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/crc32.h>
  12 #include <linux/nvme-tcp.h>
  13 #include <linux/nvme-keyring.h>
  14 #include <net/sock.h>
  15 #include <net/tcp.h>
  16 #include <net/tls.h>
  17 #include <net/tls_prot.h>
  18 #include <net/handshake.h>
  19 #include <linux/blk-mq.h>
  20 #include <net/busy_poll.h>
  21 #include <trace/events/sock.h>
  22
  23 #include "nvme.h"
  24 #include "fabrics.h"
  25
  26 struct nvme_tcp_queue;
  27
  28 /* Define the socket priority to use for connections were it is desirable
  29  * that the NIC consider performing optimized packet processing or filtering.
  30  * A non-zero value being sufficient to indicate general consideration of any
  31  * possible optimization.  Making it a module param allows for alternative
  32  * values that may be unique for some NIC implementations.
  33  */
  34 static int so_priority;
  35 module_param(so_priority, int, 0644);
  36 MODULE_PARM_DESC(so_priority, "nvme tcp socket optimize priority");
  37
  38 /*
  39  * Use the unbound workqueue for nvme_tcp_wq, then we can set the cpu affinity
  40  * from sysfs.
  41  */
  42 static bool wq_unbound;
  43 module_param(wq_unbound, bool, 0644);
  44 MODULE_PARM_DESC(wq_unbound, "Use unbound workqueue for nvme-tcp IO context (default false)");
  45
  46 /*
  47  * TLS handshake timeout
  48  */
  49 static int tls_handshake_timeout = 10;
  50 #ifdef CONFIG_NVME_TCP_TLS
  51 module_param(tls_handshake_timeout, int, 0644);
  52 MODULE_PARM_DESC(tls_handshake_timeout,
  53                  "nvme TLS handshake timeout in seconds (default 10)");
  54 #endif
  55
  56 static atomic_t nvme_tcp_cpu_queues[NR_CPUS];
  57
  58 #ifdef CONFIG_DEBUG_LOCK_ALLOC
  59 /* lockdep can detect a circular dependency of the form
  60  *   sk_lock -> mmap_lock (page fault) -> fs locks -> sk_lock
  61  * because dependencies are tracked for both nvme-tcp and user contexts. Using
  62  * a separate class prevents lockdep from conflating nvme-tcp socket use with
  63  * user-space socket API use.
  64  */
  65 static struct lock_class_key nvme_tcp_sk_key[2];
  66 static struct lock_class_key nvme_tcp_slock_key[2];
  67
  68 static void nvme_tcp_reclassify_socket(struct socket *sock)
  69 {
  70         struct sock *sk = sock->sk;
  71
  72         if (WARN_ON_ONCE(!sock_allow_reclassification(sk)))
  73                 return;
  74
  75         switch (sk->sk_family) {
  76         case AF_INET:
  77                 sock_lock_init_class_and_name(sk, "slock-AF_INET-NVME",
  78                                               &nvme_tcp_slock_key[0],
  79                                               "sk_lock-AF_INET-NVME",
  80                                               &nvme_tcp_sk_key[0]);
  81                 break;
  82         case AF_INET6:
  83                 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NVME",
  84                                               &nvme_tcp_slock_key[1],
  85                                               "sk_lock-AF_INET6-NVME",
  86                                               &nvme_tcp_sk_key[1]);
  87                 break;
  88         default:
  89                 WARN_ON_ONCE(1);
  90         }
  91 }
  92 #else
  93 static void nvme_tcp_reclassify_socket(struct socket *sock) { }
  94 #endif
  95
  96 enum nvme_tcp_send_state {
  97         NVME_TCP_SEND_CMD_PDU = 0,
  98         NVME_TCP_SEND_H2C_PDU,
  99         NVME_TCP_SEND_DATA,
 100         NVME_TCP_SEND_DDGST,
 101 };
 102
 103 struct nvme_tcp_request {
 104         struct nvme_request     req;
 105         void                    *pdu;
 106         struct nvme_tcp_queue   *queue;
 107         u32                     data_len;
 108         u32                     pdu_len;
 109         u32                     pdu_sent;
 110         u32                     h2cdata_left;
 111         u32                     h2cdata_offset;
 112         u16                     ttag;
 113         __le16                  status;
 114         struct list_head        entry;
 115         struct llist_node       lentry;
 116         __le32                  ddgst;
 117
 118         struct bio              *curr_bio;
 119         struct iov_iter         iter;
 120
 121         /* send state */
 122         size_t                  offset;
 123         size_t                  data_sent;
 124         enum nvme_tcp_send_state state;
 125 };
 126
 127 enum nvme_tcp_queue_flags {
 128         NVME_TCP_Q_ALLOCATED    = 0,
 129         NVME_TCP_Q_LIVE         = 1,
 130         NVME_TCP_Q_POLLING      = 2,
 131         NVME_TCP_Q_IO_CPU_SET   = 3,
 132 };
 133
 134 enum nvme_tcp_recv_state {
 135         NVME_TCP_RECV_PDU = 0,
 136         NVME_TCP_RECV_DATA,
 137         NVME_TCP_RECV_DDGST,
 138 };
 139
 140 struct nvme_tcp_ctrl;
 141 struct nvme_tcp_queue {
 142         struct socket           *sock;
 143         struct work_struct      io_work;
 144         int                     io_cpu;
 145
 146         struct mutex            queue_lock;
 147         struct mutex            send_mutex;
 148         struct llist_head       req_list;
 149         struct list_head        send_list;
 150
 151         /* recv state */
 152         void                    *pdu;
 153         int                     pdu_remaining;
 154         int                     pdu_offset;
 155         size_t                  data_remaining;
 156         size_t                  ddgst_remaining;
 157         unsigned int            nr_cqe;
 158
 159         /* send state */
 160         struct nvme_tcp_request *request;
 161
 162         u32                     maxh2cdata;
 163         size_t                  cmnd_capsule_len;
 164         struct nvme_tcp_ctrl    *ctrl;
 165         unsigned long           flags;
 166         bool                    rd_enabled;
 167
 168         bool                    hdr_digest;
 169         bool                    data_digest;
 170         bool                    tls_enabled;
 171         u32                     rcv_crc;
 172         u32                     snd_crc;
 173         __le32                  exp_ddgst;
 174         __le32                  recv_ddgst;
 175         struct completion       tls_complete;
 176         int                     tls_err;
 177         struct page_frag_cache  pf_cache;
 178
 179         void (*state_change)(struct sock *);
 180         void (*data_ready)(struct sock *);
 181         void (*write_space)(struct sock *);
 182 };
 183
 184 struct nvme_tcp_ctrl {
 185         /* read only in the hot path */
 186         struct nvme_tcp_queue   *queues;
 187         struct blk_mq_tag_set   tag_set;
 188
 189         /* other member variables */
 190         struct list_head        list;
 191         struct blk_mq_tag_set   admin_tag_set;
 192         struct sockaddr_storage addr;
 193         struct sockaddr_storage src_addr;
 194         struct nvme_ctrl        ctrl;
 195
 196         struct work_struct      err_work;
 197         struct delayed_work     connect_work;
 198         struct nvme_tcp_request async_req;
 199         u32                     io_queues[HCTX_MAX_TYPES];
 200 };
 201
 202 static LIST_HEAD(nvme_tcp_ctrl_list);
 203 static DEFINE_MUTEX(nvme_tcp_ctrl_mutex);
 204 static struct workqueue_struct *nvme_tcp_wq;
 205 static const struct blk_mq_ops nvme_tcp_mq_ops;
 206 static const struct blk_mq_ops nvme_tcp_admin_mq_ops;
 207 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue);
 208
 209 static inline struct nvme_tcp_ctrl *to_tcp_ctrl(struct nvme_ctrl *ctrl)
 210 {
 211         return container_of(ctrl, struct nvme_tcp_ctrl, ctrl);
 212 }
 213
 214 static inline int nvme_tcp_queue_id(struct nvme_tcp_queue *queue)
 215 {
 216         return queue - queue->ctrl->queues;
 217 }
 218
 219 static inline bool nvme_tcp_recv_pdu_supported(enum nvme_tcp_pdu_type type)
 220 {
 221         switch (type) {
 222         case nvme_tcp_c2h_term:
 223         case nvme_tcp_c2h_data:
 224         case nvme_tcp_r2t:
 225         case nvme_tcp_rsp:
 226                 return true;
 227         default:
 228                 return false;
 229         }
 230 }
 231
 232 /*
 233  * Check if the queue is TLS encrypted
 234  */
 235 static inline bool nvme_tcp_queue_tls(struct nvme_tcp_queue *queue)
 236 {
 237         if (!IS_ENABLED(CONFIG_NVME_TCP_TLS))
 238                 return 0;
 239
 240         return queue->tls_enabled;
 241 }
 242
 243 /*
 244  * Check if TLS is configured for the controller.
 245  */
 246 static inline bool nvme_tcp_tls_configured(struct nvme_ctrl *ctrl)
 247 {
 248         if (!IS_ENABLED(CONFIG_NVME_TCP_TLS))
 249                 return 0;
 250
 251         return ctrl->opts->tls || ctrl->opts->concat;
 252 }
 253
 254 static inline struct blk_mq_tags *nvme_tcp_tagset(struct nvme_tcp_queue *queue)
 255 {
 256         u32 queue_idx = nvme_tcp_queue_id(queue);
 257
 258         if (queue_idx == 0)
 259                 return queue->ctrl->admin_tag_set.tags[queue_idx];
 260         return queue->ctrl->tag_set.tags[queue_idx - 1];
 261 }
 262
 263 static inline u8 nvme_tcp_hdgst_len(struct nvme_tcp_queue *queue)
 264 {
 265         return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 266 }
 267
 268 static inline u8 nvme_tcp_ddgst_len(struct nvme_tcp_queue *queue)
 269 {
 270         return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 271 }
 272
 273 static inline void *nvme_tcp_req_cmd_pdu(struct nvme_tcp_request *req)
 274 {
 275         return req->pdu;
 276 }
 277
 278 static inline void *nvme_tcp_req_data_pdu(struct nvme_tcp_request *req)
 279 {
 280         /* use the pdu space in the back for the data pdu */
 281         return req->pdu + sizeof(struct nvme_tcp_cmd_pdu) -
 282                 sizeof(struct nvme_tcp_data_pdu);
 283 }
 284
 285 static inline size_t nvme_tcp_inline_data_size(struct nvme_tcp_request *req)
 286 {
 287         if (nvme_is_fabrics(req->req.cmd))
 288                 return NVME_TCP_ADMIN_CCSZ;
 289         return req->queue->cmnd_capsule_len - sizeof(struct nvme_command);
 290 }
 291
 292 static inline bool nvme_tcp_async_req(struct nvme_tcp_request *req)
 293 {
 294         return req == &req->queue->ctrl->async_req;
 295 }
 296
 297 static inline bool nvme_tcp_has_inline_data(struct nvme_tcp_request *req)
 298 {
 299         struct request *rq;
 300
 301         if (unlikely(nvme_tcp_async_req(req)))
 302                 return false; /* async events don't have a request */
 303
 304         rq = blk_mq_rq_from_pdu(req);
 305
 306         return rq_data_dir(rq) == WRITE && req->data_len &&
 307                 req->data_len <= nvme_tcp_inline_data_size(req);
 308 }
 309
 310 static inline struct page *nvme_tcp_req_cur_page(struct nvme_tcp_request *req)
 311 {
 312         return req->iter.bvec->bv_page;
 313 }
 314
 315 static inline size_t nvme_tcp_req_cur_offset(struct nvme_tcp_request *req)
 316 {
 317         return req->iter.bvec->bv_offset + req->iter.iov_offset;
 318 }
 319
 320 static inline size_t nvme_tcp_req_cur_length(struct nvme_tcp_request *req)
 321 {
 322         return min_t(size_t, iov_iter_single_seg_count(&req->iter),
 323                         req->pdu_len - req->pdu_sent);
 324 }
 325
 326 static inline size_t nvme_tcp_pdu_data_left(struct nvme_tcp_request *req)
 327 {
 328         return rq_data_dir(blk_mq_rq_from_pdu(req)) == WRITE ?
 329                         req->pdu_len - req->pdu_sent : 0;
 330 }
 331
 332 static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req,
 333                 int len)
 334 {
 335         return nvme_tcp_pdu_data_left(req) <= len;
 336 }
 337
 338 static void nvme_tcp_init_iter(struct nvme_tcp_request *req,
 339                 unsigned int dir)
 340 {
 341         struct request *rq = blk_mq_rq_from_pdu(req);
 342         struct bio_vec *vec;
 343         unsigned int size;
 344         int nr_bvec;
 345         size_t offset;
 346
 347         if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) {
 348                 vec = &rq->special_vec;
 349                 nr_bvec = 1;
 350                 size = blk_rq_payload_bytes(rq);
 351                 offset = 0;
 352         } else {
 353                 struct bio *bio = req->curr_bio;
 354                 struct bvec_iter bi;
 355                 struct bio_vec bv;
 356
 357                 vec = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
 358                 nr_bvec = 0;
 359                 bio_for_each_bvec(bv, bio, bi) {
 360                         nr_bvec++;
 361                 }
 362                 size = bio->bi_iter.bi_size;
 363                 offset = bio->bi_iter.bi_bvec_done;
 364         }
 365
 366         iov_iter_bvec(&req->iter, dir, vec, nr_bvec, size);
 367         req->iter.iov_offset = offset;
 368 }
 369
 370 static inline void nvme_tcp_advance_req(struct nvme_tcp_request *req,
 371                 int len)
 372 {
 373         req->data_sent += len;
 374         req->pdu_sent += len;
 375         iov_iter_advance(&req->iter, len);
 376         if (!iov_iter_count(&req->iter) &&
 377             req->data_sent < req->data_len) {
 378                 req->curr_bio = req->curr_bio->bi_next;
 379                 nvme_tcp_init_iter(req, ITER_SOURCE);
 380         }
 381 }
 382
 383 static inline void nvme_tcp_send_all(struct nvme_tcp_queue *queue)
 384 {
 385         int ret;
 386
 387         /* drain the send queue as much as we can... */
 388         do {
 389                 ret = nvme_tcp_try_send(queue);
 390         } while (ret > 0);
 391 }
 392
 393 static inline bool nvme_tcp_queue_has_pending(struct nvme_tcp_queue *queue)
 394 {
 395         return !list_empty(&queue->send_list) ||
 396                 !llist_empty(&queue->req_list);
 397 }
 398
 399 static inline bool nvme_tcp_queue_more(struct nvme_tcp_queue *queue)
 400 {
 401         return !nvme_tcp_queue_tls(queue) &&
 402                 nvme_tcp_queue_has_pending(queue);
 403 }
 404
 405 static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req,
 406                 bool last)
 407 {
 408         struct nvme_tcp_queue *queue = req->queue;
 409         bool empty;
 410
 411         empty = llist_add(&req->lentry, &queue->req_list) &&
 412                 list_empty(&queue->send_list) && !queue->request;
 413
 414         /*
 415          * if we're the first on the send_list and we can try to send
 416          * directly, otherwise queue io_work. Also, only do that if we
 417          * are on the same cpu, so we don't introduce contention.
