drivers/nvme/host/tcp.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * NVMe over Fabrics TCP host.
   4  * Copyright (c) 2018 Lightbits Labs. All rights reserved.
   5  */
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7 #include <linux/module.h>
   8 #include <linux/init.h>
   9 #include <linux/slab.h>
  10 #include <linux/err.h>
  11 #include <linux/nvme-tcp.h>
  12 #include <net/sock.h>
  13 #include <net/tcp.h>
  14 #include <linux/blk-mq.h>
  15 #include <crypto/hash.h>
  16 #include <net/busy_poll.h>
  17 #include <trace/events/sock.h>
  18
  19 #include "nvme.h"
  20 #include "fabrics.h"
  21
  22 struct nvme_tcp_queue;
  23
  24 /* Define the socket priority to use for connections were it is desirable
  25  * that the NIC consider performing optimized packet processing or filtering.
  26  * A non-zero value being sufficient to indicate general consideration of any
  27  * possible optimization.  Making it a module param allows for alternative
  28  * values that may be unique for some NIC implementations.
  29  */
  30 static int so_priority;
  31 module_param(so_priority, int, 0644);
  32 MODULE_PARM_DESC(so_priority, "nvme tcp socket optimize priority");
  33
  34 #ifdef CONFIG_DEBUG_LOCK_ALLOC
  35 /* lockdep can detect a circular dependency of the form
  36  *   sk_lock -> mmap_lock (page fault) -> fs locks -> sk_lock
  37  * because dependencies are tracked for both nvme-tcp and user contexts. Using
  38  * a separate class prevents lockdep from conflating nvme-tcp socket use with
  39  * user-space socket API use.
  40  */
  41 static struct lock_class_key nvme_tcp_sk_key[2];
  42 static struct lock_class_key nvme_tcp_slock_key[2];
  43
  44 static void nvme_tcp_reclassify_socket(struct socket *sock)
  45 {
  46         struct sock *sk = sock->sk;
  47
  48         if (WARN_ON_ONCE(!sock_allow_reclassification(sk)))
  49                 return;
  50
  51         switch (sk->sk_family) {
  52         case AF_INET:
  53                 sock_lock_init_class_and_name(sk, "slock-AF_INET-NVME",
  54                                               &nvme_tcp_slock_key[0],
  55                                               "sk_lock-AF_INET-NVME",
  56                                               &nvme_tcp_sk_key[0]);
  57                 break;
  58         case AF_INET6:
  59                 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NVME",
  60                                               &nvme_tcp_slock_key[1],
  61                                               "sk_lock-AF_INET6-NVME",
  62                                               &nvme_tcp_sk_key[1]);
  63                 break;
  64         default:
  65                 WARN_ON_ONCE(1);
  66         }
  67 }
  68 #else
  69 static void nvme_tcp_reclassify_socket(struct socket *sock) { }
  70 #endif
  71
  72 enum nvme_tcp_send_state {
  73         NVME_TCP_SEND_CMD_PDU = 0,
  74         NVME_TCP_SEND_H2C_PDU,
  75         NVME_TCP_SEND_DATA,
  76         NVME_TCP_SEND_DDGST,
  77 };
  78
  79 struct nvme_tcp_request {
  80         struct nvme_request     req;
  81         void                    *pdu;
  82         struct nvme_tcp_queue   *queue;
  83         u32                     data_len;
  84         u32                     pdu_len;
  85         u32                     pdu_sent;
  86         u32                     h2cdata_left;
  87         u32                     h2cdata_offset;
  88         u16                     ttag;
  89         __le16                  status;
  90         struct list_head        entry;
  91         struct llist_node       lentry;
  92         __le32                  ddgst;
  93
  94         struct bio              *curr_bio;
  95         struct iov_iter         iter;
  96
  97         /* send state */
  98         size_t                  offset;
  99         size_t                  data_sent;
 100         enum nvme_tcp_send_state state;
 101 };
 102
 103 enum nvme_tcp_queue_flags {
 104         NVME_TCP_Q_ALLOCATED    = 0,
 105         NVME_TCP_Q_LIVE         = 1,
 106         NVME_TCP_Q_POLLING      = 2,
 107 };
 108
 109 enum nvme_tcp_recv_state {
 110         NVME_TCP_RECV_PDU = 0,
 111         NVME_TCP_RECV_DATA,
 112         NVME_TCP_RECV_DDGST,
 113 };
 114
 115 struct nvme_tcp_ctrl;
 116 struct nvme_tcp_queue {
 117         struct socket           *sock;
 118         struct work_struct      io_work;
 119         int                     io_cpu;
 120
 121         struct mutex            queue_lock;
 122         struct mutex            send_mutex;
 123         struct llist_head       req_list;
 124         struct list_head        send_list;
 125
 126         /* recv state */
 127         void                    *pdu;
 128         int                     pdu_remaining;
 129         int                     pdu_offset;
 130         size_t                  data_remaining;
 131         size_t                  ddgst_remaining;
 132         unsigned int            nr_cqe;
 133
 134         /* send state */
 135         struct nvme_tcp_request *request;
 136
 137         u32                     maxh2cdata;
 138         size_t                  cmnd_capsule_len;
 139         struct nvme_tcp_ctrl    *ctrl;
 140         unsigned long           flags;
 141         bool                    rd_enabled;
 142
 143         bool                    hdr_digest;
 144         bool                    data_digest;
 145         struct ahash_request    *rcv_hash;
 146         struct ahash_request    *snd_hash;
 147         __le32                  exp_ddgst;
 148         __le32                  recv_ddgst;
 149
 150         struct page_frag_cache  pf_cache;
 151
 152         void (*state_change)(struct sock *);
 153         void (*data_ready)(struct sock *);
 154         void (*write_space)(struct sock *);
 155 };
 156
 157 struct nvme_tcp_ctrl {
 158         /* read only in the hot path */
 159         struct nvme_tcp_queue   *queues;
 160         struct blk_mq_tag_set   tag_set;
 161
 162         /* other member variables */
 163         struct list_head        list;
 164         struct blk_mq_tag_set   admin_tag_set;
 165         struct sockaddr_storage addr;
 166         struct sockaddr_storage src_addr;
 167         struct nvme_ctrl        ctrl;
 168
 169         struct work_struct      err_work;
 170         struct delayed_work     connect_work;
 171         struct nvme_tcp_request async_req;
 172         u32                     io_queues[HCTX_MAX_TYPES];
 173 };
 174
 175 static LIST_HEAD(nvme_tcp_ctrl_list);
 176 static DEFINE_MUTEX(nvme_tcp_ctrl_mutex);
 177 static struct workqueue_struct *nvme_tcp_wq;
 178 static const struct blk_mq_ops nvme_tcp_mq_ops;
 179 static const struct blk_mq_ops nvme_tcp_admin_mq_ops;
 180 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue);
 181
 182 static inline struct nvme_tcp_ctrl *to_tcp_ctrl(struct nvme_ctrl *ctrl)
 183 {
 184         return container_of(ctrl, struct nvme_tcp_ctrl, ctrl);
 185 }
 186
 187 static inline int nvme_tcp_queue_id(struct nvme_tcp_queue *queue)
 188 {
 189         return queue - queue->ctrl->queues;
 190 }
 191
 192 static inline struct blk_mq_tags *nvme_tcp_tagset(struct nvme_tcp_queue *queue)
 193 {
 194         u32 queue_idx = nvme_tcp_queue_id(queue);
 195
 196         if (queue_idx == 0)
 197                 return queue->ctrl->admin_tag_set.tags[queue_idx];
 198         return queue->ctrl->tag_set.tags[queue_idx - 1];
 199 }
 200
 201 static inline u8 nvme_tcp_hdgst_len(struct nvme_tcp_queue *queue)
 202 {
 203         return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 204 }
 205
 206 static inline u8 nvme_tcp_ddgst_len(struct nvme_tcp_queue *queue)
 207 {
 208         return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 209 }
 210
 211 static inline void *nvme_tcp_req_cmd_pdu(struct nvme_tcp_request *req)
 212 {
 213         return req->pdu;
 214 }
 215
 216 static inline void *nvme_tcp_req_data_pdu(struct nvme_tcp_request *req)
 217 {
 218         /* use the pdu space in the back for the data pdu */
 219         return req->pdu + sizeof(struct nvme_tcp_cmd_pdu) -
 220                 sizeof(struct nvme_tcp_data_pdu);
 221 }
 222
 223 static inline size_t nvme_tcp_inline_data_size(struct nvme_tcp_request *req)
 224 {
 225         if (nvme_is_fabrics(req->req.cmd))
 226                 return NVME_TCP_ADMIN_CCSZ;
 227         return req->queue->cmnd_capsule_len - sizeof(struct nvme_command);
 228 }
 229
 230 static inline bool nvme_tcp_async_req(struct nvme_tcp_request *req)
 231 {
 232         return req == &req->queue->ctrl->async_req;
 233 }
 234
 235 static inline bool nvme_tcp_has_inline_data(struct nvme_tcp_request *req)
 236 {
 237         struct request *rq;
 238
 239         if (unlikely(nvme_tcp_async_req(req)))
 240                 return false; /* async events don't have a request */
 241
 242         rq = blk_mq_rq_from_pdu(req);
 243
 244         return rq_data_dir(rq) == WRITE && req->data_len &&
 245                 req->data_len <= nvme_tcp_inline_data_size(req);
 246 }
 247
 248 static inline struct page *nvme_tcp_req_cur_page(struct nvme_tcp_request *req)
 249 {
 250         return req->iter.bvec->bv_page;
 251 }
 252
 253 static inline size_t nvme_tcp_req_cur_offset(struct nvme_tcp_request *req)
 254 {
 255         return req->iter.bvec->bv_offset + req->iter.iov_offset;
 256 }
 257
 258 static inline size_t nvme_tcp_req_cur_length(struct nvme_tcp_request *req)
 259 {
 260         return min_t(size_t, iov_iter_single_seg_count(&req->iter),
 261                         req->pdu_len - req->pdu_sent);
 262 }
 263
 264 static inline size_t nvme_tcp_pdu_data_left(struct nvme_tcp_request *req)
 265 {
 266         return rq_data_dir(blk_mq_rq_from_pdu(req)) == WRITE ?
 267                         req->pdu_len - req->pdu_sent : 0;
 268 }
 269
 270 static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req,
 271                 int len)
 272 {
 273         return nvme_tcp_pdu_data_left(req) <= len;
 274 }
 275
 276 static void nvme_tcp_init_iter(struct nvme_tcp_request *req,
 277                 unsigned int dir)
 278 {
 279         struct request *rq = blk_mq_rq_from_pdu(req);
 280         struct bio_vec *vec;
 281         unsigned int size;
 282         int nr_bvec;
 283         size_t offset;
 284
 285         if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) {
 286                 vec = &rq->special_vec;
 287                 nr_bvec = 1;
 288                 size = blk_rq_payload_bytes(rq);
 289                 offset = 0;
 290         } else {
 291                 struct bio *bio = req->curr_bio;
 292                 struct bvec_iter bi;
 293                 struct bio_vec bv;
 294
 295                 vec = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
 296                 nr_bvec = 0;
 297                 bio_for_each_bvec(bv, bio, bi) {
 298                         nr_bvec++;
 299                 }
 300                 size = bio->bi_iter.bi_size;
 301                 offset = bio->bi_iter.bi_bvec_done;
 302         }
 303
 304         iov_iter_bvec(&req->iter, dir, vec, nr_bvec, size);
 305         req->iter.iov_offset = offset;
 306 }
 307
 308 static inline void nvme_tcp_advance_req(struct nvme_tcp_request *req,
 309                 int len)
 310 {
 311         req->data_sent += len;
 312         req->pdu_sent += len;
 313         iov_iter_advance(&req->iter, len);
 314         if (!iov_iter_count(&req->iter) &&
 315             req->data_sent < req->data_len) {
 316                 req->curr_bio = req->curr_bio->bi_next;
 317                 nvme_tcp_init_iter(req, ITER_SOURCE);
 318         }
 319 }
 320
 321 static inline void nvme_tcp_send_all(struct nvme_tcp_queue *queue)
 322 {
 323         int ret;
 324
 325         /* drain the send queue as much as we can... */
 326         do {
 327                 ret = nvme_tcp_try_send(queue);
 328         } while (ret > 0);
 329 }
 330
 331 static inline bool nvme_tcp_queue_more(struct nvme_tcp_queue *queue)
 332 {
 333         return !list_empty(&queue->send_list) ||
 334                 !llist_empty(&queue->req_list);
 335 }
 336
 337 static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req,
 338                 bool sync, bool last)
 339 {
 340         struct nvme_tcp_queue *queue = req->queue;
 341         bool empty;
 342
 343         empty = llist_add(&req->lentry, &queue->req_list) &&
 344                 list_empty(&queue->send_list) && !queue->request;
 345
 346         /*
 347          * if we're the first on the send_list and we can try to send
 348          * directly, otherwise queue io_work. Also, only do that if we
 349          * are on the same cpu, so we don't introduce contention.
