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 = nvme_find_rq(nvme_tcp_tagset(queue), cqe->command_id);
 491         if (!rq) {
 492                 dev_err(queue->ctrl->ctrl.device,
 493                         "got bad cqe.command_id %#x on queue %d\n",
 494                         cqe->command_id, nvme_tcp_queue_id(queue));
 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 = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
 512         if (!rq) {
 513                 dev_err(queue->ctrl->ctrl.device,
 514                         "got bad c2hdata.command_id %#x on queue %d\n",
 515                         pdu->command_id, nvme_tcp_queue_id(queue));
 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 = nvme_cid(rq);
 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 = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
 623         if (!rq) {
 624                 dev_err(queue->ctrl->ctrl.device,
 625                         "got bad r2t.command_id %#x on queue %d\n",
 626                         pdu->command_id, nvme_tcp_queue_id(queue));
 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 request *rq =
 706                 nvme_cid_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
 707         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 708
 709         while (true) {
 710                 int recv_len, ret;
 711
 712                 recv_len = min_t(size_t, *len, queue->data_remaining);
 713                 if (!recv_len)
 714                         break;
 715
 716                 if (!iov_iter_count(&req->iter)) {
 717                         req->curr_bio = req->curr_bio->bi_next;
 718
 719                         /*
 720                          * If we don`t have any bios it means that controller
 721                          * sent more data than we requested, hence error
 722                          */
 723                         if (!req->curr_bio) {
 724                                 dev_err(queue->ctrl->ctrl.device,
 725                                         "queue %d no space in request %#x",
 726                                         nvme_tcp_queue_id(queue), rq->tag);
 727                                 nvme_tcp_init_recv_ctx(queue);
 728                                 return -EIO;
 729                         }
 730                         nvme_tcp_init_iter(req, READ);
 731                 }
 732
 733                 /* we can read only from what is left in this bio */
 734                 recv_len = min_t(size_t, recv_len,
 735                                 iov_iter_count(&req->iter));
 736
 737                 if (queue->data_digest)
 738                         ret = skb_copy_and_hash_datagram_iter(skb, *offset,
 739                                 &req->iter, recv_len, queue->rcv_hash);
 740                 else
 741                         ret = skb_copy_datagram_iter(skb, *offset,
 742                                         &req->iter, recv_len);
 743                 if (ret) {
 744                         dev_err(queue->ctrl->ctrl.device,
 745                                 "queue %d failed to copy request %#x data",
 746                                 nvme_tcp_queue_id(queue), rq->tag);
 747                         return ret;
 748                 }
 749
 750                 *len -= recv_len;
 751                 *offset += recv_len;
 752                 queue->data_remaining -= recv_len;
 753         }
 754
 755         if (!queue->data_remaining) {
 756                 if (queue->data_digest) {
 757                         nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst);
 758                         queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
 759                 } else {
 760                         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
 761                                 nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
 762                                 queue->nr_cqe++;
 763                         }
 764                         nvme_tcp_init_recv_ctx(queue);
 765                 }
 766         }
 767
 768         return 0;
 769 }
 770
 771 static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
 772                 struct sk_buff *skb, unsigned int *offset, size_t *len)
 773 {
 774         struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
 775         char *ddgst = (char *)&queue->recv_ddgst;
 776         size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
 777         off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
 778         int ret;
 779
 780         ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
 781         if (unlikely(ret))
 782                 return ret;
 783
 784         queue->ddgst_remaining -= recv_len;
 785         *offset += recv_len;
 786         *len -= recv_len;
 787         if (queue->ddgst_remaining)
 788                 return 0;
 789
 790         if (queue->recv_ddgst != queue->exp_ddgst) {
 791                 dev_err(queue->ctrl->ctrl.device,
 792                         "data digest error: recv %#x expected %#x\n",
 793                         le32_to_cpu(queue->recv_ddgst),
 794                         le32_to_cpu(queue->exp_ddgst));
 795                 return -EIO;
 796         }
 797
 798         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
 799                 struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
 800                                         pdu->command_id);
 801
 802                 nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
 803                 queue->nr_cqe++;
 804         }
 805
 806         nvme_tcp_init_recv_ctx(queue);
 807         return 0;
 808 }
 809
 810 static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
 811                              unsigned int offset, size_t len)
 812 {
 813         struct nvme_tcp_queue *queue = desc->arg.data;
 814         size_t consumed = len;
 815         int result;
 816
 817         while (len) {
 818                 switch (nvme_tcp_recv_state(queue)) {
 819                 case NVME_TCP_RECV_PDU:
 820                         result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
 821                         break;
 822                 case NVME_TCP_RECV_DATA:
 823                         result = nvme_tcp_recv_data(queue, skb, &offset, &len);
 824                         break;
 825                 case NVME_TCP_RECV_DDGST:
 826                         result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
 827                         break;
 828                 default:
 829                         result = -EFAULT;
 830                 }
 831                 if (result) {
 832                         dev_err(queue->ctrl->ctrl.device,
 833                                 "receive failed:  %d\n", result);
 834                         queue->rd_enabled = false;
 835                         nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 836                         return result;
 837                 }
 838         }
 839
 840         return consumed;
 841 }
 842
 843 static void nvme_tcp_data_ready(struct sock *sk)
 844 {
 845         struct nvme_tcp_queue *queue;
 846
 847         read_lock_bh(&sk->sk_callback_lock);
