
 168         struct delayed_work     connect_work;
 169         struct nvme_tcp_request async_req;
 170         u32                     io_queues[HCTX_MAX_TYPES];
 171 };
 172
 173 static LIST_HEAD(nvme_tcp_ctrl_list);
 174 static DEFINE_MUTEX(nvme_tcp_ctrl_mutex);
 175 static struct workqueue_struct *nvme_tcp_wq;
 176 static const struct blk_mq_ops nvme_tcp_mq_ops;
 177 static const struct blk_mq_ops nvme_tcp_admin_mq_ops;
 178 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue);
 179
 180 static inline struct nvme_tcp_ctrl *to_tcp_ctrl(struct nvme_ctrl *ctrl)
 181 {
 182         return container_of(ctrl, struct nvme_tcp_ctrl, ctrl);
 183 }
 184
 185 static inline int nvme_tcp_queue_id(struct nvme_tcp_queue *queue)
 186 {
 187         return queue - queue->ctrl->queues;
 188 }
 189
 190 static inline struct blk_mq_tags *nvme_tcp_tagset(struct nvme_tcp_queue *queue)
 191 {
 192         u32 queue_idx = nvme_tcp_queue_id(queue);
 193
 194         if (queue_idx == 0)
 195                 return queue->ctrl->admin_tag_set.tags[queue_idx];
 196         return queue->ctrl->tag_set.tags[queue_idx - 1];
 197 }
 198
 199 static inline u8 nvme_tcp_hdgst_len(struct nvme_tcp_queue *queue)
 200 {
 201         return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 202 }
 203
 204 static inline u8 nvme_tcp_ddgst_len(struct nvme_tcp_queue *queue)
 205 {
 206         return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 207 }
 208
 209 static inline size_t nvme_tcp_inline_data_size(struct nvme_tcp_queue *queue)
 210 {
 211         return queue->cmnd_capsule_len - sizeof(struct nvme_command);
 212 }
 213
 214 static inline bool nvme_tcp_async_req(struct nvme_tcp_request *req)
 215 {
 216         return req == &req->queue->ctrl->async_req;
 217 }
 218
 219 static inline bool nvme_tcp_has_inline_data(struct nvme_tcp_request *req)
 220 {
 221         struct request *rq;
 222
 223         if (unlikely(nvme_tcp_async_req(req)))
 224                 return false; /* async events don't have a request */
 225
 226         rq = blk_mq_rq_from_pdu(req);
 227
 228         return rq_data_dir(rq) == WRITE && req->data_len &&
 229                 req->data_len <= nvme_tcp_inline_data_size(req->queue);
 230 }
 231
 232 static inline struct page *nvme_tcp_req_cur_page(struct nvme_tcp_request *req)
 233 {
 234         return req->iter.bvec->bv_page;
 235 }
 236
 237 static inline size_t nvme_tcp_req_cur_offset(struct nvme_tcp_request *req)
 238 {
 239         return req->iter.bvec->bv_offset + req->iter.iov_offset;
 240 }
 241
 242 static inline size_t nvme_tcp_req_cur_length(struct nvme_tcp_request *req)
 243 {
 244         return min_t(size_t, iov_iter_single_seg_count(&req->iter),
 245                         req->pdu_len - req->pdu_sent);
 246 }
 247
 248 static inline size_t nvme_tcp_pdu_data_left(struct nvme_tcp_request *req)
 249 {
 250         return rq_data_dir(blk_mq_rq_from_pdu(req)) == WRITE ?
 251                         req->pdu_len - req->pdu_sent : 0;
 252 }
 253
 254 static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req,
 255                 int len)
 256 {
 257         return nvme_tcp_pdu_data_left(req) <= len;
 258 }
 259
 260 static void nvme_tcp_init_iter(struct nvme_tcp_request *req,
 261                 unsigned int dir)
 262 {
 263         struct request *rq = blk_mq_rq_from_pdu(req);
 264         struct bio_vec *vec;
 265         unsigned int size;
 266         int nr_bvec;
 267         size_t offset;
 268
 269         if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) {
 270                 vec = &rq->special_vec;
 271                 nr_bvec = 1;
 272                 size = blk_rq_payload_bytes(rq);
 273                 offset = 0;
 274         } else {
 275                 struct bio *bio = req->curr_bio;
 276                 struct bvec_iter bi;
 277                 struct bio_vec bv;
 278
 279                 vec = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
 280                 nr_bvec = 0;
 281                 bio_for_each_bvec(bv, bio, bi) {
 282                         nr_bvec++;
 283                 }
 284                 size = bio->bi_iter.bi_size;
 285                 offset = bio->bi_iter.bi_bvec_done;
 286         }
 287
 288         iov_iter_bvec(&req->iter, dir, vec, nr_bvec, size);
 289         req->iter.iov_offset = offset;
 290 }
 291
 292 static inline void nvme_tcp_advance_req(struct nvme_tcp_request *req,
 293                 int len)
 294 {
 295         req->data_sent += len;
 296         req->pdu_sent += len;
 297         iov_iter_advance(&req->iter, len);
 298         if (!iov_iter_count(&req->iter) &&
 299             req->data_sent < req->data_len) {
 300                 req->curr_bio = req->curr_bio->bi_next;
 301                 nvme_tcp_init_iter(req, WRITE);
 302         }
 303 }
 304
 305 static inline void nvme_tcp_send_all(struct nvme_tcp_queue *queue)
 306 {
 307         int ret;
 308
 309         /* drain the send queue as much as we can... */
 310         do {
 311                 ret = nvme_tcp_try_send(queue);
 312         } while (ret > 0);
 313 }
 314
 315 static inline bool nvme_tcp_queue_more(struct nvme_tcp_queue *queue)
 316 {
 317         return !list_empty(&queue->send_list) ||
 318                 !llist_empty(&queue->req_list) || queue->more_requests;
 319 }
 320
 321 static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req,
 322                 bool sync, bool last)
 323 {
 324         struct nvme_tcp_queue *queue = req->queue;
 325         bool empty;
 326
 327         empty = llist_add(&req->lentry, &queue->req_list) &&
 328                 list_empty(&queue->send_list) && !queue->request;
 329
 330         /*
 331          * if we're the first on the send_list and we can try to send
 332          * directly, otherwise queue io_work. Also, only do that if we
 333          * are on the same cpu, so we don't introduce contention.
 334          */
 335         if (queue->io_cpu == raw_smp_processor_id() &&
 336             sync && empty && mutex_trylock(&queue->send_mutex)) {
 337                 queue->more_requests = !last;
 338                 nvme_tcp_send_all(queue);
 339                 queue->more_requests = false;
 340                 mutex_unlock(&queue->send_mutex);
 341         }
 342
 343         if (last && nvme_tcp_queue_more(queue))
 344                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 345 }
 346
 347 static void nvme_tcp_process_req_list(struct nvme_tcp_queue *queue)
 348 {
 349         struct nvme_tcp_request *req;
 350         struct llist_node *node;
 351
 352         for (node = llist_del_all(&queue->req_list); node; node = node->next) {
 353                 req = llist_entry(node, struct nvme_tcp_request, lentry);
 354                 list_add(&req->entry, &queue->send_list);
 355         }
 356 }
 357
 358 static inline struct nvme_tcp_request *
 359 nvme_tcp_fetch_request(struct nvme_tcp_queue *queue)
 360 {
 361         struct nvme_tcp_request *req;
 362
 363         req = list_first_entry_or_null(&queue->send_list,
 364                         struct nvme_tcp_request, entry);
 365         if (!req) {
 366                 nvme_tcp_process_req_list(queue);
 367                 req = list_first_entry_or_null(&queue->send_list,
 368                                 struct nvme_tcp_request, entry);
 369                 if (unlikely(!req))
 370                         return NULL;
 371         }
 372
 373         list_del(&req->entry);
 374         return req;
 375 }
 376
 377 static inline void nvme_tcp_ddgst_final(struct ahash_request *hash,
 378                 __le32 *dgst)
 379 {
 380         ahash_request_set_crypt(hash, NULL, (u8 *)dgst, 0);
 381         crypto_ahash_final(hash);
 382 }
 383
 384 static inline void nvme_tcp_ddgst_update(struct ahash_request *hash,
 385                 struct page *page, off_t off, size_t len)
 386 {
 387         struct scatterlist sg;
 388
 389         sg_init_marker(&sg, 1);
 390         sg_set_page(&sg, page, len, off);
 391         ahash_request_set_crypt(hash, &sg, NULL, len);
 392         crypto_ahash_update(hash);
 393 }
 394
 395 static inline void nvme_tcp_hdgst(struct ahash_request *hash,
 396                 void *pdu, size_t len)
 397 {
 398         struct scatterlist sg;
 399
 400         sg_init_one(&sg, pdu, len);
 401         ahash_request_set_crypt(hash, &sg, pdu + len, len);
 402         crypto_ahash_digest(hash);
 403 }
 404
 405 static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue,
 406                 void *pdu, size_t pdu_len)
 407 {
 408         struct nvme_tcp_hdr *hdr = pdu;
 409         __le32 recv_digest;
 410         __le32 exp_digest;
 411
 412         if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
 413                 dev_err(queue->ctrl->ctrl.device,
 414                         "queue %d: header digest flag is cleared\n",
 415                         nvme_tcp_queue_id(queue));
 416                 return -EPROTO;
 417         }
 418
 419         recv_digest = *(__le32 *)(pdu + hdr->hlen);
 420         nvme_tcp_hdgst(queue->rcv_hash, pdu, pdu_len);
 421         exp_digest = *(__le32 *)(pdu + hdr->hlen);
 422         if (recv_digest != exp_digest) {
 423                 dev_err(queue->ctrl->ctrl.device,
 424                         "header digest error: recv %#x expected %#x\n",
 425                         le32_to_cpu(recv_digest), le32_to_cpu(exp_digest));
 426                 return -EIO;
 427         }
 428
 429         return 0;
 430 }
 431
 432 static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu)
 433 {
 434         struct nvme_tcp_hdr *hdr = pdu;
 435         u8 digest_len = nvme_tcp_hdgst_len(queue);
 436         u32 len;
 437
 438         len = le32_to_cpu(hdr->plen) - hdr->hlen -
 439                 ((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0);
 440
 441         if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
 442                 dev_err(queue->ctrl->ctrl.device,
 443                         "queue %d: data digest flag is cleared\n",
 444                 nvme_tcp_queue_id(queue));
 445                 return -EPROTO;
 446         }
 447         crypto_ahash_init(queue->rcv_hash);
 448
 449         return 0;
 450 }
 451
 452 static void nvme_tcp_exit_request(struct blk_mq_tag_set *set,
 453                 struct request *rq, unsigned int hctx_idx)
 454 {
 455         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 456
 457         page_frag_free(req->pdu);
 458 }
 459
 460 static int nvme_tcp_init_request(struct blk_mq_tag_set *set,
 461                 struct request *rq, unsigned int hctx_idx,
 462                 unsigned int numa_node)
 463 {
 464         struct nvme_tcp_ctrl *ctrl = set->driver_data;
 465         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 466         struct nvme_tcp_cmd_pdu *pdu;