 350          */
 351         if (queue->io_cpu == raw_smp_processor_id() &&
 352             sync && empty && mutex_trylock(&queue->send_mutex)) {
 353                 nvme_tcp_send_all(queue);
 354                 mutex_unlock(&queue->send_mutex);
 355         }
 356
 357         if (last && nvme_tcp_queue_more(queue))
 358                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 359 }
 360
 361 static void nvme_tcp_process_req_list(struct nvme_tcp_queue *queue)
 362 {
 363         struct nvme_tcp_request *req;
 364         struct llist_node *node;
 365
 366         for (node = llist_del_all(&queue->req_list); node; node = node->next) {
 367                 req = llist_entry(node, struct nvme_tcp_request, lentry);
 368                 list_add(&req->entry, &queue->send_list);
 369         }
 370 }
 371
 372 static inline struct nvme_tcp_request *
 373 nvme_tcp_fetch_request(struct nvme_tcp_queue *queue)
 374 {
 375         struct nvme_tcp_request *req;
 376
 377         req = list_first_entry_or_null(&queue->send_list,
 378                         struct nvme_tcp_request, entry);
 379         if (!req) {
 380                 nvme_tcp_process_req_list(queue);
 381                 req = list_first_entry_or_null(&queue->send_list,
 382                                 struct nvme_tcp_request, entry);
 383                 if (unlikely(!req))
 384                         return NULL;
 385         }
 386
 387         list_del(&req->entry);
 388         return req;
 389 }
 390
 391 static inline void nvme_tcp_ddgst_final(struct ahash_request *hash,
 392                 __le32 *dgst)
 393 {
 394         ahash_request_set_crypt(hash, NULL, (u8 *)dgst, 0);
 395         crypto_ahash_final(hash);
 396 }
 397
 398 static inline void nvme_tcp_ddgst_update(struct ahash_request *hash,
 399                 struct page *page, off_t off, size_t len)
 400 {
 401         struct scatterlist sg;
 402
 403         sg_init_table(&sg, 1);
 404         sg_set_page(&sg, page, len, off);
 405         ahash_request_set_crypt(hash, &sg, NULL, len);
 406         crypto_ahash_update(hash);
 407 }
 408
 409 static inline void nvme_tcp_hdgst(struct ahash_request *hash,
 410                 void *pdu, size_t len)
 411 {
 412         struct scatterlist sg;
 413
 414         sg_init_one(&sg, pdu, len);
 415         ahash_request_set_crypt(hash, &sg, pdu + len, len);
 416         crypto_ahash_digest(hash);
 417 }
 418
 419 static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue,
 420                 void *pdu, size_t pdu_len)
 421 {
 422         struct nvme_tcp_hdr *hdr = pdu;
 423         __le32 recv_digest;
 424         __le32 exp_digest;
 425
 426         if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
 427                 dev_err(queue->ctrl->ctrl.device,
 428                         "queue %d: header digest flag is cleared\n",
 429                         nvme_tcp_queue_id(queue));
 430                 return -EPROTO;
 431         }
 432
 433         recv_digest = *(__le32 *)(pdu + hdr->hlen);
 434         nvme_tcp_hdgst(queue->rcv_hash, pdu, pdu_len);
 435         exp_digest = *(__le32 *)(pdu + hdr->hlen);
 436         if (recv_digest != exp_digest) {
 437                 dev_err(queue->ctrl->ctrl.device,
 438                         "header digest error: recv %#x expected %#x\n",
 439                         le32_to_cpu(recv_digest), le32_to_cpu(exp_digest));
 440                 return -EIO;
 441         }
 442
 443         return 0;
 444 }
 445
 446 static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu)
 447 {
 448         struct nvme_tcp_hdr *hdr = pdu;
 449         u8 digest_len = nvme_tcp_hdgst_len(queue);
 450         u32 len;
 451
 452         len = le32_to_cpu(hdr->plen) - hdr->hlen -
 453                 ((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0);
 454
 455         if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
 456                 dev_err(queue->ctrl->ctrl.device,
 457                         "queue %d: data digest flag is cleared\n",
 458                 nvme_tcp_queue_id(queue));
 459                 return -EPROTO;
 460         }
 461         crypto_ahash_init(queue->rcv_hash);
 462
 463         return 0;
 464 }
 465
 466 static void nvme_tcp_exit_request(struct blk_mq_tag_set *set,
 467                 struct request *rq, unsigned int hctx_idx)
 468 {
 469         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 470
 471         page_frag_free(req->pdu);
 472 }
 473
 474 static int nvme_tcp_init_request(struct blk_mq_tag_set *set,
 475                 struct request *rq, unsigned int hctx_idx,
 476                 unsigned int numa_node)
 477 {
 478         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(set->driver_data);
 479         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 480         struct nvme_tcp_cmd_pdu *pdu;
 481         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 482         struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx];
 483         u8 hdgst = nvme_tcp_hdgst_len(queue);
 484
 485         req->pdu = page_frag_alloc(&queue->pf_cache,
 486                 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
 487                 GFP_KERNEL | __GFP_ZERO);
 488         if (!req->pdu)
 489                 return -ENOMEM;
 490
 491         pdu = req->pdu;
 492         req->queue = queue;
 493         nvme_req(rq)->ctrl = &ctrl->ctrl;
 494         nvme_req(rq)->cmd = &pdu->cmd;
 495
 496         return 0;
 497 }
 498
 499 static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 500                 unsigned int hctx_idx)
 501 {
 502         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(data);
 503         struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1];
 504
 505         hctx->driver_data = queue;
 506         return 0;
 507 }
 508
 509 static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 510                 unsigned int hctx_idx)
 511 {
 512         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(data);
 513         struct nvme_tcp_queue *queue = &ctrl->queues[0];
 514
 515         hctx->driver_data = queue;
 516         return 0;
 517 }
 518
 519 static enum nvme_tcp_recv_state
 520 nvme_tcp_recv_state(struct nvme_tcp_queue *queue)
 521 {
 522         return  (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
 523                 (queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST :
 524                 NVME_TCP_RECV_DATA;
 525 }
 526
 527 static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue)
 528 {
 529         queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) +
 530                                 nvme_tcp_hdgst_len(queue);
 531         queue->pdu_offset = 0;
 532         queue->data_remaining = -1;
 533         queue->ddgst_remaining = 0;
 534 }
 535
 536 static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl)
 537 {
 538         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
 539                 return;
 540
 541         dev_warn(ctrl->device, "starting error recovery\n");
 542         queue_work(nvme_reset_wq, &to_tcp_ctrl(ctrl)->err_work);
 543 }
 544
 545 static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
 546                 struct nvme_completion *cqe)
 547 {
 548         struct nvme_tcp_request *req;
 549         struct request *rq;
 550
 551         rq = nvme_find_rq(nvme_tcp_tagset(queue), cqe->command_id);
 552         if (!rq) {
 553                 dev_err(queue->ctrl->ctrl.device,
 554                         "got bad cqe.command_id %#x on queue %d\n",
 555                         cqe->command_id, nvme_tcp_queue_id(queue));
 556                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 557                 return -EINVAL;
 558         }
 559
 560         req = blk_mq_rq_to_pdu(rq);
 561         if (req->status == cpu_to_le16(NVME_SC_SUCCESS))
 562                 req->status = cqe->status;
 563
 564         if (!nvme_try_complete_req(rq, req->status, cqe->result))
 565                 nvme_complete_rq(rq);
 566         queue->nr_cqe++;
 567
 568         return 0;
 569 }
 570
 571 static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue,
 572                 struct nvme_tcp_data_pdu *pdu)
 573 {
 574         struct request *rq;
 575
 576         rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
 577         if (!rq) {
 578                 dev_err(queue->ctrl->ctrl.device,
 579                         "got bad c2hdata.command_id %#x on queue %d\n",
 580                         pdu->command_id, nvme_tcp_queue_id(queue));
 581                 return -ENOENT;
 582         }
 583
 584         if (!blk_rq_payload_bytes(rq)) {
 585                 dev_err(queue->ctrl->ctrl.device,
 586                         "queue %d tag %#x unexpected data\n",
 587                         nvme_tcp_queue_id(queue), rq->tag);
 588                 return -EIO;
 589         }
 590
 591         queue->data_remaining = le32_to_cpu(pdu->data_length);
 592
 593         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS &&
 594             unlikely(!(pdu->hdr.flags & NVME_TCP_F_DATA_LAST))) {
 595                 dev_err(queue->ctrl->ctrl.device,
 596                         "queue %d tag %#x SUCCESS set but not last PDU\n",
 597                         nvme_tcp_queue_id(queue), rq->tag);
 598                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 599                 return -EPROTO;
 600         }
 601
 602         return 0;
 603 }
 604
 605 static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue,
 606                 struct nvme_tcp_rsp_pdu *pdu)
 607 {
 608         struct nvme_completion *cqe = &pdu->cqe;
 609         int ret = 0;
 610
 611         /*
 612          * AEN requests are special as they don't time out and can
 613          * survive any kind of queue freeze and often don't respond to
 614          * aborts.  We don't even bother to allocate a struct request
 615          * for them but rather special case them here.
