drivers/nvme/host/tcp.c

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