 467         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 468         struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx];
 469         u8 hdgst = nvme_tcp_hdgst_len(queue);
 470
 471         req->pdu = page_frag_alloc(&queue->pf_cache,
 472                 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
 473                 GFP_KERNEL | __GFP_ZERO);
 474         if (!req->pdu)
 475                 return -ENOMEM;
 476
 477         pdu = req->pdu;
 478         req->queue = queue;
 479         nvme_req(rq)->ctrl = &ctrl->ctrl;
 480         nvme_req(rq)->cmd = &pdu->cmd;
 481
 482         return 0;
 483 }
 484
 485 static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 486                 unsigned int hctx_idx)
 487 {
 488         struct nvme_tcp_ctrl *ctrl = data;
 489         struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1];
 490
 491         hctx->driver_data = queue;
 492         return 0;
 493 }
 494
 495 static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 496                 unsigned int hctx_idx)
 497 {
 498         struct nvme_tcp_ctrl *ctrl = data;
 499         struct nvme_tcp_queue *queue = &ctrl->queues[0];
 500
 501         hctx->driver_data = queue;
 502         return 0;
 503 }
 504
 505 static enum nvme_tcp_recv_state
 506 nvme_tcp_recv_state(struct nvme_tcp_queue *queue)
 507 {
 508         return  (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
 509                 (queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST :
 510                 NVME_TCP_RECV_DATA;
 511 }
 512
 513 static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue)
 514 {
 515         queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) +
 516                                 nvme_tcp_hdgst_len(queue);
 517         queue->pdu_offset = 0;
 518         queue->data_remaining = -1;
 519         queue->ddgst_remaining = 0;
 520 }
 521
 522 static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl)
 523 {
 524         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
 525                 return;
 526
 527         dev_warn(ctrl->device, "starting error recovery\n");
 528         queue_work(nvme_reset_wq, &to_tcp_ctrl(ctrl)->err_work);
 529 }
 530
 531 static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
 532                 struct nvme_completion *cqe)
 533 {
 534         struct nvme_tcp_request *req;
 535         struct request *rq;
 536
 537         rq = nvme_find_rq(nvme_tcp_tagset(queue), cqe->command_id);
 538         if (!rq) {
 539                 dev_err(queue->ctrl->ctrl.device,
 540                         "got bad cqe.command_id %#x on queue %d\n",
 541                         cqe->command_id, nvme_tcp_queue_id(queue));
 542                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 543                 return -EINVAL;
 544         }
 545
 546         req = blk_mq_rq_to_pdu(rq);
 547         if (req->status == cpu_to_le16(NVME_SC_SUCCESS))
 548                 req->status = cqe->status;
 549
 550         if (!nvme_try_complete_req(rq, req->status, cqe->result))
 551                 nvme_complete_rq(rq);
 552         queue->nr_cqe++;
 553
 554         return 0;
 555 }
 556
 557 static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue,
 558                 struct nvme_tcp_data_pdu *pdu)
 559 {
 560         struct request *rq;
 561
 562         rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
 563         if (!rq) {
 564                 dev_err(queue->ctrl->ctrl.device,
 565                         "got bad c2hdata.command_id %#x on queue %d\n",
 566                         pdu->command_id, nvme_tcp_queue_id(queue));
 567                 return -ENOENT;
 568         }
 569
 570         if (!blk_rq_payload_bytes(rq)) {
 571                 dev_err(queue->ctrl->ctrl.device,
 572                         "queue %d tag %#x unexpected data\n",
 573                         nvme_tcp_queue_id(queue), rq->tag);
 574                 return -EIO;
 575         }
 576
 577         queue->data_remaining = le32_to_cpu(pdu->data_length);
 578
 579         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS &&
 580             unlikely(!(pdu->hdr.flags & NVME_TCP_F_DATA_LAST))) {
 581                 dev_err(queue->ctrl->ctrl.device,
 582                         "queue %d tag %#x SUCCESS set but not last PDU\n",
 583                         nvme_tcp_queue_id(queue), rq->tag);
 584                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 585                 return -EPROTO;
 586         }
 587
 588         return 0;
 589 }
 590
 591 static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue,
 592                 struct nvme_tcp_rsp_pdu *pdu)
 593 {
 594         struct nvme_completion *cqe = &pdu->cqe;
 595         int ret = 0;
 596
 597         /*
 598          * AEN requests are special as they don't time out and can
 599          * survive any kind of queue freeze and often don't respond to
 600          * aborts.  We don't even bother to allocate a struct request
 601          * for them but rather special case them here.
 602          */
 603         if (unlikely(nvme_is_aen_req(nvme_tcp_queue_id(queue),
 604                                      cqe->command_id)))
 605                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
 606                                 &cqe->result);
 607         else
 608                 ret = nvme_tcp_process_nvme_cqe(queue, cqe);
 609
 610         return ret;
 611 }
 612
 613 static int nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req,
 614                 struct nvme_tcp_r2t_pdu *pdu)
 615 {
 616         struct nvme_tcp_data_pdu *data = req->pdu;
 617         struct nvme_tcp_queue *queue = req->queue;
 618         struct request *rq = blk_mq_rq_from_pdu(req);
 619         u8 hdgst = nvme_tcp_hdgst_len(queue);
 620         u8 ddgst = nvme_tcp_ddgst_len(queue);
 621
 622         req->pdu_len = le32_to_cpu(pdu->r2t_length);
 623         req->pdu_sent = 0;
 624
 625         if (unlikely(!req->pdu_len)) {
 626                 dev_err(queue->ctrl->ctrl.device,
 627                         "req %d r2t len is %u, probably a bug...\n",
 628                         rq->tag, req->pdu_len);
 629                 return -EPROTO;
 630         }
 631
 632         if (unlikely(req->data_sent + req->pdu_len > req->data_len)) {
 633                 dev_err(queue->ctrl->ctrl.device,
 634                         "req %d r2t len %u exceeded data len %u (%zu sent)\n",
 635                         rq->tag, req->pdu_len, req->data_len,
 636                         req->data_sent);
 637                 return -EPROTO;
 638         }
 639
 640         if (unlikely(le32_to_cpu(pdu->r2t_offset) < req->data_sent)) {
 641                 dev_err(queue->ctrl->ctrl.device,
 642                         "req %d unexpected r2t offset %u (expected %zu)\n",
 643                         rq->tag, le32_to_cpu(pdu->r2t_offset),
 644                         req->data_sent);
 645                 return -EPROTO;
 646         }
 647
 648         memset(data, 0, sizeof(*data));
 649         data->hdr.type = nvme_tcp_h2c_data;
 650         data->hdr.flags = NVME_TCP_F_DATA_LAST;
 651         if (queue->hdr_digest)
 652                 data->hdr.flags |= NVME_TCP_F_HDGST;
 653         if (queue->data_digest)
 654                 data->hdr.flags |= NVME_TCP_F_DDGST;
 655         data->hdr.hlen = sizeof(*data);
 656         data->hdr.pdo = data->hdr.hlen + hdgst;
 657         data->hdr.plen =
 658                 cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst);
 659         data->ttag = pdu->ttag;
 660         data->command_id = nvme_cid(rq);
 661         data->data_offset = pdu->r2t_offset;
 662         data->data_length = cpu_to_le32(req->pdu_len);
 663         return 0;
 664 }
 665
 666 static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue,
 667                 struct nvme_tcp_r2t_pdu *pdu)
 668 {
 669         struct nvme_tcp_request *req;
 670         struct request *rq;
 671         int ret;
 672
 673         rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
 674         if (!rq) {
 675                 dev_err(queue->ctrl->ctrl.device,
 676                         "got bad r2t.command_id %#x on queue %d\n",
 677                         pdu->command_id, nvme_tcp_queue_id(queue));
 678                 return -ENOENT;
 679         }
 680         req = blk_mq_rq_to_pdu(rq);
 681
 682         ret = nvme_tcp_setup_h2c_data_pdu(req, pdu);
 683         if (unlikely(ret))
 684                 return ret;
 685
 686         req->state = NVME_TCP_SEND_H2C_PDU;
 687         req->offset = 0;
 688
 689         nvme_tcp_queue_request(req, false, true);
 690
 691         return 0;
 692 }
 693
 694 static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 695                 unsigned int *offset, size_t *len)
 696 {
 697         struct nvme_tcp_hdr *hdr;
 698         char *pdu = queue->pdu;
 699         size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
 700         int ret;
 701
 702         ret = skb_copy_bits(skb, *offset,
 703                 &pdu[queue->pdu_offset], rcv_len);
 704         if (unlikely(ret))
 705                 return ret;
 706
 707         queue->pdu_remaining -= rcv_len;
 708         queue->pdu_offset += rcv_len;
 709         *offset += rcv_len;
 710         *len -= rcv_len;
 711         if (queue->pdu_remaining)
 712                 return 0;
 713
 714         hdr = queue->pdu;
 715         if (queue->hdr_digest) {
 716                 ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen);
 717                 if (unlikely(ret))
 718                         return ret;
 719         }
 720
 721
 722         if (queue->data_digest) {
 723                 ret = nvme_tcp_check_ddgst(queue, queue->pdu);
 724                 if (unlikely(ret))
 725                         return ret;
 726         }
 727
 728         switch (hdr->type) {
 729         case nvme_tcp_c2h_data:
 730                 return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
 731         case nvme_tcp_rsp:
 732                 nvme_tcp_init_recv_ctx(queue);
 733                 return nvme_tcp_handle_comp(queue, (void *)queue->pdu);
 734         case nvme_tcp_r2t:
 735                 nvme_tcp_init_recv_ctx(queue);
 736                 return nvme_tcp_handle_r2t(queue, (void *)queue->pdu);
 737         default:
 738                 dev_err(queue->ctrl->ctrl.device,
 739                         "unsupported pdu type (%d)\n", hdr->type);
 740                 return -EINVAL;
 741         }
 742 }
 743
 744 static inline void nvme_tcp_end_request(struct request *rq, u16 status)
 745 {
 746         union nvme_result res = {};
 747
 748         if (!nvme_try_complete_req(rq, cpu_to_le16(status << 1), res))