 616          */
 617         if (unlikely(nvme_is_aen_req(nvme_tcp_queue_id(queue),
 618                                      cqe->command_id)))
 619                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
 620                                 &cqe->result);
 621         else
 622                 ret = nvme_tcp_process_nvme_cqe(queue, cqe);
 623
 624         return ret;
 625 }
 626
 627 static void nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req)
 628 {
 629         struct nvme_tcp_data_pdu *data = nvme_tcp_req_data_pdu(req);
 630         struct nvme_tcp_queue *queue = req->queue;
 631         struct request *rq = blk_mq_rq_from_pdu(req);
 632         u32 h2cdata_sent = req->pdu_len;
 633         u8 hdgst = nvme_tcp_hdgst_len(queue);
 634         u8 ddgst = nvme_tcp_ddgst_len(queue);
 635
 636         req->state = NVME_TCP_SEND_H2C_PDU;
 637         req->offset = 0;
 638         req->pdu_len = min(req->h2cdata_left, queue->maxh2cdata);
 639         req->pdu_sent = 0;
 640         req->h2cdata_left -= req->pdu_len;
 641         req->h2cdata_offset += h2cdata_sent;
 642
 643         memset(data, 0, sizeof(*data));
 644         data->hdr.type = nvme_tcp_h2c_data;
 645         if (!req->h2cdata_left)
 646                 data->hdr.flags = NVME_TCP_F_DATA_LAST;
 647         if (queue->hdr_digest)
 648                 data->hdr.flags |= NVME_TCP_F_HDGST;
 649         if (queue->data_digest)
 650                 data->hdr.flags |= NVME_TCP_F_DDGST;
 651         data->hdr.hlen = sizeof(*data);
 652         data->hdr.pdo = data->hdr.hlen + hdgst;
 653         data->hdr.plen =
 654                 cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst);
 655         data->ttag = req->ttag;
 656         data->command_id = nvme_cid(rq);
 657         data->data_offset = cpu_to_le32(req->h2cdata_offset);
 658         data->data_length = cpu_to_le32(req->pdu_len);
 659 }
 660
 661 static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue,
 662                 struct nvme_tcp_r2t_pdu *pdu)
 663 {
 664         struct nvme_tcp_request *req;
 665         struct request *rq;
 666         u32 r2t_length = le32_to_cpu(pdu->r2t_length);
 667         u32 r2t_offset = le32_to_cpu(pdu->r2t_offset);
 668
 669         rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
 670         if (!rq) {
 671                 dev_err(queue->ctrl->ctrl.device,
 672                         "got bad r2t.command_id %#x on queue %d\n",
 673                         pdu->command_id, nvme_tcp_queue_id(queue));
 674                 return -ENOENT;
 675         }
 676         req = blk_mq_rq_to_pdu(rq);
 677
 678         if (unlikely(!r2t_length)) {
 679                 dev_err(queue->ctrl->ctrl.device,
 680                         "req %d r2t len is %u, probably a bug...\n",
 681                         rq->tag, r2t_length);
 682                 return -EPROTO;
 683         }
 684
 685         if (unlikely(req->data_sent + r2t_length > req->data_len)) {
 686                 dev_err(queue->ctrl->ctrl.device,
 687                         "req %d r2t len %u exceeded data len %u (%zu sent)\n",
 688                         rq->tag, r2t_length, req->data_len, req->data_sent);
 689                 return -EPROTO;
 690         }
 691
 692         if (unlikely(r2t_offset < req->data_sent)) {
 693                 dev_err(queue->ctrl->ctrl.device,
 694                         "req %d unexpected r2t offset %u (expected %zu)\n",
 695                         rq->tag, r2t_offset, req->data_sent);
 696                 return -EPROTO;
 697         }
 698
 699         req->pdu_len = 0;
 700         req->h2cdata_left = r2t_length;
 701         req->h2cdata_offset = r2t_offset;
 702         req->ttag = pdu->ttag;
 703
 704         nvme_tcp_setup_h2c_data_pdu(req);
 705         nvme_tcp_queue_request(req, false, true);
 706
 707         return 0;
 708 }
 709
 710 static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 711                 unsigned int *offset, size_t *len)
 712 {
 713         struct nvme_tcp_hdr *hdr;
 714         char *pdu = queue->pdu;
 715         size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
 716         int ret;
 717
 718         ret = skb_copy_bits(skb, *offset,
 719                 &pdu[queue->pdu_offset], rcv_len);
 720         if (unlikely(ret))
 721                 return ret;
 722
 723         queue->pdu_remaining -= rcv_len;
 724         queue->pdu_offset += rcv_len;
 725         *offset += rcv_len;
 726         *len -= rcv_len;
 727         if (queue->pdu_remaining)
 728                 return 0;
 729
 730         hdr = queue->pdu;
 731         if (queue->hdr_digest) {
 732                 ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen);
 733                 if (unlikely(ret))
 734                         return ret;
 735         }
 736
 737
 738         if (queue->data_digest) {
 739                 ret = nvme_tcp_check_ddgst(queue, queue->pdu);
 740                 if (unlikely(ret))
 741                         return ret;
 742         }
 743
 744         switch (hdr->type) {
 745         case nvme_tcp_c2h_data:
 746                 return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
 747         case nvme_tcp_rsp:
 748                 nvme_tcp_init_recv_ctx(queue);
 749                 return nvme_tcp_handle_comp(queue, (void *)queue->pdu);
 750         case nvme_tcp_r2t:
 751                 nvme_tcp_init_recv_ctx(queue);
 752                 return nvme_tcp_handle_r2t(queue, (void *)queue->pdu);
 753         default:
 754                 dev_err(queue->ctrl->ctrl.device,
 755                         "unsupported pdu type (%d)\n", hdr->type);
 756                 return -EINVAL;
 757         }
 758 }
 759
 760 static inline void nvme_tcp_end_request(struct request *rq, u16 status)
 761 {
 762         union nvme_result res = {};
 763
 764         if (!nvme_try_complete_req(rq, cpu_to_le16(status << 1), res))
 765                 nvme_complete_rq(rq);
 766 }
 767
 768 static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 769                               unsigned int *offset, size_t *len)
 770 {
 771         struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
 772         struct request *rq =
 773                 nvme_cid_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
 774         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 775
 776         while (true) {
 777                 int recv_len, ret;
 778
 779                 recv_len = min_t(size_t, *len, queue->data_remaining);
 780                 if (!recv_len)
 781                         break;
 782
 783                 if (!iov_iter_count(&req->iter)) {
 784                         req->curr_bio = req->curr_bio->bi_next;
 785
 786                         /*
 787                          * If we don`t have any bios it means that controller
 788                          * sent more data than we requested, hence error
 789                          */
 790                         if (!req->curr_bio) {
 791                                 dev_err(queue->ctrl->ctrl.device,
 792                                         "queue %d no space in request %#x",
 793                                         nvme_tcp_queue_id(queue), rq->tag);
 794                                 nvme_tcp_init_recv_ctx(queue);
 795                                 return -EIO;
 796                         }
 797                         nvme_tcp_init_iter(req, ITER_DEST);
 798                 }
 799
 800                 /* we can read only from what is left in this bio */
 801                 recv_len = min_t(size_t, recv_len,
 802                                 iov_iter_count(&req->iter));
 803
 804                 if (queue->data_digest)
 805                         ret = skb_copy_and_hash_datagram_iter(skb, *offset,
 806                                 &req->iter, recv_len, queue->rcv_hash);
 807                 else
 808                         ret = skb_copy_datagram_iter(skb, *offset,
 809                                         &req->iter, recv_len);
 810                 if (ret) {
 811                         dev_err(queue->ctrl->ctrl.device,
 812                                 "queue %d failed to copy request %#x data",
 813                                 nvme_tcp_queue_id(queue), rq->tag);
 814                         return ret;
 815                 }
 816
 817                 *len -= recv_len;
 818                 *offset += recv_len;
 819                 queue->data_remaining -= recv_len;
 820         }
 821
 822         if (!queue->data_remaining) {
 823                 if (queue->data_digest) {
 824                         nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst);
 825                         queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
 826                 } else {
 827                         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
 828                                 nvme_tcp_end_request(rq,
 829                                                 le16_to_cpu(req->status));
 830                                 queue->nr_cqe++;
 831                         }
 832                         nvme_tcp_init_recv_ctx(queue);
 833                 }
 834         }
 835
 836         return 0;
 837 }
 838
 839 static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
 840                 struct sk_buff *skb, unsigned int *offset, size_t *len)
 841 {
 842         struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
 843         char *ddgst = (char *)&queue->recv_ddgst;
 844         size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
 845         off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
 846         int ret;
 847
 848         ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
 849         if (unlikely(ret))
 850                 return ret;
 851
 852         queue->ddgst_remaining -= recv_len;
 853         *offset += recv_len;
 854         *len -= recv_len;
 855         if (queue->ddgst_remaining)
 856                 return 0;
 857
 858         if (queue->recv_ddgst != queue->exp_ddgst) {
 859                 struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
 860                                         pdu->command_id);
 861                 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 862
 863                 req->status = cpu_to_le16(NVME_SC_DATA_XFER_ERROR);
 864
 865                 dev_err(queue->ctrl->ctrl.device,
 866                         "data digest error: recv %#x expected %#x\n",
 867                         le32_to_cpu(queue->recv_ddgst),
 868                         le32_to_cpu(queue->exp_ddgst));
 869         }
 870
 871         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
 872                 struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
 873                                         pdu->command_id);
 874                 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 875
 876                 nvme_tcp_end_request(rq, le16_to_cpu(req->status));
 877                 queue->nr_cqe++;
 878         }
 879
 880         nvme_tcp_init_recv_ctx(queue);
 881         return 0;
 882 }
 883
 884 static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
 885                              unsigned int offset, size_t len)
 886 {
 887         struct nvme_tcp_queue *queue = desc->arg.data;
 888         size_t consumed = len;
 889         int result;
 890
 891         if (unlikely(!queue->rd_enabled))
 892                 return -EFAULT;
 893
 894         while (len) {
 895                 switch (nvme_tcp_recv_state(queue)) {
 896                 case NVME_TCP_RECV_PDU:
 897                         result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
 898                         break;
 899                 case NVME_TCP_RECV_DATA:
 900                         result = nvme_tcp_recv_data(queue, skb, &offset, &len);
 901                         break;
 902                 case NVME_TCP_RECV_DDGST:
 903                         result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
 904                         break;
 905                 default:
 906                         result = -EFAULT;
 907                 }
 908                 if (result) {
 909                         dev_err(queue->ctrl->ctrl.device,
 910                                 "receive failed:  %d\n", result);
 911                         queue->rd_enabled = false;
 912                         nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 913                         return result;
 914                 }
 915         }
 916
 917         return consumed;
 918 }
 919
 920 static void nvme_tcp_data_ready(struct sock *sk)
 921 {
 922         struct nvme_tcp_queue *queue;
 923
 924         trace_sk_data_ready(sk);
 925
 926         read_lock_bh(&sk->sk_callback_lock);
 927         queue = sk->sk_user_data;
 928         if (likely(queue && queue->rd_enabled) &&
 929             !test_bit(NVME_TCP_Q_POLLING, &queue->flags))
 930                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 931         read_unlock_bh(&sk->sk_callback_lock);
 932 }
 933
 934 static void nvme_tcp_write_space(struct sock *sk)
 935 {
 936         struct nvme_tcp_queue *queue;
 937
 938         read_lock_bh(&sk->sk_callback_lock);
 939         queue = sk->sk_user_data;
 940         if (likely(queue && sk_stream_is_writeable(sk))) {
 941                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 942                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 943         }
 944         read_unlock_bh(&sk->sk_callback_lock);
 945 }
 946
 947 static void nvme_tcp_state_change(struct sock *sk)
 948 {
 949         struct nvme_tcp_queue *queue;
 950
 951         read_lock_bh(&sk->sk_callback_lock);
 952         queue = sk->sk_user_data;
 953         if (!queue)
 954                 goto done;
 955
 956         switch (sk->sk_state) {
 957         case TCP_CLOSE:
 958         case TCP_CLOSE_WAIT:
 959         case TCP_LAST_ACK:
 960         case TCP_FIN_WAIT1:
 961         case TCP_FIN_WAIT2:
 962                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 963                 break;
 964         default:
 965                 dev_info(queue->ctrl->ctrl.device,
 966                         "queue %d socket state %d\n",
 967                         nvme_tcp_queue_id(queue), sk->sk_state);
 968         }
 969
 970         queue->state_change(sk);
 971 done:
 972         read_unlock_bh(&sk->sk_callback_lock);
 973 }
 974
 975 static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
 976 {
 977         queue->request = NULL;
 978 }
 979
 980 static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
 981 {
 982         if (nvme_tcp_async_req(req)) {
 983                 union nvme_result res = {};
 984
 985                 nvme_complete_async_event(&req->queue->ctrl->ctrl,
 986                                 cpu_to_le16(NVME_SC_HOST_PATH_ERROR), &res);
 987         } else {
 988                 nvme_tcp_end_request(blk_mq_rq_from_pdu(req),
 989                                 NVME_SC_HOST_PATH_ERROR);
 990         }
 991 }
 992
 993 static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
 994 {
 995         struct nvme_tcp_queue *queue = req->queue;
 996         int req_data_len = req->data_len;
 997         u32 h2cdata_left = req->h2cdata_left;
 998
 999         while (true) {
1000                 struct page *page = nvme_tcp_req_cur_page(req);
1001                 size_t offset = nvme_tcp_req_cur_offset(req);
1002                 size_t len = nvme_tcp_req_cur_length(req);
1003                 bool last = nvme_tcp_pdu_last_send(req, len);
1004                 int req_data_sent = req->data_sent;
1005                 int ret, flags = MSG_DONTWAIT;
1006
1007                 if (last && !queue->data_digest && !nvme_tcp_queue_more(queue))
1008                         flags |= MSG_EOR;
1009                 else
1010                         flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
1011
1012                 if (sendpage_ok(page)) {
1013                         ret = kernel_sendpage(queue->sock, page, offset, len,
1014                                         flags);
1015                 } else {
1016                         ret = sock_no_sendpage(queue->sock, page, offset, len,
1017                                         flags);
1018                 }
1019                 if (ret <= 0)
1020                         return ret;
1021
1022                 if (queue->data_digest)
1023                         nvme_tcp_ddgst_update(queue->snd_hash, page,
1024                                         offset, ret);
1025
1026                 /*
1027                  * update the request iterator except for the last payload send
1028                  * in the request where we don't want to modify it as we may
1029                  * compete with the RX path completing the request.
