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 bio_vec bvec;
1001                 struct msghdr msg = {
1002                         .msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES,
1003                 };
1004                 struct page *page = nvme_tcp_req_cur_page(req);
1005                 size_t offset = nvme_tcp_req_cur_offset(req);
1006                 size_t len = nvme_tcp_req_cur_length(req);
1007                 bool last = nvme_tcp_pdu_last_send(req, len);
1008                 int req_data_sent = req->data_sent;
1009                 int ret;
1010
1011                 if (last && !queue->data_digest && !nvme_tcp_queue_more(queue))
1012                         msg.msg_flags |= MSG_EOR;
1013                 else
1014                         msg.msg_flags |= MSG_MORE;
1015
1016                 if (!sendpage_ok(page))
1017                         msg.msg_flags &= ~MSG_SPLICE_PAGES,
1018
1019                 bvec_set_page(&bvec, page, len, offset);
1020                 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, len);
1021                 ret = sock_sendmsg(queue->sock, &msg);
1022                 if (ret <= 0)
1023                         return ret;
1024
1025                 if (queue->data_digest)
1026                         nvme_tcp_ddgst_update(queue->snd_hash, page,
1027                                         offset, ret);
1028
1029                 /*
1030                  * update the request iterator except for the last payload send
1031                  * in the request where we don't want to modify it as we may
1032                  * compete with the RX path completing the request.
1033                  */
1034                 if (req_data_sent + ret < req_data_len)
1035                         nvme_tcp_advance_req(req, ret);
1036
1037                 /* fully successful last send in current PDU */
1038                 if (last && ret == len) {
1039                         if (queue->data_digest) {
1040                                 nvme_tcp_ddgst_final(queue->snd_hash,
1041                                         &req->ddgst);
1042                                 req->state = NVME_TCP_SEND_DDGST;
1043                                 req->offset = 0;
1044                         } else {
1045                                 if (h2cdata_left)
1046                                         nvme_tcp_setup_h2c_data_pdu(req);
1047                                 else
1048                                         nvme_tcp_done_send_req(queue);
1049                         }
1050                         return 1;
1051                 }
1052         }
1053         return -EAGAIN;
1054 }
1055
1056 static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
1057 {
1058         struct nvme_tcp_queue *queue = req->queue;
1059         struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
1060         struct bio_vec bvec;
1061         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES, };
1062         bool inline_data = nvme_tcp_has_inline_data(req);
1063         u8 hdgst = nvme_tcp_hdgst_len(queue);
1064         int len = sizeof(*pdu) + hdgst - req->offset;
1065         int ret;
1066
1067         if (inline_data || nvme_tcp_queue_more(queue))
1068                 msg.msg_flags |= MSG_MORE;
1069         else
1070                 msg.msg_flags |= MSG_EOR;
1071
1072         if (queue->hdr_digest && !req->offset)
1073                 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
1074
1075         bvec_set_virt(&bvec, (void *)pdu + req->offset, len);
1076         iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, len);
1077         ret = sock_sendmsg(queue->sock, &msg);
1078         if (unlikely(ret <= 0))
1079                 return ret;
1080
1081         len -= ret;
1082         if (!len) {
1083                 if (inline_data) {
1084                         req->state = NVME_TCP_SEND_DATA;
1085                         if (queue->data_digest)
1086                                 crypto_ahash_init(queue->snd_hash);
1087                 } else {
1088                         nvme_tcp_done_send_req(queue);
1089                 }
1090                 return 1;
1091         }
1092         req->offset += ret;
1093
1094         return -EAGAIN;
1095 }
1096
1097 static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
1098 {
1099         struct nvme_tcp_queue *queue = req->queue;
1100         struct nvme_tcp_data_pdu *pdu = nvme_tcp_req_data_pdu(req);
1101         struct bio_vec bvec;
1102         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_MORE, };
1103         u8 hdgst = nvme_tcp_hdgst_len(queue);
1104         int len = sizeof(*pdu) - req->offset + hdgst;
1105         int ret;
1106
1107         if (queue->hdr_digest && !req->offset)
1108                 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
1109
1110         if (!req->h2cdata_left)
1111                 msg.msg_flags |= MSG_SPLICE_PAGES;
1112
1113         bvec_set_virt(&bvec, (void *)pdu + req->offset, len);
1114         iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, len);
1115         ret = sock_sendmsg(queue->sock, &msg);
1116         if (unlikely(ret <= 0))
1117                 return ret;
1118
1119         len -= ret;
1120         if (!len) {
1121                 req->state = NVME_TCP_SEND_DATA;
1122                 if (queue->data_digest)
1123                         crypto_ahash_init(queue->snd_hash);
1124                 return 1;
1125         }
1126         req->offset += ret;
1127
1128         return -EAGAIN;
1129 }
1130
1131 static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
1132 {
1133         struct nvme_tcp_queue *queue = req->queue;
1134         size_t offset = req->offset;
1135         u32 h2cdata_left = req->h2cdata_left;
1136         int ret;
1137         struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1138         struct kvec iov = {
1139                 .iov_base = (u8 *)&req->ddgst + req->offset,
1140                 .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
1141         };
1142
1143         if (nvme_tcp_queue_more(queue))
1144                 msg.msg_flags |= MSG_MORE;
1145         else
1146                 msg.msg_flags |= MSG_EOR;
1147
1148         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1149         if (unlikely(ret <= 0))
1150                 return ret;
1151
1152         if (offset + ret == NVME_TCP_DIGEST_LENGTH) {
1153                 if (h2cdata_left)
1154                         nvme_tcp_setup_h2c_data_pdu(req);
1155                 else
1156                         nvme_tcp_done_send_req(queue);
1157                 return 1;
1158         }
1159
1160         req->offset += ret;
1161         return -EAGAIN;
1162 }
1163
1164 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
1165 {
1166         struct nvme_tcp_request *req;
1167         unsigned int noreclaim_flag;
1168         int ret = 1;
1169
1170         if (!queue->request) {
1171                 queue->request = nvme_tcp_fetch_request(queue);
1172                 if (!queue->request)
1173                         return 0;
1174         }
1175         req = queue->request;
1176
1177         noreclaim_flag = memalloc_noreclaim_save();
1178         if (req->state == NVME_TCP_SEND_CMD_PDU) {
1179                 ret = nvme_tcp_try_send_cmd_pdu(req);
1180                 if (ret <= 0)
1181                         goto done;
1182                 if (!nvme_tcp_has_inline_data(req))
1183                         goto out;
1184         }
1185
1186         if (req->state == NVME_TCP_SEND_H2C_PDU) {
1187                 ret = nvme_tcp_try_send_data_pdu(req);
1188                 if (ret <= 0)
1189                         goto done;
1190         }
1191
1192         if (req->state == NVME_TCP_SEND_DATA) {
1193                 ret = nvme_tcp_try_send_data(req);
1194                 if (ret <= 0)
1195                         goto done;
1196         }
1197
1198         if (req->state == NVME_TCP_SEND_DDGST)
1199                 ret = nvme_tcp_try_send_ddgst(req);
1200 done:
1201         if (ret == -EAGAIN) {
1202                 ret = 0;
1203         } else if (ret < 0) {
1204                 dev_err(queue->ctrl->ctrl.device,