 749                 nvme_complete_rq(rq);
 750 }
 751
 752 static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 753                               unsigned int *offset, size_t *len)
 754 {
 755         struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
 756         struct request *rq =
 757                 nvme_cid_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
 758         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 759
 760         while (true) {
 761                 int recv_len, ret;
 762
 763                 recv_len = min_t(size_t, *len, queue->data_remaining);
 764                 if (!recv_len)
 765                         break;
 766
 767                 if (!iov_iter_count(&req->iter)) {
 768                         req->curr_bio = req->curr_bio->bi_next;
 769
 770                         /*
 771                          * If we don`t have any bios it means that controller
 772                          * sent more data than we requested, hence error
 773                          */
 774                         if (!req->curr_bio) {
 775                                 dev_err(queue->ctrl->ctrl.device,
 776                                         "queue %d no space in request %#x",
 777                                         nvme_tcp_queue_id(queue), rq->tag);
 778                                 nvme_tcp_init_recv_ctx(queue);
 779                                 return -EIO;
 780                         }
 781                         nvme_tcp_init_iter(req, READ);
 782                 }
 783
 784                 /* we can read only from what is left in this bio */
 785                 recv_len = min_t(size_t, recv_len,
 786                                 iov_iter_count(&req->iter));
 787
 788                 if (queue->data_digest)
 789                         ret = skb_copy_and_hash_datagram_iter(skb, *offset,
 790                                 &req->iter, recv_len, queue->rcv_hash);
 791                 else
 792                         ret = skb_copy_datagram_iter(skb, *offset,
 793                                         &req->iter, recv_len);
 794                 if (ret) {
 795                         dev_err(queue->ctrl->ctrl.device,
 796                                 "queue %d failed to copy request %#x data",
 797                                 nvme_tcp_queue_id(queue), rq->tag);
 798                         return ret;
 799                 }
 800
 801                 *len -= recv_len;
 802                 *offset += recv_len;
 803                 queue->data_remaining -= recv_len;
 804         }
 805
 806         if (!queue->data_remaining) {
 807                 if (queue->data_digest) {
 808                         nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst);
 809                         queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
 810                 } else {
 811                         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
 812                                 nvme_tcp_end_request(rq,
 813                                                 le16_to_cpu(req->status));
 814                                 queue->nr_cqe++;
 815                         }
 816                         nvme_tcp_init_recv_ctx(queue);
 817                 }
 818         }
 819
 820         return 0;
 821 }
 822
 823 static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
 824                 struct sk_buff *skb, unsigned int *offset, size_t *len)
 825 {
 826         struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
 827         char *ddgst = (char *)&queue->recv_ddgst;
 828         size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
 829         off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
 830         int ret;
 831
 832         ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
 833         if (unlikely(ret))
 834                 return ret;
 835
 836         queue->ddgst_remaining -= recv_len;
 837         *offset += recv_len;
 838         *len -= recv_len;
 839         if (queue->ddgst_remaining)
 840                 return 0;
 841
 842         if (queue->recv_ddgst != queue->exp_ddgst) {
 843                 struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
 844                                         pdu->command_id);
 845                 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 846
 847                 req->status = cpu_to_le16(NVME_SC_DATA_XFER_ERROR);
 848
 849                 dev_err(queue->ctrl->ctrl.device,
 850                         "data digest error: recv %#x expected %#x\n",
 851                         le32_to_cpu(queue->recv_ddgst),
 852                         le32_to_cpu(queue->exp_ddgst));
 853         }
 854
 855         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
 856                 struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
 857                                         pdu->command_id);
 858                 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 859
 860                 nvme_tcp_end_request(rq, le16_to_cpu(req->status));
 861                 queue->nr_cqe++;
 862         }
 863
 864         nvme_tcp_init_recv_ctx(queue);
 865         return 0;
 866 }
 867
 868 static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
 869                              unsigned int offset, size_t len)
 870 {
 871         struct nvme_tcp_queue *queue = desc->arg.data;
 872         size_t consumed = len;
 873         int result;
 874
 875         while (len) {
 876                 switch (nvme_tcp_recv_state(queue)) {
 877                 case NVME_TCP_RECV_PDU:
 878                         result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
 879                         break;
 880                 case NVME_TCP_RECV_DATA:
 881                         result = nvme_tcp_recv_data(queue, skb, &offset, &len);
 882                         break;
 883                 case NVME_TCP_RECV_DDGST:
 884                         result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
 885                         break;
 886                 default:
 887                         result = -EFAULT;
 888                 }
 889                 if (result) {
 890                         dev_err(queue->ctrl->ctrl.device,
 891                                 "receive failed:  %d\n", result);
 892                         queue->rd_enabled = false;
 893                         nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 894                         return result;
 895                 }
 896         }
 897
 898         return consumed;
 899 }
 900
 901 static void nvme_tcp_data_ready(struct sock *sk)
 902 {
 903         struct nvme_tcp_queue *queue;
 904
 905         read_lock_bh(&sk->sk_callback_lock);
 906         queue = sk->sk_user_data;
 907         if (likely(queue && queue->rd_enabled) &&
 908             !test_bit(NVME_TCP_Q_POLLING, &queue->flags))
 909                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 910         read_unlock_bh(&sk->sk_callback_lock);
 911 }
 912
 913 static void nvme_tcp_write_space(struct sock *sk)
 914 {
 915         struct nvme_tcp_queue *queue;
 916
 917         read_lock_bh(&sk->sk_callback_lock);
 918         queue = sk->sk_user_data;
 919         if (likely(queue && sk_stream_is_writeable(sk))) {
 920                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 921                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 922         }
 923         read_unlock_bh(&sk->sk_callback_lock);
 924 }
 925
 926 static void nvme_tcp_state_change(struct sock *sk)
 927 {
 928         struct nvme_tcp_queue *queue;
 929
 930         read_lock_bh(&sk->sk_callback_lock);
 931         queue = sk->sk_user_data;
 932         if (!queue)
 933                 goto done;
 934
 935         switch (sk->sk_state) {
 936         case TCP_CLOSE:
 937         case TCP_CLOSE_WAIT:
 938         case TCP_LAST_ACK:
 939         case TCP_FIN_WAIT1:
 940         case TCP_FIN_WAIT2:
 941                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 942                 break;
 943         default:
 944                 dev_info(queue->ctrl->ctrl.device,
 945                         "queue %d socket state %d\n",
 946                         nvme_tcp_queue_id(queue), sk->sk_state);
 947         }
 948
 949         queue->state_change(sk);
 950 done:
 951         read_unlock_bh(&sk->sk_callback_lock);
 952 }
 953
 954 static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
 955 {
 956         queue->request = NULL;
 957 }
 958
 959 static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
 960 {
 961         if (nvme_tcp_async_req(req)) {
 962                 union nvme_result res = {};
 963
 964                 nvme_complete_async_event(&req->queue->ctrl->ctrl,
 965                                 cpu_to_le16(NVME_SC_HOST_PATH_ERROR), &res);
 966         } else {
 967                 nvme_tcp_end_request(blk_mq_rq_from_pdu(req),
 968                                 NVME_SC_HOST_PATH_ERROR);
 969         }
 970 }
 971
 972 static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
 973 {
 974         struct nvme_tcp_queue *queue = req->queue;
 975         int req_data_len = req->data_len;
 976
 977         while (true) {
 978                 struct page *page = nvme_tcp_req_cur_page(req);
 979                 size_t offset = nvme_tcp_req_cur_offset(req);
 980                 size_t len = nvme_tcp_req_cur_length(req);
 981                 bool last = nvme_tcp_pdu_last_send(req, len);
 982                 int req_data_sent = req->data_sent;
 983                 int ret, flags = MSG_DONTWAIT;
 984
 985                 if (last && !queue->data_digest && !nvme_tcp_queue_more(queue))
 986                         flags |= MSG_EOR;
 987                 else
 988                         flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
 989
 990                 if (sendpage_ok(page)) {
 991                         ret = kernel_sendpage(queue->sock, page, offset, len,
 992                                         flags);
 993                 } else {
 994                         ret = sock_no_sendpage(queue->sock, page, offset, len,
 995                                         flags);
 996                 }
 997                 if (ret <= 0)
 998                         return ret;
 999
1000                 if (queue->data_digest)
1001                         nvme_tcp_ddgst_update(queue->snd_hash, page,
1002                                         offset, ret);
1003
1004                 /*
1005                  * update the request iterator except for the last payload send
1006                  * in the request where we don't want to modify it as we may
1007                  * compete with the RX path completing the request.