 848         queue = sk->sk_user_data;
 849         if (likely(queue && queue->rd_enabled) &&
 850             !test_bit(NVME_TCP_Q_POLLING, &queue->flags))
 851                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 852         read_unlock_bh(&sk->sk_callback_lock);
 853 }
 854
 855 static void nvme_tcp_write_space(struct sock *sk)
 856 {
 857         struct nvme_tcp_queue *queue;
 858
 859         read_lock_bh(&sk->sk_callback_lock);
 860         queue = sk->sk_user_data;
 861         if (likely(queue && sk_stream_is_writeable(sk))) {
 862                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 863                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 864         }
 865         read_unlock_bh(&sk->sk_callback_lock);
 866 }
 867
 868 static void nvme_tcp_state_change(struct sock *sk)
 869 {
 870         struct nvme_tcp_queue *queue;
 871
 872         read_lock_bh(&sk->sk_callback_lock);
 873         queue = sk->sk_user_data;
 874         if (!queue)
 875                 goto done;
 876
 877         switch (sk->sk_state) {
 878         case TCP_CLOSE:
 879         case TCP_CLOSE_WAIT:
 880         case TCP_LAST_ACK:
 881         case TCP_FIN_WAIT1:
 882         case TCP_FIN_WAIT2:
 883                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 884                 break;
 885         default:
 886                 dev_info(queue->ctrl->ctrl.device,
 887                         "queue %d socket state %d\n",
 888                         nvme_tcp_queue_id(queue), sk->sk_state);
 889         }
 890
 891         queue->state_change(sk);
 892 done:
 893         read_unlock_bh(&sk->sk_callback_lock);
 894 }
 895
 896 static inline bool nvme_tcp_queue_more(struct nvme_tcp_queue *queue)
 897 {
 898         return !list_empty(&queue->send_list) ||
 899                 !llist_empty(&queue->req_list) || queue->more_requests;
 900 }
 901
 902 static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
 903 {
 904         queue->request = NULL;
 905 }
 906
 907 static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
 908 {
 909         nvme_tcp_end_request(blk_mq_rq_from_pdu(req), NVME_SC_HOST_PATH_ERROR);
 910 }
 911
 912 static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
 913 {
 914         struct nvme_tcp_queue *queue = req->queue;
 915
 916         while (true) {
 917                 struct page *page = nvme_tcp_req_cur_page(req);
 918                 size_t offset = nvme_tcp_req_cur_offset(req);
 919                 size_t len = nvme_tcp_req_cur_length(req);
 920                 bool last = nvme_tcp_pdu_last_send(req, len);
 921                 int ret, flags = MSG_DONTWAIT;
 922
 923                 if (last && !queue->data_digest && !nvme_tcp_queue_more(queue))
 924                         flags |= MSG_EOR;
 925                 else
 926                         flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
 927
 928                 if (sendpage_ok(page)) {
 929                         ret = kernel_sendpage(queue->sock, page, offset, len,
 930                                         flags);
 931                 } else {
 932                         ret = sock_no_sendpage(queue->sock, page, offset, len,
 933                                         flags);
 934                 }
 935                 if (ret <= 0)
 936                         return ret;
 937
 938                 if (queue->data_digest)
 939                         nvme_tcp_ddgst_update(queue->snd_hash, page,
 940                                         offset, ret);
 941
 942                 /* fully successful last write*/
 943                 if (last && ret == len) {
 944                         if (queue->data_digest) {
 945                                 nvme_tcp_ddgst_final(queue->snd_hash,
 946                                         &req->ddgst);
 947                                 req->state = NVME_TCP_SEND_DDGST;
 948                                 req->offset = 0;
 949                         } else {
 950                                 nvme_tcp_done_send_req(queue);
 951                         }
 952                         return 1;
 953                 }
 954                 nvme_tcp_advance_req(req, ret);
 955         }
 956         return -EAGAIN;
 957 }
 958
 959 static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
 960 {
 961         struct nvme_tcp_queue *queue = req->queue;
 962         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
 963         bool inline_data = nvme_tcp_has_inline_data(req);
 964         u8 hdgst = nvme_tcp_hdgst_len(queue);
 965         int len = sizeof(*pdu) + hdgst - req->offset;
 966         int flags = MSG_DONTWAIT;
 967         int ret;
 968
 969         if (inline_data || nvme_tcp_queue_more(queue))
 970                 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
 971         else
 972                 flags |= MSG_EOR;
 973
 974         if (queue->hdr_digest && !req->offset)
 975                 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
 976
 977         ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
 978                         offset_in_page(pdu) + req->offset, len,  flags);
 979         if (unlikely(ret <= 0))
 980                 return ret;
 981
 982         len -= ret;
 983         if (!len) {
 984                 if (inline_data) {
 985                         req->state = NVME_TCP_SEND_DATA;
 986                         if (queue->data_digest)
 987                                 crypto_ahash_init(queue->snd_hash);
 988                 } else {
 989                         nvme_tcp_done_send_req(queue);
 990                 }
 991                 return 1;
 992         }
 993         req->offset += ret;
 994
 995         return -EAGAIN;
 996 }
 997
 998 static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
 999 {
1000         struct nvme_tcp_queue *queue = req->queue;
1001         struct nvme_tcp_data_pdu *pdu = req->pdu;
1002         u8 hdgst = nvme_tcp_hdgst_len(queue);
1003         int len = sizeof(*pdu) - req->offset + hdgst;
1004         int ret;
1005
1006         if (queue->hdr_digest && !req->offset)
1007                 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
1008
1009         ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
1010                         offset_in_page(pdu) + req->offset, len,
1011                         MSG_DONTWAIT | MSG_MORE | MSG_SENDPAGE_NOTLAST);
1012         if (unlikely(ret <= 0))
1013                 return ret;
1014
1015         len -= ret;
1016         if (!len) {
1017                 req->state = NVME_TCP_SEND_DATA;
1018                 if (queue->data_digest)
1019                         crypto_ahash_init(queue->snd_hash);
1020                 return 1;
1021         }
1022         req->offset += ret;
1023
1024         return -EAGAIN;
1025 }
1026
1027 static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
1028 {
1029         struct nvme_tcp_queue *queue = req->queue;
1030         int ret;
1031         struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1032         struct kvec iov = {
1033                 .iov_base = &req->ddgst + req->offset,
1034                 .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