 418          */
 419         if (queue->io_cpu == raw_smp_processor_id() &&
 420             empty && mutex_trylock(&queue->send_mutex)) {
 421                 nvme_tcp_send_all(queue);
 422                 mutex_unlock(&queue->send_mutex);
 423         }
 424
 425         if (last && nvme_tcp_queue_has_pending(queue))
 426                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 427 }
 428
 429 static void nvme_tcp_process_req_list(struct nvme_tcp_queue *queue)
 430 {
 431         struct nvme_tcp_request *req;
 432         struct llist_node *node;
 433
 434         for (node = llist_del_all(&queue->req_list); node; node = node->next) {
 435                 req = llist_entry(node, struct nvme_tcp_request, lentry);
 436                 list_add(&req->entry, &queue->send_list);
 437         }
 438 }
 439
 440 static inline struct nvme_tcp_request *
 441 nvme_tcp_fetch_request(struct nvme_tcp_queue *queue)
 442 {
 443         struct nvme_tcp_request *req;
 444
 445         req = list_first_entry_or_null(&queue->send_list,
 446                         struct nvme_tcp_request, entry);
 447         if (!req) {
 448                 nvme_tcp_process_req_list(queue);
 449                 req = list_first_entry_or_null(&queue->send_list,
 450                                 struct nvme_tcp_request, entry);
 451                 if (unlikely(!req))
 452                         return NULL;
 453         }
 454
 455         list_del_init(&req->entry);
 456         init_llist_node(&req->lentry);
 457         return req;
 458 }
 459
 460 #define NVME_TCP_CRC_SEED (~0)
 461
 462 static inline void nvme_tcp_ddgst_update(u32 *crcp,
 463                 struct page *page, size_t off, size_t len)
 464 {
 465         page += off / PAGE_SIZE;
 466         off %= PAGE_SIZE;
 467         while (len) {
 468                 const void *vaddr = kmap_local_page(page);
 469                 size_t n = min(len, (size_t)PAGE_SIZE - off);
 470
 471                 *crcp = crc32c(*crcp, vaddr + off, n);
 472                 kunmap_local(vaddr);
 473                 page++;
 474                 off = 0;
 475                 len -= n;
 476         }
 477 }
 478
 479 static inline __le32 nvme_tcp_ddgst_final(u32 crc)
 480 {
 481         return cpu_to_le32(~crc);
 482 }
 483
 484 static inline __le32 nvme_tcp_hdgst(const void *pdu, size_t len)
 485 {
 486         return cpu_to_le32(~crc32c(NVME_TCP_CRC_SEED, pdu, len));
 487 }
 488
 489 static inline void nvme_tcp_set_hdgst(void *pdu, size_t len)
 490 {
 491         *(__le32 *)(pdu + len) = nvme_tcp_hdgst(pdu, len);
 492 }
 493
 494 static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue,
 495                 void *pdu, size_t pdu_len)
 496 {
 497         struct nvme_tcp_hdr *hdr = pdu;
 498         __le32 recv_digest;
 499         __le32 exp_digest;
 500
 501         if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
 502                 dev_err(queue->ctrl->ctrl.device,
 503                         "queue %d: header digest flag is cleared\n",
 504                         nvme_tcp_queue_id(queue));
 505                 return -EPROTO;
 506         }
 507
 508         recv_digest = *(__le32 *)(pdu + hdr->hlen);
 509         exp_digest = nvme_tcp_hdgst(pdu, pdu_len);
 510         if (recv_digest != exp_digest) {
 511                 dev_err(queue->ctrl->ctrl.device,
 512                         "header digest error: recv %#x expected %#x\n",
 513                         le32_to_cpu(recv_digest), le32_to_cpu(exp_digest));
 514                 return -EIO;
 515         }
 516
 517         return 0;
 518 }
 519
 520 static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu)
 521 {
 522         struct nvme_tcp_hdr *hdr = pdu;
 523         u8 digest_len = nvme_tcp_hdgst_len(queue);
 524         u32 len;
 525
 526         len = le32_to_cpu(hdr->plen) - hdr->hlen -
 527                 ((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0);
 528
 529         if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
 530                 dev_err(queue->ctrl->ctrl.device,
 531                         "queue %d: data digest flag is cleared\n",
 532                 nvme_tcp_queue_id(queue));
 533                 return -EPROTO;
 534         }
 535         queue->rcv_crc = NVME_TCP_CRC_SEED;
 536
 537         return 0;
 538 }
 539
 540 static void nvme_tcp_exit_request(struct blk_mq_tag_set *set,
 541                 struct request *rq, unsigned int hctx_idx)
 542 {
 543         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 544
 545         page_frag_free(req->pdu);
 546 }
 547
 548 static int nvme_tcp_init_request(struct blk_mq_tag_set *set,
 549                 struct request *rq, unsigned int hctx_idx,
 550                 unsigned int numa_node)
 551 {
 552         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(set->driver_data);
 553         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 554         struct nvme_tcp_cmd_pdu *pdu;
 555         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 556         struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx];
 557         u8 hdgst = nvme_tcp_hdgst_len(queue);
 558
 559         req->pdu = page_frag_alloc(&queue->pf_cache,
 560                 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
 561                 GFP_KERNEL | __GFP_ZERO);
 562         if (!req->pdu)
 563                 return -ENOMEM;
 564
 565         pdu = req->pdu;
 566         req->queue = queue;
 567         nvme_req(rq)->ctrl = &ctrl->ctrl;
 568         nvme_req(rq)->cmd = &pdu->cmd;
 569         init_llist_node(&req->lentry);
 570         INIT_LIST_HEAD(&req->entry);
 571
 572         return 0;
 573 }
 574
 575 static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 576                 unsigned int hctx_idx)
 577 {
 578         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(data);
 579         struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1];
 580
 581         hctx->driver_data = queue;
 582         return 0;
 583 }
 584
 585 static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 586                 unsigned int hctx_idx)
 587 {
 588         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(data);
 589         struct nvme_tcp_queue *queue = &ctrl->queues[0];
 590
 591         hctx->driver_data = queue;
 592         return 0;
 593 }
 594
 595 static enum nvme_tcp_recv_state
 596 nvme_tcp_recv_state(struct nvme_tcp_queue *queue)
 597 {
 598         return  (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
 599                 (queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST :
 600                 NVME_TCP_RECV_DATA;
 601 }
 602
 603 static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue)
 604 {
 605         queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) +
 606                                 nvme_tcp_hdgst_len(queue);
 607         queue->pdu_offset = 0;
 608         queue->data_remaining = -1;
 609         queue->ddgst_remaining = 0;
 610 }
 611
 612 static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl)
 613 {
 614         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
 615                 return;
 616
 617         dev_warn(ctrl->device, "starting error recovery\n");
 618         queue_work(nvme_reset_wq, &to_tcp_ctrl(ctrl)->err_work);
 619 }
 620
 621 static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
 622                 struct nvme_completion *cqe)
 623 {
 624         struct nvme_tcp_request *req;
 625         struct request *rq;
 626
 627         rq = nvme_find_rq(nvme_tcp_tagset(queue), cqe->command_id);
 628         if (!rq) {
 629                 dev_err(queue->ctrl->ctrl.device,
 630                         "got bad cqe.command_id %#x on queue %d\n",
 631                         cqe->command_id, nvme_tcp_queue_id(queue));
 632                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 633                 return -EINVAL;
 634         }
 635
 636         req = blk_mq_rq_to_pdu(rq);
 637         if (req->status == cpu_to_le16(NVME_SC_SUCCESS))
 638                 req->status = cqe->status;
 639
 640         if (!nvme_try_complete_req(rq, req->status, cqe->result))
 641                 nvme_complete_rq(rq);
 642         queue->nr_cqe++;
 643
 644         return 0;
 645 }
 646
 647 static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue,
 648                 struct nvme_tcp_data_pdu *pdu)
 649 {
 650         struct request *rq;
 651
 652         rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
 653         if (!rq) {
 654                 dev_err(queue->ctrl->ctrl.device,
 655                         "got bad c2hdata.command_id %#x on queue %d\n",
 656                         pdu->command_id, nvme_tcp_queue_id(queue));
 657                 return -ENOENT;
 658         }
 659
 660         if (!blk_rq_payload_bytes(rq)) {
 661                 dev_err(queue->ctrl->ctrl.device,
 662                         "queue %d tag %#x unexpected data\n",
 663                         nvme_tcp_queue_id(queue), rq->tag);
 664                 return -EIO;
 665         }
 666
 667         queue->data_remaining = le32_to_cpu(pdu->data_length);
 668
 669         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS &&
 670             unlikely(!(pdu->hdr.flags & NVME_TCP_F_DATA_LAST))) {
 671                 dev_err(queue->ctrl->ctrl.device,
 672                         "queue %d tag %#x SUCCESS set but not last PDU\n",
 673                         nvme_tcp_queue_id(queue), rq->tag);
 674                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 675                 return -EPROTO;
 676         }
 677
 678         return 0;
 679 }
 680
 681 static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue,
 682                 struct nvme_tcp_rsp_pdu *pdu)
 683 {
 684         struct nvme_completion *cqe = &pdu->cqe;
 685         int ret = 0;
 686
 687         /*
 688          * AEN requests are special as they don't time out and can
 689          * survive any kind of queue freeze and often don't respond to
 690          * aborts.  We don't even bother to allocate a struct request
 691          * for them but rather special case them here.