1008                  */
1009                 if (req_data_sent + ret < req_data_len)
1010                         nvme_tcp_advance_req(req, ret);
1011
1012                 /* fully successful last send in current PDU */
1013                 if (last && ret == len) {
1014                         if (queue->data_digest) {
1015                                 nvme_tcp_ddgst_final(queue->snd_hash,
1016                                         &req->ddgst);
1017                                 req->state = NVME_TCP_SEND_DDGST;
1018                                 req->offset = 0;
1019                         } else {
1020                                 nvme_tcp_done_send_req(queue);
1021                         }
1022                         return 1;
1023                 }
1024         }
1025         return -EAGAIN;
1026 }
1027
1028 static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
1029 {
1030         struct nvme_tcp_queue *queue = req->queue;
1031         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
1032         bool inline_data = nvme_tcp_has_inline_data(req);
1033         u8 hdgst = nvme_tcp_hdgst_len(queue);
1034         int len = sizeof(*pdu) + hdgst - req->offset;
1035         int flags = MSG_DONTWAIT;
1036         int ret;
1037
1038         if (inline_data || nvme_tcp_queue_more(queue))
1039                 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
1040         else
1041                 flags |= MSG_EOR;
1042
1043         if (queue->hdr_digest && !req->offset)
1044                 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
1045
1046         ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
1047                         offset_in_page(pdu) + req->offset, len,  flags);
1048         if (unlikely(ret <= 0))
1049                 return ret;
1050
1051         len -= ret;
1052         if (!len) {
1053                 if (inline_data) {
1054                         req->state = NVME_TCP_SEND_DATA;
1055                         if (queue->data_digest)
1056                                 crypto_ahash_init(queue->snd_hash);
1057                 } else {
1058                         nvme_tcp_done_send_req(queue);
1059                 }
1060                 return 1;
1061         }
1062         req->offset += ret;
1063
1064         return -EAGAIN;
1065 }
1066
1067 static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
1068 {
1069         struct nvme_tcp_queue *queue = req->queue;
1070         struct nvme_tcp_data_pdu *pdu = req->pdu;
1071         u8 hdgst = nvme_tcp_hdgst_len(queue);
1072         int len = sizeof(*pdu) - req->offset + hdgst;
1073         int ret;
1074
1075         if (queue->hdr_digest && !req->offset)
1076                 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
1077
1078         ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
1079                         offset_in_page(pdu) + req->offset, len,
1080                         MSG_DONTWAIT | MSG_MORE | MSG_SENDPAGE_NOTLAST);
1081         if (unlikely(ret <= 0))
1082                 return ret;
1083
1084         len -= ret;
1085         if (!len) {
1086                 req->state = NVME_TCP_SEND_DATA;
1087                 if (queue->data_digest)
1088                         crypto_ahash_init(queue->snd_hash);
1089                 return 1;
1090         }
1091         req->offset += ret;
1092
1093         return -EAGAIN;
1094 }
1095
1096 static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
1097 {
1098         struct nvme_tcp_queue *queue = req->queue;
1099         size_t offset = req->offset;
1100         int ret;
1101         struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1102         struct kvec iov = {
1103                 .iov_base = (u8 *)&req->ddgst + req->offset,
1104                 .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
1105         };
1106
1107         if (nvme_tcp_queue_more(queue))
1108                 msg.msg_flags |= MSG_MORE;
1109         else
1110                 msg.msg_flags |= MSG_EOR;
1111
1112         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1113         if (unlikely(ret <= 0))
1114                 return ret;
1115
1116         if (offset + ret == NVME_TCP_DIGEST_LENGTH) {
1117                 nvme_tcp_done_send_req(queue);
1118                 return 1;
1119         }
1120
1121         req->offset += ret;
1122         return -EAGAIN;
1123 }
1124
1125 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
1126 {
1127         struct nvme_tcp_request *req;
1128         int ret = 1;
1129
1130         if (!queue->request) {
1131                 queue->request = nvme_tcp_fetch_request(queue);
1132                 if (!queue->request)
1133                         return 0;
1134         }
1135         req = queue->request;
1136
1137         if (req->state == NVME_TCP_SEND_CMD_PDU) {
1138                 ret = nvme_tcp_try_send_cmd_pdu(req);
1139                 if (ret <= 0)
1140                         goto done;
1141                 if (!nvme_tcp_has_inline_data(req))
1142                         return ret;
1143         }
1144
1145         if (req->state == NVME_TCP_SEND_H2C_PDU) {
1146                 ret = nvme_tcp_try_send_data_pdu(req);
1147                 if (ret <= 0)
1148                         goto done;
1149         }
1150
1151         if (req->state == NVME_TCP_SEND_DATA) {
1152                 ret = nvme_tcp_try_send_data(req);
1153                 if (ret <= 0)
1154                         goto done;
1155         }
1156
1157         if (req->state == NVME_TCP_SEND_DDGST)
1158                 ret = nvme_tcp_try_send_ddgst(req);
1159 done:
1160         if (ret == -EAGAIN) {
1161                 ret = 0;
1162         } else if (ret < 0) {
1163                 dev_err(queue->ctrl->ctrl.device,
1164                         "failed to send request %d\n", ret);
1165                 if (ret != -EPIPE && ret != -ECONNRESET)
1166                         nvme_tcp_fail_request(queue->request);
1167                 nvme_tcp_done_send_req(queue);
1168         }
1169         return ret;
1170 }
1171
1172 static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1173 {
1174         struct socket *sock = queue->sock;
1175         struct sock *sk = sock->sk;
1176         read_descriptor_t rd_desc;
1177         int consumed;
1178
1179         rd_desc.arg.data = queue;
1180         rd_desc.count = 1;
1181         lock_sock(sk);
1182         queue->nr_cqe = 0;
1183         consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1184         release_sock(sk);
1185         return consumed;
1186 }
1187
1188 static void nvme_tcp_io_work(struct work_struct *w)
1189 {
1190         struct nvme_tcp_queue *queue =
1191                 container_of(w, struct nvme_tcp_queue, io_work);
1192         unsigned long deadline = jiffies + msecs_to_jiffies(1);
1193
1194         do {
1195                 bool pending = false;
1196                 int result;
1197
1198                 if (mutex_trylock(&queue->send_mutex)) {
1199                         result = nvme_tcp_try_send(queue);
1200                         mutex_unlock(&queue->send_mutex);
1201                         if (result > 0)
1202                                 pending = true;
1203                         else if (unlikely(result < 0))
1204                                 break;
1205                 }
1206
1207                 result = nvme_tcp_try_recv(queue);
1208                 if (result > 0)
1209                         pending = true;
1210                 else if (unlikely(result < 0))
1211                         return;
1212
1213                 if (!pending)
1214                         return;
1215
1216         } while (!time_after(jiffies, deadline)); /* quota is exhausted */
1217
1218         queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1219 }
1220
1221 static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue)
1222 {
1223         struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
1224
1225         ahash_request_free(queue->rcv_hash);
1226         ahash_request_free(queue->snd_hash);
1227         crypto_free_ahash(tfm);
1228 }
1229
1230 static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue)
1231 {
1232         struct crypto_ahash *tfm;
1233
1234         tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
1235         if (IS_ERR(tfm))
1236                 return PTR_ERR(tfm);
1237
1238         queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1239         if (!queue->snd_hash)
1240                 goto free_tfm;
1241         ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
1242
1243         queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1244         if (!queue->rcv_hash)
1245                 goto free_snd_hash;
1246         ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
1247
1248         return 0;
1249 free_snd_hash:
1250         ahash_request_free(queue->snd_hash);
1251 free_tfm:
1252         crypto_free_ahash(tfm);
1253         return -ENOMEM;
1254 }
1255
1256 static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1257 {
1258         struct nvme_tcp_request *async = &ctrl->async_req;
1259
1260         page_frag_free(async->pdu);
1261 }
1262
1263 static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1264 {
1265         struct nvme_tcp_queue *queue = &ctrl->queues[0];
1266         struct nvme_tcp_request *async = &ctrl->async_req;
1267         u8 hdgst = nvme_tcp_hdgst_len(queue);
1268
1269         async->pdu = page_frag_alloc(&queue->pf_cache,
1270                 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1271                 GFP_KERNEL | __GFP_ZERO);
1272         if (!async->pdu)
1273                 return -ENOMEM;
1274
1275         async->queue = &ctrl->queues[0];
1276         return 0;
1277 }
1278
1279 static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1280 {
1281         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1282         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1283
1284         if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1285                 return;
1286
1287         if (queue->hdr_digest || queue->data_digest)
1288                 nvme_tcp_free_crypto(queue);
1289
1290         sock_release(queue->sock);
1291         kfree(queue->pdu);
1292         mutex_destroy(&queue->send_mutex);
1293         mutex_destroy(&queue->queue_lock);
1294 }
1295
1296 static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1297 {
1298         struct nvme_tcp_icreq_pdu *icreq;
1299         struct nvme_tcp_icresp_pdu *icresp;
1300         struct msghdr msg = {};
1301         struct kvec iov;
1302         bool ctrl_hdgst, ctrl_ddgst;
1303         int ret;
1304
1305         icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1306         if (!icreq)
1307                 return -ENOMEM;
1308
1309         icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1310         if (!icresp) {
1311                 ret = -ENOMEM;
1312                 goto free_icreq;
1313         }
1314
1315         icreq->hdr.type = nvme_tcp_icreq;
1316         icreq->hdr.hlen = sizeof(*icreq);
1317         icreq->hdr.pdo = 0;
1318         icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1319         icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1320         icreq->maxr2t = 0; /* single inflight r2t supported */
1321         icreq->hpda = 0; /* no alignment constraint */
1322         if (queue->hdr_digest)
1323                 icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1324         if (queue->data_digest)
1325                 icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1326
1327         iov.iov_base = icreq;
1328         iov.iov_len = sizeof(*icreq);
1329         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1330         if (ret < 0)
1331                 goto free_icresp;
1332
1333         memset(&msg, 0, sizeof(msg));
1334         iov.iov_base = icresp;
1335         iov.iov_len = sizeof(*icresp);
1336         ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1337                         iov.iov_len, msg.msg_flags);
1338         if (ret < 0)
1339                 goto free_icresp;
1340
1341         ret = -EINVAL;
1342         if (icresp->hdr.type != nvme_tcp_icresp) {
1343                 pr_err("queue %d: bad type returned %d\n",
1344                         nvme_tcp_queue_id(queue), icresp->hdr.type);
1345                 goto free_icresp;
1346         }
1347
1348         if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1349                 pr_err("queue %d: bad pdu length returned %d\n",
1350                         nvme_tcp_queue_id(queue), icresp->hdr.plen);
1351                 goto free_icresp;
1352         }
1353
1354         if (icresp->pfv != NVME_TCP_PFV_1_0) {
1355                 pr_err("queue %d: bad pfv returned %d\n",
1356                         nvme_tcp_queue_id(queue), icresp->pfv);
1357                 goto free_icresp;
1358         }
1359
1360         ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1361         if ((queue->data_digest && !ctrl_ddgst) ||
1362             (!queue->data_digest && ctrl_ddgst)) {
1363                 pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1364                         nvme_tcp_queue_id(queue),
1365                         queue->data_digest ? "enabled" : "disabled",
1366                         ctrl_ddgst ? "enabled" : "disabled");
1367                 goto free_icresp;
1368         }
1369
1370         ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1371         if ((queue->hdr_digest && !ctrl_hdgst) ||
1372             (!queue->hdr_digest && ctrl_hdgst)) {
1373                 pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1374                         nvme_tcp_queue_id(queue),
1375                         queue->hdr_digest ? "enabled" : "disabled",
1376                         ctrl_hdgst ? "enabled" : "disabled");
1377                 goto free_icresp;
1378         }
1379
1380         if (icresp->cpda != 0) {
1381                 pr_err("queue %d: unsupported cpda returned %d\n",
1382                         nvme_tcp_queue_id(queue), icresp->cpda);
1383                 goto free_icresp;
1384         }
1385
1386         ret = 0;
1387 free_icresp:
1388         kfree(icresp);
1389 free_icreq:
1390         kfree(icreq);
1391         return ret;
1392 }
1393
1394 static bool nvme_tcp_admin_queue(struct nvme_tcp_queue *queue)
1395 {
1396         return nvme_tcp_queue_id(queue) == 0;
1397 }
1398
1399 static bool nvme_tcp_default_queue(struct nvme_tcp_queue *queue)
1400 {
1401         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1402         int qid = nvme_tcp_queue_id(queue);
1403
1404         return !nvme_tcp_admin_queue(queue) &&
1405                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT];
1406 }
1407
1408 static bool nvme_tcp_read_queue(struct nvme_tcp_queue *queue)
1409 {
1410         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1411         int qid = nvme_tcp_queue_id(queue);
1412
1413         return !nvme_tcp_admin_queue(queue) &&
1414                 !nvme_tcp_default_queue(queue) &&
1415                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1416                           ctrl->io_queues[HCTX_TYPE_READ];
1417 }
1418
1419 static bool nvme_tcp_poll_queue(struct nvme_tcp_queue *queue)
1420 {
1421         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1422         int qid = nvme_tcp_queue_id(queue);
1423
1424         return !nvme_tcp_admin_queue(queue) &&
1425                 !nvme_tcp_default_queue(queue) &&
1426                 !nvme_tcp_read_queue(queue) &&
1427                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1428                           ctrl->io_queues[HCTX_TYPE_READ] +
1429                           ctrl->io_queues[HCTX_TYPE_POLL];
1430 }
1431
1432 static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
1433 {
1434         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1435         int qid = nvme_tcp_queue_id(queue);
1436         int n = 0;
1437
1438         if (nvme_tcp_default_queue(queue))
1439                 n = qid - 1;
1440         else if (nvme_tcp_read_queue(queue))
1441                 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] - 1;
1442         else if (nvme_tcp_poll_queue(queue))
1443                 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] -
1444                                 ctrl->io_queues[HCTX_TYPE_READ] - 1;
1445         queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
1446 }
1447
1448 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl,
1449                 int qid, size_t queue_size)
1450 {
1451         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1452         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1453         int ret, rcv_pdu_size;
1454
1455         mutex_init(&queue->queue_lock);
1456         queue->ctrl = ctrl;
1457         init_llist_head(&queue->req_list);
1458         INIT_LIST_HEAD(&queue->send_list);
1459         mutex_init(&queue->send_mutex);
1460         INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1461         queue->queue_size = queue_size;
1462
1463         if (qid > 0)
1464                 queue->cmnd_capsule_len = nctrl->ioccsz * 16;
1465         else
1466                 queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1467                                                 NVME_TCP_ADMIN_CCSZ;
1468
1469         ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM,
1470                         IPPROTO_TCP, &queue->sock);
1471         if (ret) {
1472                 dev_err(nctrl->device,
1473                         "failed to create socket: %d\n", ret);
1474                 goto err_destroy_mutex;
1475         }
1476
1477         nvme_tcp_reclassify_socket(queue->sock);
1478
1479         /* Single syn retry */
1480         tcp_sock_set_syncnt(queue->sock->sk, 1);
1481
1482         /* Set TCP no delay */
1483         tcp_sock_set_nodelay(queue->sock->sk);
1484
1485         /*
1486          * Cleanup whatever is sitting in the TCP transmit queue on socket
1487          * close. This is done to prevent stale data from being sent should
1488          * the network connection be restored before TCP times out.