1035         };
1036
1037         if (nvme_tcp_queue_more(queue))
1038                 msg.msg_flags |= MSG_MORE;
1039         else
1040                 msg.msg_flags |= MSG_EOR;
1041
1042         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1043         if (unlikely(ret <= 0))
1044                 return ret;
1045
1046         if (req->offset + ret == NVME_TCP_DIGEST_LENGTH) {
1047                 nvme_tcp_done_send_req(queue);
1048                 return 1;
1049         }
1050
1051         req->offset += ret;
1052         return -EAGAIN;
1053 }
1054
1055 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
1056 {
1057         struct nvme_tcp_request *req;
1058         int ret = 1;
1059
1060         if (!queue->request) {
1061                 queue->request = nvme_tcp_fetch_request(queue);
1062                 if (!queue->request)
1063                         return 0;
1064         }
1065         req = queue->request;
1066
1067         if (req->state == NVME_TCP_SEND_CMD_PDU) {
1068                 ret = nvme_tcp_try_send_cmd_pdu(req);
1069                 if (ret <= 0)
1070                         goto done;
1071                 if (!nvme_tcp_has_inline_data(req))
1072                         return ret;
1073         }
1074
1075         if (req->state == NVME_TCP_SEND_H2C_PDU) {
1076                 ret = nvme_tcp_try_send_data_pdu(req);
1077                 if (ret <= 0)
1078                         goto done;
1079         }
1080
1081         if (req->state == NVME_TCP_SEND_DATA) {
1082                 ret = nvme_tcp_try_send_data(req);
1083                 if (ret <= 0)
1084                         goto done;
1085         }
1086
1087         if (req->state == NVME_TCP_SEND_DDGST)
1088                 ret = nvme_tcp_try_send_ddgst(req);
1089 done:
1090         if (ret == -EAGAIN) {
1091                 ret = 0;
1092         } else if (ret < 0) {
1093                 dev_err(queue->ctrl->ctrl.device,
1094                         "failed to send request %d\n", ret);
1095                 if (ret != -EPIPE && ret != -ECONNRESET)
1096                         nvme_tcp_fail_request(queue->request);
1097                 nvme_tcp_done_send_req(queue);
1098         }
1099         return ret;
1100 }
1101
1102 static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1103 {
1104         struct socket *sock = queue->sock;
1105         struct sock *sk = sock->sk;
1106         read_descriptor_t rd_desc;
1107         int consumed;
1108
1109         rd_desc.arg.data = queue;
1110         rd_desc.count = 1;
1111         lock_sock(sk);
1112         queue->nr_cqe = 0;
1113         consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1114         release_sock(sk);
1115         return consumed;
1116 }
1117
1118 static void nvme_tcp_io_work(struct work_struct *w)
1119 {
1120         struct nvme_tcp_queue *queue =
1121                 container_of(w, struct nvme_tcp_queue, io_work);
1122         unsigned long deadline = jiffies + msecs_to_jiffies(1);
1123
1124         do {
1125                 bool pending = false;
1126                 int result;
1127
1128                 if (mutex_trylock(&queue->send_mutex)) {
1129                         result = nvme_tcp_try_send(queue);
1130                         mutex_unlock(&queue->send_mutex);
1131                         if (result > 0)
1132                                 pending = true;
1133                         else if (unlikely(result < 0))
1134                                 break;
1135                 } else
1136                         pending = !llist_empty(&queue->req_list);
1137
1138                 result = nvme_tcp_try_recv(queue);
1139                 if (result > 0)
1140                         pending = true;
1141                 else if (unlikely(result < 0))
1142                         return;
1143
1144                 if (!pending)
1145                         return;
1146
1147         } while (!time_after(jiffies, deadline)); /* quota is exhausted */
1148
1149         queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1150 }
1151
1152 static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue)
1153 {
1154         struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
1155
1156         ahash_request_free(queue->rcv_hash);
1157         ahash_request_free(queue->snd_hash);
1158         crypto_free_ahash(tfm);
1159 }
1160
1161 static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue)
1162 {
1163         struct crypto_ahash *tfm;
1164
1165         tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
1166         if (IS_ERR(tfm))
1167                 return PTR_ERR(tfm);
1168
1169         queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1170         if (!queue->snd_hash)
1171                 goto free_tfm;
1172         ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
1173
1174         queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1175         if (!queue->rcv_hash)
1176                 goto free_snd_hash;
1177         ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
1178
1179         return 0;
1180 free_snd_hash:
1181         ahash_request_free(queue->snd_hash);
1182 free_tfm:
1183         crypto_free_ahash(tfm);
1184         return -ENOMEM;
1185 }
1186
1187 static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1188 {
1189         struct nvme_tcp_request *async = &ctrl->async_req;
1190
1191         page_frag_free(async->pdu);
1192 }
1193
1194 static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1195 {
1196         struct nvme_tcp_queue *queue = &ctrl->queues[0];
1197         struct nvme_tcp_request *async = &ctrl->async_req;
1198         u8 hdgst = nvme_tcp_hdgst_len(queue);
1199
1200         async->pdu = page_frag_alloc(&queue->pf_cache,
1201                 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1202                 GFP_KERNEL | __GFP_ZERO);
1203         if (!async->pdu)
1204                 return -ENOMEM;
1205
1206         async->queue = &ctrl->queues[0];
1207         return 0;
1208 }
1209
1210 static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1211 {
1212         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1213         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1214
1215         if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1216                 return;
1217
1218         if (queue->hdr_digest || queue->data_digest)
1219                 nvme_tcp_free_crypto(queue);
1220
1221         sock_release(queue->sock);
1222         kfree(queue->pdu);
1223         mutex_destroy(&queue->send_mutex);
1224         mutex_destroy(&queue->queue_lock);
1225 }
1226
1227 static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1228 {
1229         struct nvme_tcp_icreq_pdu *icreq;
1230         struct nvme_tcp_icresp_pdu *icresp;
1231         struct msghdr msg = {};
1232         struct kvec iov;
1233         bool ctrl_hdgst, ctrl_ddgst;
1234         int ret;
1235
1236         icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1237         if (!icreq)
1238                 return -ENOMEM;
1239
1240         icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1241         if (!icresp) {
1242                 ret = -ENOMEM;
1243                 goto free_icreq;