1030                  */
1031                 if (req_data_sent + ret < req_data_len)
1032                         nvme_tcp_advance_req(req, ret);
1033
1034                 /* fully successful last send in current PDU */
1035                 if (last && ret == len) {
1036                         if (queue->data_digest) {
1037                                 nvme_tcp_ddgst_final(queue->snd_hash,
1038                                         &req->ddgst);
1039                                 req->state = NVME_TCP_SEND_DDGST;
1040                                 req->offset = 0;
1041                         } else {
1042                                 if (h2cdata_left)
1043                                         nvme_tcp_setup_h2c_data_pdu(req);
1044                                 else
1045                                         nvme_tcp_done_send_req(queue);
1046                         }
1047                         return 1;
1048                 }
1049         }
1050         return -EAGAIN;
1051 }
1052
1053 static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
1054 {
1055         struct nvme_tcp_queue *queue = req->queue;
1056         struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
1057         bool inline_data = nvme_tcp_has_inline_data(req);
1058         u8 hdgst = nvme_tcp_hdgst_len(queue);
1059         int len = sizeof(*pdu) + hdgst - req->offset;
1060         int flags = MSG_DONTWAIT;
1061         int ret;
1062
1063         if (inline_data || nvme_tcp_queue_more(queue))
1064                 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
1065         else
1066                 flags |= MSG_EOR;
1067
1068         if (queue->hdr_digest && !req->offset)
1069                 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
1070
1071         ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
1072                         offset_in_page(pdu) + req->offset, len,  flags);
1073         if (unlikely(ret <= 0))
1074                 return ret;
1075
1076         len -= ret;
1077         if (!len) {
1078                 if (inline_data) {
1079                         req->state = NVME_TCP_SEND_DATA;
1080                         if (queue->data_digest)
1081                                 crypto_ahash_init(queue->snd_hash);
1082                 } else {
1083                         nvme_tcp_done_send_req(queue);
1084                 }
1085                 return 1;
1086         }
1087         req->offset += ret;
1088
1089         return -EAGAIN;
1090 }
1091
1092 static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
1093 {
1094         struct nvme_tcp_queue *queue = req->queue;
1095         struct nvme_tcp_data_pdu *pdu = nvme_tcp_req_data_pdu(req);
1096         u8 hdgst = nvme_tcp_hdgst_len(queue);
1097         int len = sizeof(*pdu) - req->offset + hdgst;
1098         int ret;
1099
1100         if (queue->hdr_digest && !req->offset)
1101                 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
1102
1103         if (!req->h2cdata_left)
1104                 ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
1105                                 offset_in_page(pdu) + req->offset, len,
1106                                 MSG_DONTWAIT | MSG_MORE | MSG_SENDPAGE_NOTLAST);
1107         else
1108                 ret = sock_no_sendpage(queue->sock, virt_to_page(pdu),
1109                                 offset_in_page(pdu) + req->offset, len,
1110                                 MSG_DONTWAIT | MSG_MORE);
1111         if (unlikely(ret <= 0))
1112                 return ret;
1113
1114         len -= ret;
1115         if (!len) {
1116                 req->state = NVME_TCP_SEND_DATA;
1117                 if (queue->data_digest)
1118                         crypto_ahash_init(queue->snd_hash);
1119                 return 1;
1120         }
1121         req->offset += ret;
1122
1123         return -EAGAIN;
1124 }
1125
1126 static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
1127 {
1128         struct nvme_tcp_queue *queue = req->queue;
1129         size_t offset = req->offset;
1130         u32 h2cdata_left = req->h2cdata_left;
1131         int ret;
1132         struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1133         struct kvec iov = {
1134                 .iov_base = (u8 *)&req->ddgst + req->offset,
1135                 .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
1136         };
1137
1138         if (nvme_tcp_queue_more(queue))
1139                 msg.msg_flags |= MSG_MORE;
1140         else
1141                 msg.msg_flags |= MSG_EOR;
1142
1143         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1144         if (unlikely(ret <= 0))
1145                 return ret;
1146
1147         if (offset + ret == NVME_TCP_DIGEST_LENGTH) {
1148                 if (h2cdata_left)
1149                         nvme_tcp_setup_h2c_data_pdu(req);
1150                 else
1151                         nvme_tcp_done_send_req(queue);
1152                 return 1;
1153         }
1154
1155         req->offset += ret;
1156         return -EAGAIN;
1157 }
1158
1159 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
1160 {
1161         struct nvme_tcp_request *req;
1162         unsigned int noreclaim_flag;
1163         int ret = 1;
1164
1165         if (!queue->request) {
1166                 queue->request = nvme_tcp_fetch_request(queue);
1167                 if (!queue->request)
1168                         return 0;
1169         }
1170         req = queue->request;
1171
1172         noreclaim_flag = memalloc_noreclaim_save();
1173         if (req->state == NVME_TCP_SEND_CMD_PDU) {
1174                 ret = nvme_tcp_try_send_cmd_pdu(req);
1175                 if (ret <= 0)
1176                         goto done;
1177                 if (!nvme_tcp_has_inline_data(req))
1178                         goto out;
1179         }
1180
1181         if (req->state == NVME_TCP_SEND_H2C_PDU) {
1182                 ret = nvme_tcp_try_send_data_pdu(req);
1183                 if (ret <= 0)
1184                         goto done;
1185         }
1186
1187         if (req->state == NVME_TCP_SEND_DATA) {
1188                 ret = nvme_tcp_try_send_data(req);
1189                 if (ret <= 0)
1190                         goto done;
1191         }
1192
1193         if (req->state == NVME_TCP_SEND_DDGST)
1194                 ret = nvme_tcp_try_send_ddgst(req);
1195 done:
1196         if (ret == -EAGAIN) {
1197                 ret = 0;
1198         } else if (ret < 0) {
1199                 dev_err(queue->ctrl->ctrl.device,
1200                         "failed to send request %d\n", ret);
1201                 nvme_tcp_fail_request(queue->request);
1202                 nvme_tcp_done_send_req(queue);
1203         }
1204 out:
1205         memalloc_noreclaim_restore(noreclaim_flag);
1206         return ret;
1207 }
1208
1209 static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1210 {
1211         struct socket *sock = queue->sock;
1212         struct sock *sk = sock->sk;
1213         read_descriptor_t rd_desc;
1214         int consumed;
1215
1216         rd_desc.arg.data = queue;
1217         rd_desc.count = 1;
1218         lock_sock(sk);
1219         queue->nr_cqe = 0;
1220         consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1221         release_sock(sk);
1222         return consumed;
1223 }
1224
1225 static void nvme_tcp_io_work(struct work_struct *w)
1226 {
1227         struct nvme_tcp_queue *queue =
1228                 container_of(w, struct nvme_tcp_queue, io_work);
1229         unsigned long deadline = jiffies + msecs_to_jiffies(1);
1230
1231         do {
1232                 bool pending = false;
1233                 int result;
1234
1235                 if (mutex_trylock(&queue->send_mutex)) {
1236                         result = nvme_tcp_try_send(queue);
1237                         mutex_unlock(&queue->send_mutex);
1238                         if (result > 0)
1239                                 pending = true;
1240                         else if (unlikely(result < 0))
1241                                 break;
1242                 }
1243
1244                 result = nvme_tcp_try_recv(queue);
1245                 if (result > 0)
1246                         pending = true;
1247                 else if (unlikely(result < 0))
1248                         return;
1249
1250                 if (!pending || !queue->rd_enabled)
1251                         return;
1252
1253         } while (!time_after(jiffies, deadline)); /* quota is exhausted */
1254
1255         queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1256 }
1257
1258 static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue)
1259 {
1260         struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
1261
1262         ahash_request_free(queue->rcv_hash);
1263         ahash_request_free(queue->snd_hash);
1264         crypto_free_ahash(tfm);
1265 }
1266
1267 static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue)
1268 {
1269         struct crypto_ahash *tfm;
1270
1271         tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
1272         if (IS_ERR(tfm))
1273                 return PTR_ERR(tfm);
1274
1275         queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1276         if (!queue->snd_hash)
1277                 goto free_tfm;
1278         ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
1279
1280         queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1281         if (!queue->rcv_hash)
1282                 goto free_snd_hash;
1283         ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
1284
1285         return 0;
1286 free_snd_hash:
1287         ahash_request_free(queue->snd_hash);
1288 free_tfm:
1289         crypto_free_ahash(tfm);
1290         return -ENOMEM;
1291 }
1292
1293 static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1294 {
1295         struct nvme_tcp_request *async = &ctrl->async_req;
1296
1297         page_frag_free(async->pdu);
1298 }
1299
1300 static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1301 {
1302         struct nvme_tcp_queue *queue = &ctrl->queues[0];
1303         struct nvme_tcp_request *async = &ctrl->async_req;
1304         u8 hdgst = nvme_tcp_hdgst_len(queue);
1305
1306         async->pdu = page_frag_alloc(&queue->pf_cache,
1307                 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1308                 GFP_KERNEL | __GFP_ZERO);
1309         if (!async->pdu)
1310                 return -ENOMEM;
1311
1312         async->queue = &ctrl->queues[0];
1313         return 0;
1314 }
1315
1316 static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1317 {
1318         struct page *page;
1319         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1320         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1321         unsigned int noreclaim_flag;
1322
1323         if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1324                 return;
1325
1326         if (queue->hdr_digest || queue->data_digest)
1327                 nvme_tcp_free_crypto(queue);
1328
1329         if (queue->pf_cache.va) {
1330                 page = virt_to_head_page(queue->pf_cache.va);
1331                 __page_frag_cache_drain(page, queue->pf_cache.pagecnt_bias);
1332                 queue->pf_cache.va = NULL;
1333         }
1334
1335         noreclaim_flag = memalloc_noreclaim_save();
1336         sock_release(queue->sock);
1337         memalloc_noreclaim_restore(noreclaim_flag);
1338
1339         kfree(queue->pdu);
1340         mutex_destroy(&queue->send_mutex);
1341         mutex_destroy(&queue->queue_lock);
1342 }
1343
1344 static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1345 {
1346         struct nvme_tcp_icreq_pdu *icreq;
1347         struct nvme_tcp_icresp_pdu *icresp;
1348         struct msghdr msg = {};
1349         struct kvec iov;
1350         bool ctrl_hdgst, ctrl_ddgst;
1351         u32 maxh2cdata;
1352         int ret;
1353
1354         icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1355         if (!icreq)
1356                 return -ENOMEM;
1357
1358         icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1359         if (!icresp) {
1360                 ret = -ENOMEM;
1361                 goto free_icreq;
1362         }
1363
1364         icreq->hdr.type = nvme_tcp_icreq;
1365         icreq->hdr.hlen = sizeof(*icreq);
1366         icreq->hdr.pdo = 0;
1367         icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1368         icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1369         icreq->maxr2t = 0; /* single inflight r2t supported */
1370         icreq->hpda = 0; /* no alignment constraint */
1371         if (queue->hdr_digest)
1372                 icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1373         if (queue->data_digest)
1374                 icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1375
1376         iov.iov_base = icreq;
1377         iov.iov_len = sizeof(*icreq);
1378         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1379         if (ret < 0)
1380                 goto free_icresp;
1381
1382         memset(&msg, 0, sizeof(msg));
1383         iov.iov_base = icresp;
1384         iov.iov_len = sizeof(*icresp);
1385         ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1386                         iov.iov_len, msg.msg_flags);
1387         if (ret < 0)
1388                 goto free_icresp;
1389
1390         ret = -EINVAL;
