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