1244         }
1245
1246         icreq->hdr.type = nvme_tcp_icreq;
1247         icreq->hdr.hlen = sizeof(*icreq);
1248         icreq->hdr.pdo = 0;
1249         icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1250         icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1251         icreq->maxr2t = 0; /* single inflight r2t supported */
1252         icreq->hpda = 0; /* no alignment constraint */
1253         if (queue->hdr_digest)
1254                 icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1255         if (queue->data_digest)
1256                 icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1257
1258         iov.iov_base = icreq;
1259         iov.iov_len = sizeof(*icreq);
1260         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1261         if (ret < 0)
1262                 goto free_icresp;
1263
1264         memset(&msg, 0, sizeof(msg));
1265         iov.iov_base = icresp;
1266         iov.iov_len = sizeof(*icresp);
1267         ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1268                         iov.iov_len, msg.msg_flags);
1269         if (ret < 0)
1270                 goto free_icresp;
1271
1272         ret = -EINVAL;
1273         if (icresp->hdr.type != nvme_tcp_icresp) {
1274                 pr_err("queue %d: bad type returned %d\n",
1275                         nvme_tcp_queue_id(queue), icresp->hdr.type);
1276                 goto free_icresp;
1277         }
1278
1279         if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1280                 pr_err("queue %d: bad pdu length returned %d\n",
1281                         nvme_tcp_queue_id(queue), icresp->hdr.plen);
1282                 goto free_icresp;
1283         }
1284
1285         if (icresp->pfv != NVME_TCP_PFV_1_0) {
1286                 pr_err("queue %d: bad pfv returned %d\n",
1287                         nvme_tcp_queue_id(queue), icresp->pfv);
1288                 goto free_icresp;
1289         }
1290
1291         ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1292         if ((queue->data_digest && !ctrl_ddgst) ||
1293             (!queue->data_digest && ctrl_ddgst)) {
1294                 pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1295                         nvme_tcp_queue_id(queue),
1296                         queue->data_digest ? "enabled" : "disabled",
1297                         ctrl_ddgst ? "enabled" : "disabled");
1298                 goto free_icresp;
1299         }
1300
1301         ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1302         if ((queue->hdr_digest && !ctrl_hdgst) ||
1303             (!queue->hdr_digest && ctrl_hdgst)) {
1304                 pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1305                         nvme_tcp_queue_id(queue),
1306                         queue->hdr_digest ? "enabled" : "disabled",
1307                         ctrl_hdgst ? "enabled" : "disabled");
1308                 goto free_icresp;
1309         }
1310
1311         if (icresp->cpda != 0) {
1312                 pr_err("queue %d: unsupported cpda returned %d\n",
1313                         nvme_tcp_queue_id(queue), icresp->cpda);
1314                 goto free_icresp;
1315         }
1316
1317         ret = 0;
1318 free_icresp:
1319         kfree(icresp);
1320 free_icreq:
1321         kfree(icreq);
1322         return ret;
1323 }
1324
1325 static bool nvme_tcp_admin_queue(struct nvme_tcp_queue *queue)
1326 {
1327         return nvme_tcp_queue_id(queue) == 0;
1328 }
1329
1330 static bool nvme_tcp_default_queue(struct nvme_tcp_queue *queue)
1331 {
1332         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1333         int qid = nvme_tcp_queue_id(queue);
1334
1335         return !nvme_tcp_admin_queue(queue) &&
1336                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT];
1337 }
1338
1339 static bool nvme_tcp_read_queue(struct nvme_tcp_queue *queue)
1340 {
1341         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1342         int qid = nvme_tcp_queue_id(queue);
1343
1344         return !nvme_tcp_admin_queue(queue) &&
1345                 !nvme_tcp_default_queue(queue) &&
1346                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1347                           ctrl->io_queues[HCTX_TYPE_READ];
1348 }
1349
1350 static bool nvme_tcp_poll_queue(struct nvme_tcp_queue *queue)
1351 {
1352         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1353         int qid = nvme_tcp_queue_id(queue);
1354
1355         return !nvme_tcp_admin_queue(queue) &&
1356                 !nvme_tcp_default_queue(queue) &&
1357                 !nvme_tcp_read_queue(queue) &&
1358                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1359                           ctrl->io_queues[HCTX_TYPE_READ] +
1360                           ctrl->io_queues[HCTX_TYPE_POLL];
1361 }
1362
1363 static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
1364 {
1365         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1366         int qid = nvme_tcp_queue_id(queue);
1367         int n = 0;
1368
1369         if (nvme_tcp_default_queue(queue))
1370                 n = qid - 1;
1371         else if (nvme_tcp_read_queue(queue))
1372                 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] - 1;
1373         else if (nvme_tcp_poll_queue(queue))
1374                 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] -
1375                                 ctrl->io_queues[HCTX_TYPE_READ] - 1;
1376         queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
1377 }
1378
1379 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl,
1380                 int qid, size_t queue_size)
1381 {
1382         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1383         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1384         int ret, rcv_pdu_size;
1385
1386         mutex_init(&queue->queue_lock);
1387         queue->ctrl = ctrl;
1388         init_llist_head(&queue->req_list);
1389         INIT_LIST_HEAD(&queue->send_list);
1390         mutex_init(&queue->send_mutex);
1391         INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1392         queue->queue_size = queue_size;
1393
1394         if (qid > 0)
1395                 queue->cmnd_capsule_len = nctrl->ioccsz * 16;
1396         else
1397                 queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1398                                                 NVME_TCP_ADMIN_CCSZ;
1399
1400         ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM,
1401                         IPPROTO_TCP, &queue->sock);
1402         if (ret) {
1403                 dev_err(nctrl->device,
1404                         "failed to create socket: %d\n", ret);
1405                 goto err_destroy_mutex;
1406         }
1407
1408         /* Single syn retry */
1409         tcp_sock_set_syncnt(queue->sock->sk, 1);
1410
1411         /* Set TCP no delay */
1412         tcp_sock_set_nodelay(queue->sock->sk);
1413
1414         /*
1415          * Cleanup whatever is sitting in the TCP transmit queue on socket
1416          * close. This is done to prevent stale data from being sent should
1417          * the network connection be restored before TCP times out.