1205                         "failed to send request %d\n", ret);
1206                 nvme_tcp_fail_request(queue->request);
1207                 nvme_tcp_done_send_req(queue);
1208         }
1209 out:
1210         memalloc_noreclaim_restore(noreclaim_flag);
1211         return ret;
1212 }
1213
1214 static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1215 {
1216         struct socket *sock = queue->sock;
1217         struct sock *sk = sock->sk;
1218         read_descriptor_t rd_desc;
1219         int consumed;
1220
1221         rd_desc.arg.data = queue;
1222         rd_desc.count = 1;
1223         lock_sock(sk);
1224         queue->nr_cqe = 0;
1225         consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1226         release_sock(sk);
1227         return consumed;
1228 }
1229
1230 static void nvme_tcp_io_work(struct work_struct *w)
1231 {
1232         struct nvme_tcp_queue *queue =
1233                 container_of(w, struct nvme_tcp_queue, io_work);
1234         unsigned long deadline = jiffies + msecs_to_jiffies(1);
1235
1236         do {
1237                 bool pending = false;
1238                 int result;
1239
1240                 if (mutex_trylock(&queue->send_mutex)) {
1241                         result = nvme_tcp_try_send(queue);
1242                         mutex_unlock(&queue->send_mutex);
1243                         if (result > 0)
1244                                 pending = true;
1245                         else if (unlikely(result < 0))
1246                                 break;
1247                 }
1248
1249                 result = nvme_tcp_try_recv(queue);
1250                 if (result > 0)
1251                         pending = true;
1252                 else if (unlikely(result < 0))
1253                         return;
1254
1255                 if (!pending || !queue->rd_enabled)
1256                         return;
1257
1258         } while (!time_after(jiffies, deadline)); /* quota is exhausted */
1259
1260         queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1261 }
1262
1263 static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue)
1264 {
1265         struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
1266
1267         ahash_request_free(queue->rcv_hash);
1268         ahash_request_free(queue->snd_hash);
1269         crypto_free_ahash(tfm);
1270 }
1271
1272 static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue)
1273 {
1274         struct crypto_ahash *tfm;
1275
1276         tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
1277         if (IS_ERR(tfm))
1278                 return PTR_ERR(tfm);
1279
1280         queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1281         if (!queue->snd_hash)
1282                 goto free_tfm;
1283         ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
1284
1285         queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1286         if (!queue->rcv_hash)
1287                 goto free_snd_hash;
1288         ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
1289
1290         return 0;
1291 free_snd_hash:
1292         ahash_request_free(queue->snd_hash);
1293 free_tfm:
1294         crypto_free_ahash(tfm);
1295         return -ENOMEM;
1296 }
1297
1298 static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1299 {
1300         struct nvme_tcp_request *async = &ctrl->async_req;
1301
1302         page_frag_free(async->pdu);
1303 }
1304
1305 static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1306 {
1307         struct nvme_tcp_queue *queue = &ctrl->queues[0];
1308         struct nvme_tcp_request *async = &ctrl->async_req;
1309         u8 hdgst = nvme_tcp_hdgst_len(queue);
1310
1311         async->pdu = page_frag_alloc(&queue->pf_cache,
1312                 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1313                 GFP_KERNEL | __GFP_ZERO);
1314         if (!async->pdu)
1315                 return -ENOMEM;
1316
1317         async->queue = &ctrl->queues[0];
1318         return 0;
1319 }
1320
1321 static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1322 {
1323         struct page *page;
1324         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1325         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1326         unsigned int noreclaim_flag;
1327
1328         if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1329                 return;
1330
1331         if (queue->hdr_digest || queue->data_digest)
1332                 nvme_tcp_free_crypto(queue);
1333
1334         if (queue->pf_cache.va) {
1335                 page = virt_to_head_page(queue->pf_cache.va);
1336                 __page_frag_cache_drain(page, queue->pf_cache.pagecnt_bias);
1337                 queue->pf_cache.va = NULL;
1338         }
1339
1340         noreclaim_flag = memalloc_noreclaim_save();
1341         sock_release(queue->sock);
1342         memalloc_noreclaim_restore(noreclaim_flag);
1343
1344         kfree(queue->pdu);
1345         mutex_destroy(&queue->send_mutex);
1346         mutex_destroy(&queue->queue_lock);
1347 }
1348
1349 static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1350 {
1351         struct nvme_tcp_icreq_pdu *icreq;
1352         struct nvme_tcp_icresp_pdu *icresp;
1353         struct msghdr msg = {};
1354         struct kvec iov;
1355         bool ctrl_hdgst, ctrl_ddgst;
1356         u32 maxh2cdata;
1357         int ret;
1358
1359         icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1360         if (!icreq)
1361                 return -ENOMEM;
1362
1363         icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1364         if (!icresp) {
1365                 ret = -ENOMEM;
1366                 goto free_icreq;
1367         }
1368
1369         icreq->hdr.type = nvme_tcp_icreq;
1370         icreq->hdr.hlen = sizeof(*icreq);
1371         icreq->hdr.pdo = 0;
1372         icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1373         icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1374         icreq->maxr2t = 0; /* single inflight r2t supported */
1375         icreq->hpda = 0; /* no alignment constraint */
1376         if (queue->hdr_digest)
1377                 icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1378         if (queue->data_digest)
1379                 icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1380
1381         iov.iov_base = icreq;
1382         iov.iov_len = sizeof(*icreq);
1383         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1384         if (ret < 0)
1385                 goto free_icresp;
1386
1387         memset(&msg, 0, sizeof(msg));
1388         iov.iov_base = icresp;
1389         iov.iov_len = sizeof(*icresp);
1390         ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1391                         iov.iov_len, msg.msg_flags);
1392         if (ret < 0)
1393                 goto free_icresp;
1394
1395         ret = -EINVAL;
1396         if (icresp->hdr.type != nvme_tcp_icresp) {
1397                 pr_err("queue %d: bad type returned %d\n",
1398                         nvme_tcp_queue_id(queue), icresp->hdr.type);
1399                 goto free_icresp;
1400         }
1401
1402         if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1403                 pr_err("queue %d: bad pdu length returned %d\n",
1404                         nvme_tcp_queue_id(queue), icresp->hdr.plen);
1405                 goto free_icresp;
1406         }
1407
1408         if (icresp->pfv != NVME_TCP_PFV_1_0) {
1409                 pr_err("queue %d: bad pfv returned %d\n",
1410                         nvme_tcp_queue_id(queue), icresp->pfv);
1411                 goto free_icresp;
1412         }
1413
1414         ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1415         if ((queue->data_digest && !ctrl_ddgst) ||
1416             (!queue->data_digest && ctrl_ddgst)) {
1417                 pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1418                         nvme_tcp_queue_id(queue),
1419                         queue->data_digest ? "enabled" : "disabled",
1420                         ctrl_ddgst ? "enabled" : "disabled");
1421                 goto free_icresp;
1422         }
1423