1391         if (icresp->hdr.type != nvme_tcp_icresp) {
1392                 pr_err("queue %d: bad type returned %d\n",
1393                         nvme_tcp_queue_id(queue), icresp->hdr.type);
1394                 goto free_icresp;
1395         }
1396
1397         if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1398                 pr_err("queue %d: bad pdu length returned %d\n",
1399                         nvme_tcp_queue_id(queue), icresp->hdr.plen);
1400                 goto free_icresp;
1401         }
1402
1403         if (icresp->pfv != NVME_TCP_PFV_1_0) {
1404                 pr_err("queue %d: bad pfv returned %d\n",
1405                         nvme_tcp_queue_id(queue), icresp->pfv);
1406                 goto free_icresp;
1407         }
1408
1409         ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1410         if ((queue->data_digest && !ctrl_ddgst) ||
1411             (!queue->data_digest && ctrl_ddgst)) {
1412                 pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1413                         nvme_tcp_queue_id(queue),
1414                         queue->data_digest ? "enabled" : "disabled",
1415                         ctrl_ddgst ? "enabled" : "disabled");
1416                 goto free_icresp;
1417         }
1418
1419         ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1420         if ((queue->hdr_digest && !ctrl_hdgst) ||
1421             (!queue->hdr_digest && ctrl_hdgst)) {
1422                 pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1423                         nvme_tcp_queue_id(queue),
1424                         queue->hdr_digest ? "enabled" : "disabled",
1425                         ctrl_hdgst ? "enabled" : "disabled");
1426                 goto free_icresp;
1427         }
1428
1429         if (icresp->cpda != 0) {
1430                 pr_err("queue %d: unsupported cpda returned %d\n",
1431                         nvme_tcp_queue_id(queue), icresp->cpda);
1432                 goto free_icresp;
1433         }
1434
1435         maxh2cdata = le32_to_cpu(icresp->maxdata);
1436         if ((maxh2cdata % 4) || (maxh2cdata < NVME_TCP_MIN_MAXH2CDATA)) {
1437                 pr_err("queue %d: invalid maxh2cdata returned %u\n",
1438                        nvme_tcp_queue_id(queue), maxh2cdata);
1439                 goto free_icresp;
1440         }
1441         queue->maxh2cdata = maxh2cdata;
1442
1443         ret = 0;
1444 free_icresp:
1445         kfree(icresp);
1446 free_icreq:
1447         kfree(icreq);
1448         return ret;
1449 }
1450
1451 static bool nvme_tcp_admin_queue(struct nvme_tcp_queue *queue)
1452 {
1453         return nvme_tcp_queue_id(queue) == 0;
1454 }
1455
1456 static bool nvme_tcp_default_queue(struct nvme_tcp_queue *queue)
1457 {
1458         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1459         int qid = nvme_tcp_queue_id(queue);
1460
1461         return !nvme_tcp_admin_queue(queue) &&
1462                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT];
1463 }
1464
1465 static bool nvme_tcp_read_queue(struct nvme_tcp_queue *queue)
1466 {
1467         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1468         int qid = nvme_tcp_queue_id(queue);
1469
1470         return !nvme_tcp_admin_queue(queue) &&
1471                 !nvme_tcp_default_queue(queue) &&
1472                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1473                           ctrl->io_queues[HCTX_TYPE_READ];
1474 }
1475
1476 static bool nvme_tcp_poll_queue(struct nvme_tcp_queue *queue)
1477 {
1478         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1479         int qid = nvme_tcp_queue_id(queue);
1480
1481         return !nvme_tcp_admin_queue(queue) &&
1482                 !nvme_tcp_default_queue(queue) &&
1483                 !nvme_tcp_read_queue(queue) &&
1484                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1485                           ctrl->io_queues[HCTX_TYPE_READ] +
1486                           ctrl->io_queues[HCTX_TYPE_POLL];
1487 }
1488
1489 static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
1490 {
1491         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1492         int qid = nvme_tcp_queue_id(queue);
1493         int n = 0;
1494
1495         if (nvme_tcp_default_queue(queue))
1496                 n = qid - 1;
1497         else if (nvme_tcp_read_queue(queue))
1498                 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] - 1;
1499         else if (nvme_tcp_poll_queue(queue))
1500                 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] -
1501                                 ctrl->io_queues[HCTX_TYPE_READ] - 1;
1502         queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
1503 }
1504
1505 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl, int qid)
1506 {
1507         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1508         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1509         int ret, rcv_pdu_size;
1510
1511         mutex_init(&queue->queue_lock);
1512         queue->ctrl = ctrl;
1513         init_llist_head(&queue->req_list);
1514         INIT_LIST_HEAD(&queue->send_list);
1515         mutex_init(&queue->send_mutex);
1516         INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1517
1518         if (qid > 0)
1519                 queue->cmnd_capsule_len = nctrl->ioccsz * 16;
1520         else
1521                 queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1522                                                 NVME_TCP_ADMIN_CCSZ;
1523
1524         ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM,
1525                         IPPROTO_TCP, &queue->sock);
1526         if (ret) {
1527                 dev_err(nctrl->device,
1528                         "failed to create socket: %d\n", ret);
1529                 goto err_destroy_mutex;
1530         }
1531
1532         nvme_tcp_reclassify_socket(queue->sock);
1533
1534         /* Single syn retry */
1535         tcp_sock_set_syncnt(queue->sock->sk, 1);
1536
1537         /* Set TCP no delay */
1538         tcp_sock_set_nodelay(queue->sock->sk);
1539
1540         /*
1541          * Cleanup whatever is sitting in the TCP transmit queue on socket
1542          * close. This is done to prevent stale data from being sent should
1543          * the network connection be restored before TCP times out.
1544          */
1545         sock_no_linger(queue->sock->sk);
1546
1547         if (so_priority > 0)
1548                 sock_set_priority(queue->sock->sk, so_priority);
1549
1550         /* Set socket type of service */
1551         if (nctrl->opts->tos >= 0)
1552                 ip_sock_set_tos(queue->sock->sk, nctrl->opts->tos);
1553
1554         /* Set 10 seconds timeout for icresp recvmsg */
1555         queue->sock->sk->sk_rcvtimeo = 10 * HZ;
1556
1557         queue->sock->sk->sk_allocation = GFP_ATOMIC;
1558         queue->sock->sk->sk_use_task_frag = false;
1559         nvme_tcp_set_queue_io_cpu(queue);
1560         queue->request = NULL;
1561         queue->data_remaining = 0;
1562         queue->ddgst_remaining = 0;
1563         queue->pdu_remaining = 0;
1564         queue->pdu_offset = 0;
1565         sk_set_memalloc(queue->sock->sk);
1566
1567         if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
1568                 ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1569                         sizeof(ctrl->src_addr));
1570                 if (ret) {
1571                         dev_err(nctrl->device,
1572                                 "failed to bind queue %d socket %d\n",
1573                                 qid, ret);
1574                         goto err_sock;
1575                 }
1576         }
1577
1578         if (nctrl->opts->mask & NVMF_OPT_HOST_IFACE) {
1579                 char *iface = nctrl->opts->host_iface;
1580                 sockptr_t optval = KERNEL_SOCKPTR(iface);
1581
1582                 ret = sock_setsockopt(queue->sock, SOL_SOCKET, SO_BINDTODEVICE,
1583                                       optval, strlen(iface));
1584                 if (ret) {
1585                         dev_err(nctrl->device,
1586                           "failed to bind to interface %s queue %d err %d\n",
1587                           iface, qid, ret);
1588                         goto err_sock;
1589                 }
1590         }
1591
1592         queue->hdr_digest = nctrl->opts->hdr_digest;
1593         queue->data_digest = nctrl->opts->data_digest;
1594         if (queue->hdr_digest || queue->data_digest) {
1595                 ret = nvme_tcp_alloc_crypto(queue);
1596                 if (ret) {
1597                         dev_err(nctrl->device,
1598                                 "failed to allocate queue %d crypto\n", qid);
1599                         goto err_sock;
1600                 }
1601         }
1602
1603         rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1604                         nvme_tcp_hdgst_len(queue);
1605         queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1606         if (!queue->pdu) {
1607                 ret = -ENOMEM;
1608                 goto err_crypto;
1609         }
1610
1611         dev_dbg(nctrl->device, "connecting queue %d\n",
1612                         nvme_tcp_queue_id(queue));
1613
1614         ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1615                 sizeof(ctrl->addr), 0);
1616         if (ret) {
1617                 dev_err(nctrl->device,
1618                         "failed to connect socket: %d\n", ret);
1619                 goto err_rcv_pdu;
1620         }
1621
1622         ret = nvme_tcp_init_connection(queue);
1623         if (ret)
1624                 goto err_init_connect;
1625
1626         set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1627
1628         return 0;
1629
1630 err_init_connect:
1631         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1632 err_rcv_pdu:
1633         kfree(queue->pdu);
1634 err_crypto:
1635         if (queue->hdr_digest || queue->data_digest)
1636                 nvme_tcp_free_crypto(queue);
1637 err_sock:
1638         sock_release(queue->sock);
1639         queue->sock = NULL;
1640 err_destroy_mutex:
1641         mutex_destroy(&queue->send_mutex);
1642         mutex_destroy(&queue->queue_lock);
1643         return ret;
1644 }
1645
1646 static void nvme_tcp_restore_sock_ops(struct nvme_tcp_queue *queue)
1647 {
1648         struct socket *sock = queue->sock;
1649
1650         write_lock_bh(&sock->sk->sk_callback_lock);
1651         sock->sk->sk_user_data  = NULL;
1652         sock->sk->sk_data_ready = queue->data_ready;
1653         sock->sk->sk_state_change = queue->state_change;
1654         sock->sk->sk_write_space  = queue->write_space;
1655         write_unlock_bh(&sock->sk->sk_callback_lock);
1656 }
1657
1658 static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1659 {
1660         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1661         nvme_tcp_restore_sock_ops(queue);
1662         cancel_work_sync(&queue->io_work);
1663 }
1664
1665 static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1666 {
1667         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1668         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1669
1670         if (!test_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1671                 return;
1672
1673         mutex_lock(&queue->queue_lock);
1674         if (test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1675                 __nvme_tcp_stop_queue(queue);
1676         mutex_unlock(&queue->queue_lock);
1677 }
1678
1679 static void nvme_tcp_setup_sock_ops(struct nvme_tcp_queue *queue)
1680 {
1681         write_lock_bh(&queue->sock->sk->sk_callback_lock);
1682         queue->sock->sk->sk_user_data = queue;
1683         queue->state_change = queue->sock->sk->sk_state_change;
1684         queue->data_ready = queue->sock->sk->sk_data_ready;
1685         queue->write_space = queue->sock->sk->sk_write_space;
1686         queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1687         queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1688         queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1689 #ifdef CONFIG_NET_RX_BUSY_POLL
1690         queue->sock->sk->sk_ll_usec = 1;
1691 #endif
1692         write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1693 }
1694
1695 static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1696 {
1697         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1698         struct nvme_tcp_queue *queue = &ctrl->queues[idx];
1699         int ret;
1700
1701         queue->rd_enabled = true;
1702         nvme_tcp_init_recv_ctx(queue);
1703         nvme_tcp_setup_sock_ops(queue);
1704
1705         if (idx)
1706                 ret = nvmf_connect_io_queue(nctrl, idx);
1707         else
1708                 ret = nvmf_connect_admin_queue(nctrl);
1709
1710         if (!ret) {
1711                 set_bit(NVME_TCP_Q_LIVE, &queue->flags);
1712         } else {
1713                 if (test_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1714                         __nvme_tcp_stop_queue(queue);
1715                 dev_err(nctrl->device,
1716                         "failed to connect queue: %d ret=%d\n", idx, ret);
1717         }
1718         return ret;
1719 }
1720
1721 static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
1722 {
1723         if (to_tcp_ctrl(ctrl)->async_req.pdu) {