 692          */
 693         if (unlikely(nvme_is_aen_req(nvme_tcp_queue_id(queue),
 694                                      cqe->command_id)))
 695                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
 696                                 &cqe->result);
 697         else
 698                 ret = nvme_tcp_process_nvme_cqe(queue, cqe);
 699
 700         return ret;
 701 }
 702
 703 static void nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req)
 704 {
 705         struct nvme_tcp_data_pdu *data = nvme_tcp_req_data_pdu(req);
 706         struct nvme_tcp_queue *queue = req->queue;
 707         struct request *rq = blk_mq_rq_from_pdu(req);
 708         u32 h2cdata_sent = req->pdu_len;
 709         u8 hdgst = nvme_tcp_hdgst_len(queue);
 710         u8 ddgst = nvme_tcp_ddgst_len(queue);
 711
 712         req->state = NVME_TCP_SEND_H2C_PDU;
 713         req->offset = 0;
 714         req->pdu_len = min(req->h2cdata_left, queue->maxh2cdata);
 715         req->pdu_sent = 0;
 716         req->h2cdata_left -= req->pdu_len;
 717         req->h2cdata_offset += h2cdata_sent;
 718
 719         memset(data, 0, sizeof(*data));
 720         data->hdr.type = nvme_tcp_h2c_data;
 721         if (!req->h2cdata_left)
 722                 data->hdr.flags = NVME_TCP_F_DATA_LAST;
 723         if (queue->hdr_digest)
 724                 data->hdr.flags |= NVME_TCP_F_HDGST;
 725         if (queue->data_digest)
 726                 data->hdr.flags |= NVME_TCP_F_DDGST;
 727         data->hdr.hlen = sizeof(*data);
 728         data->hdr.pdo = data->hdr.hlen + hdgst;
 729         data->hdr.plen =
 730                 cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst);
 731         data->ttag = req->ttag;
 732         data->command_id = nvme_cid(rq);
 733         data->data_offset = cpu_to_le32(req->h2cdata_offset);
 734         data->data_length = cpu_to_le32(req->pdu_len);
 735 }
 736
 737 static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue,
 738                 struct nvme_tcp_r2t_pdu *pdu)
 739 {
 740         struct nvme_tcp_request *req;
 741         struct request *rq;
 742         u32 r2t_length = le32_to_cpu(pdu->r2t_length);
 743         u32 r2t_offset = le32_to_cpu(pdu->r2t_offset);
 744
 745         rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
 746         if (!rq) {
 747                 dev_err(queue->ctrl->ctrl.device,
 748                         "got bad r2t.command_id %#x on queue %d\n",
 749                         pdu->command_id, nvme_tcp_queue_id(queue));
 750                 return -ENOENT;
 751         }
 752         req = blk_mq_rq_to_pdu(rq);
 753
 754         if (unlikely(!r2t_length)) {
 755                 dev_err(queue->ctrl->ctrl.device,
 756                         "req %d r2t len is %u, probably a bug...\n",
 757                         rq->tag, r2t_length);
 758                 return -EPROTO;
 759         }
 760
 761         if (unlikely(req->data_sent + r2t_length > req->data_len)) {
 762                 dev_err(queue->ctrl->ctrl.device,
 763                         "req %d r2t len %u exceeded data len %u (%zu sent)\n",
 764                         rq->tag, r2t_length, req->data_len, req->data_sent);
 765                 return -EPROTO;
 766         }
 767
 768         if (unlikely(r2t_offset < req->data_sent)) {
 769                 dev_err(queue->ctrl->ctrl.device,
 770                         "req %d unexpected r2t offset %u (expected %zu)\n",
 771                         rq->tag, r2t_offset, req->data_sent);
 772                 return -EPROTO;
 773         }
 774
 775         if (llist_on_list(&req->lentry) ||
 776             !list_empty(&req->entry)) {
 777                 dev_err(queue->ctrl->ctrl.device,
 778                         "req %d unexpected r2t while processing request\n",
 779                         rq->tag);
 780                 return -EPROTO;
 781         }
 782
 783         req->pdu_len = 0;
 784         req->h2cdata_left = r2t_length;
 785         req->h2cdata_offset = r2t_offset;
 786         req->ttag = pdu->ttag;
 787
 788         nvme_tcp_setup_h2c_data_pdu(req);
 789
 790         llist_add(&req->lentry, &queue->req_list);
 791         queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 792
 793         return 0;
 794 }
 795
 796 static void nvme_tcp_handle_c2h_term(struct nvme_tcp_queue *queue,
 797                 struct nvme_tcp_term_pdu *pdu)
 798 {
 799         u16 fes;
 800         const char *msg;
 801         u32 plen = le32_to_cpu(pdu->hdr.plen);
 802
 803         static const char * const msg_table[] = {
 804                 [NVME_TCP_FES_INVALID_PDU_HDR] = "Invalid PDU Header Field",
 805                 [NVME_TCP_FES_PDU_SEQ_ERR] = "PDU Sequence Error",
 806                 [NVME_TCP_FES_HDR_DIGEST_ERR] = "Header Digest Error",
 807                 [NVME_TCP_FES_DATA_OUT_OF_RANGE] = "Data Transfer Out Of Range",
 808                 [NVME_TCP_FES_DATA_LIMIT_EXCEEDED] = "Data Transfer Limit Exceeded",
 809                 [NVME_TCP_FES_UNSUPPORTED_PARAM] = "Unsupported Parameter",
 810         };
 811
 812         if (plen < NVME_TCP_MIN_C2HTERM_PLEN ||
 813             plen > NVME_TCP_MAX_C2HTERM_PLEN) {
 814                 dev_err(queue->ctrl->ctrl.device,
 815                         "Received a malformed C2HTermReq PDU (plen = %u)\n",
 816                         plen);
 817                 return;
 818         }
 819
 820         fes = le16_to_cpu(pdu->fes);
 821         if (fes && fes < ARRAY_SIZE(msg_table))
 822                 msg = msg_table[fes];
 823         else
 824                 msg = "Unknown";
 825
 826         dev_err(queue->ctrl->ctrl.device,
 827                 "Received C2HTermReq (FES = %s)\n", msg);
 828 }
 829
 830 static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 831                 unsigned int *offset, size_t *len)
 832 {
 833         struct nvme_tcp_hdr *hdr;
 834         char *pdu = queue->pdu;
 835         size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
 836         int ret;
 837
 838         ret = skb_copy_bits(skb, *offset,
 839                 &pdu[queue->pdu_offset], rcv_len);
 840         if (unlikely(ret))
 841                 return ret;
 842
 843         queue->pdu_remaining -= rcv_len;
 844         queue->pdu_offset += rcv_len;
 845         *offset += rcv_len;
 846         *len -= rcv_len;
 847         if (queue->pdu_remaining)
 848                 return 0;
 849
 850         hdr = queue->pdu;
 851         if (unlikely(hdr->hlen != sizeof(struct nvme_tcp_rsp_pdu))) {
 852                 if (!nvme_tcp_recv_pdu_supported(hdr->type))
 853                         goto unsupported_pdu;
 854
 855                 dev_err(queue->ctrl->ctrl.device,
 856                         "pdu type %d has unexpected header length (%d)\n",
 857                         hdr->type, hdr->hlen);
 858                 return -EPROTO;
 859         }
 860
 861         if (unlikely(hdr->type == nvme_tcp_c2h_term)) {
 862                 /*
 863                  * C2HTermReq never includes Header or Data digests.
 864                  * Skip the checks.
 865                  */
 866                 nvme_tcp_handle_c2h_term(queue, (void *)queue->pdu);
 867                 return -EINVAL;
 868         }
 869
 870         if (queue->hdr_digest) {
 871                 ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen);
 872                 if (unlikely(ret))
 873                         return ret;
 874         }
 875
 876
 877         if (queue->data_digest) {
 878                 ret = nvme_tcp_check_ddgst(queue, queue->pdu);
 879                 if (unlikely(ret))
 880                         return ret;
 881         }
 882
 883         switch (hdr->type) {
 884         case nvme_tcp_c2h_data:
 885                 return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
 886         case nvme_tcp_rsp:
 887                 nvme_tcp_init_recv_ctx(queue);
 888                 return nvme_tcp_handle_comp(queue, (void *)queue->pdu);
 889         case nvme_tcp_r2t:
 890                 nvme_tcp_init_recv_ctx(queue);
 891                 return nvme_tcp_handle_r2t(queue, (void *)queue->pdu);
 892         default:
 893                 goto unsupported_pdu;
 894         }
 895
 896 unsupported_pdu:
 897         dev_err(queue->ctrl->ctrl.device,
 898                 "unsupported pdu type (%d)\n", hdr->type);
 899         return -EINVAL;
 900 }
 901
 902 static inline void nvme_tcp_end_request(struct request *rq, u16 status)
 903 {
 904         union nvme_result res = {};
 905
 906         if (!nvme_try_complete_req(rq, cpu_to_le16(status << 1), res))
 907                 nvme_complete_rq(rq);
 908 }
 909
 910 static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 911                               unsigned int *offset, size_t *len)
 912 {
 913         struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
 914         struct request *rq =
 915                 nvme_cid_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
 916         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 917
 918         while (true) {
 919                 int recv_len, ret;
 920
 921                 recv_len = min_t(size_t, *len, queue->data_remaining);
 922                 if (!recv_len)
 923                         break;
 924
 925                 if (!iov_iter_count(&req->iter)) {
 926                         req->curr_bio = req->curr_bio->bi_next;
 927
 928                         /*
 929                          * If we don`t have any bios it means that controller
 930                          * sent more data than we requested, hence error
 931                          */
 932                         if (!req->curr_bio) {
 933                                 dev_err(queue->ctrl->ctrl.device,
 934                                         "queue %d no space in request %#x",
 935                                         nvme_tcp_queue_id(queue), rq->tag);
 936                                 nvme_tcp_init_recv_ctx(queue);
 937                                 return -EIO;
 938                         }
 939                         nvme_tcp_init_iter(req, ITER_DEST);
 940                 }
 941
 942                 /* we can read only from what is left in this bio */
 943                 recv_len = min_t(size_t, recv_len,
 944                                 iov_iter_count(&req->iter));
 945
 946                 if (queue->data_digest)
 947                         ret = skb_copy_and_crc32c_datagram_iter(skb, *offset,
 948                                 &req->iter, recv_len, &queue->rcv_crc);
 949                 else
 950                         ret = skb_copy_datagram_iter(skb, *offset,
 951                                         &req->iter, recv_len);
 952                 if (ret) {
 953                         dev_err(queue->ctrl->ctrl.device,
 954                                 "queue %d failed to copy request %#x data",
 955                                 nvme_tcp_queue_id(queue), rq->tag);
 956                         return ret;
 957                 }
 958
 959                 *len -= recv_len;
 960                 *offset += recv_len;
 961                 queue->data_remaining -= recv_len;
 962         }
 963
 964         if (!queue->data_remaining) {
 965                 if (queue->data_digest) {
 966                         queue->exp_ddgst = nvme_tcp_ddgst_final(queue->rcv_crc);
 967                         queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
 968                 } else {
 969                         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
 970                                 nvme_tcp_end_request(rq,
 971                                                 le16_to_cpu(req->status));
 972                                 queue->nr_cqe++;
 973                         }
 974                         nvme_tcp_init_recv_ctx(queue);
 975                 }
 976         }
 977
 978         return 0;
 979 }
 980
 981 static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
 982                 struct sk_buff *skb, unsigned int *offset, size_t *len)
 983 {
 984         struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
 985         char *ddgst = (char *)&queue->recv_ddgst;
 986         size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
 987         off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
 988         int ret;
 989
 990         ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
 991         if (unlikely(ret))
 992                 return ret;
 993
 994         queue->ddgst_remaining -= recv_len;
 995         *offset += recv_len;
 996         *len -= recv_len;
 997         if (queue->ddgst_remaining)
 998                 return 0;
 999
1000         if (queue->recv_ddgst != queue->exp_ddgst) {
1001                 struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
1002                                         pdu->command_id);
1003                 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
1004
1005                 req->status = cpu_to_le16(NVME_SC_DATA_XFER_ERROR);
1006
1007                 dev_err(queue->ctrl->ctrl.device,
1008                         "data digest error: recv %#x expected %#x\n",
1009                         le32_to_cpu(queue->recv_ddgst),
1010                         le32_to_cpu(queue->exp_ddgst));
1011         }
1012
1013         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
1014                 struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
1015                                         pdu->command_id);
1016                 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
1017
1018                 nvme_tcp_end_request(rq, le16_to_cpu(req->status));
1019                 queue->nr_cqe++;
1020         }
1021
1022         nvme_tcp_init_recv_ctx(queue);
1023         return 0;
1024 }
1025
1026 static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
1027                              unsigned int offset, size_t len)
1028 {
1029         struct nvme_tcp_queue *queue = desc->arg.data;
1030         size_t consumed = len;
1031         int result;
1032
1033         if (unlikely(!queue->rd_enabled))
1034                 return -EFAULT;
1035
1036         while (len) {
1037                 switch (nvme_tcp_recv_state(queue)) {
1038                 case NVME_TCP_RECV_PDU:
1039                         result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
1040                         break;
1041                 case NVME_TCP_RECV_DATA:
1042                         result = nvme_tcp_recv_data(queue, skb, &offset, &len);
1043                         break;
1044                 case NVME_TCP_RECV_DDGST:
1045                         result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
1046                         break;
1047                 default:
1048                         result = -EFAULT;
1049                 }
1050                 if (result) {
1051                         dev_err(queue->ctrl->ctrl.device,
1052                                 "receive failed:  %d\n", result);
1053                         queue->rd_enabled = false;
1054                         nvme_tcp_error_recovery(&queue->ctrl->ctrl);
1055                         return result;
1056                 }
1057         }
1058
1059         return consumed;
1060 }
1061
1062 static void nvme_tcp_data_ready(struct sock *sk)
1063 {
1064         struct nvme_tcp_queue *queue;
1065
1066         trace_sk_data_ready(sk);
1067
1068         read_lock_bh(&sk->sk_callback_lock);
1069         queue = sk->sk_user_data;
1070         if (likely(queue && queue->rd_enabled) &&
1071             !test_bit(NVME_TCP_Q_POLLING, &queue->flags))
1072                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1073         read_unlock_bh(&sk->sk_callback_lock);
1074 }
1075
1076 static void nvme_tcp_write_space(struct sock *sk)
1077 {
1078         struct nvme_tcp_queue *queue;
1079
1080         read_lock_bh(&sk->sk_callback_lock);
1081         queue = sk->sk_user_data;
1082         if (likely(queue && sk_stream_is_writeable(sk))) {
1083                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1084                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1085         }
1086         read_unlock_bh(&sk->sk_callback_lock);
1087 }
1088
1089 static void nvme_tcp_state_change(struct sock *sk)
1090 {
1091         struct nvme_tcp_queue *queue;
1092
1093         read_lock_bh(&sk->sk_callback_lock);
1094         queue = sk->sk_user_data;
1095         if (!queue)
1096                 goto done;
1097
1098         switch (sk->sk_state) {
1099         case TCP_CLOSE:
1100         case TCP_CLOSE_WAIT:
1101         case TCP_LAST_ACK:
1102         case TCP_FIN_WAIT1:
1103         case TCP_FIN_WAIT2:
1104                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
1105                 break;
1106         default:
1107                 dev_info(queue->ctrl->ctrl.device,
1108                         "queue %d socket state %d\n",
1109                         nvme_tcp_queue_id(queue), sk->sk_state);
1110         }
1111
1112         queue->state_change(sk);
1113 done:
1114         read_unlock_bh(&sk->sk_callback_lock);
1115 }
1116
1117 static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
1118 {
1119         queue->request = NULL;
1120 }
1121
1122 static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
1123 {
1124         if (nvme_tcp_async_req(req)) {
1125                 union nvme_result res = {};
1126
1127                 nvme_complete_async_event(&req->queue->ctrl->ctrl,
1128                                 cpu_to_le16(NVME_SC_HOST_PATH_ERROR), &res);
1129         } else {
1130                 nvme_tcp_end_request(blk_mq_rq_from_pdu(req),
1131                                 NVME_SC_HOST_PATH_ERROR);
1132         }
1133 }
1134
1135 static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
1136 {
1137         struct nvme_tcp_queue *queue = req->queue;
1138         int req_data_len = req->data_len;
1139         u32 h2cdata_left = req->h2cdata_left;
1140
1141         while (true) {
1142                 struct bio_vec bvec;
1143                 struct msghdr msg = {
1144                         .msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES,
1145                 };
1146                 struct page *page = nvme_tcp_req_cur_page(req);
1147                 size_t offset = nvme_tcp_req_cur_offset(req);
1148                 size_t len = nvme_tcp_req_cur_length(req);
1149                 bool last = nvme_tcp_pdu_last_send(req, len);
1150                 int req_data_sent = req->data_sent;
1151                 int ret;
1152
1153                 if (last && !queue->data_digest && !nvme_tcp_queue_more(queue))
1154                         msg.msg_flags |= MSG_EOR;
1155                 else
1156                         msg.msg_flags |= MSG_MORE;
1157
1158                 if (!sendpages_ok(page, len, offset))
1159                         msg.msg_flags &= ~MSG_SPLICE_PAGES;
1160
1161                 bvec_set_page(&bvec, page, len, offset);
1162                 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, len);
1163                 ret = sock_sendmsg(queue->sock, &msg);
1164                 if (ret <= 0)
1165                         return ret;
1166
1167                 if (queue->data_digest)
1168                         nvme_tcp_ddgst_update(&queue->snd_crc, page,
1169                                         offset, ret);
1170
1171                 /*
1172                  * update the request iterator except for the last payload send
1173                  * in the request where we don't want to modify it as we may
1174                  * compete with the RX path completing the request.
