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