1418          */
1419         sock_no_linger(queue->sock->sk);
1420
1421         if (so_priority > 0)
1422                 sock_set_priority(queue->sock->sk, so_priority);
1423
1424         /* Set socket type of service */
1425         if (nctrl->opts->tos >= 0)
1426                 ip_sock_set_tos(queue->sock->sk, nctrl->opts->tos);
1427
1428         /* Set 10 seconds timeout for icresp recvmsg */
1429         queue->sock->sk->sk_rcvtimeo = 10 * HZ;
1430
1431         queue->sock->sk->sk_allocation = GFP_ATOMIC;
1432         nvme_tcp_set_queue_io_cpu(queue);
1433         queue->request = NULL;
1434         queue->data_remaining = 0;
1435         queue->ddgst_remaining = 0;
1436         queue->pdu_remaining = 0;
1437         queue->pdu_offset = 0;
1438         sk_set_memalloc(queue->sock->sk);
1439
1440         if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
1441                 ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1442                         sizeof(ctrl->src_addr));
1443                 if (ret) {
1444                         dev_err(nctrl->device,
1445                                 "failed to bind queue %d socket %d\n",
1446                                 qid, ret);
1447                         goto err_sock;
1448                 }
1449         }
1450
1451         if (nctrl->opts->mask & NVMF_OPT_HOST_IFACE) {
1452                 char *iface = nctrl->opts->host_iface;
1453                 sockptr_t optval = KERNEL_SOCKPTR(iface);
1454
1455                 ret = sock_setsockopt(queue->sock, SOL_SOCKET, SO_BINDTODEVICE,
1456                                       optval, strlen(iface));
1457                 if (ret) {
1458                         dev_err(nctrl->device,
1459                           "failed to bind to interface %s queue %d err %d\n",
1460                           iface, qid, ret);
1461                         goto err_sock;
1462                 }
1463         }
1464
1465         queue->hdr_digest = nctrl->opts->hdr_digest;
1466         queue->data_digest = nctrl->opts->data_digest;
1467         if (queue->hdr_digest || queue->data_digest) {
1468                 ret = nvme_tcp_alloc_crypto(queue);
1469                 if (ret) {
1470                         dev_err(nctrl->device,
1471                                 "failed to allocate queue %d crypto\n", qid);
1472                         goto err_sock;
1473                 }
1474         }
1475
1476         rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1477                         nvme_tcp_hdgst_len(queue);
1478         queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1479         if (!queue->pdu) {
1480                 ret = -ENOMEM;
1481                 goto err_crypto;
1482         }
1483
1484         dev_dbg(nctrl->device, "connecting queue %d\n",
1485                         nvme_tcp_queue_id(queue));
1486
1487         ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1488                 sizeof(ctrl->addr), 0);
1489         if (ret) {
1490                 dev_err(nctrl->device,
1491                         "failed to connect socket: %d\n", ret);
1492                 goto err_rcv_pdu;
1493         }
1494
1495         ret = nvme_tcp_init_connection(queue);
1496         if (ret)
1497                 goto err_init_connect;
1498
1499         queue->rd_enabled = true;
1500         set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1501         nvme_tcp_init_recv_ctx(queue);
1502
1503         write_lock_bh(&queue->sock->sk->sk_callback_lock);
1504         queue->sock->sk->sk_user_data = queue;
1505         queue->state_change = queue->sock->sk->sk_state_change;
1506         queue->data_ready = queue->sock->sk->sk_data_ready;
1507         queue->write_space = queue->sock->sk->sk_write_space;
1508         queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1509         queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1510         queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1511 #ifdef CONFIG_NET_RX_BUSY_POLL
1512         queue->sock->sk->sk_ll_usec = 1;
1513 #endif
1514         write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1515
1516         return 0;
1517
1518 err_init_connect:
1519         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1520 err_rcv_pdu:
1521         kfree(queue->pdu);
1522 err_crypto:
1523         if (queue->hdr_digest || queue->data_digest)
1524                 nvme_tcp_free_crypto(queue);
1525 err_sock:
1526         sock_release(queue->sock);
1527         queue->sock = NULL;
1528 err_destroy_mutex:
1529         mutex_destroy(&queue->send_mutex);
1530         mutex_destroy(&queue->queue_lock);
1531         return ret;
1532 }
1533
1534 static void nvme_tcp_restore_sock_calls(struct nvme_tcp_queue *queue)
1535 {
1536         struct socket *sock = queue->sock;
1537
1538         write_lock_bh(&sock->sk->sk_callback_lock);
1539         sock->sk->sk_user_data  = NULL;
1540         sock->sk->sk_data_ready = queue->data_ready;
1541         sock->sk->sk_state_change = queue->state_change;
1542         sock->sk->sk_write_space  = queue->write_space;
1543         write_unlock_bh(&sock->sk->sk_callback_lock);
1544 }
1545
1546 static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1547 {
1548         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1549         nvme_tcp_restore_sock_calls(queue);
1550         cancel_work_sync(&queue->io_work);
1551 }
1552
1553 static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1554 {
1555         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1556         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1557
1558         mutex_lock(&queue->queue_lock);
1559         if (test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1560                 __nvme_tcp_stop_queue(queue);
1561         mutex_unlock(&queue->queue_lock);
1562 }
1563
1564 static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1565 {
1566         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1567         int ret;
1568
1569         if (idx)
1570                 ret = nvmf_connect_io_queue(nctrl, idx);
1571         else
1572                 ret = nvmf_connect_admin_queue(nctrl);
1573
1574         if (!ret) {