1175                  */
1176                 if (req_data_sent + ret < req_data_len)
1177                         nvme_tcp_advance_req(req, ret);
1178
1179                 /* fully successful last send in current PDU */
1180                 if (last && ret == len) {
1181                         if (queue->data_digest) {
1182                                 req->ddgst =
1183                                         nvme_tcp_ddgst_final(queue->snd_crc);
1184                                 req->state = NVME_TCP_SEND_DDGST;
1185                                 req->offset = 0;
1186                         } else {
1187                                 if (h2cdata_left)
1188                                         nvme_tcp_setup_h2c_data_pdu(req);
1189                                 else
1190                                         nvme_tcp_done_send_req(queue);
1191                         }
1192                         return 1;
1193                 }
1194         }
1195         return -EAGAIN;
1196 }
1197
1198 static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
1199 {
1200         struct nvme_tcp_queue *queue = req->queue;
1201         struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
1202         struct bio_vec bvec;
1203         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES, };
1204         bool inline_data = nvme_tcp_has_inline_data(req);
1205         u8 hdgst = nvme_tcp_hdgst_len(queue);
1206         int len = sizeof(*pdu) + hdgst - req->offset;
1207         int ret;
1208
1209         if (inline_data || nvme_tcp_queue_more(queue))
1210                 msg.msg_flags |= MSG_MORE;
1211         else
1212                 msg.msg_flags |= MSG_EOR;
1213
1214         if (queue->hdr_digest && !req->offset)
1215                 nvme_tcp_set_hdgst(pdu, sizeof(*pdu));
1216
1217         bvec_set_virt(&bvec, (void *)pdu + req->offset, len);
1218         iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, len);
1219         ret = sock_sendmsg(queue->sock, &msg);
1220         if (unlikely(ret <= 0))
1221                 return ret;
1222
1223         len -= ret;
1224         if (!len) {
1225                 if (inline_data) {
1226                         req->state = NVME_TCP_SEND_DATA;
1227                         if (queue->data_digest)
1228                                 queue->snd_crc = NVME_TCP_CRC_SEED;
1229                 } else {
1230                         nvme_tcp_done_send_req(queue);
1231                 }
1232                 return 1;
1233         }
1234         req->offset += ret;
1235
1236         return -EAGAIN;
1237 }
1238
1239 static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
1240 {
1241         struct nvme_tcp_queue *queue = req->queue;
1242         struct nvme_tcp_data_pdu *pdu = nvme_tcp_req_data_pdu(req);
1243         struct bio_vec bvec;
1244         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_MORE, };
1245         u8 hdgst = nvme_tcp_hdgst_len(queue);
1246         int len = sizeof(*pdu) - req->offset + hdgst;
1247         int ret;
1248
1249         if (queue->hdr_digest && !req->offset)
1250                 nvme_tcp_set_hdgst(pdu, sizeof(*pdu));
1251
1252         if (!req->h2cdata_left)
1253                 msg.msg_flags |= MSG_SPLICE_PAGES;
1254
1255         bvec_set_virt(&bvec, (void *)pdu + req->offset, len);
1256         iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, len);
1257         ret = sock_sendmsg(queue->sock, &msg);
1258         if (unlikely(ret <= 0))
1259                 return ret;
1260
1261         len -= ret;
1262         if (!len) {
1263                 req->state = NVME_TCP_SEND_DATA;
1264                 if (queue->data_digest)
1265                         queue->snd_crc = NVME_TCP_CRC_SEED;
1266                 return 1;
1267         }
1268         req->offset += ret;
1269
1270         return -EAGAIN;
1271 }
1272
1273 static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
1274 {
1275         struct nvme_tcp_queue *queue = req->queue;
1276         size_t offset = req->offset;
1277         u32 h2cdata_left = req->h2cdata_left;
1278         int ret;
1279         struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1280         struct kvec iov = {
1281                 .iov_base = (u8 *)&req->ddgst + req->offset,
1282                 .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
1283         };
1284
1285         if (nvme_tcp_queue_more(queue))
1286                 msg.msg_flags |= MSG_MORE;
1287         else
1288                 msg.msg_flags |= MSG_EOR;
1289
1290         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1291         if (unlikely(ret <= 0))
1292                 return ret;
1293
1294         if (offset + ret == NVME_TCP_DIGEST_LENGTH) {
1295                 if (h2cdata_left)
1296                         nvme_tcp_setup_h2c_data_pdu(req);
1297                 else
1298                         nvme_tcp_done_send_req(queue);
1299                 return 1;
1300         }
1301
1302         req->offset += ret;
1303         return -EAGAIN;
1304 }
1305
1306 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
1307 {
1308         struct nvme_tcp_request *req;
1309         unsigned int noreclaim_flag;
1310         int ret = 1;
1311
1312         if (!queue->request) {
1313                 queue->request = nvme_tcp_fetch_request(queue);
1314                 if (!queue->request)
1315                         return 0;
1316         }
1317         req = queue->request;
1318
1319         noreclaim_flag = memalloc_noreclaim_save();
1320         if (req->state == NVME_TCP_SEND_CMD_PDU) {
1321                 ret = nvme_tcp_try_send_cmd_pdu(req);
1322                 if (ret <= 0)
1323                         goto done;
1324                 if (!nvme_tcp_has_inline_data(req))
1325                         goto out;
1326         }
1327
1328         if (req->state == NVME_TCP_SEND_H2C_PDU) {
1329                 ret = nvme_tcp_try_send_data_pdu(req);
1330                 if (ret <= 0)
1331                         goto done;
1332         }
1333
1334         if (req->state == NVME_TCP_SEND_DATA) {
1335                 ret = nvme_tcp_try_send_data(req);
1336                 if (ret <= 0)
1337                         goto done;
1338         }
1339
1340         if (req->state == NVME_TCP_SEND_DDGST)
1341                 ret = nvme_tcp_try_send_ddgst(req);
1342 done:
1343         if (ret == -EAGAIN) {
1344                 ret = 0;
1345         } else if (ret < 0) {
1346                 dev_err(queue->ctrl->ctrl.device,
1347                         "failed to send request %d\n", ret);
1348                 nvme_tcp_fail_request(queue->request);
1349                 nvme_tcp_done_send_req(queue);
1350         }
1351 out:
1352         memalloc_noreclaim_restore(noreclaim_flag);
1353         return ret;
1354 }
1355
1356 static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1357 {
1358         struct socket *sock = queue->sock;
1359         struct sock *sk = sock->sk;
1360         read_descriptor_t rd_desc;
1361         int consumed;
1362
1363         rd_desc.arg.data = queue;
1364         rd_desc.count = 1;
1365         lock_sock(sk);
1366         queue->nr_cqe = 0;
1367         consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1368         release_sock(sk);
1369         return consumed == -EAGAIN ? 0 : consumed;
1370 }
1371
1372 static void nvme_tcp_io_work(struct work_struct *w)
1373 {
1374         struct nvme_tcp_queue *queue =
1375                 container_of(w, struct nvme_tcp_queue, io_work);
1376         unsigned long deadline = jiffies + msecs_to_jiffies(1);
1377
1378         do {
1379                 bool pending = false;
1380                 int result;
1381
1382                 if (mutex_trylock(&queue->send_mutex)) {
1383                         result = nvme_tcp_try_send(queue);
1384                         mutex_unlock(&queue->send_mutex);
1385                         if (result > 0)
1386                                 pending = true;
1387                         else if (unlikely(result < 0))
1388                                 break;
1389                 }
1390
1391                 result = nvme_tcp_try_recv(queue);
1392                 if (result > 0)
1393                         pending = true;
1394                 else if (unlikely(result < 0))
1395                         return;
1396
1397                 /* did we get some space after spending time in recv? */
1398                 if (nvme_tcp_queue_has_pending(queue) &&
1399                     sk_stream_is_writeable(queue->sock->sk))
1400                         pending = true;
1401
1402                 if (!pending || !queue->rd_enabled)
1403                         return;
1404
1405         } while (!time_after(jiffies, deadline)); /* quota is exhausted */
1406
1407         queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1408 }
1409
1410 static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1411 {
1412         struct nvme_tcp_request *async = &ctrl->async_req;
1413
1414         page_frag_free(async->pdu);
1415 }
1416
1417 static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1418 {
1419         struct nvme_tcp_queue *queue = &ctrl->queues[0];
1420         struct nvme_tcp_request *async = &ctrl->async_req;
1421         u8 hdgst = nvme_tcp_hdgst_len(queue);
1422
1423         async->pdu = page_frag_alloc(&queue->pf_cache,
1424                 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1425                 GFP_KERNEL | __GFP_ZERO);
1426         if (!async->pdu)
1427                 return -ENOMEM;
1428
1429         async->queue = &ctrl->queues[0];
1430         return 0;
1431 }
1432
1433 static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1434 {
1435         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1436         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1437         unsigned int noreclaim_flag;
1438
1439         if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1440                 return;
1441
1442         page_frag_cache_drain(&queue->pf_cache);
1443
1444         noreclaim_flag = memalloc_noreclaim_save();
1445         /* ->sock will be released by fput() */
1446         fput(queue->sock->file);
1447         queue->sock = NULL;
1448         memalloc_noreclaim_restore(noreclaim_flag);
1449
1450         kfree(queue->pdu);
1451         mutex_destroy(&queue->send_mutex);
1452         mutex_destroy(&queue->queue_lock);
1453 }
1454
1455 static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1456 {
1457         struct nvme_tcp_icreq_pdu *icreq;
1458         struct nvme_tcp_icresp_pdu *icresp;
1459         char cbuf[CMSG_LEN(sizeof(char))] = {};
1460         u8 ctype;
1461         struct msghdr msg = {};
1462         struct kvec iov;
1463         bool ctrl_hdgst, ctrl_ddgst;
1464         u32 maxh2cdata;
1465         int ret;
1466
1467         icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1468         if (!icreq)
1469                 return -ENOMEM;
1470
1471         icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1472         if (!icresp) {
1473                 ret = -ENOMEM;
1474                 goto free_icreq;
1475         }
1476
1477         icreq->hdr.type = nvme_tcp_icreq;
1478         icreq->hdr.hlen = sizeof(*icreq);
1479         icreq->hdr.pdo = 0;
1480         icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1481         icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1482         icreq->maxr2t = 0; /* single inflight r2t supported */
1483         icreq->hpda = 0; /* no alignment constraint */
1484         if (queue->hdr_digest)
1485                 icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1486         if (queue->data_digest)
1487                 icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1488
1489         iov.iov_base = icreq;
1490         iov.iov_len = sizeof(*icreq);
1491         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1492         if (ret < 0) {
1493                 pr_warn("queue %d: failed to send icreq, error %d\n",
1494                         nvme_tcp_queue_id(queue), ret);
1495                 goto free_icresp;
1496         }
1497
1498         memset(&msg, 0, sizeof(msg));
1499         iov.iov_base = icresp;
1500         iov.iov_len = sizeof(*icresp);
1501         if (nvme_tcp_queue_tls(queue)) {
1502                 msg.msg_control = cbuf;
1503                 msg.msg_controllen = sizeof(cbuf);
1504         }
1505         msg.msg_flags = MSG_WAITALL;
1506         ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1507                         iov.iov_len, msg.msg_flags);
1508         if (ret >= 0 && ret < sizeof(*icresp))
1509                 ret = -ECONNRESET;
1510         if (ret < 0) {
1511                 pr_warn("queue %d: failed to receive icresp, error %d\n",
1512                         nvme_tcp_queue_id(queue), ret);
1513                 goto free_icresp;
1514         }
1515         ret = -ENOTCONN;
1516         if (nvme_tcp_queue_tls(queue)) {
1517                 ctype = tls_get_record_type(queue->sock->sk,
1518                                             (struct cmsghdr *)cbuf);
1519                 if (ctype != TLS_RECORD_TYPE_DATA) {
1520                         pr_err("queue %d: unhandled TLS record %d\n",
1521                                nvme_tcp_queue_id(queue), ctype);
1522                         goto free_icresp;
1523                 }
1524         }
1525         ret = -EINVAL;
1526         if (icresp->hdr.type != nvme_tcp_icresp) {
1527                 pr_err("queue %d: bad type returned %d\n",
1528                         nvme_tcp_queue_id(queue), icresp->hdr.type);
1529                 goto free_icresp;
1530         }
1531
1532         if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1533                 pr_err("queue %d: bad pdu length returned %d\n",
1534                         nvme_tcp_queue_id(queue), icresp->hdr.plen);
1535                 goto free_icresp;
1536         }
1537
1538         if (icresp->pfv != NVME_TCP_PFV_1_0) {
1539                 pr_err("queue %d: bad pfv returned %d\n",
1540                         nvme_tcp_queue_id(queue), icresp->pfv);
1541                 goto free_icresp;
1542         }
1543
1544         ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1545         if ((queue->data_digest && !ctrl_ddgst) ||
1546             (!queue->data_digest && ctrl_ddgst)) {
1547                 pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1548                         nvme_tcp_queue_id(queue),
1549                         queue->data_digest ? "enabled" : "disabled",
1550                         ctrl_ddgst ? "enabled" : "disabled");
1551                 goto free_icresp;
1552         }
1553
1554         ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1555         if ((queue->hdr_digest && !ctrl_hdgst) ||
1556             (!queue->hdr_digest && ctrl_hdgst)) {
1557                 pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1558                         nvme_tcp_queue_id(queue),
1559                         queue->hdr_digest ? "enabled" : "disabled",
1560                         ctrl_hdgst ? "enabled" : "disabled");
1561                 goto free_icresp;
1562         }
1563
1564         if (icresp->cpda != 0) {
1565                 pr_err("queue %d: unsupported cpda returned %d\n",
1566                         nvme_tcp_queue_id(queue), icresp->cpda);
1567                 goto free_icresp;
1568         }
1569
1570         maxh2cdata = le32_to_cpu(icresp->maxdata);
1571         if ((maxh2cdata % 4) || (maxh2cdata < NVME_TCP_MIN_MAXH2CDATA)) {
1572                 pr_err("queue %d: invalid maxh2cdata returned %u\n",
1573                        nvme_tcp_queue_id(queue), maxh2cdata);
1574                 goto free_icresp;
1575         }
1576         queue->maxh2cdata = maxh2cdata;
1577
1578         ret = 0;
1579 free_icresp:
1580         kfree(icresp);
1581 free_icreq:
1582         kfree(icreq);
1583         return ret;
1584 }
1585
1586 static bool nvme_tcp_admin_queue(struct nvme_tcp_queue *queue)
1587 {
1588         return nvme_tcp_queue_id(queue) == 0;
1589 }
1590
1591 static bool nvme_tcp_default_queue(struct nvme_tcp_queue *queue)
1592 {
1593         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1594         int qid = nvme_tcp_queue_id(queue);
1595
1596         return !nvme_tcp_admin_queue(queue) &&
1597                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT];
1598 }
1599
1600 static bool nvme_tcp_read_queue(struct nvme_tcp_queue *queue)
1601 {
1602         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1603         int qid = nvme_tcp_queue_id(queue);
1604
1605         return !nvme_tcp_admin_queue(queue) &&
1606                 !nvme_tcp_default_queue(queue) &&
1607                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1608                           ctrl->io_queues[HCTX_TYPE_READ];
1609 }
1610
1611 static bool nvme_tcp_poll_queue(struct nvme_tcp_queue *queue)
1612 {
1613         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1614         int qid = nvme_tcp_queue_id(queue);
1615
1616         return !nvme_tcp_admin_queue(queue) &&
1617                 !nvme_tcp_default_queue(queue) &&
1618                 !nvme_tcp_read_queue(queue) &&
1619                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1620                           ctrl->io_queues[HCTX_TYPE_READ] +
1621                           ctrl->io_queues[HCTX_TYPE_POLL];
1622 }
1623
1624 /*
1625  * Track the number of queues assigned to each cpu using a global per-cpu
1626  * counter and select the least used cpu from the mq_map. Our goal is to spread
1627  * different controllers I/O threads across different cpu cores.
