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