drivers/nvme/host/tcp.c

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