1628  *
1629  * Note that the accounting is not 100% perfect, but we don't need to be, we're
1630  * simply putting our best effort to select the best candidate cpu core that we
1631  * find at any given point.
1632  */
1633 static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
1634 {
1635         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1636         struct blk_mq_tag_set *set = &ctrl->tag_set;
1637         int qid = nvme_tcp_queue_id(queue) - 1;
1638         unsigned int *mq_map = NULL;
1639         int cpu, min_queues = INT_MAX, io_cpu;
1640
1641         if (wq_unbound)
1642                 goto out;
1643
1644         if (nvme_tcp_default_queue(queue))
1645                 mq_map = set->map[HCTX_TYPE_DEFAULT].mq_map;
1646         else if (nvme_tcp_read_queue(queue))
1647                 mq_map = set->map[HCTX_TYPE_READ].mq_map;
1648         else if (nvme_tcp_poll_queue(queue))
1649                 mq_map = set->map[HCTX_TYPE_POLL].mq_map;
1650
1651         if (WARN_ON(!mq_map))
1652                 goto out;
1653
1654         /* Search for the least used cpu from the mq_map */
1655         io_cpu = WORK_CPU_UNBOUND;
1656         for_each_online_cpu(cpu) {
1657                 int num_queues = atomic_read(&nvme_tcp_cpu_queues[cpu]);
1658
1659                 if (mq_map[cpu] != qid)
1660                         continue;
1661                 if (num_queues < min_queues) {
1662                         io_cpu = cpu;
1663                         min_queues = num_queues;
1664                 }
1665         }
1666         if (io_cpu != WORK_CPU_UNBOUND) {
1667                 queue->io_cpu = io_cpu;
1668                 atomic_inc(&nvme_tcp_cpu_queues[io_cpu]);
1669                 set_bit(NVME_TCP_Q_IO_CPU_SET, &queue->flags);
1670         }
1671 out:
1672         dev_dbg(ctrl->ctrl.device, "queue %d: using cpu %d\n",
1673                 qid, queue->io_cpu);
1674 }
1675
1676 static void nvme_tcp_tls_done(void *data, int status, key_serial_t pskid)
1677 {
1678         struct nvme_tcp_queue *queue = data;
1679         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1680         int qid = nvme_tcp_queue_id(queue);
1681         struct key *tls_key;
1682
1683         dev_dbg(ctrl->ctrl.device, "queue %d: TLS handshake done, key %x, status %d\n",
1684                 qid, pskid, status);
1685
1686         if (status) {
1687                 queue->tls_err = -status;
1688                 goto out_complete;
1689         }
1690
1691         tls_key = nvme_tls_key_lookup(pskid);
1692         if (IS_ERR(tls_key)) {
1693                 dev_warn(ctrl->ctrl.device, "queue %d: Invalid key %x\n",
1694                          qid, pskid);
1695                 queue->tls_err = -ENOKEY;
1696         } else {
1697                 queue->tls_enabled = true;
1698                 if (qid == 0)
1699                         ctrl->ctrl.tls_pskid = key_serial(tls_key);
1700                 key_put(tls_key);
1701                 queue->tls_err = 0;
1702         }
1703
1704 out_complete:
1705         complete(&queue->tls_complete);
1706 }
1707
1708 static int nvme_tcp_start_tls(struct nvme_ctrl *nctrl,
1709                               struct nvme_tcp_queue *queue,
1710                               key_serial_t pskid)
1711 {
1712         int qid = nvme_tcp_queue_id(queue);
1713         int ret;
1714         struct tls_handshake_args args;
1715         unsigned long tmo = tls_handshake_timeout * HZ;
1716         key_serial_t keyring = nvme_keyring_id();
1717
1718         dev_dbg(nctrl->device, "queue %d: start TLS with key %x\n",
1719                 qid, pskid);
1720         memset(&args, 0, sizeof(args));
1721         args.ta_sock = queue->sock;
1722         args.ta_done = nvme_tcp_tls_done;
1723         args.ta_data = queue;
1724         args.ta_my_peerids[0] = pskid;
1725         args.ta_num_peerids = 1;
1726         if (nctrl->opts->keyring)
1727                 keyring = key_serial(nctrl->opts->keyring);
1728         args.ta_keyring = keyring;
1729         args.ta_timeout_ms = tls_handshake_timeout * 1000;
1730         queue->tls_err = -EOPNOTSUPP;
1731         init_completion(&queue->tls_complete);
1732         ret = tls_client_hello_psk(&args, GFP_KERNEL);
1733         if (ret) {
1734                 dev_err(nctrl->device, "queue %d: failed to start TLS: %d\n",
1735                         qid, ret);
1736                 return ret;
1737         }
1738         ret = wait_for_completion_interruptible_timeout(&queue->tls_complete, tmo);
1739         if (ret <= 0) {
1740                 if (ret == 0)
1741                         ret = -ETIMEDOUT;
1742
1743                 dev_err(nctrl->device,
1744                         "queue %d: TLS handshake failed, error %d\n",
1745                         qid, ret);
1746                 tls_handshake_cancel(queue->sock->sk);
1747         } else {
1748                 if (queue->tls_err) {
1749                         dev_err(nctrl->device,
1750                                 "queue %d: TLS handshake complete, error %d\n",
1751                                 qid, queue->tls_err);
1752                 } else {
1753                         dev_dbg(nctrl->device,
1754                                 "queue %d: TLS handshake complete\n", qid);
1755                 }
1756                 ret = queue->tls_err;
1757         }
1758         return ret;
1759 }
1760
1761 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl, int qid,
1762                                 key_serial_t pskid)
1763 {
1764         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1765         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1766         int ret, rcv_pdu_size;
1767         struct file *sock_file;
1768
1769         mutex_init(&queue->queue_lock);
1770         queue->ctrl = ctrl;
1771         init_llist_head(&queue->req_list);
1772         INIT_LIST_HEAD(&queue->send_list);
1773         mutex_init(&queue->send_mutex);
1774         INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1775
1776         if (qid > 0)
1777                 queue->cmnd_capsule_len = nctrl->ioccsz * 16;
1778         else
1779                 queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1780                                                 NVME_TCP_ADMIN_CCSZ;
1781
1782         ret = sock_create_kern(current->nsproxy->net_ns,
1783                         ctrl->addr.ss_family, SOCK_STREAM,
1784                         IPPROTO_TCP, &queue->sock);
1785         if (ret) {
1786                 dev_err(nctrl->device,
1787                         "failed to create socket: %d\n", ret);
1788                 goto err_destroy_mutex;
1789         }
1790
1791         sock_file = sock_alloc_file(queue->sock, O_CLOEXEC, NULL);
1792         if (IS_ERR(sock_file)) {
1793                 ret = PTR_ERR(sock_file);
1794                 goto err_destroy_mutex;
1795         }
1796
1797         sk_net_refcnt_upgrade(queue->sock->sk);
1798         nvme_tcp_reclassify_socket(queue->sock);
1799
1800         /* Single syn retry */
1801         tcp_sock_set_syncnt(queue->sock->sk, 1);
1802
1803         /* Set TCP no delay */
1804         tcp_sock_set_nodelay(queue->sock->sk);
1805
1806         /*
1807          * Cleanup whatever is sitting in the TCP transmit queue on socket
1808          * close. This is done to prevent stale data from being sent should
1809          * the network connection be restored before TCP times out.