1424         ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1425         if ((queue->hdr_digest && !ctrl_hdgst) ||
1426             (!queue->hdr_digest && ctrl_hdgst)) {
1427                 pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1428                         nvme_tcp_queue_id(queue),
1429                         queue->hdr_digest ? "enabled" : "disabled",
1430                         ctrl_hdgst ? "enabled" : "disabled");
1431                 goto free_icresp;
1432         }
1433
1434         if (icresp->cpda != 0) {
1435                 pr_err("queue %d: unsupported cpda returned %d\n",
1436                         nvme_tcp_queue_id(queue), icresp->cpda);
1437                 goto free_icresp;
1438         }
1439
1440         maxh2cdata = le32_to_cpu(icresp->maxdata);
1441         if ((maxh2cdata % 4) || (maxh2cdata < NVME_TCP_MIN_MAXH2CDATA)) {
1442                 pr_err("queue %d: invalid maxh2cdata returned %u\n",
1443                        nvme_tcp_queue_id(queue), maxh2cdata);
1444                 goto free_icresp;
1445         }
1446         queue->maxh2cdata = maxh2cdata;
1447
1448         ret = 0;
1449 free_icresp:
1450         kfree(icresp);
1451 free_icreq:
1452         kfree(icreq);
1453         return ret;
1454 }
1455
1456 static bool nvme_tcp_admin_queue(struct nvme_tcp_queue *queue)
1457 {
1458         return nvme_tcp_queue_id(queue) == 0;
1459 }
1460
1461 static bool nvme_tcp_default_queue(struct nvme_tcp_queue *queue)
1462 {
1463         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1464         int qid = nvme_tcp_queue_id(queue);
1465
1466         return !nvme_tcp_admin_queue(queue) &&
1467                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT];
1468 }
1469
1470 static bool nvme_tcp_read_queue(struct nvme_tcp_queue *queue)
1471 {
1472         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1473         int qid = nvme_tcp_queue_id(queue);
1474
1475         return !nvme_tcp_admin_queue(queue) &&
1476                 !nvme_tcp_default_queue(queue) &&
1477                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1478                           ctrl->io_queues[HCTX_TYPE_READ];
1479 }
1480
1481 static bool nvme_tcp_poll_queue(struct nvme_tcp_queue *queue)
1482 {
1483         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1484         int qid = nvme_tcp_queue_id(queue);
1485
1486         return !nvme_tcp_admin_queue(queue) &&
1487                 !nvme_tcp_default_queue(queue) &&
1488                 !nvme_tcp_read_queue(queue) &&
1489                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1490                           ctrl->io_queues[HCTX_TYPE_READ] +
1491                           ctrl->io_queues[HCTX_TYPE_POLL];
1492 }
1493
1494 static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
1495 {
1496         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1497         int qid = nvme_tcp_queue_id(queue);
1498         int n = 0;
1499
1500         if (nvme_tcp_default_queue(queue))
1501                 n = qid - 1;
1502         else if (nvme_tcp_read_queue(queue))
1503                 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] - 1;
1504         else if (nvme_tcp_poll_queue(queue))
1505                 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] -
1506                                 ctrl->io_queues[HCTX_TYPE_READ] - 1;
1507         queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
1508 }
1509
1510 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl, int qid)
1511 {
1512         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1513         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1514         int ret, rcv_pdu_size;
1515
1516         mutex_init(&queue->queue_lock);
1517         queue->ctrl = ctrl;
1518         init_llist_head(&queue->req_list);
1519         INIT_LIST_HEAD(&queue->send_list);
1520         mutex_init(&queue->send_mutex);
1521         INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1522
1523         if (qid > 0)
1524                 queue->cmnd_capsule_len = nctrl->ioccsz * 16;
1525         else
1526                 queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1527                                                 NVME_TCP_ADMIN_CCSZ;
1528
1529         ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM,
1530                         IPPROTO_TCP, &queue->sock);
1531         if (ret) {
1532                 dev_err(nctrl->device,
1533                         "failed to create socket: %d\n", ret);
1534                 goto err_destroy_mutex;
1535         }
1536
1537         nvme_tcp_reclassify_socket(queue->sock);
1538
1539         /* Single syn retry */
1540         tcp_sock_set_syncnt(queue->sock->sk, 1);
1541
1542         /* Set TCP no delay */
1543         tcp_sock_set_nodelay(queue->sock->sk);
1544
1545         /*
1546          * Cleanup whatever is sitting in the TCP transmit queue on socket
1547          * close. This is done to prevent stale data from being sent should
1548          * the network connection be restored before TCP times out.
1549          */
1550         sock_no_linger(queue->sock->sk);
1551
1552         if (so_priority > 0)
1553                 sock_set_priority(queue->sock->sk, so_priority);
1554
1555         /* Set socket type of service */
1556         if (nctrl->opts->tos >= 0)
1557                 ip_sock_set_tos(queue->sock->sk, nctrl->opts->tos);
1558
1559         /* Set 10 seconds timeout for icresp recvmsg */
1560         queue->sock->sk->sk_rcvtimeo = 10 * HZ;
1561
1562         queue->sock->sk->sk_allocation = GFP_ATOMIC;
1563         queue->sock->sk->sk_use_task_frag = false;
1564         nvme_tcp_set_queue_io_cpu(queue);
1565         queue->request = NULL;
1566         queue->data_remaining = 0;
1567         queue->ddgst_remaining = 0;
1568         queue->pdu_remaining = 0;
1569         queue->pdu_offset = 0;
1570         sk_set_memalloc(queue->sock->sk);
1571
1572         if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
1573                 ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1574                         sizeof(ctrl->src_addr));
1575                 if (ret) {
1576                         dev_err(nctrl->device,
1577                                 "failed to bind queue %d socket %d\n",
1578                                 qid, ret);
1579                         goto err_sock;
1580                 }
1581         }
1582
1583         if (nctrl->opts->mask & NVMF_OPT_HOST_IFACE) {
1584                 char *iface = nctrl->opts->host_iface;
1585                 sockptr_t optval = KERNEL_SOCKPTR(iface);
1586
1587                 ret = sock_setsockopt(queue->sock, SOL_SOCKET, SO_BINDTODEVICE,
1588                                       optval, strlen(iface));
1589                 if (ret) {
1590                         dev_err(nctrl->device,
1591                           "failed to bind to interface %s queue %d err %d\n",
1592                           iface, qid, ret);
1593                         goto err_sock;
1594                 }
1595         }
1596
1597         queue->hdr_digest = nctrl->opts->hdr_digest;
1598         queue->data_digest = nctrl->opts->data_digest;
1599         if (queue->hdr_digest || queue->data_digest) {
1600                 ret = nvme_tcp_alloc_crypto(queue);
1601                 if (ret) {
1602                         dev_err(nctrl->device,
1603                                 "failed to allocate queue %d crypto\n", qid);
1604                         goto err_sock;
1605                 }
1606         }
1607
1608         rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1609                         nvme_tcp_hdgst_len(queue);
1610         queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1611         if (!queue->pdu) {
1612                 ret = -ENOMEM;
1613                 goto err_crypto;
1614         }
1615
1616         dev_dbg(nctrl->device, "connecting queue %d\n",
1617                         nvme_tcp_queue_id(queue));
1618
1619         ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1620                 sizeof(ctrl->addr), 0);
1621         if (ret) {
1622                 dev_err(nctrl->device,