1489          */
1490         sock_no_linger(queue->sock->sk);
1491
1492         if (so_priority > 0)
1493                 sock_set_priority(queue->sock->sk, so_priority);
1494
1495         /* Set socket type of service */
1496         if (nctrl->opts->tos >= 0)
1497                 ip_sock_set_tos(queue->sock->sk, nctrl->opts->tos);
1498
1499         /* Set 10 seconds timeout for icresp recvmsg */
1500         queue->sock->sk->sk_rcvtimeo = 10 * HZ;
1501
1502         queue->sock->sk->sk_allocation = GFP_ATOMIC;
1503         nvme_tcp_set_queue_io_cpu(queue);
1504         queue->request = NULL;
1505         queue->data_remaining = 0;
1506         queue->ddgst_remaining = 0;
1507         queue->pdu_remaining = 0;
1508         queue->pdu_offset = 0;
1509         sk_set_memalloc(queue->sock->sk);
1510
1511         if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
1512                 ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1513                         sizeof(ctrl->src_addr));
1514                 if (ret) {
1515                         dev_err(nctrl->device,
1516                                 "failed to bind queue %d socket %d\n",
1517                                 qid, ret);
1518                         goto err_sock;
1519                 }
1520         }
1521
1522         if (nctrl->opts->mask & NVMF_OPT_HOST_IFACE) {
1523                 char *iface = nctrl->opts->host_iface;
1524                 sockptr_t optval = KERNEL_SOCKPTR(iface);
1525
1526                 ret = sock_setsockopt(queue->sock, SOL_SOCKET, SO_BINDTODEVICE,
1527                                       optval, strlen(iface));
1528                 if (ret) {
1529                         dev_err(nctrl->device,
1530                           "failed to bind to interface %s queue %d err %d\n",
1531                           iface, qid, ret);
1532                         goto err_sock;
1533                 }
1534         }
1535
1536         queue->hdr_digest = nctrl->opts->hdr_digest;
1537         queue->data_digest = nctrl->opts->data_digest;
1538         if (queue->hdr_digest || queue->data_digest) {
1539                 ret = nvme_tcp_alloc_crypto(queue);
1540                 if (ret) {
1541                         dev_err(nctrl->device,
1542                                 "failed to allocate queue %d crypto\n", qid);
1543                         goto err_sock;
1544                 }
1545         }
1546
1547         rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1548                         nvme_tcp_hdgst_len(queue);
1549         queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1550         if (!queue->pdu) {
1551                 ret = -ENOMEM;
1552                 goto err_crypto;
1553         }
1554
1555         dev_dbg(nctrl->device, "connecting queue %d\n",
1556                         nvme_tcp_queue_id(queue));
1557
1558         ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1559                 sizeof(ctrl->addr), 0);
1560         if (ret) {
1561                 dev_err(nctrl->device,
1562                         "failed to connect socket: %d\n", ret);
1563                 goto err_rcv_pdu;
1564         }
1565
1566         ret = nvme_tcp_init_connection(queue);
1567         if (ret)
1568                 goto err_init_connect;
1569
1570         queue->rd_enabled = true;
1571         set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1572         nvme_tcp_init_recv_ctx(queue);
1573
1574         write_lock_bh(&queue->sock->sk->sk_callback_lock);
1575         queue->sock->sk->sk_user_data = queue;
1576         queue->state_change = queue->sock->sk->sk_state_change;
1577         queue->data_ready = queue->sock->sk->sk_data_ready;
1578         queue->write_space = queue->sock->sk->sk_write_space;
1579         queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1580         queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1581         queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1582 #ifdef CONFIG_NET_RX_BUSY_POLL
1583         queue->sock->sk->sk_ll_usec = 1;
1584 #endif
1585         write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1586
1587         return 0;
1588
1589 err_init_connect:
1590         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1591 err_rcv_pdu:
1592         kfree(queue->pdu);
1593 err_crypto:
1594         if (queue->hdr_digest || queue->data_digest)
1595                 nvme_tcp_free_crypto(queue);
1596 err_sock:
1597         sock_release(queue->sock);
1598         queue->sock = NULL;
1599 err_destroy_mutex:
1600         mutex_destroy(&queue->send_mutex);
1601         mutex_destroy(&queue->queue_lock);
1602         return ret;
1603 }
1604
1605 static void nvme_tcp_restore_sock_calls(struct nvme_tcp_queue *queue)
1606 {
1607         struct socket *sock = queue->sock;
1608
1609         write_lock_bh(&sock->sk->sk_callback_lock);
1610         sock->sk->sk_user_data  = NULL;
1611         sock->sk->sk_data_ready = queue->data_ready;
1612         sock->sk->sk_state_change = queue->state_change;
1613         sock->sk->sk_write_space  = queue->write_space;
1614         write_unlock_bh(&sock->sk->sk_callback_lock);
1615 }
1616
1617 static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1618 {
1619         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1620         nvme_tcp_restore_sock_calls(queue);
1621         cancel_work_sync(&queue->io_work);
1622 }
1623
1624 static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1625 {
1626         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1627         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1628
1629         mutex_lock(&queue->queue_lock);
1630         if (test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1631                 __nvme_tcp_stop_queue(queue);
1632         mutex_unlock(&queue->queue_lock);
1633 }
1634
1635 static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1636 {
1637         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1638         int ret;
1639
1640         if (idx)
1641                 ret = nvmf_connect_io_queue(nctrl, idx);
1642         else
1643                 ret = nvmf_connect_admin_queue(nctrl);
1644
1645         if (!ret) {
1646                 set_bit(NVME_TCP_Q_LIVE, &ctrl->queues[idx].flags);
1647         } else {
1648                 if (test_bit(NVME_TCP_Q_ALLOCATED, &ctrl->queues[idx].flags))
1649                         __nvme_tcp_stop_queue(&ctrl->queues[idx]);
1650                 dev_err(nctrl->device,
1651                         "failed to connect queue: %d ret=%d\n", idx, ret);
1652         }
1653         return ret;
1654 }
1655
1656 static struct blk_mq_tag_set *nvme_tcp_alloc_tagset(struct nvme_ctrl *nctrl,
1657                 bool admin)
1658 {
1659         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1660         struct blk_mq_tag_set *set;
1661         int ret;
1662
1663         if (admin) {
1664                 set = &ctrl->admin_tag_set;
1665                 memset(set, 0, sizeof(*set));
1666                 set->ops = &nvme_tcp_admin_mq_ops;
1667                 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1668                 set->reserved_tags = NVMF_RESERVED_TAGS;
1669                 set->numa_node = nctrl->numa_node;
1670                 set->flags = BLK_MQ_F_BLOCKING;
1671                 set->cmd_size = sizeof(struct nvme_tcp_request);
1672                 set->driver_data = ctrl;
1673                 set->nr_hw_queues = 1;
1674                 set->timeout = NVME_ADMIN_TIMEOUT;
1675         } else {
1676                 set = &ctrl->tag_set;
1677                 memset(set, 0, sizeof(*set));
1678                 set->ops = &nvme_tcp_mq_ops;
1679                 set->queue_depth = nctrl->sqsize + 1;
1680                 set->reserved_tags = NVMF_RESERVED_TAGS;
1681                 set->numa_node = nctrl->numa_node;
1682                 set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_BLOCKING;
1683                 set->cmd_size = sizeof(struct nvme_tcp_request);
1684                 set->driver_data = ctrl;
1685                 set->nr_hw_queues = nctrl->queue_count - 1;
1686                 set->timeout = NVME_IO_TIMEOUT;
1687                 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
1688         }
1689
1690         ret = blk_mq_alloc_tag_set(set);
1691         if (ret)
1692                 return ERR_PTR(ret);
1693
1694         return set;
1695 }
1696
1697 static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
1698 {
1699         if (to_tcp_ctrl(ctrl)->async_req.pdu) {
1700                 cancel_work_sync(&ctrl->async_event_work);
1701                 nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
1702                 to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
1703         }
1704
1705         nvme_tcp_free_queue(ctrl, 0);
1706 }
1707
1708 static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
1709 {
1710         int i;
1711
1712         for (i = 1; i < ctrl->queue_count; i++)
1713                 nvme_tcp_free_queue(ctrl, i);
1714 }
1715
1716 static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
1717 {
1718         int i;
1719
1720         for (i = 1; i < ctrl->queue_count; i++)
1721                 nvme_tcp_stop_queue(ctrl, i);
1722 }
1723
1724 static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl)
1725 {
1726         int i, ret = 0;
1727
1728         for (i = 1; i < ctrl->queue_count; i++) {
1729                 ret = nvme_tcp_start_queue(ctrl, i);
1730                 if (ret)
1731                         goto out_stop_queues;
1732         }
1733
1734         return 0;
1735
1736 out_stop_queues:
1737         for (i--; i >= 1; i--)
1738                 nvme_tcp_stop_queue(ctrl, i);
1739         return ret;
1740 }
1741
1742 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
1743 {
1744         int ret;
1745
1746         ret = nvme_tcp_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
1747         if (ret)
1748                 return ret;
1749
1750         ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
1751         if (ret)
1752                 goto out_free_queue;
1753
1754         return 0;
1755
1756 out_free_queue:
1757         nvme_tcp_free_queue(ctrl, 0);
1758         return ret;
1759 }
1760
1761 static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1762 {
1763         int i, ret;
1764
1765         for (i = 1; i < ctrl->queue_count; i++) {
1766                 ret = nvme_tcp_alloc_queue(ctrl, i,
1767                                 ctrl->sqsize + 1);
1768                 if (ret)
1769                         goto out_free_queues;
1770         }
1771
1772         return 0;
1773
1774 out_free_queues:
1775         for (i--; i >= 1; i--)
1776                 nvme_tcp_free_queue(ctrl, i);
1777
1778         return ret;
1779 }
1780
1781 static unsigned int nvme_tcp_nr_io_queues(struct nvme_ctrl *ctrl)
1782 {
1783         unsigned int nr_io_queues;
1784
1785         nr_io_queues = min(ctrl->opts->nr_io_queues, num_online_cpus());
1786         nr_io_queues += min(ctrl->opts->nr_write_queues, num_online_cpus());
1787         nr_io_queues += min(ctrl->opts->nr_poll_queues, num_online_cpus());
1788
1789         return nr_io_queues;
1790 }
1791
1792 static void nvme_tcp_set_io_queues(struct nvme_ctrl *nctrl,
1793                 unsigned int nr_io_queues)
1794 {
1795         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1796         struct nvmf_ctrl_options *opts = nctrl->opts;
1797
1798         if (opts->nr_write_queues && opts->nr_io_queues < nr_io_queues) {
1799                 /*
1800                  * separate read/write queues
1801                  * hand out dedicated default queues only after we have
1802                  * sufficient read queues.