1810          */
1811         sock_no_linger(queue->sock->sk);
1812
1813         if (so_priority > 0)
1814                 sock_set_priority(queue->sock->sk, so_priority);
1815
1816         /* Set socket type of service */
1817         if (nctrl->opts->tos >= 0)
1818                 ip_sock_set_tos(queue->sock->sk, nctrl->opts->tos);
1819
1820         /* Set 10 seconds timeout for icresp recvmsg */
1821         queue->sock->sk->sk_rcvtimeo = 10 * HZ;
1822
1823         queue->sock->sk->sk_allocation = GFP_ATOMIC;
1824         queue->sock->sk->sk_use_task_frag = false;
1825         queue->io_cpu = WORK_CPU_UNBOUND;
1826         queue->request = NULL;
1827         queue->data_remaining = 0;
1828         queue->ddgst_remaining = 0;
1829         queue->pdu_remaining = 0;
1830         queue->pdu_offset = 0;
1831         sk_set_memalloc(queue->sock->sk);
1832
1833         if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
1834                 ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1835                         sizeof(ctrl->src_addr));
1836                 if (ret) {
1837                         dev_err(nctrl->device,
1838                                 "failed to bind queue %d socket %d\n",
1839                                 qid, ret);
1840                         goto err_sock;
1841                 }
1842         }
1843
1844         if (nctrl->opts->mask & NVMF_OPT_HOST_IFACE) {
1845                 char *iface = nctrl->opts->host_iface;
1846                 sockptr_t optval = KERNEL_SOCKPTR(iface);
1847
1848                 ret = sock_setsockopt(queue->sock, SOL_SOCKET, SO_BINDTODEVICE,
1849                                       optval, strlen(iface));
1850                 if (ret) {
1851                         dev_err(nctrl->device,
1852                           "failed to bind to interface %s queue %d err %d\n",
1853                           iface, qid, ret);
1854                         goto err_sock;
1855                 }
1856         }
1857
1858         queue->hdr_digest = nctrl->opts->hdr_digest;
1859         queue->data_digest = nctrl->opts->data_digest;
1860
1861         rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1862                         nvme_tcp_hdgst_len(queue);
1863         queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1864         if (!queue->pdu) {
1865                 ret = -ENOMEM;
1866                 goto err_sock;
1867         }
1868
1869         dev_dbg(nctrl->device, "connecting queue %d\n",
1870                         nvme_tcp_queue_id(queue));
1871
1872         ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1873                 sizeof(ctrl->addr), 0);
1874         if (ret) {
1875                 dev_err(nctrl->device,
1876                         "failed to connect socket: %d\n", ret);
1877                 goto err_rcv_pdu;
1878         }
1879
1880         /* If PSKs are configured try to start TLS */
1881         if (nvme_tcp_tls_configured(nctrl) && pskid) {
1882                 ret = nvme_tcp_start_tls(nctrl, queue, pskid);
1883                 if (ret)
1884                         goto err_init_connect;
1885         }
1886
1887         ret = nvme_tcp_init_connection(queue);
1888         if (ret)
1889                 goto err_init_connect;
1890
1891         set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1892
1893         return 0;
1894
1895 err_init_connect:
1896         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1897 err_rcv_pdu:
1898         kfree(queue->pdu);
1899 err_sock:
1900         /* ->sock will be released by fput() */
1901         fput(queue->sock->file);
1902         queue->sock = NULL;
1903 err_destroy_mutex:
1904         mutex_destroy(&queue->send_mutex);
1905         mutex_destroy(&queue->queue_lock);
1906         return ret;
1907 }
1908
1909 static void nvme_tcp_restore_sock_ops(struct nvme_tcp_queue *queue)
1910 {
1911         struct socket *sock = queue->sock;
1912
1913         write_lock_bh(&sock->sk->sk_callback_lock);
1914         sock->sk->sk_user_data  = NULL;
1915         sock->sk->sk_data_ready = queue->data_ready;
1916         sock->sk->sk_state_change = queue->state_change;
1917         sock->sk->sk_write_space  = queue->write_space;
1918         write_unlock_bh(&sock->sk->sk_callback_lock);
1919 }
1920
1921 static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1922 {
1923         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1924         nvme_tcp_restore_sock_ops(queue);
1925         cancel_work_sync(&queue->io_work);
1926 }
1927
1928 static void nvme_tcp_stop_queue_nowait(struct nvme_ctrl *nctrl, int qid)
1929 {
1930         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1931         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1932
1933         if (!test_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1934                 return;
1935
1936         if (test_and_clear_bit(NVME_TCP_Q_IO_CPU_SET, &queue->flags))
1937                 atomic_dec(&nvme_tcp_cpu_queues[queue->io_cpu]);
1938
1939         mutex_lock(&queue->queue_lock);
1940         if (test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1941                 __nvme_tcp_stop_queue(queue);
1942         /* Stopping the queue will disable TLS */
1943         queue->tls_enabled = false;
1944         mutex_unlock(&queue->queue_lock);
1945 }
1946
1947 static void nvme_tcp_wait_queue(struct nvme_ctrl *nctrl, int qid)
1948 {
1949         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1950         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1951         int timeout = 100;
1952
1953         while (timeout > 0) {
1954                 if (!test_bit(NVME_TCP_Q_ALLOCATED, &queue->flags) ||
1955                     !sk_wmem_alloc_get(queue->sock->sk))
1956                         return;
1957                 msleep(2);
1958                 timeout -= 2;
1959         }
1960         dev_warn(nctrl->device,
1961                  "qid %d: timeout draining sock wmem allocation expired\n",
1962                  qid);
1963 }
1964
1965 static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1966 {
1967         nvme_tcp_stop_queue_nowait(nctrl, qid);
1968         nvme_tcp_wait_queue(nctrl, qid);
1969 }
1970
1971
1972 static void nvme_tcp_setup_sock_ops(struct nvme_tcp_queue *queue)
1973 {
1974         write_lock_bh(&queue->sock->sk->sk_callback_lock);
1975         queue->sock->sk->sk_user_data = queue;
1976         queue->state_change = queue->sock->sk->sk_state_change;
1977         queue->data_ready = queue->sock->sk->sk_data_ready;
1978         queue->write_space = queue->sock->sk->sk_write_space;
1979         queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1980         queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1981         queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1982 #ifdef CONFIG_NET_RX_BUSY_POLL
1983         queue->sock->sk->sk_ll_usec = 1;
1984 #endif
1985         write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1986 }
1987
1988 static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1989 {
1990         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1991         struct nvme_tcp_queue *queue = &ctrl->queues[idx];
1992         int ret;
1993
1994         queue->rd_enabled = true;
1995         nvme_tcp_init_recv_ctx(queue);
1996         nvme_tcp_setup_sock_ops(queue);
1997
1998         if (idx) {
1999                 nvme_tcp_set_queue_io_cpu(queue);
2000                 ret = nvmf_connect_io_queue(nctrl, idx);
2001         } else
2002                 ret = nvmf_connect_admin_queue(nctrl);
2003
2004         if (!ret) {
2005                 set_bit(NVME_TCP_Q_LIVE, &queue->flags);
2006         } else {
2007                 if (test_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
2008                         __nvme_tcp_stop_queue(queue);
2009                 dev_err(nctrl->device,
2010                         "failed to connect queue: %d ret=%d\n", idx, ret);
2011         }
2012         return ret;
2013 }
2014
2015 static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
2016 {
2017         if (to_tcp_ctrl(ctrl)->async_req.pdu) {
2018                 cancel_work_sync(&ctrl->async_event_work);
2019                 nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
2020                 to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
2021         }
2022
2023         nvme_tcp_free_queue(ctrl, 0);
2024 }
2025
2026 static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
2027 {
2028         int i;
2029
2030         for (i = 1; i < ctrl->queue_count; i++)
2031                 nvme_tcp_free_queue(ctrl, i);
2032 }
2033
2034 static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
2035 {
2036         int i;
2037
2038         for (i = 1; i < ctrl->queue_count; i++)
2039                 nvme_tcp_stop_queue_nowait(ctrl, i);
2040         for (i = 1; i < ctrl->queue_count; i++)
2041                 nvme_tcp_wait_queue(ctrl, i);
2042 }
2043
2044 static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl,
2045                                     int first, int last)
2046 {
2047         int i, ret;
2048
2049         for (i = first; i < last; i++) {
2050                 ret = nvme_tcp_start_queue(ctrl, i);
2051                 if (ret)
2052                         goto out_stop_queues;
2053         }
2054
2055         return 0;
2056
2057 out_stop_queues:
2058         for (i--; i >= first; i--)
2059                 nvme_tcp_stop_queue(ctrl, i);
2060         return ret;
2061 }
2062
2063 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
2064 {
2065         int ret;
2066         key_serial_t pskid = 0;
2067
2068         if (nvme_tcp_tls_configured(ctrl)) {
2069                 if (ctrl->opts->tls_key)
2070                         pskid = key_serial(ctrl->opts->tls_key);
2071                 else if (ctrl->opts->tls) {
2072                         pskid = nvme_tls_psk_default(ctrl->opts->keyring,
2073                                                       ctrl->opts->host->nqn,
2074                                                       ctrl->opts->subsysnqn);
2075                         if (!pskid) {
2076                                 dev_err(ctrl->device, "no valid PSK found\n");
2077                                 return -ENOKEY;
2078                         }
2079                 }
2080         }
2081
2082         ret = nvme_tcp_alloc_queue(ctrl, 0, pskid);
2083         if (ret)
2084                 return ret;
2085
2086         ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
2087         if (ret)
2088                 goto out_free_queue;
2089
2090         return 0;
2091
2092 out_free_queue:
2093         nvme_tcp_free_queue(ctrl, 0);
2094         return ret;
2095 }
2096
2097 static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
2098 {
2099         int i, ret;
2100
2101         if (nvme_tcp_tls_configured(ctrl)) {
2102                 if (ctrl->opts->concat) {
2103                         /*
2104                          * The generated PSK is stored in the
2105                          * fabric options
2106                          */
2107                         if (!ctrl->opts->tls_key) {
2108                                 dev_err(ctrl->device, "no PSK generated\n");
2109                                 return -ENOKEY;
2110                         }
2111                         if (ctrl->tls_pskid &&
2112                             ctrl->tls_pskid != key_serial(ctrl->opts->tls_key)) {
2113                                 dev_err(ctrl->device, "Stale PSK id %08x\n", ctrl->tls_pskid);
2114                                 ctrl->tls_pskid = 0;
2115                         }
2116                 } else if (!ctrl->tls_pskid) {
2117                         dev_err(ctrl->device, "no PSK negotiated\n");
2118                         return -ENOKEY;
2119                 }
2120         }
2121
2122         for (i = 1; i < ctrl->queue_count; i++) {
2123                 ret = nvme_tcp_alloc_queue(ctrl, i,
2124                                 ctrl->tls_pskid);
2125                 if (ret)
2126                         goto out_free_queues;
2127         }
2128
2129         return 0;
2130
2131 out_free_queues:
2132         for (i--; i >= 1; i--)
2133                 nvme_tcp_free_queue(ctrl, i);
2134
2135         return ret;
2136 }
2137
2138 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
2139 {
2140         unsigned int nr_io_queues;
2141         int ret;
2142
2143         nr_io_queues = nvmf_nr_io_queues(ctrl->opts);
2144         ret = nvme_set_queue_count(ctrl, &nr_io_queues);
2145         if (ret)
2146                 return ret;
2147
2148         if (nr_io_queues == 0) {
2149                 dev_err(ctrl->device,
2150                         "unable to set any I/O queues\n");
2151                 return -ENOMEM;
2152         }
2153
2154         ctrl->queue_count = nr_io_queues + 1;
2155         dev_info(ctrl->device,
2156                 "creating %d I/O queues.\n", nr_io_queues);
2157
2158         nvmf_set_io_queues(ctrl->opts, nr_io_queues,
2159                            to_tcp_ctrl(ctrl)->io_queues);
2160         return __nvme_tcp_alloc_io_queues(ctrl);
2161 }
2162
2163 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
2164 {
2165         int ret, nr_queues;
2166
2167         ret = nvme_tcp_alloc_io_queues(ctrl);
2168         if (ret)
2169                 return ret;
2170
2171         if (new) {
2172                 ret = nvme_alloc_io_tag_set(ctrl, &to_tcp_ctrl(ctrl)->tag_set,
2173                                 &nvme_tcp_mq_ops,
2174                                 ctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2,
2175                                 sizeof(struct nvme_tcp_request));
2176                 if (ret)
2177                         goto out_free_io_queues;
2178         }
2179
2180         /*
2181          * Only start IO queues for which we have allocated the tagset
2182          * and limited it to the available queues. On reconnects, the
2183          * queue number might have changed.
2184          */
2185         nr_queues = min(ctrl->tagset->nr_hw_queues + 1, ctrl->queue_count);
2186         ret = nvme_tcp_start_io_queues(ctrl, 1, nr_queues);
2187         if (ret)
2188                 goto out_cleanup_connect_q;
2189
2190         if (!new) {
2191                 nvme_start_freeze(ctrl);
2192                 nvme_unquiesce_io_queues(ctrl);
2193                 if (!nvme_wait_freeze_timeout(ctrl, NVME_IO_TIMEOUT)) {
2194                         /*
2195                          * If we timed out waiting for freeze we are likely to
2196                          * be stuck.  Fail the controller initialization just
2197                          * to be safe.
2198                          */
2199                         ret = -ENODEV;
2200                         nvme_unfreeze(ctrl);
2201                         goto out_wait_freeze_timed_out;
2202                 }
2203                 blk_mq_update_nr_hw_queues(ctrl->tagset,
2204                         ctrl->queue_count - 1);
2205                 nvme_unfreeze(ctrl);
2206         }
2207
2208         /*
2209          * If the number of queues has increased (reconnect case)
2210          * start all new queues now.
2211          */
2212         ret = nvme_tcp_start_io_queues(ctrl, nr_queues,
2213                                        ctrl->tagset->nr_hw_queues + 1);
2214         if (ret)
2215                 goto out_wait_freeze_timed_out;
2216
2217         return 0;
2218
2219 out_wait_freeze_timed_out:
2220         nvme_quiesce_io_queues(ctrl);
2221         nvme_sync_io_queues(ctrl);
2222         nvme_tcp_stop_io_queues(ctrl);
2223 out_cleanup_connect_q:
2224         nvme_cancel_tagset(ctrl);
2225         if (new)
2226                 nvme_remove_io_tag_set(ctrl);
2227 out_free_io_queues:
2228         nvme_tcp_free_io_queues(ctrl);
2229         return ret;
2230 }
2231
2232 static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
2233 {
2234         int error;
2235
2236         error = nvme_tcp_alloc_admin_queue(ctrl);
2237         if (error)
2238                 return error;
2239
2240         if (new) {
2241                 error = nvme_alloc_admin_tag_set(ctrl,
2242                                 &to_tcp_ctrl(ctrl)->admin_tag_set,
2243                                 &nvme_tcp_admin_mq_ops,
2244                                 sizeof(struct nvme_tcp_request));
2245                 if (error)
2246                         goto out_free_queue;
2247         }
2248
2249         error = nvme_tcp_start_queue(ctrl, 0);
2250         if (error)
2251                 goto out_cleanup_tagset;
2252
2253         error = nvme_enable_ctrl(ctrl);
2254         if (error)
2255                 goto out_stop_queue;
2256
2257         nvme_unquiesce_admin_queue(ctrl);
2258
2259         error = nvme_init_ctrl_finish(ctrl, false);
2260         if (error)
2261                 goto out_quiesce_queue;
2262
2263         return 0;
2264
2265 out_quiesce_queue:
2266         nvme_quiesce_admin_queue(ctrl);
2267         blk_sync_queue(ctrl->admin_q);
2268 out_stop_queue:
2269         nvme_tcp_stop_queue(ctrl, 0);
2270         nvme_cancel_admin_tagset(ctrl);
2271 out_cleanup_tagset:
2272         if (new)
2273                 nvme_remove_admin_tag_set(ctrl);
2274 out_free_queue:
2275         nvme_tcp_free_admin_queue(ctrl);
2276         return error;
2277 }
2278
2279 static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
2280                 bool remove)
2281 {
2282         nvme_quiesce_admin_queue(ctrl);
2283         blk_sync_queue(ctrl->admin_q);
2284         nvme_tcp_stop_queue(ctrl, 0);
2285         nvme_cancel_admin_tagset(ctrl);
2286         if (remove) {
2287                 nvme_unquiesce_admin_queue(ctrl);
2288                 nvme_remove_admin_tag_set(ctrl);
2289         }
2290         nvme_tcp_free_admin_queue(ctrl);
2291         if (ctrl->tls_pskid) {
2292                 dev_dbg(ctrl->device, "Wipe negotiated TLS_PSK %08x\n",
2293                         ctrl->tls_pskid);
2294                 ctrl->tls_pskid = 0;
2295         }
2296 }
2297
2298 static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
2299                 bool remove)
2300 {
2301         if (ctrl->queue_count <= 1)
2302                 return;
2303         nvme_quiesce_io_queues(ctrl);
2304         nvme_sync_io_queues(ctrl);
2305         nvme_tcp_stop_io_queues(ctrl);
2306         nvme_cancel_tagset(ctrl);
2307         if (remove) {
2308                 nvme_unquiesce_io_queues(ctrl);
2309                 nvme_remove_io_tag_set(ctrl);
2310         }
2311         nvme_tcp_free_io_queues(ctrl);
2312 }
2313
2314 static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl,
2315                 int status)
2316 {
2317         enum nvme_ctrl_state state = nvme_ctrl_state(ctrl);
2318
2319         /* If we are resetting/deleting then do nothing */
2320         if (state != NVME_CTRL_CONNECTING) {
2321                 WARN_ON_ONCE(state == NVME_CTRL_NEW || state == NVME_CTRL_LIVE);
2322                 return;
2323         }
2324
2325         if (nvmf_should_reconnect(ctrl, status)) {
2326                 dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
2327                         ctrl->opts->reconnect_delay);
2328                 queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
2329                                 ctrl->opts->reconnect_delay * HZ);
2330         } else {
2331                 dev_info(ctrl->device, "Removing controller (%d)...\n",
2332                          status);
2333                 nvme_delete_ctrl(ctrl);
2334         }
2335 }
2336
2337 /*
2338  * The TLS key is set by secure concatenation after negotiation has been
2339  * completed on the admin queue. We need to revoke the key when:
2340  * - concatenation is enabled (otherwise it's a static key set by the user)
2341  * and
2342  * - the generated key is present in ctrl->tls_key (otherwise there's nothing
2343  *   to revoke)
2344  * and
2345  * - a valid PSK key ID has been set in ctrl->tls_pskid (otherwise TLS
2346  *   negotiation has not run).