1623                         "failed to connect socket: %d\n", ret);
1624                 goto err_rcv_pdu;
1625         }
1626
1627         ret = nvme_tcp_init_connection(queue);
1628         if (ret)
1629                 goto err_init_connect;
1630
1631         set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1632
1633         return 0;
1634
1635 err_init_connect:
1636         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1637 err_rcv_pdu:
1638         kfree(queue->pdu);
1639 err_crypto:
1640         if (queue->hdr_digest || queue->data_digest)
1641                 nvme_tcp_free_crypto(queue);
1642 err_sock:
1643         sock_release(queue->sock);
1644         queue->sock = NULL;
1645 err_destroy_mutex:
1646         mutex_destroy(&queue->send_mutex);
1647         mutex_destroy(&queue->queue_lock);
1648         return ret;
1649 }
1650
1651 static void nvme_tcp_restore_sock_ops(struct nvme_tcp_queue *queue)
1652 {
1653         struct socket *sock = queue->sock;
1654
1655         write_lock_bh(&sock->sk->sk_callback_lock);
1656         sock->sk->sk_user_data  = NULL;
1657         sock->sk->sk_data_ready = queue->data_ready;
1658         sock->sk->sk_state_change = queue->state_change;
1659         sock->sk->sk_write_space  = queue->write_space;
1660         write_unlock_bh(&sock->sk->sk_callback_lock);
1661 }
1662
1663 static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1664 {
1665         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1666         nvme_tcp_restore_sock_ops(queue);
1667         cancel_work_sync(&queue->io_work);
1668 }
1669
1670 static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1671 {
1672         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1673         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1674
1675         if (!test_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1676                 return;
1677
1678         mutex_lock(&queue->queue_lock);
1679         if (test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1680                 __nvme_tcp_stop_queue(queue);
1681         mutex_unlock(&queue->queue_lock);
1682 }
1683
1684 static void nvme_tcp_setup_sock_ops(struct nvme_tcp_queue *queue)
1685 {
1686         write_lock_bh(&queue->sock->sk->sk_callback_lock);
1687         queue->sock->sk->sk_user_data = queue;
1688         queue->state_change = queue->sock->sk->sk_state_change;
1689         queue->data_ready = queue->sock->sk->sk_data_ready;
1690         queue->write_space = queue->sock->sk->sk_write_space;
1691         queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1692         queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1693         queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1694 #ifdef CONFIG_NET_RX_BUSY_POLL
1695         queue->sock->sk->sk_ll_usec = 1;
1696 #endif
1697         write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1698 }
1699
1700 static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1701 {
1702         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1703         struct nvme_tcp_queue *queue = &ctrl->queues[idx];
1704         int ret;
1705
1706         queue->rd_enabled = true;
1707         nvme_tcp_init_recv_ctx(queue);
1708         nvme_tcp_setup_sock_ops(queue);
1709
1710         if (idx)
1711                 ret = nvmf_connect_io_queue(nctrl, idx);
1712         else
1713                 ret = nvmf_connect_admin_queue(nctrl);
1714
1715         if (!ret) {
1716                 set_bit(NVME_TCP_Q_LIVE, &queue->flags);
1717         } else {
1718                 if (test_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1719                         __nvme_tcp_stop_queue(queue);
1720                 dev_err(nctrl->device,
1721                         "failed to connect queue: %d ret=%d\n", idx, ret);
1722         }
1723         return ret;
1724 }
1725
1726 static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
1727 {
1728         if (to_tcp_ctrl(ctrl)->async_req.pdu) {
1729                 cancel_work_sync(&ctrl->async_event_work);
1730                 nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
1731                 to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
1732         }
1733
1734         nvme_tcp_free_queue(ctrl, 0);
1735 }
1736
1737 static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
1738 {
1739         int i;
1740
1741         for (i = 1; i < ctrl->queue_count; i++)
1742                 nvme_tcp_free_queue(ctrl, i);
1743 }
1744
1745 static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
1746 {
1747         int i;
1748
1749         for (i = 1; i < ctrl->queue_count; i++)
1750                 nvme_tcp_stop_queue(ctrl, i);
1751 }
1752
1753 static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl,
1754                                     int first, int last)
1755 {
1756         int i, ret;
1757
1758         for (i = first; i < last; i++) {
1759                 ret = nvme_tcp_start_queue(ctrl, i);
1760                 if (ret)
1761                         goto out_stop_queues;
1762         }
1763
1764         return 0;
1765
1766 out_stop_queues:
1767         for (i--; i >= first; i--)
1768                 nvme_tcp_stop_queue(ctrl, i);
1769         return ret;
1770 }
1771
1772 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
1773 {
1774         int ret;
1775
1776         ret = nvme_tcp_alloc_queue(ctrl, 0);
1777         if (ret)
1778                 return ret;
1779
1780         ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
1781         if (ret)
1782                 goto out_free_queue;
1783
1784         return 0;
1785
1786 out_free_queue:
1787         nvme_tcp_free_queue(ctrl, 0);
1788         return ret;
1789 }
1790
1791 static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1792 {
1793         int i, ret;
1794
1795         for (i = 1; i < ctrl->queue_count; i++) {
1796                 ret = nvme_tcp_alloc_queue(ctrl, i);
1797                 if (ret)
1798                         goto out_free_queues;
1799         }
1800
1801         return 0;
1802
1803 out_free_queues:
1804         for (i--; i >= 1; i--)
1805                 nvme_tcp_free_queue(ctrl, i);
1806
1807         return ret;
1808 }
1809
1810 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1811 {
1812         unsigned int nr_io_queues;
1813         int ret;
1814
1815         nr_io_queues = nvmf_nr_io_queues(ctrl->opts);
1816         ret = nvme_set_queue_count(ctrl, &nr_io_queues);
1817         if (ret)
1818                 return ret;
1819
1820         if (nr_io_queues == 0) {
1821                 dev_err(ctrl->device,
1822                         "unable to set any I/O queues\n");
1823                 return -ENOMEM;
1824         }
1825
1826         ctrl->queue_count = nr_io_queues + 1;
1827         dev_info(ctrl->device,
1828                 "creating %d I/O queues.\n", nr_io_queues);
1829
1830         nvmf_set_io_queues(ctrl->opts, nr_io_queues,
1831                            to_tcp_ctrl(ctrl)->io_queues);
1832         return __nvme_tcp_alloc_io_queues(ctrl);
1833 }
1834
1835 static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove)
1836 {
1837         nvme_tcp_stop_io_queues(ctrl);
1838         if (remove)
1839                 nvme_remove_io_tag_set(ctrl);
1840         nvme_tcp_free_io_queues(ctrl);
1841 }
1842
1843 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
1844 {
1845         int ret, nr_queues;
1846
1847         ret = nvme_tcp_alloc_io_queues(ctrl);
1848         if (ret)
1849                 return ret;
1850
1851         if (new) {
1852                 ret = nvme_alloc_io_tag_set(ctrl, &to_tcp_ctrl(ctrl)->tag_set,
1853                                 &nvme_tcp_mq_ops,
1854                                 ctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2,
1855                                 sizeof(struct nvme_tcp_request));
1856                 if (ret)
1857                         goto out_free_io_queues;
1858         }
1859
1860         /*
1861          * Only start IO queues for which we have allocated the tagset
1862          * and limitted it to the available queues. On reconnects, the
1863          * queue number might have changed.