1803                  */
1804                 ctrl->io_queues[HCTX_TYPE_READ] = opts->nr_io_queues;
1805                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
1806                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1807                         min(opts->nr_write_queues, nr_io_queues);
1808                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1809         } else {
1810                 /*
1811                  * shared read/write queues
1812                  * either no write queues were requested, or we don't have
1813                  * sufficient queue count to have dedicated default queues.
1814                  */
1815                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1816                         min(opts->nr_io_queues, nr_io_queues);
1817                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1818         }
1819
1820         if (opts->nr_poll_queues && nr_io_queues) {
1821                 /* map dedicated poll queues only if we have queues left */
1822                 ctrl->io_queues[HCTX_TYPE_POLL] =
1823                         min(opts->nr_poll_queues, nr_io_queues);
1824         }
1825 }
1826
1827 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1828 {
1829         unsigned int nr_io_queues;
1830         int ret;
1831
1832         nr_io_queues = nvme_tcp_nr_io_queues(ctrl);
1833         ret = nvme_set_queue_count(ctrl, &nr_io_queues);
1834         if (ret)
1835                 return ret;
1836
1837         if (nr_io_queues == 0) {
1838                 dev_err(ctrl->device,
1839                         "unable to set any I/O queues\n");
1840                 return -ENOMEM;
1841         }
1842
1843         ctrl->queue_count = nr_io_queues + 1;
1844         dev_info(ctrl->device,
1845                 "creating %d I/O queues.\n", nr_io_queues);
1846
1847         nvme_tcp_set_io_queues(ctrl, nr_io_queues);
1848
1849         return __nvme_tcp_alloc_io_queues(ctrl);
1850 }
1851
1852 static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove)
1853 {
1854         nvme_tcp_stop_io_queues(ctrl);
1855         if (remove) {
1856                 blk_cleanup_queue(ctrl->connect_q);
1857                 blk_mq_free_tag_set(ctrl->tagset);
1858         }
1859         nvme_tcp_free_io_queues(ctrl);
1860 }
1861
1862 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
1863 {
1864         int ret;
1865
1866         ret = nvme_tcp_alloc_io_queues(ctrl);
1867         if (ret)
1868                 return ret;
1869
1870         if (new) {
1871                 ctrl->tagset = nvme_tcp_alloc_tagset(ctrl, false);
1872                 if (IS_ERR(ctrl->tagset)) {
1873                         ret = PTR_ERR(ctrl->tagset);
1874                         goto out_free_io_queues;
1875                 }
1876
1877                 ctrl->connect_q = blk_mq_init_queue(ctrl->tagset);
1878                 if (IS_ERR(ctrl->connect_q)) {
1879                         ret = PTR_ERR(ctrl->connect_q);
1880                         goto out_free_tag_set;
1881                 }
1882         }
1883
1884         ret = nvme_tcp_start_io_queues(ctrl);
1885         if (ret)
1886                 goto out_cleanup_connect_q;
1887
1888         if (!new) {
1889                 nvme_start_queues(ctrl);
1890                 if (!nvme_wait_freeze_timeout(ctrl, NVME_IO_TIMEOUT)) {
1891                         /*
1892                          * If we timed out waiting for freeze we are likely to
1893                          * be stuck.  Fail the controller initialization just
1894                          * to be safe.
1895                          */
1896                         ret = -ENODEV;
1897                         goto out_wait_freeze_timed_out;
1898                 }
1899                 blk_mq_update_nr_hw_queues(ctrl->tagset,
1900                         ctrl->queue_count - 1);
1901                 nvme_unfreeze(ctrl);
1902         }
1903
1904         return 0;
1905
1906 out_wait_freeze_timed_out:
1907         nvme_stop_queues(ctrl);
1908         nvme_sync_io_queues(ctrl);
1909         nvme_tcp_stop_io_queues(ctrl);
1910 out_cleanup_connect_q:
1911         nvme_cancel_tagset(ctrl);
1912         if (new)
1913                 blk_cleanup_queue(ctrl->connect_q);
1914 out_free_tag_set:
1915         if (new)
1916                 blk_mq_free_tag_set(ctrl->tagset);
1917 out_free_io_queues:
1918         nvme_tcp_free_io_queues(ctrl);
1919         return ret;
1920 }
1921
1922 static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove)
1923 {
1924         nvme_tcp_stop_queue(ctrl, 0);
1925         if (remove) {
1926                 blk_cleanup_queue(ctrl->admin_q);
1927                 blk_cleanup_queue(ctrl->fabrics_q);
1928                 blk_mq_free_tag_set(ctrl->admin_tagset);
1929         }
1930         nvme_tcp_free_admin_queue(ctrl);
1931 }
1932
1933 static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
1934 {
1935         int error;
1936
1937         error = nvme_tcp_alloc_admin_queue(ctrl);
1938         if (error)
1939                 return error;
1940
1941         if (new) {
1942                 ctrl->admin_tagset = nvme_tcp_alloc_tagset(ctrl, true);
1943                 if (IS_ERR(ctrl->admin_tagset)) {
1944                         error = PTR_ERR(ctrl->admin_tagset);
1945                         goto out_free_queue;
1946                 }
1947
1948                 ctrl->fabrics_q = blk_mq_init_queue(ctrl->admin_tagset);
1949                 if (IS_ERR(ctrl->fabrics_q)) {
1950                         error = PTR_ERR(ctrl->fabrics_q);
1951                         goto out_free_tagset;
1952                 }
1953
1954                 ctrl->admin_q = blk_mq_init_queue(ctrl->admin_tagset);
1955                 if (IS_ERR(ctrl->admin_q)) {
1956                         error = PTR_ERR(ctrl->admin_q);
1957                         goto out_cleanup_fabrics_q;
1958                 }
1959         }
1960
1961         error = nvme_tcp_start_queue(ctrl, 0);
1962         if (error)
1963                 goto out_cleanup_queue;
1964
1965         error = nvme_enable_ctrl(ctrl);
1966         if (error)
1967                 goto out_stop_queue;
1968
1969         blk_mq_unquiesce_queue(ctrl->admin_q);
1970
1971         error = nvme_init_ctrl_finish(ctrl);
1972         if (error)
1973                 goto out_quiesce_queue;
1974
1975         return 0;
1976
1977 out_quiesce_queue:
1978         blk_mq_quiesce_queue(ctrl->admin_q);
1979         blk_sync_queue(ctrl->admin_q);
1980 out_stop_queue:
1981         nvme_tcp_stop_queue(ctrl, 0);
1982         nvme_cancel_admin_tagset(ctrl);
1983 out_cleanup_queue:
1984         if (new)
1985                 blk_cleanup_queue(ctrl->admin_q);
1986 out_cleanup_fabrics_q:
1987         if (new)
1988                 blk_cleanup_queue(ctrl->fabrics_q);
1989 out_free_tagset:
1990         if (new)
1991                 blk_mq_free_tag_set(ctrl->admin_tagset);
1992 out_free_queue:
1993         nvme_tcp_free_admin_queue(ctrl);
1994         return error;
1995 }
1996
1997 static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
1998                 bool remove)
1999 {
2000         blk_mq_quiesce_queue(ctrl->admin_q);
2001         blk_sync_queue(ctrl->admin_q);
2002         nvme_tcp_stop_queue(ctrl, 0);
2003         nvme_cancel_admin_tagset(ctrl);
2004         if (remove)
2005                 blk_mq_unquiesce_queue(ctrl->admin_q);
2006         nvme_tcp_destroy_admin_queue(ctrl, remove);
2007 }
2008
2009 static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
2010                 bool remove)
2011 {
2012         if (ctrl->queue_count <= 1)
2013                 return;
2014         blk_mq_quiesce_queue(ctrl->admin_q);
2015         nvme_start_freeze(ctrl);
2016         nvme_stop_queues(ctrl);
2017         nvme_sync_io_queues(ctrl);
2018         nvme_tcp_stop_io_queues(ctrl);
2019         nvme_cancel_tagset(ctrl);
2020         if (remove)
2021                 nvme_start_queues(ctrl);
2022         nvme_tcp_destroy_io_queues(ctrl, remove);
2023 }
2024
2025 static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
2026 {
2027         /* If we are resetting/deleting then do nothing */
2028         if (ctrl->state != NVME_CTRL_CONNECTING) {
2029                 WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW ||
2030                         ctrl->state == NVME_CTRL_LIVE);
2031                 return;
2032         }
2033
2034         if (nvmf_should_reconnect(ctrl)) {
2035                 dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
2036                         ctrl->opts->reconnect_delay);
2037                 queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
2038                                 ctrl->opts->reconnect_delay * HZ);
2039         } else {
2040                 dev_info(ctrl->device, "Removing controller...\n");
2041                 nvme_delete_ctrl(ctrl);
2042         }
2043 }
2044
2045 static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
2046 {
2047         struct nvmf_ctrl_options *opts = ctrl->opts;
2048         int ret;
2049
2050         ret = nvme_tcp_configure_admin_queue(ctrl, new);
2051         if (ret)
2052                 return ret;
2053
2054         if (ctrl->icdoff) {
2055                 ret = -EOPNOTSUPP;
2056                 dev_err(ctrl->device, "icdoff is not supported!\n");
2057                 goto destroy_admin;
2058         }
2059
2060         if (!nvme_ctrl_sgl_supported(ctrl)) {
2061                 ret = -EOPNOTSUPP;
2062                 dev_err(ctrl->device, "Mandatory sgls are not supported!\n");
2063                 goto destroy_admin;
2064         }
2065
2066         if (opts->queue_size > ctrl->sqsize + 1)
2067                 dev_warn(ctrl->device,
2068                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
2069                         opts->queue_size, ctrl->sqsize + 1);
2070
2071         if (ctrl->sqsize + 1 > ctrl->maxcmd) {
2072                 dev_warn(ctrl->device,
2073                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
2074                         ctrl->sqsize + 1, ctrl->maxcmd);
2075                 ctrl->sqsize = ctrl->maxcmd - 1;
2076         }
2077
2078         if (ctrl->queue_count > 1) {
2079                 ret = nvme_tcp_configure_io_queues(ctrl, new);
2080                 if (ret)
2081                         goto destroy_admin;
2082         }
2083
2084         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
2085                 /*
2086                  * state change failure is ok if we started ctrl delete,
2087                  * unless we're during creation of a new controller to
2088                  * avoid races with teardown flow.