2347  *
2348  * We cannot always revoke the key as nvme_tcp_alloc_admin_queue() is called
2349  * twice during secure concatenation, once on a 'normal' connection to run the
2350  * DH-HMAC-CHAP negotiation (which generates the key, so it _must not_ be set),
2351  * and once after the negotiation (which uses the key, so it _must_ be set).
2352  */
2353 static bool nvme_tcp_key_revoke_needed(struct nvme_ctrl *ctrl)
2354 {
2355         return ctrl->opts->concat && ctrl->opts->tls_key && ctrl->tls_pskid;
2356 }
2357
2358 static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
2359 {
2360         struct nvmf_ctrl_options *opts = ctrl->opts;
2361         int ret;
2362
2363         ret = nvme_tcp_configure_admin_queue(ctrl, new);
2364         if (ret)
2365                 return ret;
2366
2367         if (ctrl->opts->concat && !ctrl->tls_pskid) {
2368                 /* See comments for nvme_tcp_key_revoke_needed() */
2369                 dev_dbg(ctrl->device, "restart admin queue for secure concatenation\n");
2370                 nvme_stop_keep_alive(ctrl);
2371                 nvme_tcp_teardown_admin_queue(ctrl, false);
2372                 ret = nvme_tcp_configure_admin_queue(ctrl, false);
2373                 if (ret)
2374                         goto destroy_admin;
2375         }
2376
2377         if (ctrl->icdoff) {
2378                 ret = -EOPNOTSUPP;
2379                 dev_err(ctrl->device, "icdoff is not supported!\n");
2380                 goto destroy_admin;
2381         }
2382
2383         if (!nvme_ctrl_sgl_supported(ctrl)) {
2384                 ret = -EOPNOTSUPP;
2385                 dev_err(ctrl->device, "Mandatory sgls are not supported!\n");
2386                 goto destroy_admin;
2387         }
2388
2389         if (opts->queue_size > ctrl->sqsize + 1)
2390                 dev_warn(ctrl->device,
2391                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
2392                         opts->queue_size, ctrl->sqsize + 1);
2393
2394         if (ctrl->sqsize + 1 > ctrl->maxcmd) {
2395                 dev_warn(ctrl->device,
2396                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
2397                         ctrl->sqsize + 1, ctrl->maxcmd);
2398                 ctrl->sqsize = ctrl->maxcmd - 1;
2399         }
2400
2401         if (ctrl->queue_count > 1) {
2402                 ret = nvme_tcp_configure_io_queues(ctrl, new);
2403                 if (ret)
2404                         goto destroy_admin;
2405         }
2406
2407         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
2408                 /*
2409                  * state change failure is ok if we started ctrl delete,
2410                  * unless we're during creation of a new controller to
2411                  * avoid races with teardown flow.
2412                  */
2413                 enum nvme_ctrl_state state = nvme_ctrl_state(ctrl);
2414
2415                 WARN_ON_ONCE(state != NVME_CTRL_DELETING &&
2416                              state != NVME_CTRL_DELETING_NOIO);
2417                 WARN_ON_ONCE(new);
2418                 ret = -EINVAL;
2419                 goto destroy_io;
2420         }
2421
2422         nvme_start_ctrl(ctrl);
2423         return 0;
2424
2425 destroy_io:
2426         if (ctrl->queue_count > 1) {
2427                 nvme_quiesce_io_queues(ctrl);
2428                 nvme_sync_io_queues(ctrl);
2429                 nvme_tcp_stop_io_queues(ctrl);
2430                 nvme_cancel_tagset(ctrl);
2431                 if (new)
2432                         nvme_remove_io_tag_set(ctrl);
2433                 nvme_tcp_free_io_queues(ctrl);
2434         }
2435 destroy_admin:
2436         nvme_stop_keep_alive(ctrl);
2437         nvme_tcp_teardown_admin_queue(ctrl, new);
2438         return ret;
2439 }
2440
2441 static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
2442 {
2443         struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
2444                         struct nvme_tcp_ctrl, connect_work);
2445         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2446         int ret;
2447
2448         ++ctrl->nr_reconnects;
2449
2450         ret = nvme_tcp_setup_ctrl(ctrl, false);
2451         if (ret)
2452                 goto requeue;
2453
2454         dev_info(ctrl->device, "Successfully reconnected (attempt %d/%d)\n",
2455                  ctrl->nr_reconnects, ctrl->opts->max_reconnects);
2456
2457         ctrl->nr_reconnects = 0;
2458
2459         return;
2460
2461 requeue:
2462         dev_info(ctrl->device, "Failed reconnect attempt %d/%d\n",
2463                  ctrl->nr_reconnects, ctrl->opts->max_reconnects);
2464         nvme_tcp_reconnect_or_remove(ctrl, ret);
2465 }
2466
2467 static void nvme_tcp_error_recovery_work(struct work_struct *work)
2468 {
2469         struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
2470                                 struct nvme_tcp_ctrl, err_work);
2471         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2472
2473         if (nvme_tcp_key_revoke_needed(ctrl))
2474                 nvme_auth_revoke_tls_key(ctrl);
2475         nvme_stop_keep_alive(ctrl);
2476         flush_work(&ctrl->async_event_work);
2477         nvme_tcp_teardown_io_queues(ctrl, false);
2478         /* unquiesce to fail fast pending requests */
2479         nvme_unquiesce_io_queues(ctrl);
2480         nvme_tcp_teardown_admin_queue(ctrl, false);
2481         nvme_unquiesce_admin_queue(ctrl);
2482         nvme_auth_stop(ctrl);
2483
2484         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2485                 /* state change failure is ok if we started ctrl delete */
2486                 enum nvme_ctrl_state state = nvme_ctrl_state(ctrl);
2487
2488                 WARN_ON_ONCE(state != NVME_CTRL_DELETING &&
2489                              state != NVME_CTRL_DELETING_NOIO);
2490                 return;
2491         }
2492
2493         nvme_tcp_reconnect_or_remove(ctrl, 0);
2494 }
2495
2496 static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
2497 {
2498         nvme_tcp_teardown_io_queues(ctrl, shutdown);
2499         nvme_quiesce_admin_queue(ctrl);
2500         nvme_disable_ctrl(ctrl, shutdown);
2501         nvme_tcp_teardown_admin_queue(ctrl, shutdown);
2502 }
2503
2504 static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
2505 {
2506         nvme_tcp_teardown_ctrl(ctrl, true);
2507 }
2508
2509 static void nvme_reset_ctrl_work(struct work_struct *work)
2510 {
2511         struct nvme_ctrl *ctrl =
2512                 container_of(work, struct nvme_ctrl, reset_work);
2513         int ret;
2514
2515         if (nvme_tcp_key_revoke_needed(ctrl))
2516                 nvme_auth_revoke_tls_key(ctrl);
2517         nvme_stop_ctrl(ctrl);
2518         nvme_tcp_teardown_ctrl(ctrl, false);
2519
2520         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2521                 /* state change failure is ok if we started ctrl delete */
2522                 enum nvme_ctrl_state state = nvme_ctrl_state(ctrl);
2523
2524                 WARN_ON_ONCE(state != NVME_CTRL_DELETING &&
2525                              state != NVME_CTRL_DELETING_NOIO);
2526                 return;
2527         }
2528
2529         ret = nvme_tcp_setup_ctrl(ctrl, false);
2530         if (ret)
2531                 goto out_fail;
2532
2533         return;
2534
2535 out_fail:
2536         ++ctrl->nr_reconnects;
2537         nvme_tcp_reconnect_or_remove(ctrl, ret);
2538 }
2539
2540 static void nvme_tcp_stop_ctrl(struct nvme_ctrl *ctrl)
2541 {
2542         flush_work(&to_tcp_ctrl(ctrl)->err_work);
2543         cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
2544 }
2545
2546 static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
2547 {
2548         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
2549
2550         if (list_empty(&ctrl->list))
2551                 goto free_ctrl;
2552
2553         mutex_lock(&nvme_tcp_ctrl_mutex);
2554         list_del(&ctrl->list);
2555         mutex_unlock(&nvme_tcp_ctrl_mutex);
2556
2557         nvmf_free_options(nctrl->opts);
2558 free_ctrl:
2559         kfree(ctrl->queues);
2560         kfree(ctrl);
2561 }
2562
2563 static void nvme_tcp_set_sg_null(struct nvme_command *c)
2564 {
2565         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2566
2567         sg->addr = 0;
2568         sg->length = 0;
2569         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2570                         NVME_SGL_FMT_TRANSPORT_A;
2571 }
2572
2573 static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
2574                 struct nvme_command *c, u32 data_len)
2575 {
2576         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2577
2578         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
2579         sg->length = cpu_to_le32(data_len);
2580         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
2581 }
2582
2583 static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
2584                 u32 data_len)
2585 {
2586         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2587
2588         sg->addr = 0;
2589         sg->length = cpu_to_le32(data_len);
2590         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2591                         NVME_SGL_FMT_TRANSPORT_A;
2592 }
2593
2594 static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
2595 {
2596         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2597         struct nvme_tcp_queue *queue = &ctrl->queues[0];
2598         struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2599         struct nvme_command *cmd = &pdu->cmd;
2600         u8 hdgst = nvme_tcp_hdgst_len(queue);
2601
2602         memset(pdu, 0, sizeof(*pdu));
2603         pdu->hdr.type = nvme_tcp_cmd;
2604         if (queue->hdr_digest)
2605                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2606         pdu->hdr.hlen = sizeof(*pdu);
2607         pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2608
2609         cmd->common.opcode = nvme_admin_async_event;
2610         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2611         cmd->common.flags |= NVME_CMD_SGL_METABUF;
2612         nvme_tcp_set_sg_null(cmd);
2613
2614         ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2615         ctrl->async_req.offset = 0;
2616         ctrl->async_req.curr_bio = NULL;
2617         ctrl->async_req.data_len = 0;
2618         init_llist_node(&ctrl->async_req.lentry);
2619         INIT_LIST_HEAD(&ctrl->async_req.entry);
2620
2621         nvme_tcp_queue_request(&ctrl->async_req, true);
2622 }
2623
2624 static void nvme_tcp_complete_timed_out(struct request *rq)
2625 {
2626         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2627         struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2628
2629         nvme_tcp_stop_queue(ctrl, nvme_tcp_queue_id(req->queue));
2630         nvmf_complete_timed_out_request(rq);
2631 }
2632
2633 static enum blk_eh_timer_return nvme_tcp_timeout(struct request *rq)
2634 {
2635         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2636         struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2637         struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
2638         struct nvme_command *cmd = &pdu->cmd;
2639         int qid = nvme_tcp_queue_id(req->queue);
2640
2641         dev_warn(ctrl->device,
2642                  "I/O tag %d (%04x) type %d opcode %#x (%s) QID %d timeout\n",
2643                  rq->tag, nvme_cid(rq), pdu->hdr.type, cmd->common.opcode,
2644                  nvme_fabrics_opcode_str(qid, cmd), qid);
2645
2646         if (nvme_ctrl_state(ctrl) != NVME_CTRL_LIVE) {
2647                 /*
2648                  * If we are resetting, connecting or deleting we should
2649                  * complete immediately because we may block controller
2650                  * teardown or setup sequence
2651                  * - ctrl disable/shutdown fabrics requests
2652                  * - connect requests
2653                  * - initialization admin requests
2654                  * - I/O requests that entered after unquiescing and
2655                  *   the controller stopped responding
2656                  *
2657                  * All other requests should be cancelled by the error
2658                  * recovery work, so it's fine that we fail it here.
2659                  */
2660                 nvme_tcp_complete_timed_out(rq);
2661                 return BLK_EH_DONE;
2662         }
2663
2664         /*
2665          * LIVE state should trigger the normal error recovery which will
2666          * handle completing this request.
