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