1575                 set_bit(NVME_TCP_Q_LIVE, &ctrl->queues[idx].flags);
1576         } else {
1577                 if (test_bit(NVME_TCP_Q_ALLOCATED, &ctrl->queues[idx].flags))
1578                         __nvme_tcp_stop_queue(&ctrl->queues[idx]);
1579                 dev_err(nctrl->device,
1580                         "failed to connect queue: %d ret=%d\n", idx, ret);
1581         }
1582         return ret;
1583 }
1584
1585 static struct blk_mq_tag_set *nvme_tcp_alloc_tagset(struct nvme_ctrl *nctrl,
1586                 bool admin)
1587 {
1588         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1589         struct blk_mq_tag_set *set;
1590         int ret;
1591
1592         if (admin) {
1593                 set = &ctrl->admin_tag_set;
1594                 memset(set, 0, sizeof(*set));
1595                 set->ops = &nvme_tcp_admin_mq_ops;
1596                 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1597                 set->reserved_tags = NVMF_RESERVED_TAGS;
1598                 set->numa_node = nctrl->numa_node;
1599                 set->flags = BLK_MQ_F_BLOCKING;
1600                 set->cmd_size = sizeof(struct nvme_tcp_request);
1601                 set->driver_data = ctrl;
1602                 set->nr_hw_queues = 1;
1603                 set->timeout = NVME_ADMIN_TIMEOUT;
1604         } else {
1605                 set = &ctrl->tag_set;
1606                 memset(set, 0, sizeof(*set));
1607                 set->ops = &nvme_tcp_mq_ops;
1608                 set->queue_depth = nctrl->sqsize + 1;
1609                 set->reserved_tags = NVMF_RESERVED_TAGS;
1610                 set->numa_node = nctrl->numa_node;
1611                 set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_BLOCKING;
1612                 set->cmd_size = sizeof(struct nvme_tcp_request);
1613                 set->driver_data = ctrl;
1614                 set->nr_hw_queues = nctrl->queue_count - 1;
1615                 set->timeout = NVME_IO_TIMEOUT;
1616                 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
1617         }
1618
1619         ret = blk_mq_alloc_tag_set(set);
1620         if (ret)
1621                 return ERR_PTR(ret);
1622
1623         return set;
1624 }
1625
1626 static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
1627 {
1628         if (to_tcp_ctrl(ctrl)->async_req.pdu) {
1629                 cancel_work_sync(&ctrl->async_event_work);
1630                 nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
1631                 to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
1632         }
1633
1634         nvme_tcp_free_queue(ctrl, 0);
1635 }
1636
1637 static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
1638 {
1639         int i;
1640
1641         for (i = 1; i < ctrl->queue_count; i++)
1642                 nvme_tcp_free_queue(ctrl, i);
1643 }
1644
1645 static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
1646 {
1647         int i;
1648
1649         for (i = 1; i < ctrl->queue_count; i++)
1650                 nvme_tcp_stop_queue(ctrl, i);
1651 }
1652
1653 static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl)
1654 {
1655         int i, ret = 0;
1656
1657         for (i = 1; i < ctrl->queue_count; i++) {
1658                 ret = nvme_tcp_start_queue(ctrl, i);
1659                 if (ret)
1660                         goto out_stop_queues;
1661         }
1662
1663         return 0;
1664
1665 out_stop_queues:
1666         for (i--; i >= 1; i--)
1667                 nvme_tcp_stop_queue(ctrl, i);
1668         return ret;
1669 }
1670
1671 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
1672 {
1673         int ret;
1674
1675         ret = nvme_tcp_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
1676         if (ret)
1677                 return ret;
1678
1679         ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
1680         if (ret)
1681                 goto out_free_queue;
1682
1683         return 0;
1684
1685 out_free_queue:
1686         nvme_tcp_free_queue(ctrl, 0);
1687         return ret;
1688 }
1689
1690 static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1691 {
1692         int i, ret;
1693
1694         for (i = 1; i < ctrl->queue_count; i++) {
1695                 ret = nvme_tcp_alloc_queue(ctrl, i,
1696                                 ctrl->sqsize + 1);
1697                 if (ret)
1698                         goto out_free_queues;
1699         }
1700
1701         return 0;
1702
1703 out_free_queues:
1704         for (i--; i >= 1; i--)
1705                 nvme_tcp_free_queue(ctrl, i);
1706
1707         return ret;
1708 }
1709
1710 static unsigned int nvme_tcp_nr_io_queues(struct nvme_ctrl *ctrl)
1711 {
1712         unsigned int nr_io_queues;
1713
1714         nr_io_queues = min(ctrl->opts->nr_io_queues, num_online_cpus());
1715         nr_io_queues += min(ctrl->opts->nr_write_queues, num_online_cpus());
1716         nr_io_queues += min(ctrl->opts->nr_poll_queues, num_online_cpus());
1717
1718         return nr_io_queues;
1719 }
1720
1721 static void nvme_tcp_set_io_queues(struct nvme_ctrl *nctrl,
1722                 unsigned int nr_io_queues)
1723 {
1724         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1725         struct nvmf_ctrl_options *opts = nctrl->opts;
1726
1727         if (opts->nr_write_queues && opts->nr_io_queues < nr_io_queues) {
1728                 /*
1729                  * separate read/write queues
1730                  * hand out dedicated default queues only after we have
1731                  * sufficient read queues.
1732                  */
1733                 ctrl->io_queues[HCTX_TYPE_READ] = opts->nr_io_queues;
1734                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
1735                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1736                         min(opts->nr_write_queues, nr_io_queues);
1737                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1738         } else {
1739                 /*
1740                  * shared read/write queues
1741                  * either no write queues were requested, or we don't have
1742                  * sufficient queue count to have dedicated default queues.
