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