1864          */
1865         nr_queues = min(ctrl->tagset->nr_hw_queues + 1, ctrl->queue_count);
1866         ret = nvme_tcp_start_io_queues(ctrl, 1, nr_queues);
1867         if (ret)
1868                 goto out_cleanup_connect_q;
1869
1870         if (!new) {
1871                 nvme_unquiesce_io_queues(ctrl);
1872                 if (!nvme_wait_freeze_timeout(ctrl, NVME_IO_TIMEOUT)) {
1873                         /*
1874                          * If we timed out waiting for freeze we are likely to
1875                          * be stuck.  Fail the controller initialization just
1876                          * to be safe.
1877                          */
1878                         ret = -ENODEV;
1879                         goto out_wait_freeze_timed_out;
1880                 }
1881                 blk_mq_update_nr_hw_queues(ctrl->tagset,
1882                         ctrl->queue_count - 1);
1883                 nvme_unfreeze(ctrl);
1884         }
1885
1886         /*
1887          * If the number of queues has increased (reconnect case)
1888          * start all new queues now.
1889          */
1890         ret = nvme_tcp_start_io_queues(ctrl, nr_queues,
1891                                        ctrl->tagset->nr_hw_queues + 1);
1892         if (ret)
1893                 goto out_wait_freeze_timed_out;
1894
1895         return 0;
1896
1897 out_wait_freeze_timed_out:
1898         nvme_quiesce_io_queues(ctrl);
1899         nvme_sync_io_queues(ctrl);
1900         nvme_tcp_stop_io_queues(ctrl);
1901 out_cleanup_connect_q:
1902         nvme_cancel_tagset(ctrl);
1903         if (new)
1904                 nvme_remove_io_tag_set(ctrl);
1905 out_free_io_queues:
1906         nvme_tcp_free_io_queues(ctrl);
1907         return ret;
1908 }
1909
1910 static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove)
1911 {
1912         nvme_tcp_stop_queue(ctrl, 0);
1913         if (remove)
1914                 nvme_remove_admin_tag_set(ctrl);
1915         nvme_tcp_free_admin_queue(ctrl);
1916 }
1917
1918 static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
1919 {
1920         int error;
1921
1922         error = nvme_tcp_alloc_admin_queue(ctrl);
1923         if (error)
1924                 return error;
1925
1926         if (new) {
1927                 error = nvme_alloc_admin_tag_set(ctrl,
1928                                 &to_tcp_ctrl(ctrl)->admin_tag_set,
1929                                 &nvme_tcp_admin_mq_ops,
1930                                 sizeof(struct nvme_tcp_request));
1931                 if (error)
1932                         goto out_free_queue;
1933         }
1934
1935         error = nvme_tcp_start_queue(ctrl, 0);
1936         if (error)
1937                 goto out_cleanup_tagset;
1938
1939         error = nvme_enable_ctrl(ctrl);
1940         if (error)
1941                 goto out_stop_queue;
1942
1943         nvme_unquiesce_admin_queue(ctrl);
1944
1945         error = nvme_init_ctrl_finish(ctrl, false);
1946         if (error)
1947                 goto out_quiesce_queue;
1948
1949         return 0;
1950
1951 out_quiesce_queue:
1952         nvme_quiesce_admin_queue(ctrl);
1953         blk_sync_queue(ctrl->admin_q);
1954 out_stop_queue:
1955         nvme_tcp_stop_queue(ctrl, 0);
1956         nvme_cancel_admin_tagset(ctrl);
1957 out_cleanup_tagset:
1958         if (new)
1959                 nvme_remove_admin_tag_set(ctrl);
1960 out_free_queue:
1961         nvme_tcp_free_admin_queue(ctrl);
1962         return error;
1963 }
1964
1965 static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
1966                 bool remove)
1967 {
1968         nvme_quiesce_admin_queue(ctrl);
1969         blk_sync_queue(ctrl->admin_q);
1970         nvme_tcp_stop_queue(ctrl, 0);
1971         nvme_cancel_admin_tagset(ctrl);
1972         if (remove)
1973                 nvme_unquiesce_admin_queue(ctrl);
1974         nvme_tcp_destroy_admin_queue(ctrl, remove);
1975 }
1976
1977 static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
1978                 bool remove)
1979 {
1980         if (ctrl->queue_count <= 1)
1981                 return;
1982         nvme_quiesce_admin_queue(ctrl);
1983         nvme_start_freeze(ctrl);
1984         nvme_quiesce_io_queues(ctrl);
1985         nvme_sync_io_queues(ctrl);
1986         nvme_tcp_stop_io_queues(ctrl);
1987         nvme_cancel_tagset(ctrl);
1988         if (remove)
1989                 nvme_unquiesce_io_queues(ctrl);
1990         nvme_tcp_destroy_io_queues(ctrl, remove);
1991 }
1992
1993 static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
1994 {
1995         /* If we are resetting/deleting then do nothing */
1996         if (ctrl->state != NVME_CTRL_CONNECTING) {
1997                 WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW ||
1998                         ctrl->state == NVME_CTRL_LIVE);
1999                 return;
2000         }
2001
2002         if (nvmf_should_reconnect(ctrl)) {
2003                 dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
2004                         ctrl->opts->reconnect_delay);
2005                 queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
2006                                 ctrl->opts->reconnect_delay * HZ);
2007         } else {
2008                 dev_info(ctrl->device, "Removing controller...\n");
2009                 nvme_delete_ctrl(ctrl);
2010         }
2011 }
2012
2013 static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
2014 {
2015         struct nvmf_ctrl_options *opts = ctrl->opts;
2016         int ret;
2017
2018         ret = nvme_tcp_configure_admin_queue(ctrl, new);
2019         if (ret)
2020                 return ret;
2021
2022         if (ctrl->icdoff) {
2023                 ret = -EOPNOTSUPP;
2024                 dev_err(ctrl->device, "icdoff is not supported!\n");
2025                 goto destroy_admin;
2026         }
2027
2028         if (!nvme_ctrl_sgl_supported(ctrl)) {
2029                 ret = -EOPNOTSUPP;
2030                 dev_err(ctrl->device, "Mandatory sgls are not supported!\n");
2031                 goto destroy_admin;
2032         }
2033
2034         if (opts->queue_size > ctrl->sqsize + 1)
2035                 dev_warn(ctrl->device,
2036                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
2037                         opts->queue_size, ctrl->sqsize + 1);
2038
2039         if (ctrl->sqsize + 1 > ctrl->maxcmd) {
2040                 dev_warn(ctrl->device,
2041                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
2042                         ctrl->sqsize + 1, ctrl->maxcmd);
2043                 ctrl->sqsize = ctrl->maxcmd - 1;
2044         }
2045
2046         if (ctrl->queue_count > 1) {
2047                 ret = nvme_tcp_configure_io_queues(ctrl, new);
2048                 if (ret)
2049                         goto destroy_admin;
2050         }
2051
2052         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
2053                 /*
2054                  * state change failure is ok if we started ctrl delete,
2055                  * unless we're during creation of a new controller to
2056                  * avoid races with teardown flow.