1724                 cancel_work_sync(&ctrl->async_event_work);
1725                 nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
1726                 to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
1727         }
1728
1729         nvme_tcp_free_queue(ctrl, 0);
1730 }
1731
1732 static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
1733 {
1734         int i;
1735
1736         for (i = 1; i < ctrl->queue_count; i++)
1737                 nvme_tcp_free_queue(ctrl, i);
1738 }
1739
1740 static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
1741 {
1742         int i;
1743
1744         for (i = 1; i < ctrl->queue_count; i++)
1745                 nvme_tcp_stop_queue(ctrl, i);
1746 }
1747
1748 static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl,
1749                                     int first, int last)
1750 {
1751         int i, ret;
1752
1753         for (i = first; i < last; i++) {
1754                 ret = nvme_tcp_start_queue(ctrl, i);
1755                 if (ret)
1756                         goto out_stop_queues;
1757         }
1758
1759         return 0;
1760
1761 out_stop_queues:
1762         for (i--; i >= first; i--)
1763                 nvme_tcp_stop_queue(ctrl, i);
1764         return ret;
1765 }
1766
1767 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
1768 {
1769         int ret;
1770
1771         ret = nvme_tcp_alloc_queue(ctrl, 0);
1772         if (ret)
1773                 return ret;
1774
1775         ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
1776         if (ret)
1777                 goto out_free_queue;
1778
1779         return 0;
1780
1781 out_free_queue:
1782         nvme_tcp_free_queue(ctrl, 0);
1783         return ret;
1784 }
1785
1786 static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1787 {
1788         int i, ret;
1789
1790         for (i = 1; i < ctrl->queue_count; i++) {
1791                 ret = nvme_tcp_alloc_queue(ctrl, i);
1792                 if (ret)
1793                         goto out_free_queues;
1794         }
1795
1796         return 0;
1797
1798 out_free_queues:
1799         for (i--; i >= 1; i--)
1800                 nvme_tcp_free_queue(ctrl, i);
1801
1802         return ret;
1803 }
1804
1805 static unsigned int nvme_tcp_nr_io_queues(struct nvme_ctrl *ctrl)
1806 {
1807         unsigned int nr_io_queues;
1808
1809         nr_io_queues = min(ctrl->opts->nr_io_queues, num_online_cpus());
1810         nr_io_queues += min(ctrl->opts->nr_write_queues, num_online_cpus());
1811         nr_io_queues += min(ctrl->opts->nr_poll_queues, num_online_cpus());
1812
1813         return nr_io_queues;
1814 }
1815
1816 static void nvme_tcp_set_io_queues(struct nvme_ctrl *nctrl,
1817                 unsigned int nr_io_queues)
1818 {
1819         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1820         struct nvmf_ctrl_options *opts = nctrl->opts;
1821
1822         if (opts->nr_write_queues && opts->nr_io_queues < nr_io_queues) {
1823                 /*
1824                  * separate read/write queues
1825                  * hand out dedicated default queues only after we have
1826                  * sufficient read queues.
1827                  */
1828                 ctrl->io_queues[HCTX_TYPE_READ] = opts->nr_io_queues;
1829                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
1830                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1831                         min(opts->nr_write_queues, nr_io_queues);
1832                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1833         } else {
1834                 /*
1835                  * shared read/write queues
1836                  * either no write queues were requested, or we don't have
1837                  * sufficient queue count to have dedicated default queues.
1838                  */
1839                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1840                         min(opts->nr_io_queues, nr_io_queues);
1841                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1842         }
1843
1844         if (opts->nr_poll_queues && nr_io_queues) {
1845                 /* map dedicated poll queues only if we have queues left */
1846                 ctrl->io_queues[HCTX_TYPE_POLL] =
1847                         min(opts->nr_poll_queues, nr_io_queues);
1848         }
1849 }
1850
1851 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1852 {
1853         unsigned int nr_io_queues;
1854         int ret;
1855
1856         nr_io_queues = nvme_tcp_nr_io_queues(ctrl);
1857         ret = nvme_set_queue_count(ctrl, &nr_io_queues);
1858         if (ret)
1859                 return ret;
1860
1861         if (nr_io_queues == 0) {
1862                 dev_err(ctrl->device,
1863                         "unable to set any I/O queues\n");
1864                 return -ENOMEM;
1865         }
1866
1867         ctrl->queue_count = nr_io_queues + 1;
1868         dev_info(ctrl->device,
1869                 "creating %d I/O queues.\n", nr_io_queues);
1870
1871         nvme_tcp_set_io_queues(ctrl, nr_io_queues);
1872
1873         return __nvme_tcp_alloc_io_queues(ctrl);
1874 }
1875
1876 static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove)
1877 {
1878         nvme_tcp_stop_io_queues(ctrl);
1879         if (remove)
1880                 nvme_remove_io_tag_set(ctrl);
1881         nvme_tcp_free_io_queues(ctrl);
1882 }
1883
1884 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
1885 {
1886         int ret, nr_queues;
1887
1888         ret = nvme_tcp_alloc_io_queues(ctrl);
1889         if (ret)
1890                 return ret;
1891
1892         if (new) {
1893                 ret = nvme_alloc_io_tag_set(ctrl, &to_tcp_ctrl(ctrl)->tag_set,
1894                                 &nvme_tcp_mq_ops,
1895                                 ctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2,
1896                                 sizeof(struct nvme_tcp_request));
1897                 if (ret)
1898                         goto out_free_io_queues;
1899         }
1900
1901         /*
1902          * Only start IO queues for which we have allocated the tagset
1903          * and limitted it to the available queues. On reconnects, the
1904          * queue number might have changed.
1905          */
1906         nr_queues = min(ctrl->tagset->nr_hw_queues + 1, ctrl->queue_count);
1907         ret = nvme_tcp_start_io_queues(ctrl, 1, nr_queues);
1908         if (ret)
1909                 goto out_cleanup_connect_q;
1910
1911         if (!new) {
1912                 nvme_unquiesce_io_queues(ctrl);
1913                 if (!nvme_wait_freeze_timeout(ctrl, NVME_IO_TIMEOUT)) {
1914                         /*
1915                          * If we timed out waiting for freeze we are likely to
1916                          * be stuck.  Fail the controller initialization just
1917                          * to be safe.
1918                          */
1919                         ret = -ENODEV;
1920                         goto out_wait_freeze_timed_out;
1921                 }
1922                 blk_mq_update_nr_hw_queues(ctrl->tagset,
1923                         ctrl->queue_count - 1);
1924                 nvme_unfreeze(ctrl);
1925         }
1926
1927         /*
1928          * If the number of queues has increased (reconnect case)
1929          * start all new queues now.
1930          */
1931         ret = nvme_tcp_start_io_queues(ctrl, nr_queues,
1932                                        ctrl->tagset->nr_hw_queues + 1);
1933         if (ret)
1934                 goto out_wait_freeze_timed_out;
1935
1936         return 0;
1937
1938 out_wait_freeze_timed_out:
1939         nvme_quiesce_io_queues(ctrl);
1940         nvme_sync_io_queues(ctrl);
1941         nvme_tcp_stop_io_queues(ctrl);
1942 out_cleanup_connect_q:
1943         nvme_cancel_tagset(ctrl);
1944         if (new)
1945                 nvme_remove_io_tag_set(ctrl);
1946 out_free_io_queues:
1947         nvme_tcp_free_io_queues(ctrl);
1948         return ret;
1949 }
1950
1951 static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove)
1952 {
1953         nvme_tcp_stop_queue(ctrl, 0);
1954         if (remove)
1955                 nvme_remove_admin_tag_set(ctrl);
1956         nvme_tcp_free_admin_queue(ctrl);
1957 }
1958
1959 static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
1960 {
1961         int error;
1962
1963         error = nvme_tcp_alloc_admin_queue(ctrl);
1964         if (error)
1965                 return error;
1966
1967         if (new) {
1968                 error = nvme_alloc_admin_tag_set(ctrl,
1969                                 &to_tcp_ctrl(ctrl)->admin_tag_set,
1970                                 &nvme_tcp_admin_mq_ops,
1971                                 sizeof(struct nvme_tcp_request));
1972                 if (error)
1973                         goto out_free_queue;
1974         }
1975
1976         error = nvme_tcp_start_queue(ctrl, 0);
1977         if (error)
1978                 goto out_cleanup_tagset;
1979
1980         error = nvme_enable_ctrl(ctrl);
1981         if (error)
1982                 goto out_stop_queue;
1983
1984         nvme_unquiesce_admin_queue(ctrl);
1985
1986         error = nvme_init_ctrl_finish(ctrl, false);
1987         if (error)
1988                 goto out_quiesce_queue;
1989
1990         return 0;
1991
1992 out_quiesce_queue:
1993         nvme_quiesce_admin_queue(ctrl);
1994         blk_sync_queue(ctrl->admin_q);
1995 out_stop_queue:
1996         nvme_tcp_stop_queue(ctrl, 0);
1997         nvme_cancel_admin_tagset(ctrl);
1998 out_cleanup_tagset:
1999         if (new)
2000                 nvme_remove_admin_tag_set(ctrl);
2001 out_free_queue:
2002         nvme_tcp_free_admin_queue(ctrl);
2003         return error;
2004 }
2005
2006 static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
2007                 bool remove)
2008 {
2009         nvme_quiesce_admin_queue(ctrl);
2010         blk_sync_queue(ctrl->admin_q);
2011         nvme_tcp_stop_queue(ctrl, 0);
2012         nvme_cancel_admin_tagset(ctrl);
2013         if (remove)
2014                 nvme_unquiesce_admin_queue(ctrl);
2015         nvme_tcp_destroy_admin_queue(ctrl, remove);
2016 }
2017
2018 static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
2019                 bool remove)
2020 {
2021         if (ctrl->queue_count <= 1)
2022                 return;
2023         nvme_quiesce_admin_queue(ctrl);
2024         nvme_start_freeze(ctrl);
2025         nvme_quiesce_io_queues(ctrl);
2026         nvme_sync_io_queues(ctrl);
2027         nvme_tcp_stop_io_queues(ctrl);
2028         nvme_cancel_tagset(ctrl);
2029         if (remove)
2030                 nvme_unquiesce_io_queues(ctrl);
2031         nvme_tcp_destroy_io_queues(ctrl, remove);
2032 }
2033
2034 static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
2035 {
2036         /* If we are resetting/deleting then do nothing */
2037         if (ctrl->state != NVME_CTRL_CONNECTING) {
2038                 WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW ||
2039                         ctrl->state == NVME_CTRL_LIVE);
2040                 return;
2041         }
2042
2043         if (nvmf_should_reconnect(ctrl)) {
2044                 dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
2045                         ctrl->opts->reconnect_delay);
2046                 queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
2047                                 ctrl->opts->reconnect_delay * HZ);
2048         } else {
2049                 dev_info(ctrl->device, "Removing controller...\n");
2050                 nvme_delete_ctrl(ctrl);
2051         }
2052 }
2053
2054 static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
2055 {
2056         struct nvmf_ctrl_options *opts = ctrl->opts;
2057         int ret;
2058
2059         ret = nvme_tcp_configure_admin_queue(ctrl, new);
2060         if (ret)
2061                 return ret;
2062
2063         if (ctrl->icdoff) {
2064                 ret = -EOPNOTSUPP;
2065                 dev_err(ctrl->device, "icdoff is not supported!\n");
2066                 goto destroy_admin;
2067         }
2068
2069         if (!nvme_ctrl_sgl_supported(ctrl)) {
2070                 ret = -EOPNOTSUPP;
2071                 dev_err(ctrl->device, "Mandatory sgls are not supported!\n");
2072                 goto destroy_admin;
2073         }
2074
2075         if (opts->queue_size > ctrl->sqsize + 1)
2076                 dev_warn(ctrl->device,
2077                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
2078                         opts->queue_size, ctrl->sqsize + 1);
2079
2080         if (ctrl->sqsize + 1 > ctrl->maxcmd) {
2081                 dev_warn(ctrl->device,
2082                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
2083                         ctrl->sqsize + 1, ctrl->maxcmd);
2084                 ctrl->sqsize = ctrl->maxcmd - 1;
2085         }
2086
2087         if (ctrl->queue_count > 1) {
2088                 ret = nvme_tcp_configure_io_queues(ctrl, new);
2089                 if (ret)
2090                         goto destroy_admin;
2091         }
2092
2093         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
2094                 /*
2095                  * state change failure is ok if we started ctrl delete,
2096                  * unless we're during creation of a new controller to
2097                  * avoid races with teardown flow.