1743                  */
1744                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1745                         min(opts->nr_io_queues, nr_io_queues);
1746                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1747         }
1748
1749         if (opts->nr_poll_queues && nr_io_queues) {
1750                 /* map dedicated poll queues only if we have queues left */
1751                 ctrl->io_queues[HCTX_TYPE_POLL] =
1752                         min(opts->nr_poll_queues, nr_io_queues);
1753         }
1754 }
1755
1756 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1757 {
1758         unsigned int nr_io_queues;
1759         int ret;
1760
1761         nr_io_queues = nvme_tcp_nr_io_queues(ctrl);
1762         ret = nvme_set_queue_count(ctrl, &nr_io_queues);
1763         if (ret)
1764                 return ret;
1765
1766         if (nr_io_queues == 0) {
1767                 dev_err(ctrl->device,
1768                         "unable to set any I/O queues\n");
1769                 return -ENOMEM;
1770         }
1771
1772         ctrl->queue_count = nr_io_queues + 1;
1773         dev_info(ctrl->device,
1774                 "creating %d I/O queues.\n", nr_io_queues);
1775
1776         nvme_tcp_set_io_queues(ctrl, nr_io_queues);
1777
1778         return __nvme_tcp_alloc_io_queues(ctrl);
1779 }
1780
1781 static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove)
1782 {
1783         nvme_tcp_stop_io_queues(ctrl);
1784         if (remove) {
1785                 blk_cleanup_queue(ctrl->connect_q);
1786                 blk_mq_free_tag_set(ctrl->tagset);
1787         }
1788         nvme_tcp_free_io_queues(ctrl);
1789 }
1790
1791 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
1792 {
1793         int ret;
1794
1795         ret = nvme_tcp_alloc_io_queues(ctrl);
1796         if (ret)
1797                 return ret;
1798
1799         if (new) {
1800                 ctrl->tagset = nvme_tcp_alloc_tagset(ctrl, false);
1801                 if (IS_ERR(ctrl->tagset)) {
1802                         ret = PTR_ERR(ctrl->tagset);
1803                         goto out_free_io_queues;
1804                 }
1805
1806                 ctrl->connect_q = blk_mq_init_queue(ctrl->tagset);
1807                 if (IS_ERR(ctrl->connect_q)) {
1808                         ret = PTR_ERR(ctrl->connect_q);
1809                         goto out_free_tag_set;
1810                 }
1811         }
1812
1813         ret = nvme_tcp_start_io_queues(ctrl);
1814         if (ret)
1815                 goto out_cleanup_connect_q;
1816
1817         if (!new) {
1818                 nvme_start_queues(ctrl);
1819                 if (!nvme_wait_freeze_timeout(ctrl, NVME_IO_TIMEOUT)) {
1820                         /*
1821                          * If we timed out waiting for freeze we are likely to
1822                          * be stuck.  Fail the controller initialization just
1823                          * to be safe.
1824                          */
1825                         ret = -ENODEV;
1826                         goto out_wait_freeze_timed_out;
1827                 }
1828                 blk_mq_update_nr_hw_queues(ctrl->tagset,
1829                         ctrl->queue_count - 1);
1830                 nvme_unfreeze(ctrl);
1831         }
1832
1833         return 0;
1834
1835 out_wait_freeze_timed_out:
1836         nvme_stop_queues(ctrl);
1837         nvme_sync_io_queues(ctrl);
1838         nvme_tcp_stop_io_queues(ctrl);
1839 out_cleanup_connect_q:
1840         nvme_cancel_tagset(ctrl);
1841         if (new)
1842                 blk_cleanup_queue(ctrl->connect_q);
1843 out_free_tag_set:
1844         if (new)
1845                 blk_mq_free_tag_set(ctrl->tagset);
1846 out_free_io_queues:
1847         nvme_tcp_free_io_queues(ctrl);
1848         return ret;
1849 }
1850
1851 static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove)
1852 {
1853         nvme_tcp_stop_queue(ctrl, 0);
1854         if (remove) {
1855                 blk_cleanup_queue(ctrl->admin_q);
1856                 blk_cleanup_queue(ctrl->fabrics_q);
1857                 blk_mq_free_tag_set(ctrl->admin_tagset);
1858         }
1859         nvme_tcp_free_admin_queue(ctrl);
1860 }
1861
1862 static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
1863 {
1864         int error;
1865
1866         error = nvme_tcp_alloc_admin_queue(ctrl);
1867         if (error)
1868                 return error;
1869
1870         if (new) {
1871                 ctrl->admin_tagset = nvme_tcp_alloc_tagset(ctrl, true);
1872                 if (IS_ERR(ctrl->admin_tagset)) {
1873                         error = PTR_ERR(ctrl->admin_tagset);
1874                         goto out_free_queue;
1875                 }
1876
1877                 ctrl->fabrics_q = blk_mq_init_queue(ctrl->admin_tagset);
1878                 if (IS_ERR(ctrl->fabrics_q)) {
1879                         error = PTR_ERR(ctrl->fabrics_q);
1880                         goto out_free_tagset;
1881                 }
1882
1883                 ctrl->admin_q = blk_mq_init_queue(ctrl->admin_tagset);
1884                 if (IS_ERR(ctrl->admin_q)) {
1885                         error = PTR_ERR(ctrl->admin_q);
1886                         goto out_cleanup_fabrics_q;
1887                 }
1888         }
1889
1890         error = nvme_tcp_start_queue(ctrl, 0);
1891         if (error)
1892                 goto out_cleanup_queue;
1893
1894         error = nvme_enable_ctrl(ctrl);
1895         if (error)
1896                 goto out_stop_queue;
1897
1898         blk_mq_unquiesce_queue(ctrl->admin_q);
1899
1900         error = nvme_init_ctrl_finish(ctrl);
1901         if (error)
1902                 goto out_quiesce_queue;
1903
1904         return 0;
1905
1906 out_quiesce_queue:
1907         blk_mq_quiesce_queue(ctrl->admin_q);
1908         blk_sync_queue(ctrl->admin_q);
1909 out_stop_queue:
1910         nvme_tcp_stop_queue(ctrl, 0);
1911         nvme_cancel_admin_tagset(ctrl);
1912 out_cleanup_queue:
1913         if (new)
1914                 blk_cleanup_queue(ctrl->admin_q);
1915 out_cleanup_fabrics_q:
1916         if (new)
1917                 blk_cleanup_queue(ctrl->fabrics_q);
1918 out_free_tagset:
1919         if (new)
1920                 blk_mq_free_tag_set(ctrl->admin_tagset);
1921 out_free_queue:
1922         nvme_tcp_free_admin_queue(ctrl);
1923         return error;
1924 }
1925
1926 static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
1927                 bool remove)
1928 {
1929         blk_mq_quiesce_queue(ctrl->admin_q);
1930         blk_sync_queue(ctrl->admin_q);
1931         nvme_tcp_stop_queue(ctrl, 0);
1932         nvme_cancel_admin_tagset(ctrl);