2089                  */
2090                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2091                              ctrl->state != NVME_CTRL_DELETING_NOIO);
2092                 WARN_ON_ONCE(new);
2093                 ret = -EINVAL;
2094                 goto destroy_io;
2095         }
2096
2097         nvme_start_ctrl(ctrl);
2098         return 0;
2099
2100 destroy_io:
2101         if (ctrl->queue_count > 1) {
2102                 nvme_stop_queues(ctrl);
2103                 nvme_sync_io_queues(ctrl);
2104                 nvme_tcp_stop_io_queues(ctrl);
2105                 nvme_cancel_tagset(ctrl);
2106                 nvme_tcp_destroy_io_queues(ctrl, new);
2107         }
2108 destroy_admin:
2109         blk_mq_quiesce_queue(ctrl->admin_q);
2110         blk_sync_queue(ctrl->admin_q);
2111         nvme_tcp_stop_queue(ctrl, 0);
2112         nvme_cancel_admin_tagset(ctrl);
2113         nvme_tcp_destroy_admin_queue(ctrl, new);
2114         return ret;
2115 }
2116
2117 static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
2118 {
2119         struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
2120                         struct nvme_tcp_ctrl, connect_work);
2121         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2122
2123         ++ctrl->nr_reconnects;
2124
2125         if (nvme_tcp_setup_ctrl(ctrl, false))
2126                 goto requeue;
2127
2128         dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n",
2129                         ctrl->nr_reconnects);
2130
2131         ctrl->nr_reconnects = 0;
2132
2133         return;
2134
2135 requeue:
2136         dev_info(ctrl->device, "Failed reconnect attempt %d\n",
2137                         ctrl->nr_reconnects);
2138         nvme_tcp_reconnect_or_remove(ctrl);
2139 }
2140
2141 static void nvme_tcp_error_recovery_work(struct work_struct *work)
2142 {
2143         struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
2144                                 struct nvme_tcp_ctrl, err_work);
2145         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2146
2147         nvme_stop_keep_alive(ctrl);
2148         flush_work(&ctrl->async_event_work);
2149         nvme_tcp_teardown_io_queues(ctrl, false);
2150         /* unquiesce to fail fast pending requests */
2151         nvme_start_queues(ctrl);
2152         nvme_tcp_teardown_admin_queue(ctrl, false);
2153         blk_mq_unquiesce_queue(ctrl->admin_q);
2154
2155         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2156                 /* state change failure is ok if we started ctrl delete */
2157                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2158                              ctrl->state != NVME_CTRL_DELETING_NOIO);
2159                 return;
2160         }
2161
2162         nvme_tcp_reconnect_or_remove(ctrl);
2163 }
2164
2165 static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
2166 {
2167         cancel_work_sync(&to_tcp_ctrl(ctrl)->err_work);
2168         cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
2169
2170         nvme_tcp_teardown_io_queues(ctrl, shutdown);
2171         blk_mq_quiesce_queue(ctrl->admin_q);
2172         if (shutdown)
2173                 nvme_shutdown_ctrl(ctrl);
2174         else
2175                 nvme_disable_ctrl(ctrl);
2176         nvme_tcp_teardown_admin_queue(ctrl, shutdown);
2177 }
2178
2179 static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
2180 {
2181         nvme_tcp_teardown_ctrl(ctrl, true);
2182 }
2183
2184 static void nvme_reset_ctrl_work(struct work_struct *work)
2185 {
2186         struct nvme_ctrl *ctrl =
2187                 container_of(work, struct nvme_ctrl, reset_work);
2188
2189         nvme_stop_ctrl(ctrl);
2190         nvme_tcp_teardown_ctrl(ctrl, false);
2191
2192         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2193                 /* state change failure is ok if we started ctrl delete */
2194                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2195                              ctrl->state != NVME_CTRL_DELETING_NOIO);
2196                 return;
2197         }
2198
2199         if (nvme_tcp_setup_ctrl(ctrl, false))
2200                 goto out_fail;
2201
2202         return;
2203
2204 out_fail:
2205         ++ctrl->nr_reconnects;
2206         nvme_tcp_reconnect_or_remove(ctrl);
2207 }
2208
2209 static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
2210 {
2211         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
2212
2213         if (list_empty(&ctrl->list))
2214                 goto free_ctrl;
2215
2216         mutex_lock(&nvme_tcp_ctrl_mutex);
2217         list_del(&ctrl->list);
2218         mutex_unlock(&nvme_tcp_ctrl_mutex);
2219
2220         nvmf_free_options(nctrl->opts);
2221 free_ctrl:
2222         kfree(ctrl->queues);
2223         kfree(ctrl);
2224 }
2225
2226 static void nvme_tcp_set_sg_null(struct nvme_command *c)
2227 {
2228         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2229
2230         sg->addr = 0;
2231         sg->length = 0;
2232         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2233                         NVME_SGL_FMT_TRANSPORT_A;
2234 }
2235
2236 static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
2237                 struct nvme_command *c, u32 data_len)
2238 {
2239         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2240
2241         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
2242         sg->length = cpu_to_le32(data_len);
2243         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
2244 }
2245
2246 static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
2247                 u32 data_len)
2248 {
2249         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2250
2251         sg->addr = 0;
2252         sg->length = cpu_to_le32(data_len);
2253         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2254                         NVME_SGL_FMT_TRANSPORT_A;
2255 }
2256
2257 static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
2258 {
2259         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2260         struct nvme_tcp_queue *queue = &ctrl->queues[0];
2261         struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2262         struct nvme_command *cmd = &pdu->cmd;
2263         u8 hdgst = nvme_tcp_hdgst_len(queue);
2264
2265         memset(pdu, 0, sizeof(*pdu));
2266         pdu->hdr.type = nvme_tcp_cmd;
2267         if (queue->hdr_digest)
2268                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2269         pdu->hdr.hlen = sizeof(*pdu);
2270         pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2271
2272         cmd->common.opcode = nvme_admin_async_event;
2273         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2274         cmd->common.flags |= NVME_CMD_SGL_METABUF;
2275         nvme_tcp_set_sg_null(cmd);
2276
2277         ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2278         ctrl->async_req.offset = 0;
2279         ctrl->async_req.curr_bio = NULL;
2280         ctrl->async_req.data_len = 0;
2281
2282         nvme_tcp_queue_request(&ctrl->async_req, true, true);
2283 }
2284
2285 static void nvme_tcp_complete_timed_out(struct request *rq)
2286 {
2287         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2288         struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2289
2290         nvme_tcp_stop_queue(ctrl, nvme_tcp_queue_id(req->queue));
2291         if (blk_mq_request_started(rq) && !blk_mq_request_completed(rq)) {
2292                 nvme_req(rq)->status = NVME_SC_HOST_ABORTED_CMD;
2293                 blk_mq_complete_request(rq);
2294         }
2295 }
2296
2297 static enum blk_eh_timer_return
2298 nvme_tcp_timeout(struct request *rq, bool reserved)
2299 {
2300         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2301         struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2302         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2303
2304         dev_warn(ctrl->device,
2305                 "queue %d: timeout request %#x type %d\n",
2306                 nvme_tcp_queue_id(req->queue), rq->tag, pdu->hdr.type);
2307
2308         if (ctrl->state != NVME_CTRL_LIVE) {
2309                 /*
2310                  * If we are resetting, connecting or deleting we should
2311                  * complete immediately because we may block controller
2312                  * teardown or setup sequence
2313                  * - ctrl disable/shutdown fabrics requests
2314                  * - connect requests
2315                  * - initialization admin requests
2316                  * - I/O requests that entered after unquiescing and
2317                  *   the controller stopped responding
2318                  *
2319                  * All other requests should be cancelled by the error
2320                  * recovery work, so it's fine that we fail it here.
2321                  */
2322                 nvme_tcp_complete_timed_out(rq);
2323                 return BLK_EH_DONE;
2324         }
2325
2326         /*
2327          * LIVE state should trigger the normal error recovery which will
2328          * handle completing this request.
2329          */
2330         nvme_tcp_error_recovery(ctrl);
2331         return BLK_EH_RESET_TIMER;
2332 }
2333
2334 static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2335                         struct request *rq)
2336 {
2337         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2338         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2339         struct nvme_command *c = &pdu->cmd;
2340
2341         c->common.flags |= NVME_CMD_SGL_METABUF;
2342
2343         if (!blk_rq_nr_phys_segments(rq))
2344                 nvme_tcp_set_sg_null(c);
2345         else if (rq_data_dir(rq) == WRITE &&
2346             req->data_len <= nvme_tcp_inline_data_size(queue))
2347                 nvme_tcp_set_sg_inline(queue, c, req->data_len);
2348         else
2349                 nvme_tcp_set_sg_host_data(c, req->data_len);
2350
2351         return 0;
2352 }
2353
2354 static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2355                 struct request *rq)
2356 {
2357         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2358         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2359         struct nvme_tcp_queue *queue = req->queue;
2360         u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2361         blk_status_t ret;
2362
2363         ret = nvme_setup_cmd(ns, rq);
2364         if (ret)
2365                 return ret;
2366
2367         req->state = NVME_TCP_SEND_CMD_PDU;
2368         req->status = cpu_to_le16(NVME_SC_SUCCESS);
2369         req->offset = 0;
2370         req->data_sent = 0;
2371         req->pdu_len = 0;
2372         req->pdu_sent = 0;
2373         req->data_len = blk_rq_nr_phys_segments(rq) ?