2057                  */
2058                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2059                              ctrl->state != NVME_CTRL_DELETING_NOIO);
2060                 WARN_ON_ONCE(new);
2061                 ret = -EINVAL;
2062                 goto destroy_io;
2063         }
2064
2065         nvme_start_ctrl(ctrl);
2066         return 0;
2067
2068 destroy_io:
2069         if (ctrl->queue_count > 1) {
2070                 nvme_quiesce_io_queues(ctrl);
2071                 nvme_sync_io_queues(ctrl);
2072                 nvme_tcp_stop_io_queues(ctrl);
2073                 nvme_cancel_tagset(ctrl);
2074                 nvme_tcp_destroy_io_queues(ctrl, new);
2075         }
2076 destroy_admin:
2077         nvme_quiesce_admin_queue(ctrl);
2078         blk_sync_queue(ctrl->admin_q);
2079         nvme_tcp_stop_queue(ctrl, 0);
2080         nvme_cancel_admin_tagset(ctrl);
2081         nvme_tcp_destroy_admin_queue(ctrl, new);
2082         return ret;
2083 }
2084
2085 static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
2086 {
2087         struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
2088                         struct nvme_tcp_ctrl, connect_work);
2089         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2090
2091         ++ctrl->nr_reconnects;
2092
2093         if (nvme_tcp_setup_ctrl(ctrl, false))
2094                 goto requeue;
2095
2096         dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n",
2097                         ctrl->nr_reconnects);
2098
2099         ctrl->nr_reconnects = 0;
2100
2101         return;
2102
2103 requeue:
2104         dev_info(ctrl->device, "Failed reconnect attempt %d\n",
2105                         ctrl->nr_reconnects);
2106         nvme_tcp_reconnect_or_remove(ctrl);
2107 }
2108
2109 static void nvme_tcp_error_recovery_work(struct work_struct *work)
2110 {
2111         struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
2112                                 struct nvme_tcp_ctrl, err_work);
2113         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2114
2115         nvme_stop_keep_alive(ctrl);
2116         flush_work(&ctrl->async_event_work);
2117         nvme_tcp_teardown_io_queues(ctrl, false);
2118         /* unquiesce to fail fast pending requests */
2119         nvme_unquiesce_io_queues(ctrl);
2120         nvme_tcp_teardown_admin_queue(ctrl, false);
2121         nvme_unquiesce_admin_queue(ctrl);
2122         nvme_auth_stop(ctrl);
2123
2124         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2125                 /* state change failure is ok if we started ctrl delete */
2126                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2127                              ctrl->state != NVME_CTRL_DELETING_NOIO);
2128                 return;
2129         }
2130
2131         nvme_tcp_reconnect_or_remove(ctrl);
2132 }
2133
2134 static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
2135 {
2136         nvme_tcp_teardown_io_queues(ctrl, shutdown);
2137         nvme_quiesce_admin_queue(ctrl);
2138         nvme_disable_ctrl(ctrl, shutdown);
2139         nvme_tcp_teardown_admin_queue(ctrl, shutdown);
2140 }
2141
2142 static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
2143 {
2144         nvme_tcp_teardown_ctrl(ctrl, true);
2145 }
2146
2147 static void nvme_reset_ctrl_work(struct work_struct *work)
2148 {
2149         struct nvme_ctrl *ctrl =
2150                 container_of(work, struct nvme_ctrl, reset_work);
2151
2152         nvme_stop_ctrl(ctrl);
2153         nvme_tcp_teardown_ctrl(ctrl, false);
2154
2155         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2156                 /* state change failure is ok if we started ctrl delete */
2157                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2158                              ctrl->state != NVME_CTRL_DELETING_NOIO);
2159                 return;
2160         }
2161
2162         if (nvme_tcp_setup_ctrl(ctrl, false))
2163                 goto out_fail;
2164
2165         return;
2166
2167 out_fail:
2168         ++ctrl->nr_reconnects;
2169         nvme_tcp_reconnect_or_remove(ctrl);
2170 }
2171
2172 static void nvme_tcp_stop_ctrl(struct nvme_ctrl *ctrl)
2173 {
2174         flush_work(&to_tcp_ctrl(ctrl)->err_work);
2175         cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
2176 }
2177
2178 static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
2179 {
2180         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
2181
2182         if (list_empty(&ctrl->list))
2183                 goto free_ctrl;
2184
2185         mutex_lock(&nvme_tcp_ctrl_mutex);
2186         list_del(&ctrl->list);
2187         mutex_unlock(&nvme_tcp_ctrl_mutex);
2188
2189         nvmf_free_options(nctrl->opts);
2190 free_ctrl:
2191         kfree(ctrl->queues);
2192         kfree(ctrl);
2193 }
2194
2195 static void nvme_tcp_set_sg_null(struct nvme_command *c)
2196 {
2197         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2198
2199         sg->addr = 0;
2200         sg->length = 0;
2201         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2202                         NVME_SGL_FMT_TRANSPORT_A;
2203 }
2204
2205 static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
2206                 struct nvme_command *c, u32 data_len)
2207 {
2208         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2209
2210         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
2211         sg->length = cpu_to_le32(data_len);
2212         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
2213 }
2214
2215 static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
2216                 u32 data_len)
2217 {
2218         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2219
2220         sg->addr = 0;
2221         sg->length = cpu_to_le32(data_len);
2222         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2223                         NVME_SGL_FMT_TRANSPORT_A;
2224 }
2225
2226 static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
2227 {
2228         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2229         struct nvme_tcp_queue *queue = &ctrl->queues[0];
2230         struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2231         struct nvme_command *cmd = &pdu->cmd;
2232         u8 hdgst = nvme_tcp_hdgst_len(queue);
2233
2234         memset(pdu, 0, sizeof(*pdu));
2235         pdu->hdr.type = nvme_tcp_cmd;
2236         if (queue->hdr_digest)
2237                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2238         pdu->hdr.hlen = sizeof(*pdu);
2239         pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2240
2241         cmd->common.opcode = nvme_admin_async_event;
2242         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2243         cmd->common.flags |= NVME_CMD_SGL_METABUF;
2244         nvme_tcp_set_sg_null(cmd);
2245
2246         ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2247         ctrl->async_req.offset = 0;
2248         ctrl->async_req.curr_bio = NULL;
2249         ctrl->async_req.data_len = 0;
2250
2251         nvme_tcp_queue_request(&ctrl->async_req, true, true);
2252 }
2253
2254 static void nvme_tcp_complete_timed_out(struct request *rq)
2255 {
2256         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2257         struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2258
2259         nvme_tcp_stop_queue(ctrl, nvme_tcp_queue_id(req->queue));
2260         nvmf_complete_timed_out_request(rq);
2261 }
2262
2263 static enum blk_eh_timer_return nvme_tcp_timeout(struct request *rq)
2264 {
2265         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2266         struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2267         struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
2268         u8 opc = pdu->cmd.common.opcode, fctype = pdu->cmd.fabrics.fctype;
2269         int qid = nvme_tcp_queue_id(req->queue);
2270
2271         dev_warn(ctrl->device,
2272                 "queue %d: timeout cid %#x type %d opcode %#x (%s)\n",
2273                 nvme_tcp_queue_id(req->queue), nvme_cid(rq), pdu->hdr.type,
2274                 opc, nvme_opcode_str(qid, opc, fctype));
2275
2276         if (ctrl->state != NVME_CTRL_LIVE) {
2277                 /*
2278                  * If we are resetting, connecting or deleting we should
2279                  * complete immediately because we may block controller
2280                  * teardown or setup sequence
2281                  * - ctrl disable/shutdown fabrics requests
2282                  * - connect requests
2283                  * - initialization admin requests
2284                  * - I/O requests that entered after unquiescing and
2285                  *   the controller stopped responding
2286                  *
2287                  * All other requests should be cancelled by the error
2288                  * recovery work, so it's fine that we fail it here.