2098                  */
2099                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2100                              ctrl->state != NVME_CTRL_DELETING_NOIO);
2101                 WARN_ON_ONCE(new);
2102                 ret = -EINVAL;
2103                 goto destroy_io;
2104         }
2105
2106         nvme_start_ctrl(ctrl);
2107         return 0;
2108
2109 destroy_io:
2110         if (ctrl->queue_count > 1) {
2111                 nvme_quiesce_io_queues(ctrl);
2112                 nvme_sync_io_queues(ctrl);
2113                 nvme_tcp_stop_io_queues(ctrl);
2114                 nvme_cancel_tagset(ctrl);
2115                 nvme_tcp_destroy_io_queues(ctrl, new);
2116         }
2117 destroy_admin:
2118         nvme_quiesce_admin_queue(ctrl);
2119         blk_sync_queue(ctrl->admin_q);
2120         nvme_tcp_stop_queue(ctrl, 0);
2121         nvme_cancel_admin_tagset(ctrl);
2122         nvme_tcp_destroy_admin_queue(ctrl, new);
2123         return ret;
2124 }
2125
2126 static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
2127 {
2128         struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
2129                         struct nvme_tcp_ctrl, connect_work);
2130         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2131
2132         ++ctrl->nr_reconnects;
2133
2134         if (nvme_tcp_setup_ctrl(ctrl, false))
2135                 goto requeue;
2136
2137         dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n",
2138                         ctrl->nr_reconnects);
2139
2140         ctrl->nr_reconnects = 0;
2141
2142         return;
2143
2144 requeue:
2145         dev_info(ctrl->device, "Failed reconnect attempt %d\n",
2146                         ctrl->nr_reconnects);
2147         nvme_tcp_reconnect_or_remove(ctrl);
2148 }
2149
2150 static void nvme_tcp_error_recovery_work(struct work_struct *work)
2151 {
2152         struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
2153                                 struct nvme_tcp_ctrl, err_work);
2154         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2155
2156         nvme_stop_keep_alive(ctrl);
2157         flush_work(&ctrl->async_event_work);
2158         nvme_tcp_teardown_io_queues(ctrl, false);
2159         /* unquiesce to fail fast pending requests */
2160         nvme_unquiesce_io_queues(ctrl);
2161         nvme_tcp_teardown_admin_queue(ctrl, false);
2162         nvme_unquiesce_admin_queue(ctrl);
2163         nvme_auth_stop(ctrl);
2164
2165         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2166                 /* state change failure is ok if we started ctrl delete */
2167                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2168                              ctrl->state != NVME_CTRL_DELETING_NOIO);
2169                 return;
2170         }
2171
2172         nvme_tcp_reconnect_or_remove(ctrl);
2173 }
2174
2175 static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
2176 {
2177         nvme_tcp_teardown_io_queues(ctrl, shutdown);
2178         nvme_quiesce_admin_queue(ctrl);
2179         nvme_disable_ctrl(ctrl, shutdown);
2180         nvme_tcp_teardown_admin_queue(ctrl, shutdown);
2181 }
2182
2183 static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
2184 {
2185         nvme_tcp_teardown_ctrl(ctrl, true);
2186 }
2187
2188 static void nvme_reset_ctrl_work(struct work_struct *work)
2189 {
2190         struct nvme_ctrl *ctrl =
2191                 container_of(work, struct nvme_ctrl, reset_work);
2192
2193         nvme_stop_ctrl(ctrl);
2194         nvme_tcp_teardown_ctrl(ctrl, false);
2195
2196         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2197                 /* state change failure is ok if we started ctrl delete */
2198                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2199                              ctrl->state != NVME_CTRL_DELETING_NOIO);
2200                 return;
2201         }
2202
2203         if (nvme_tcp_setup_ctrl(ctrl, false))
2204                 goto out_fail;
2205
2206         return;
2207
2208 out_fail:
2209         ++ctrl->nr_reconnects;
2210         nvme_tcp_reconnect_or_remove(ctrl);
2211 }
2212
2213 static void nvme_tcp_stop_ctrl(struct nvme_ctrl *ctrl)
2214 {
2215         flush_work(&to_tcp_ctrl(ctrl)->err_work);
2216         cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
2217 }
2218
2219 static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
2220 {
2221         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
2222
2223         if (list_empty(&ctrl->list))
2224                 goto free_ctrl;
2225
2226         mutex_lock(&nvme_tcp_ctrl_mutex);
2227         list_del(&ctrl->list);
2228         mutex_unlock(&nvme_tcp_ctrl_mutex);
2229
2230         nvmf_free_options(nctrl->opts);
2231 free_ctrl:
2232         kfree(ctrl->queues);
2233         kfree(ctrl);
2234 }
2235
2236 static void nvme_tcp_set_sg_null(struct nvme_command *c)
2237 {
2238         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2239
2240         sg->addr = 0;
2241         sg->length = 0;
2242         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2243                         NVME_SGL_FMT_TRANSPORT_A;
2244 }
2245
2246 static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
2247                 struct nvme_command *c, u32 data_len)
2248 {
2249         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2250
2251         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
2252         sg->length = cpu_to_le32(data_len);
2253         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
2254 }
2255
2256 static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
2257                 u32 data_len)
2258 {
2259         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2260
2261         sg->addr = 0;
2262         sg->length = cpu_to_le32(data_len);
2263         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2264                         NVME_SGL_FMT_TRANSPORT_A;
2265 }
2266
2267 static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
2268 {
2269         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2270         struct nvme_tcp_queue *queue = &ctrl->queues[0];
2271         struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2272         struct nvme_command *cmd = &pdu->cmd;
2273         u8 hdgst = nvme_tcp_hdgst_len(queue);
2274
2275         memset(pdu, 0, sizeof(*pdu));
2276         pdu->hdr.type = nvme_tcp_cmd;
2277         if (queue->hdr_digest)
2278                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2279         pdu->hdr.hlen = sizeof(*pdu);
2280         pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2281
2282         cmd->common.opcode = nvme_admin_async_event;
2283         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2284         cmd->common.flags |= NVME_CMD_SGL_METABUF;
2285         nvme_tcp_set_sg_null(cmd);
2286
2287         ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2288         ctrl->async_req.offset = 0;
2289         ctrl->async_req.curr_bio = NULL;
2290         ctrl->async_req.data_len = 0;
2291
2292         nvme_tcp_queue_request(&ctrl->async_req, true, true);
2293 }
2294
2295 static void nvme_tcp_complete_timed_out(struct request *rq)
2296 {
2297         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2298         struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2299
2300         nvme_tcp_stop_queue(ctrl, nvme_tcp_queue_id(req->queue));
2301         nvmf_complete_timed_out_request(rq);
2302 }
2303
2304 static enum blk_eh_timer_return nvme_tcp_timeout(struct request *rq)
2305 {
2306         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2307         struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2308         struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
2309         u8 opc = pdu->cmd.common.opcode, fctype = pdu->cmd.fabrics.fctype;
2310         int qid = nvme_tcp_queue_id(req->queue);
2311
2312         dev_warn(ctrl->device,
2313                 "queue %d: timeout cid %#x type %d opcode %#x (%s)\n",
2314                 nvme_tcp_queue_id(req->queue), nvme_cid(rq), pdu->hdr.type,
2315                 opc, nvme_opcode_str(qid, opc, fctype));
2316
2317         if (ctrl->state != NVME_CTRL_LIVE) {
2318                 /*
2319                  * If we are resetting, connecting or deleting we should
2320                  * complete immediately because we may block controller
2321                  * teardown or setup sequence
2322                  * - ctrl disable/shutdown fabrics requests
2323                  * - connect requests
2324                  * - initialization admin requests
2325                  * - I/O requests that entered after unquiescing and
2326                  *   the controller stopped responding
2327                  *
2328                  * All other requests should be cancelled by the error
2329                  * recovery work, so it's fine that we fail it here.
2330                  */
2331                 nvme_tcp_complete_timed_out(rq);
2332                 return BLK_EH_DONE;
2333         }
2334
2335         /*
2336          * LIVE state should trigger the normal error recovery which will
2337          * handle completing this request.
2338          */
2339         nvme_tcp_error_recovery(ctrl);
2340         return BLK_EH_RESET_TIMER;
2341 }
2342
2343 static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2344                         struct request *rq)
2345 {
2346         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2347         struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
2348         struct nvme_command *c = &pdu->cmd;
2349
2350         c->common.flags |= NVME_CMD_SGL_METABUF;
2351
2352         if (!blk_rq_nr_phys_segments(rq))
2353                 nvme_tcp_set_sg_null(c);
2354         else if (rq_data_dir(rq) == WRITE &&
2355             req->data_len <= nvme_tcp_inline_data_size(req))
2356                 nvme_tcp_set_sg_inline(queue, c, req->data_len);
2357         else
2358                 nvme_tcp_set_sg_host_data(c, req->data_len);
2359
2360         return 0;
2361 }
2362
2363 static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2364                 struct request *rq)
2365 {
2366         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2367         struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
2368         struct nvme_tcp_queue *queue = req->queue;
2369         u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2370         blk_status_t ret;
2371
2372         ret = nvme_setup_cmd(ns, rq);
2373         if (ret)
2374                 return ret;
2375
2376         req->state = NVME_TCP_SEND_CMD_PDU;
2377         req->status = cpu_to_le16(NVME_SC_SUCCESS);
2378         req->offset = 0;
2379         req->data_sent = 0;
2380         req->pdu_len = 0;
2381         req->pdu_sent = 0;
2382         req->h2cdata_left = 0;
2383         req->data_len = blk_rq_nr_phys_segments(rq) ?
