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