2289                  */
2290                 nvme_tcp_complete_timed_out(rq);
2291                 return BLK_EH_DONE;
2292         }
2293
2294         /*
2295          * LIVE state should trigger the normal error recovery which will
2296          * handle completing this request.
2297          */
2298         nvme_tcp_error_recovery(ctrl);
2299         return BLK_EH_RESET_TIMER;
2300 }
2301
2302 static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2303                         struct request *rq)
2304 {
2305         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2306         struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
2307         struct nvme_command *c = &pdu->cmd;
2308
2309         c->common.flags |= NVME_CMD_SGL_METABUF;
2310
2311         if (!blk_rq_nr_phys_segments(rq))
2312                 nvme_tcp_set_sg_null(c);
2313         else if (rq_data_dir(rq) == WRITE &&
2314             req->data_len <= nvme_tcp_inline_data_size(req))
2315                 nvme_tcp_set_sg_inline(queue, c, req->data_len);
2316         else
2317                 nvme_tcp_set_sg_host_data(c, req->data_len);
2318
2319         return 0;
2320 }
2321
2322 static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2323                 struct request *rq)
2324 {
2325         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2326         struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
2327         struct nvme_tcp_queue *queue = req->queue;
2328         u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2329         blk_status_t ret;
2330
2331         ret = nvme_setup_cmd(ns, rq);
2332         if (ret)
2333                 return ret;
2334
2335         req->state = NVME_TCP_SEND_CMD_PDU;
2336         req->status = cpu_to_le16(NVME_SC_SUCCESS);
2337         req->offset = 0;
2338         req->data_sent = 0;
2339         req->pdu_len = 0;
2340         req->pdu_sent = 0;
2341         req->h2cdata_left = 0;
2342         req->data_len = blk_rq_nr_phys_segments(rq) ?
2343                                 blk_rq_payload_bytes(rq) : 0;
2344         req->curr_bio = rq->bio;
2345         if (req->curr_bio && req->data_len)
2346                 nvme_tcp_init_iter(req, rq_data_dir(rq));
2347
2348         if (rq_data_dir(rq) == WRITE &&
2349             req->data_len <= nvme_tcp_inline_data_size(req))
2350                 req->pdu_len = req->data_len;
2351
2352         pdu->hdr.type = nvme_tcp_cmd;
2353         pdu->hdr.flags = 0;
2354         if (queue->hdr_digest)
2355                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2356         if (queue->data_digest && req->pdu_len) {
2357                 pdu->hdr.flags |= NVME_TCP_F_DDGST;
2358                 ddgst = nvme_tcp_ddgst_len(queue);
2359         }
2360         pdu->hdr.hlen = sizeof(*pdu);
2361         pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2362         pdu->hdr.plen =
2363                 cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2364
2365         ret = nvme_tcp_map_data(queue, rq);
2366         if (unlikely(ret)) {
2367                 nvme_cleanup_cmd(rq);
2368                 dev_err(queue->ctrl->ctrl.device,
2369                         "Failed to map data (%d)\n", ret);
2370                 return ret;
2371         }
2372
2373         return 0;
2374 }
2375
2376 static void nvme_tcp_commit_rqs(struct blk_mq_hw_ctx *hctx)
2377 {
2378         struct nvme_tcp_queue *queue = hctx->driver_data;
2379
2380         if (!llist_empty(&queue->req_list))
2381                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
2382 }
2383
2384 static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2385                 const struct blk_mq_queue_data *bd)
2386 {
2387         struct nvme_ns *ns = hctx->queue->queuedata;
2388         struct nvme_tcp_queue *queue = hctx->driver_data;
2389         struct request *rq = bd->rq;
2390         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2391         bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2392         blk_status_t ret;
2393
2394         if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2395                 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
2396
2397         ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2398         if (unlikely(ret))
2399                 return ret;
2400
2401         nvme_start_request(rq);
2402
2403         nvme_tcp_queue_request(req, true, bd->last);
2404
2405         return BLK_STS_OK;
2406 }
2407
2408 static void nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2409 {
2410         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(set->driver_data);
2411
2412         nvmf_map_queues(set, &ctrl->ctrl, ctrl->io_queues);
2413 }
2414
2415 static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
2416 {
2417         struct nvme_tcp_queue *queue = hctx->driver_data;
2418         struct sock *sk = queue->sock->sk;
2419
2420         if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2421                 return 0;
2422
2423         set_bit(NVME_TCP_Q_POLLING, &queue->flags);
2424         if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue))
2425                 sk_busy_loop(sk, true);
2426         nvme_tcp_try_recv(queue);
2427         clear_bit(NVME_TCP_Q_POLLING, &queue->flags);
2428         return queue->nr_cqe;
2429 }
2430
2431 static int nvme_tcp_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2432 {
2433         struct nvme_tcp_queue *queue = &to_tcp_ctrl(ctrl)->queues[0];
2434         struct sockaddr_storage src_addr;
2435         int ret, len;
2436
2437         len = nvmf_get_address(ctrl, buf, size);
2438
2439         mutex_lock(&queue->queue_lock);
2440
2441         if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2442                 goto done;
2443         ret = kernel_getsockname(queue->sock, (struct sockaddr *)&src_addr);
2444         if (ret > 0) {
2445                 if (len > 0)
2446                         len--; /* strip trailing newline */
2447                 len += scnprintf(buf + len, size - len, "%ssrc_addr=%pISc\n",
2448                                 (len) ? "," : "", &src_addr);
2449         }
2450 done:
2451         mutex_unlock(&queue->queue_lock);
2452
2453         return len;
2454 }
2455
2456 static const struct blk_mq_ops nvme_tcp_mq_ops = {
2457         .queue_rq       = nvme_tcp_queue_rq,
2458         .commit_rqs     = nvme_tcp_commit_rqs,
2459         .complete       = nvme_complete_rq,
2460         .init_request   = nvme_tcp_init_request,
2461         .exit_request   = nvme_tcp_exit_request,
2462         .init_hctx      = nvme_tcp_init_hctx,
2463         .timeout        = nvme_tcp_timeout,
2464         .map_queues     = nvme_tcp_map_queues,
2465         .poll           = nvme_tcp_poll,
2466 };
2467
2468 static const struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2469         .queue_rq       = nvme_tcp_queue_rq,
2470         .complete       = nvme_complete_rq,
2471         .init_request   = nvme_tcp_init_request,
2472         .exit_request   = nvme_tcp_exit_request,
2473         .init_hctx      = nvme_tcp_init_admin_hctx,
2474         .timeout        = nvme_tcp_timeout,
2475 };
2476
2477 static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2478         .name                   = "tcp",
2479         .module                 = THIS_MODULE,