2384                                 blk_rq_payload_bytes(rq) : 0;
2385         req->curr_bio = rq->bio;
2386         if (req->curr_bio && req->data_len)
2387                 nvme_tcp_init_iter(req, rq_data_dir(rq));
2388
2389         if (rq_data_dir(rq) == WRITE &&
2390             req->data_len <= nvme_tcp_inline_data_size(req))
2391                 req->pdu_len = req->data_len;
2392
2393         pdu->hdr.type = nvme_tcp_cmd;
2394         pdu->hdr.flags = 0;
2395         if (queue->hdr_digest)
2396                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2397         if (queue->data_digest && req->pdu_len) {
2398                 pdu->hdr.flags |= NVME_TCP_F_DDGST;
2399                 ddgst = nvme_tcp_ddgst_len(queue);
2400         }
2401         pdu->hdr.hlen = sizeof(*pdu);
2402         pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2403         pdu->hdr.plen =
2404                 cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2405
2406         ret = nvme_tcp_map_data(queue, rq);
2407         if (unlikely(ret)) {
2408                 nvme_cleanup_cmd(rq);
2409                 dev_err(queue->ctrl->ctrl.device,
2410                         "Failed to map data (%d)\n", ret);
2411                 return ret;
2412         }
2413
2414         return 0;
2415 }
2416
2417 static void nvme_tcp_commit_rqs(struct blk_mq_hw_ctx *hctx)
2418 {
2419         struct nvme_tcp_queue *queue = hctx->driver_data;
2420
2421         if (!llist_empty(&queue->req_list))
2422                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
2423 }
2424
2425 static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2426                 const struct blk_mq_queue_data *bd)
2427 {
2428         struct nvme_ns *ns = hctx->queue->queuedata;
2429         struct nvme_tcp_queue *queue = hctx->driver_data;
2430         struct request *rq = bd->rq;
2431         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2432         bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2433         blk_status_t ret;
2434
2435         if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2436                 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
2437
2438         ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2439         if (unlikely(ret))
2440                 return ret;
2441
2442         nvme_start_request(rq);
2443
2444         nvme_tcp_queue_request(req, true, bd->last);
2445
2446         return BLK_STS_OK;
2447 }
2448
2449 static void nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2450 {
2451         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(set->driver_data);
2452         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2453
2454         if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2455                 /* separate read/write queues */
2456                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2457                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2458                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2459                 set->map[HCTX_TYPE_READ].nr_queues =
2460                         ctrl->io_queues[HCTX_TYPE_READ];
2461                 set->map[HCTX_TYPE_READ].queue_offset =
2462                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2463         } else {
2464                 /* shared read/write queues */
2465                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2466                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2467                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2468                 set->map[HCTX_TYPE_READ].nr_queues =
2469                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2470                 set->map[HCTX_TYPE_READ].queue_offset = 0;
2471         }
2472         blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2473         blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
2474
2475         if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2476                 /* map dedicated poll queues only if we have queues left */
2477                 set->map[HCTX_TYPE_POLL].nr_queues =
2478                                 ctrl->io_queues[HCTX_TYPE_POLL];
2479                 set->map[HCTX_TYPE_POLL].queue_offset =
2480                         ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2481                         ctrl->io_queues[HCTX_TYPE_READ];
2482                 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2483         }
2484
2485         dev_info(ctrl->ctrl.device,
2486                 "mapped %d/%d/%d default/read/poll queues.\n",
2487                 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2488                 ctrl->io_queues[HCTX_TYPE_READ],
2489                 ctrl->io_queues[HCTX_TYPE_POLL]);
2490 }
2491
2492 static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
2493 {
2494         struct nvme_tcp_queue *queue = hctx->driver_data;
2495         struct sock *sk = queue->sock->sk;
2496
2497         if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2498                 return 0;
2499
2500         set_bit(NVME_TCP_Q_POLLING, &queue->flags);
2501         if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue))
2502                 sk_busy_loop(sk, true);
2503         nvme_tcp_try_recv(queue);
2504         clear_bit(NVME_TCP_Q_POLLING, &queue->flags);
2505         return queue->nr_cqe;
2506 }
2507
2508 static int nvme_tcp_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2509 {
2510         struct nvme_tcp_queue *queue = &to_tcp_ctrl(ctrl)->queues[0];
2511         struct sockaddr_storage src_addr;
2512         int ret, len;
2513
2514         len = nvmf_get_address(ctrl, buf, size);
2515
2516         mutex_lock(&queue->queue_lock);
2517
2518         if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2519                 goto done;
2520         ret = kernel_getsockname(queue->sock, (struct sockaddr *)&src_addr);
2521         if (ret > 0) {
2522                 if (len > 0)
2523                         len--; /* strip trailing newline */
2524                 len += scnprintf(buf + len, size - len, "%ssrc_addr=%pISc\n",
2525                                 (len) ? "," : "", &src_addr);
2526         }
2527 done:
2528         mutex_unlock(&queue->queue_lock);
2529
2530         return len;
2531 }
2532
2533 static const struct blk_mq_ops nvme_tcp_mq_ops = {
2534         .queue_rq       = nvme_tcp_queue_rq,
2535         .commit_rqs     = nvme_tcp_commit_rqs,
2536         .complete       = nvme_complete_rq,
2537         .init_request   = nvme_tcp_init_request,
2538         .exit_request   = nvme_tcp_exit_request,
2539         .init_hctx      = nvme_tcp_init_hctx,
2540         .timeout        = nvme_tcp_timeout,
2541         .map_queues     = nvme_tcp_map_queues,
2542         .poll           = nvme_tcp_poll,
2543 };
2544
2545 static const struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2546         .queue_rq       = nvme_tcp_queue_rq,
2547         .complete       = nvme_complete_rq,
2548         .init_request   = nvme_tcp_init_request,
2549         .exit_request   = nvme_tcp_exit_request,
2550         .init_hctx      = nvme_tcp_init_admin_hctx,
2551         .timeout        = nvme_tcp_timeout,
2552 };
2553
2554 static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2555         .name                   = "tcp",
2556         .module                 = THIS_MODULE,
2557         .flags                  = NVME_F_FABRICS | NVME_F_BLOCKING,
2558         .reg_read32             = nvmf_reg_read32,
2559         .reg_read64             = nvmf_reg_read64,
2560         .reg_write32            = nvmf_reg_write32,
2561         .free_ctrl              = nvme_tcp_free_ctrl,
2562         .submit_async_event     = nvme_tcp_submit_async_event,
2563         .delete_ctrl            = nvme_tcp_delete_ctrl,
2564         .get_address            = nvme_tcp_get_address,
2565         .stop_ctrl              = nvme_tcp_stop_ctrl,
2566 };
2567
2568 static bool
2569 nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2570 {
2571         struct nvme_tcp_ctrl *ctrl;
2572         bool found = false;
2573
2574         mutex_lock(&nvme_tcp_ctrl_mutex);
2575         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2576                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2577                 if (found)
2578                         break;
2579         }
2580         mutex_unlock(&nvme_tcp_ctrl_mutex);
2581
2582         return found;
2583 }
2584
2585 static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2586                 struct nvmf_ctrl_options *opts)
2587 {
2588         struct nvme_tcp_ctrl *ctrl;
2589         int ret;
2590
2591         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2592         if (!ctrl)
2593                 return ERR_PTR(-ENOMEM);
2594
2595         INIT_LIST_HEAD(&ctrl->list);
2596         ctrl->ctrl.opts = opts;
2597         ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2598                                 opts->nr_poll_queues + 1;
2599         ctrl->ctrl.sqsize = opts->queue_size - 1;
2600         ctrl->ctrl.kato = opts->kato;
2601
2602         INIT_DELAYED_WORK(&ctrl->connect_work,
2603                         nvme_tcp_reconnect_ctrl_work);
2604         INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2605         INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2606
2607         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2608                 opts->trsvcid =
2609                         kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2610                 if (!opts->trsvcid) {
2611                         ret = -ENOMEM;
2612                         goto out_free_ctrl;
2613                 }
2614                 opts->mask |= NVMF_OPT_TRSVCID;
2615         }
2616
2617         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2618                         opts->traddr, opts->trsvcid, &ctrl->addr);
2619         if (ret) {
2620                 pr_err("malformed address passed: %s:%s\n",
2621                         opts->traddr, opts->trsvcid);
2622                 goto out_free_ctrl;
2623         }
2624
2625         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2626                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2627                         opts->host_traddr, NULL, &ctrl->src_addr);
2628                 if (ret) {
2629                         pr_err("malformed src address passed: %s\n",
2630                                opts->host_traddr);
2631                         goto out_free_ctrl;
2632                 }
2633         }
2634
2635         if (opts->mask & NVMF_OPT_HOST_IFACE) {
2636                 if (!__dev_get_by_name(&init_net, opts->host_iface)) {
2637                         pr_err("invalid interface passed: %s\n",
2638                                opts->host_iface);
2639                         ret = -ENODEV;
2640                         goto out_free_ctrl;
2641                 }
2642         }
2643
2644         if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2645                 ret = -EALREADY;
2646                 goto out_free_ctrl;
2647         }
2648
2649         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2650                                 GFP_KERNEL);
2651         if (!ctrl->queues) {
2652                 ret = -ENOMEM;
2653                 goto out_free_ctrl;
2654         }
2655
2656         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2657         if (ret)
2658                 goto out_kfree_queues;
2659
2660         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2661                 WARN_ON_ONCE(1);
2662                 ret = -EINTR;
2663                 goto out_uninit_ctrl;
2664         }
2665
2666         ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2667         if (ret)
2668                 goto out_uninit_ctrl;
2669
2670         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
2671                 nvmf_ctrl_subsysnqn(&ctrl->ctrl), &ctrl->addr);
2672
2673         mutex_lock(&nvme_tcp_ctrl_mutex);
2674         list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2675         mutex_unlock(&nvme_tcp_ctrl_mutex);
2676
2677         return &ctrl->ctrl;
2678
2679 out_uninit_ctrl:
2680         nvme_uninit_ctrl(&ctrl->ctrl);
2681         nvme_put_ctrl(&ctrl->ctrl);
2682         if (ret > 0)
2683                 ret = -EIO;
2684         return ERR_PTR(ret);
2685 out_kfree_queues:
2686         kfree(ctrl->queues);
2687 out_free_ctrl:
2688         kfree(ctrl);
2689         return ERR_PTR(ret);
2690 }
2691
2692 static struct nvmf_transport_ops nvme_tcp_transport = {
2693         .name           = "tcp",
2694         .module         = THIS_MODULE,
2695         .required_opts  = NVMF_OPT_TRADDR,
2696         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2697                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2698                           NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
2699                           NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2700                           NVMF_OPT_TOS | NVMF_OPT_HOST_IFACE,
2701         .create_ctrl    = nvme_tcp_create_ctrl,
2702 };
2703
2704 static int __init nvme_tcp_init_module(void)
2705 {
2706         BUILD_BUG_ON(sizeof(struct nvme_tcp_hdr) != 8);
2707         BUILD_BUG_ON(sizeof(struct nvme_tcp_cmd_pdu) != 72);
2708         BUILD_BUG_ON(sizeof(struct nvme_tcp_data_pdu) != 24);
2709         BUILD_BUG_ON(sizeof(struct nvme_tcp_rsp_pdu) != 24);
2710         BUILD_BUG_ON(sizeof(struct nvme_tcp_r2t_pdu) != 24);
2711         BUILD_BUG_ON(sizeof(struct nvme_tcp_icreq_pdu) != 128);
2712         BUILD_BUG_ON(sizeof(struct nvme_tcp_icresp_pdu) != 128);
2713         BUILD_BUG_ON(sizeof(struct nvme_tcp_term_pdu) != 24);
2714
2715         nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
2716                         WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2717         if (!nvme_tcp_wq)
2718                 return -ENOMEM;
2719
2720         nvmf_register_transport(&nvme_tcp_transport);
2721         return 0;
2722 }
2723
2724 static void __exit nvme_tcp_cleanup_module(void)
2725 {
2726         struct nvme_tcp_ctrl *ctrl;
2727
2728         nvmf_unregister_transport(&nvme_tcp_transport);
2729
2730         mutex_lock(&nvme_tcp_ctrl_mutex);
2731         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
2732                 nvme_delete_ctrl(&ctrl->ctrl);
2733         mutex_unlock(&nvme_tcp_ctrl_mutex);
2734         flush_workqueue(nvme_delete_wq);
2735
2736         destroy_workqueue(nvme_tcp_wq);
2737 }
2738
2739 module_init(nvme_tcp_init_module);
2740 module_exit(nvme_tcp_cleanup_module);
2741
2742 MODULE_LICENSE("GPL v2");