2374                                 blk_rq_payload_bytes(rq) : 0;
2375         req->curr_bio = rq->bio;
2376         if (req->curr_bio && req->data_len)
2377                 nvme_tcp_init_iter(req, rq_data_dir(rq));
2378
2379         if (rq_data_dir(rq) == WRITE &&
2380             req->data_len <= nvme_tcp_inline_data_size(queue))
2381                 req->pdu_len = req->data_len;
2382
2383         pdu->hdr.type = nvme_tcp_cmd;
2384         pdu->hdr.flags = 0;
2385         if (queue->hdr_digest)
2386                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2387         if (queue->data_digest && req->pdu_len) {
2388                 pdu->hdr.flags |= NVME_TCP_F_DDGST;
2389                 ddgst = nvme_tcp_ddgst_len(queue);
2390         }
2391         pdu->hdr.hlen = sizeof(*pdu);
2392         pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2393         pdu->hdr.plen =
2394                 cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2395
2396         ret = nvme_tcp_map_data(queue, rq);
2397         if (unlikely(ret)) {
2398                 nvme_cleanup_cmd(rq);
2399                 dev_err(queue->ctrl->ctrl.device,
2400                         "Failed to map data (%d)\n", ret);
2401                 return ret;
2402         }
2403
2404         return 0;
2405 }
2406
2407 static void nvme_tcp_commit_rqs(struct blk_mq_hw_ctx *hctx)
2408 {
2409         struct nvme_tcp_queue *queue = hctx->driver_data;
2410
2411         if (!llist_empty(&queue->req_list))
2412                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
2413 }
2414
2415 static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2416                 const struct blk_mq_queue_data *bd)
2417 {
2418         struct nvme_ns *ns = hctx->queue->queuedata;
2419         struct nvme_tcp_queue *queue = hctx->driver_data;
2420         struct request *rq = bd->rq;
2421         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2422         bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2423         blk_status_t ret;
2424
2425         if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2426                 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
2427
2428         ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2429         if (unlikely(ret))
2430                 return ret;
2431
2432         blk_mq_start_request(rq);
2433
2434         nvme_tcp_queue_request(req, true, bd->last);
2435
2436         return BLK_STS_OK;
2437 }
2438
2439 static int nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2440 {
2441         struct nvme_tcp_ctrl *ctrl = set->driver_data;
2442         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2443
2444         if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2445                 /* separate read/write queues */
2446                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2447                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2448                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2449                 set->map[HCTX_TYPE_READ].nr_queues =
2450                         ctrl->io_queues[HCTX_TYPE_READ];
2451                 set->map[HCTX_TYPE_READ].queue_offset =
2452                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2453         } else {
2454                 /* shared read/write queues */
2455                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2456                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2457                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2458                 set->map[HCTX_TYPE_READ].nr_queues =
2459                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2460                 set->map[HCTX_TYPE_READ].queue_offset = 0;
2461         }
2462         blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2463         blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
2464
2465         if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2466                 /* map dedicated poll queues only if we have queues left */
2467                 set->map[HCTX_TYPE_POLL].nr_queues =
2468                                 ctrl->io_queues[HCTX_TYPE_POLL];
2469                 set->map[HCTX_TYPE_POLL].queue_offset =
2470                         ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2471                         ctrl->io_queues[HCTX_TYPE_READ];
2472                 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2473         }
2474
2475         dev_info(ctrl->ctrl.device,
2476                 "mapped %d/%d/%d default/read/poll queues.\n",
2477                 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2478                 ctrl->io_queues[HCTX_TYPE_READ],
2479                 ctrl->io_queues[HCTX_TYPE_POLL]);
2480
2481         return 0;
2482 }
2483
2484 static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx)
2485 {
2486         struct nvme_tcp_queue *queue = hctx->driver_data;
2487         struct sock *sk = queue->sock->sk;
2488
2489         if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2490                 return 0;
2491
2492         set_bit(NVME_TCP_Q_POLLING, &queue->flags);
2493         if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue))
2494                 sk_busy_loop(sk, true);
2495         nvme_tcp_try_recv(queue);
2496         clear_bit(NVME_TCP_Q_POLLING, &queue->flags);
2497         return queue->nr_cqe;
2498 }
2499
2500 static const struct blk_mq_ops nvme_tcp_mq_ops = {
2501         .queue_rq       = nvme_tcp_queue_rq,
2502         .commit_rqs     = nvme_tcp_commit_rqs,
2503         .complete       = nvme_complete_rq,
2504         .init_request   = nvme_tcp_init_request,
2505         .exit_request   = nvme_tcp_exit_request,
2506         .init_hctx      = nvme_tcp_init_hctx,
2507         .timeout        = nvme_tcp_timeout,
2508         .map_queues     = nvme_tcp_map_queues,
2509         .poll           = nvme_tcp_poll,
2510 };
2511
2512 static const struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2513         .queue_rq       = nvme_tcp_queue_rq,
2514         .complete       = nvme_complete_rq,
2515         .init_request   = nvme_tcp_init_request,
2516         .exit_request   = nvme_tcp_exit_request,
2517         .init_hctx      = nvme_tcp_init_admin_hctx,
2518         .timeout        = nvme_tcp_timeout,
2519 };
2520
2521 static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2522         .name                   = "tcp",
2523         .module                 = THIS_MODULE,
2524         .flags                  = NVME_F_FABRICS,
2525         .reg_read32             = nvmf_reg_read32,
2526         .reg_read64             = nvmf_reg_read64,
2527         .reg_write32            = nvmf_reg_write32,
2528         .free_ctrl              = nvme_tcp_free_ctrl,
2529         .submit_async_event     = nvme_tcp_submit_async_event,
2530         .delete_ctrl            = nvme_tcp_delete_ctrl,
2531         .get_address            = nvmf_get_address,
2532 };
2533
2534 static bool
2535 nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2536 {
2537         struct nvme_tcp_ctrl *ctrl;
2538         bool found = false;
2539
2540         mutex_lock(&nvme_tcp_ctrl_mutex);
2541         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2542                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2543                 if (found)
2544                         break;
2545         }
2546         mutex_unlock(&nvme_tcp_ctrl_mutex);
2547
2548         return found;
2549 }
2550
2551 static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2552                 struct nvmf_ctrl_options *opts)
2553 {
2554         struct nvme_tcp_ctrl *ctrl;
2555         int ret;
2556
2557         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2558         if (!ctrl)
2559                 return ERR_PTR(-ENOMEM);
2560
2561         INIT_LIST_HEAD(&ctrl->list);
2562         ctrl->ctrl.opts = opts;
2563         ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2564                                 opts->nr_poll_queues + 1;
2565         ctrl->ctrl.sqsize = opts->queue_size - 1;
2566         ctrl->ctrl.kato = opts->kato;
2567
2568         INIT_DELAYED_WORK(&ctrl->connect_work,
2569                         nvme_tcp_reconnect_ctrl_work);
2570         INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2571         INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2572
2573         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2574                 opts->trsvcid =
2575                         kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2576                 if (!opts->trsvcid) {
2577                         ret = -ENOMEM;
2578                         goto out_free_ctrl;
2579                 }
2580                 opts->mask |= NVMF_OPT_TRSVCID;
2581         }
2582
2583         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2584                         opts->traddr, opts->trsvcid, &ctrl->addr);
2585         if (ret) {
2586                 pr_err("malformed address passed: %s:%s\n",
2587                         opts->traddr, opts->trsvcid);
2588                 goto out_free_ctrl;
2589         }
2590
2591         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2592                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2593                         opts->host_traddr, NULL, &ctrl->src_addr);
2594                 if (ret) {
2595                         pr_err("malformed src address passed: %s\n",
2596                                opts->host_traddr);
2597                         goto out_free_ctrl;
2598                 }
2599         }
2600
2601         if (opts->mask & NVMF_OPT_HOST_IFACE) {
2602                 if (!__dev_get_by_name(&init_net, opts->host_iface)) {
2603                         pr_err("invalid interface passed: %s\n",
2604                                opts->host_iface);
2605                         ret = -ENODEV;
2606                         goto out_free_ctrl;
2607                 }
2608         }
2609
2610         if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2611                 ret = -EALREADY;
2612                 goto out_free_ctrl;
2613         }
2614
2615         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2616                                 GFP_KERNEL);
2617         if (!ctrl->queues) {
2618                 ret = -ENOMEM;
2619                 goto out_free_ctrl;
2620         }
2621
2622         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2623         if (ret)
2624                 goto out_kfree_queues;
2625
2626         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2627                 WARN_ON_ONCE(1);
2628                 ret = -EINTR;
2629                 goto out_uninit_ctrl;
2630         }
2631
2632         ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2633         if (ret)
2634                 goto out_uninit_ctrl;
2635
2636         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
2637                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2638
2639         mutex_lock(&nvme_tcp_ctrl_mutex);
2640         list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2641         mutex_unlock(&nvme_tcp_ctrl_mutex);
2642
2643         return &ctrl->ctrl;
2644
2645 out_uninit_ctrl:
2646         nvme_uninit_ctrl(&ctrl->ctrl);
2647         nvme_put_ctrl(&ctrl->ctrl);
2648         if (ret > 0)
2649                 ret = -EIO;
2650         return ERR_PTR(ret);
2651 out_kfree_queues:
2652         kfree(ctrl->queues);
2653 out_free_ctrl:
2654         kfree(ctrl);
2655         return ERR_PTR(ret);
2656 }
2657
2658 static struct nvmf_transport_ops nvme_tcp_transport = {
2659         .name           = "tcp",
2660         .module         = THIS_MODULE,
2661         .required_opts  = NVMF_OPT_TRADDR,
2662         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2663                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2664                           NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
2665                           NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2666                           NVMF_OPT_TOS | NVMF_OPT_HOST_IFACE,
2667         .create_ctrl    = nvme_tcp_create_ctrl,
2668 };
2669
2670 static int __init nvme_tcp_init_module(void)
2671 {
2672         nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
2673                         WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2674         if (!nvme_tcp_wq)
2675                 return -ENOMEM;
2676
2677         nvmf_register_transport(&nvme_tcp_transport);
2678         return 0;
2679 }
2680
2681 static void __exit nvme_tcp_cleanup_module(void)
2682 {
2683         struct nvme_tcp_ctrl *ctrl;
2684
2685         nvmf_unregister_transport(&nvme_tcp_transport);
2686
2687         mutex_lock(&nvme_tcp_ctrl_mutex);
2688         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
2689                 nvme_delete_ctrl(&ctrl->ctrl);
2690         mutex_unlock(&nvme_tcp_ctrl_mutex);
2691         flush_workqueue(nvme_delete_wq);
2692
2693         destroy_workqueue(nvme_tcp_wq);
2694 }
2695
2696 module_init(nvme_tcp_init_module);
2697 module_exit(nvme_tcp_cleanup_module);
2698
2699 MODULE_LICENSE("GPL v2");