2480         .flags                  = NVME_F_FABRICS | NVME_F_BLOCKING,
2481         .reg_read32             = nvmf_reg_read32,
2482         .reg_read64             = nvmf_reg_read64,
2483         .reg_write32            = nvmf_reg_write32,
2484         .free_ctrl              = nvme_tcp_free_ctrl,
2485         .submit_async_event     = nvme_tcp_submit_async_event,
2486         .delete_ctrl            = nvme_tcp_delete_ctrl,
2487         .get_address            = nvme_tcp_get_address,
2488         .stop_ctrl              = nvme_tcp_stop_ctrl,
2489 };
2490
2491 static bool
2492 nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2493 {
2494         struct nvme_tcp_ctrl *ctrl;
2495         bool found = false;
2496
2497         mutex_lock(&nvme_tcp_ctrl_mutex);
2498         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2499                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2500                 if (found)
2501                         break;
2502         }
2503         mutex_unlock(&nvme_tcp_ctrl_mutex);
2504
2505         return found;
2506 }
2507
2508 static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2509                 struct nvmf_ctrl_options *opts)
2510 {
2511         struct nvme_tcp_ctrl *ctrl;
2512         int ret;
2513
2514         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2515         if (!ctrl)
2516                 return ERR_PTR(-ENOMEM);
2517
2518         INIT_LIST_HEAD(&ctrl->list);
2519         ctrl->ctrl.opts = opts;
2520         ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2521                                 opts->nr_poll_queues + 1;
2522         ctrl->ctrl.sqsize = opts->queue_size - 1;
2523         ctrl->ctrl.kato = opts->kato;
2524
2525         INIT_DELAYED_WORK(&ctrl->connect_work,
2526                         nvme_tcp_reconnect_ctrl_work);
2527         INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2528         INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2529
2530         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2531                 opts->trsvcid =
2532                         kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2533                 if (!opts->trsvcid) {
2534                         ret = -ENOMEM;
2535                         goto out_free_ctrl;
2536                 }
2537                 opts->mask |= NVMF_OPT_TRSVCID;
2538         }
2539
2540         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2541                         opts->traddr, opts->trsvcid, &ctrl->addr);
2542         if (ret) {
2543                 pr_err("malformed address passed: %s:%s\n",
2544                         opts->traddr, opts->trsvcid);
2545                 goto out_free_ctrl;
2546         }
2547
2548         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2549                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2550                         opts->host_traddr, NULL, &ctrl->src_addr);
2551                 if (ret) {
2552                         pr_err("malformed src address passed: %s\n",
2553                                opts->host_traddr);
2554                         goto out_free_ctrl;
2555                 }
2556         }
2557
2558         if (opts->mask & NVMF_OPT_HOST_IFACE) {
2559                 if (!__dev_get_by_name(&init_net, opts->host_iface)) {
2560                         pr_err("invalid interface passed: %s\n",
2561                                opts->host_iface);
2562                         ret = -ENODEV;
2563                         goto out_free_ctrl;
2564                 }
2565         }
2566
2567         if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2568                 ret = -EALREADY;
2569                 goto out_free_ctrl;
2570         }
2571
2572         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2573                                 GFP_KERNEL);
2574         if (!ctrl->queues) {
2575                 ret = -ENOMEM;
2576                 goto out_free_ctrl;
2577         }
2578
2579         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2580         if (ret)
2581                 goto out_kfree_queues;
2582
2583         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2584                 WARN_ON_ONCE(1);
2585                 ret = -EINTR;
2586                 goto out_uninit_ctrl;
2587         }
2588
2589         ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2590         if (ret)
2591                 goto out_uninit_ctrl;
2592
2593         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
2594                 nvmf_ctrl_subsysnqn(&ctrl->ctrl), &ctrl->addr);
2595
2596         mutex_lock(&nvme_tcp_ctrl_mutex);
2597         list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2598         mutex_unlock(&nvme_tcp_ctrl_mutex);
2599
2600         return &ctrl->ctrl;
2601
2602 out_uninit_ctrl:
2603         nvme_uninit_ctrl(&ctrl->ctrl);
2604         nvme_put_ctrl(&ctrl->ctrl);
2605         if (ret > 0)
2606                 ret = -EIO;
2607         return ERR_PTR(ret);
2608 out_kfree_queues:
2609         kfree(ctrl->queues);
2610 out_free_ctrl:
2611         kfree(ctrl);
2612         return ERR_PTR(ret);
2613 }
2614
2615 static struct nvmf_transport_ops nvme_tcp_transport = {
2616         .name           = "tcp",
2617         .module         = THIS_MODULE,
2618         .required_opts  = NVMF_OPT_TRADDR,
2619         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2620                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2621                           NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
2622                           NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2623                           NVMF_OPT_TOS | NVMF_OPT_HOST_IFACE,
2624         .create_ctrl    = nvme_tcp_create_ctrl,
2625 };
2626
2627 static int __init nvme_tcp_init_module(void)
2628 {
2629         BUILD_BUG_ON(sizeof(struct nvme_tcp_hdr) != 8);
2630         BUILD_BUG_ON(sizeof(struct nvme_tcp_cmd_pdu) != 72);
2631         BUILD_BUG_ON(sizeof(struct nvme_tcp_data_pdu) != 24);
2632         BUILD_BUG_ON(sizeof(struct nvme_tcp_rsp_pdu) != 24);
2633         BUILD_BUG_ON(sizeof(struct nvme_tcp_r2t_pdu) != 24);
2634         BUILD_BUG_ON(sizeof(struct nvme_tcp_icreq_pdu) != 128);
2635         BUILD_BUG_ON(sizeof(struct nvme_tcp_icresp_pdu) != 128);
2636         BUILD_BUG_ON(sizeof(struct nvme_tcp_term_pdu) != 24);
2637
2638         nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
2639                         WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2640         if (!nvme_tcp_wq)
2641                 return -ENOMEM;
2642
2643         nvmf_register_transport(&nvme_tcp_transport);
2644         return 0;
2645 }
2646
2647 static void __exit nvme_tcp_cleanup_module(void)
2648 {
2649         struct nvme_tcp_ctrl *ctrl;
2650
2651         nvmf_unregister_transport(&nvme_tcp_transport);
2652
2653         mutex_lock(&nvme_tcp_ctrl_mutex);
2654         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
2655                 nvme_delete_ctrl(&ctrl->ctrl);
2656         mutex_unlock(&nvme_tcp_ctrl_mutex);
2657         flush_workqueue(nvme_delete_wq);
2658
2659         destroy_workqueue(nvme_tcp_wq);
2660 }
2661
2662 module_init(nvme_tcp_init_module);
2663 module_exit(nvme_tcp_cleanup_module);
2664
2665 MODULE_LICENSE("GPL v2");