2667          */
2668         nvme_tcp_error_recovery(ctrl);
2669         return BLK_EH_RESET_TIMER;
2670 }
2671
2672 static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2673                         struct request *rq)
2674 {
2675         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2676         struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
2677         struct nvme_command *c = &pdu->cmd;
2678
2679         c->common.flags |= NVME_CMD_SGL_METABUF;
2680
2681         if (!blk_rq_nr_phys_segments(rq))
2682                 nvme_tcp_set_sg_null(c);
2683         else if (rq_data_dir(rq) == WRITE &&
2684             req->data_len <= nvme_tcp_inline_data_size(req))
2685                 nvme_tcp_set_sg_inline(queue, c, req->data_len);
2686         else
2687                 nvme_tcp_set_sg_host_data(c, req->data_len);
2688
2689         return 0;
2690 }
2691
2692 static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2693                 struct request *rq)
2694 {
2695         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2696         struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
2697         struct nvme_tcp_queue *queue = req->queue;
2698         u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2699         blk_status_t ret;
2700
2701         ret = nvme_setup_cmd(ns, rq);
2702         if (ret)
2703                 return ret;
2704
2705         req->state = NVME_TCP_SEND_CMD_PDU;
2706         req->status = cpu_to_le16(NVME_SC_SUCCESS);
2707         req->offset = 0;
2708         req->data_sent = 0;
2709         req->pdu_len = 0;
2710         req->pdu_sent = 0;
2711         req->h2cdata_left = 0;
2712         req->data_len = blk_rq_nr_phys_segments(rq) ?
2713                                 blk_rq_payload_bytes(rq) : 0;
2714         req->curr_bio = rq->bio;
2715         if (req->curr_bio && req->data_len)
2716                 nvme_tcp_init_iter(req, rq_data_dir(rq));
2717
2718         if (rq_data_dir(rq) == WRITE &&
2719             req->data_len <= nvme_tcp_inline_data_size(req))
2720                 req->pdu_len = req->data_len;
2721
2722         pdu->hdr.type = nvme_tcp_cmd;
2723         pdu->hdr.flags = 0;
2724         if (queue->hdr_digest)
2725                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2726         if (queue->data_digest && req->pdu_len) {
2727                 pdu->hdr.flags |= NVME_TCP_F_DDGST;
2728                 ddgst = nvme_tcp_ddgst_len(queue);
2729         }
2730         pdu->hdr.hlen = sizeof(*pdu);
2731         pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2732         pdu->hdr.plen =
2733                 cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2734
2735         ret = nvme_tcp_map_data(queue, rq);
2736         if (unlikely(ret)) {
2737                 nvme_cleanup_cmd(rq);
2738                 dev_err(queue->ctrl->ctrl.device,
2739                         "Failed to map data (%d)\n", ret);
2740                 return ret;
2741         }
2742
2743         return 0;
2744 }
2745
2746 static void nvme_tcp_commit_rqs(struct blk_mq_hw_ctx *hctx)
2747 {
2748         struct nvme_tcp_queue *queue = hctx->driver_data;
2749
2750         if (!llist_empty(&queue->req_list))
2751                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
2752 }
2753
2754 static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2755                 const struct blk_mq_queue_data *bd)
2756 {
2757         struct nvme_ns *ns = hctx->queue->queuedata;
2758         struct nvme_tcp_queue *queue = hctx->driver_data;
2759         struct request *rq = bd->rq;
2760         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2761         bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2762         blk_status_t ret;
2763
2764         if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2765                 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
2766
2767         ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2768         if (unlikely(ret))
2769                 return ret;
2770
2771         nvme_start_request(rq);
2772
2773         nvme_tcp_queue_request(req, bd->last);
2774
2775         return BLK_STS_OK;
2776 }
2777
2778 static void nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2779 {
2780         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(set->driver_data);
2781
2782         nvmf_map_queues(set, &ctrl->ctrl, ctrl->io_queues);
2783 }
2784
2785 static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
2786 {
2787         struct nvme_tcp_queue *queue = hctx->driver_data;
2788         struct sock *sk = queue->sock->sk;
2789         int ret;
2790
2791         if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2792                 return 0;
2793
2794         set_bit(NVME_TCP_Q_POLLING, &queue->flags);
2795         if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue))
2796                 sk_busy_loop(sk, true);
2797         ret = nvme_tcp_try_recv(queue);
2798         clear_bit(NVME_TCP_Q_POLLING, &queue->flags);
2799         return ret < 0 ? ret : queue->nr_cqe;
2800 }
2801
2802 static int nvme_tcp_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2803 {
2804         struct nvme_tcp_queue *queue = &to_tcp_ctrl(ctrl)->queues[0];
2805         struct sockaddr_storage src_addr;
2806         int ret, len;
2807
2808         len = nvmf_get_address(ctrl, buf, size);
2809
2810         if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2811                 return len;
2812
2813         mutex_lock(&queue->queue_lock);
2814
2815         ret = kernel_getsockname(queue->sock, (struct sockaddr *)&src_addr);
2816         if (ret > 0) {
2817                 if (len > 0)
2818                         len--; /* strip trailing newline */
2819                 len += scnprintf(buf + len, size - len, "%ssrc_addr=%pISc\n",
2820                                 (len) ? "," : "", &src_addr);
2821         }
2822
2823         mutex_unlock(&queue->queue_lock);
2824
2825         return len;
2826 }
2827
2828 static const struct blk_mq_ops nvme_tcp_mq_ops = {
2829         .queue_rq       = nvme_tcp_queue_rq,
2830         .commit_rqs     = nvme_tcp_commit_rqs,
2831         .complete       = nvme_complete_rq,
2832         .init_request   = nvme_tcp_init_request,
2833         .exit_request   = nvme_tcp_exit_request,
2834         .init_hctx      = nvme_tcp_init_hctx,
2835         .timeout        = nvme_tcp_timeout,
2836         .map_queues     = nvme_tcp_map_queues,
2837         .poll           = nvme_tcp_poll,
2838 };
2839
2840 static const struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2841         .queue_rq       = nvme_tcp_queue_rq,
2842         .complete       = nvme_complete_rq,
2843         .init_request   = nvme_tcp_init_request,
2844         .exit_request   = nvme_tcp_exit_request,
2845         .init_hctx      = nvme_tcp_init_admin_hctx,
2846         .timeout        = nvme_tcp_timeout,
2847 };
2848
2849 static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2850         .name                   = "tcp",
2851         .module                 = THIS_MODULE,
2852         .flags                  = NVME_F_FABRICS | NVME_F_BLOCKING,
2853         .reg_read32             = nvmf_reg_read32,
2854         .reg_read64             = nvmf_reg_read64,
2855         .reg_write32            = nvmf_reg_write32,
2856         .subsystem_reset        = nvmf_subsystem_reset,
2857         .free_ctrl              = nvme_tcp_free_ctrl,
2858         .submit_async_event     = nvme_tcp_submit_async_event,
2859         .delete_ctrl            = nvme_tcp_delete_ctrl,
2860         .get_address            = nvme_tcp_get_address,
2861         .stop_ctrl              = nvme_tcp_stop_ctrl,
2862 };
2863
2864 static bool
2865 nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2866 {
2867         struct nvme_tcp_ctrl *ctrl;
2868         bool found = false;
2869
2870         mutex_lock(&nvme_tcp_ctrl_mutex);
2871         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2872                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2873                 if (found)
2874                         break;
2875         }
2876         mutex_unlock(&nvme_tcp_ctrl_mutex);
2877
2878         return found;
2879 }
2880
2881 static struct nvme_tcp_ctrl *nvme_tcp_alloc_ctrl(struct device *dev,
2882                 struct nvmf_ctrl_options *opts)
2883 {
2884         struct nvme_tcp_ctrl *ctrl;
2885         int ret;
2886
2887         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2888         if (!ctrl)
2889                 return ERR_PTR(-ENOMEM);
2890
2891         INIT_LIST_HEAD(&ctrl->list);
2892         ctrl->ctrl.opts = opts;
2893         ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2894                                 opts->nr_poll_queues + 1;
2895         ctrl->ctrl.sqsize = opts->queue_size - 1;
2896         ctrl->ctrl.kato = opts->kato;
2897
2898         INIT_DELAYED_WORK(&ctrl->connect_work,
2899                         nvme_tcp_reconnect_ctrl_work);
2900         INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2901         INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2902
2903         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2904                 opts->trsvcid =
2905                         kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2906                 if (!opts->trsvcid) {
2907                         ret = -ENOMEM;
2908                         goto out_free_ctrl;
2909                 }
2910                 opts->mask |= NVMF_OPT_TRSVCID;
2911         }
2912
2913         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2914                         opts->traddr, opts->trsvcid, &ctrl->addr);
2915         if (ret) {
2916                 pr_err("malformed address passed: %s:%s\n",
2917                         opts->traddr, opts->trsvcid);
2918                 goto out_free_ctrl;
2919         }
2920
2921         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2922                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2923                         opts->host_traddr, NULL, &ctrl->src_addr);
2924                 if (ret) {
2925                         pr_err("malformed src address passed: %s\n",
2926                                opts->host_traddr);
2927                         goto out_free_ctrl;
2928                 }
2929         }
2930
2931         if (opts->mask & NVMF_OPT_HOST_IFACE) {
2932                 if (!__dev_get_by_name(&init_net, opts->host_iface)) {
2933                         pr_err("invalid interface passed: %s\n",
2934                                opts->host_iface);
2935                         ret = -ENODEV;
2936                         goto out_free_ctrl;
2937                 }
2938         }
2939
2940         if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2941                 ret = -EALREADY;
2942                 goto out_free_ctrl;
2943         }
2944
2945         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2946                                 GFP_KERNEL);
2947         if (!ctrl->queues) {
2948                 ret = -ENOMEM;
2949                 goto out_free_ctrl;
2950         }
2951
2952         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2953         if (ret)
2954                 goto out_kfree_queues;
2955
2956         return ctrl;
2957 out_kfree_queues:
2958         kfree(ctrl->queues);
2959 out_free_ctrl:
2960         kfree(ctrl);
2961         return ERR_PTR(ret);
2962 }
2963
2964 static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2965                 struct nvmf_ctrl_options *opts)
2966 {
2967         struct nvme_tcp_ctrl *ctrl;
2968         int ret;
2969
2970         ctrl = nvme_tcp_alloc_ctrl(dev, opts);
2971         if (IS_ERR(ctrl))
2972                 return ERR_CAST(ctrl);
2973
2974         ret = nvme_add_ctrl(&ctrl->ctrl);
2975         if (ret)
2976                 goto out_put_ctrl;
2977
2978         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2979                 WARN_ON_ONCE(1);
2980                 ret = -EINTR;
2981                 goto out_uninit_ctrl;
2982         }
2983
2984         ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2985         if (ret)
2986                 goto out_uninit_ctrl;
2987
2988         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp, hostnqn: %s\n",
2989                 nvmf_ctrl_subsysnqn(&ctrl->ctrl), &ctrl->addr, opts->host->nqn);
2990
2991         mutex_lock(&nvme_tcp_ctrl_mutex);
2992         list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2993         mutex_unlock(&nvme_tcp_ctrl_mutex);
2994
2995         return &ctrl->ctrl;
2996
2997 out_uninit_ctrl:
2998         nvme_uninit_ctrl(&ctrl->ctrl);
2999 out_put_ctrl:
3000         nvme_put_ctrl(&ctrl->ctrl);
3001         if (ret > 0)
3002                 ret = -EIO;
3003         return ERR_PTR(ret);
3004 }
3005
3006 static struct nvmf_transport_ops nvme_tcp_transport = {
3007         .name           = "tcp",
3008         .module         = THIS_MODULE,
3009         .required_opts  = NVMF_OPT_TRADDR,
3010         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
3011                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
3012                           NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
3013                           NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
3014                           NVMF_OPT_TOS | NVMF_OPT_HOST_IFACE | NVMF_OPT_TLS |
3015                           NVMF_OPT_KEYRING | NVMF_OPT_TLS_KEY | NVMF_OPT_CONCAT,
3016         .create_ctrl    = nvme_tcp_create_ctrl,
3017 };
3018
3019 static int __init nvme_tcp_init_module(void)
3020 {
3021         unsigned int wq_flags = WQ_MEM_RECLAIM | WQ_HIGHPRI | WQ_SYSFS;
3022         int cpu;
3023
3024         BUILD_BUG_ON(sizeof(struct nvme_tcp_hdr) != 8);
3025         BUILD_BUG_ON(sizeof(struct nvme_tcp_cmd_pdu) != 72);
3026         BUILD_BUG_ON(sizeof(struct nvme_tcp_data_pdu) != 24);
3027         BUILD_BUG_ON(sizeof(struct nvme_tcp_rsp_pdu) != 24);
3028         BUILD_BUG_ON(sizeof(struct nvme_tcp_r2t_pdu) != 24);
3029         BUILD_BUG_ON(sizeof(struct nvme_tcp_icreq_pdu) != 128);
3030         BUILD_BUG_ON(sizeof(struct nvme_tcp_icresp_pdu) != 128);
3031         BUILD_BUG_ON(sizeof(struct nvme_tcp_term_pdu) != 24);
3032
3033         if (wq_unbound)
3034                 wq_flags |= WQ_UNBOUND;
3035
3036         nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq", wq_flags, 0);
3037         if (!nvme_tcp_wq)
3038                 return -ENOMEM;
3039
3040         for_each_possible_cpu(cpu)
3041                 atomic_set(&nvme_tcp_cpu_queues[cpu], 0);
3042
3043         nvmf_register_transport(&nvme_tcp_transport);
3044         return 0;
3045 }
3046
3047 static void __exit nvme_tcp_cleanup_module(void)
3048 {
3049         struct nvme_tcp_ctrl *ctrl;
3050
3051         nvmf_unregister_transport(&nvme_tcp_transport);
3052
3053         mutex_lock(&nvme_tcp_ctrl_mutex);
3054         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
3055                 nvme_delete_ctrl(&ctrl->ctrl);
3056         mutex_unlock(&nvme_tcp_ctrl_mutex);
3057         flush_workqueue(nvme_delete_wq);
3058
3059         destroy_workqueue(nvme_tcp_wq);
3060 }
3061
3062 module_init(nvme_tcp_init_module);
3063 module_exit(nvme_tcp_cleanup_module);
3064
3065 MODULE_DESCRIPTION("NVMe host TCP transport driver");
3066 MODULE_LICENSE("GPL v2");