1933         if (remove)
1934                 blk_mq_unquiesce_queue(ctrl->admin_q);
1935         nvme_tcp_destroy_admin_queue(ctrl, remove);
1936 }
1937
1938 static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
1939                 bool remove)
1940 {
1941         if (ctrl->queue_count <= 1)
1942                 return;
1943         blk_mq_quiesce_queue(ctrl->admin_q);
1944         nvme_start_freeze(ctrl);
1945         nvme_stop_queues(ctrl);
1946         nvme_sync_io_queues(ctrl);
1947         nvme_tcp_stop_io_queues(ctrl);
1948         nvme_cancel_tagset(ctrl);
1949         if (remove)
1950                 nvme_start_queues(ctrl);
1951         nvme_tcp_destroy_io_queues(ctrl, remove);
1952 }
1953
1954 static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
1955 {
1956         /* If we are resetting/deleting then do nothing */
1957         if (ctrl->state != NVME_CTRL_CONNECTING) {
1958                 WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW ||
1959                         ctrl->state == NVME_CTRL_LIVE);
1960                 return;
1961         }
1962
1963         if (nvmf_should_reconnect(ctrl)) {
1964                 dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
1965                         ctrl->opts->reconnect_delay);
1966                 queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
1967                                 ctrl->opts->reconnect_delay * HZ);
1968         } else {
1969                 dev_info(ctrl->device, "Removing controller...\n");
1970                 nvme_delete_ctrl(ctrl);
1971         }
1972 }
1973
1974 static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
1975 {
1976         struct nvmf_ctrl_options *opts = ctrl->opts;
1977         int ret;
1978
1979         ret = nvme_tcp_configure_admin_queue(ctrl, new);
1980         if (ret)
1981                 return ret;
1982
1983         if (ctrl->icdoff) {
1984                 ret = -EOPNOTSUPP;
1985                 dev_err(ctrl->device, "icdoff is not supported!\n");
1986                 goto destroy_admin;
1987         }
1988
1989         if (!nvme_ctrl_sgl_supported(ctrl)) {
1990                 ret = -EOPNOTSUPP;
1991                 dev_err(ctrl->device, "Mandatory sgls are not supported!\n");
1992                 goto destroy_admin;
1993         }
1994
1995         if (opts->queue_size > ctrl->sqsize + 1)
1996                 dev_warn(ctrl->device,
1997                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
1998                         opts->queue_size, ctrl->sqsize + 1);
1999
2000         if (ctrl->sqsize + 1 > ctrl->maxcmd) {
2001                 dev_warn(ctrl->device,
2002                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
2003                         ctrl->sqsize + 1, ctrl->maxcmd);
2004                 ctrl->sqsize = ctrl->maxcmd - 1;
2005         }
2006
2007         if (ctrl->queue_count > 1) {
2008                 ret = nvme_tcp_configure_io_queues(ctrl, new);
2009                 if (ret)
2010                         goto destroy_admin;
2011         }
2012
2013         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
2014                 /*
2015                  * state change failure is ok if we started ctrl delete,
2016                  * unless we're during creation of a new controller to
2017                  * avoid races with teardown flow.
2018                  */
2019                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2020                              ctrl->state != NVME_CTRL_DELETING_NOIO);
2021                 WARN_ON_ONCE(new);
2022                 ret = -EINVAL;
2023                 goto destroy_io;
2024         }
2025
2026         nvme_start_ctrl(ctrl);
2027         return 0;
2028
2029 destroy_io:
2030         if (ctrl->queue_count > 1) {
2031                 nvme_stop_queues(ctrl);
2032                 nvme_sync_io_queues(ctrl);
2033                 nvme_tcp_stop_io_queues(ctrl);
2034                 nvme_cancel_tagset(ctrl);
2035                 nvme_tcp_destroy_io_queues(ctrl, new);
2036         }
2037 destroy_admin:
2038         blk_mq_quiesce_queue(ctrl->admin_q);
2039         blk_sync_queue(ctrl->admin_q);
2040         nvme_tcp_stop_queue(ctrl, 0);
2041         nvme_cancel_admin_tagset(ctrl);
2042         nvme_tcp_destroy_admin_queue(ctrl, new);
2043         return ret;
2044 }
2045
2046 static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
2047 {
2048         struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
2049                         struct nvme_tcp_ctrl, connect_work);
2050         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2051
2052         ++ctrl->nr_reconnects;
2053
2054         if (nvme_tcp_setup_ctrl(ctrl, false))
2055                 goto requeue;
2056
2057         dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n",
2058                         ctrl->nr_reconnects);
2059
2060         ctrl->nr_reconnects = 0;
2061
2062         return;
2063
2064 requeue:
2065         dev_info(ctrl->device, "Failed reconnect attempt %d\n",
2066                         ctrl->nr_reconnects);
2067         nvme_tcp_reconnect_or_remove(ctrl);
2068 }
2069
2070 static void nvme_tcp_error_recovery_work(struct work_struct *work)
2071 {
2072         struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
2073                                 struct nvme_tcp_ctrl, err_work);
2074         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2075
2076         nvme_stop_keep_alive(ctrl);
2077         nvme_tcp_teardown_io_queues(ctrl, false);
2078         /* unquiesce to fail fast pending requests */
2079         nvme_start_queues(ctrl);
2080         nvme_tcp_teardown_admin_queue(ctrl, false);
2081         blk_mq_unquiesce_queue(ctrl->admin_q);
2082
2083         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2084                 /* state change failure is ok if we started ctrl delete */
2085                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2086                              ctrl->state != NVME_CTRL_DELETING_NOIO);
2087                 return;
2088         }
2089
2090         nvme_tcp_reconnect_or_remove(ctrl);
2091 }
2092
2093 static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
2094 {
2095         cancel_work_sync(&to_tcp_ctrl(ctrl)->err_work);
2096         cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
2097
2098         nvme_tcp_teardown_io_queues(ctrl, shutdown);
2099         blk_mq_quiesce_queue(ctrl->admin_q);
2100         if (shutdown)
2101                 nvme_shutdown_ctrl(ctrl);
2102         else
2103                 nvme_disable_ctrl(ctrl);
2104         nvme_tcp_teardown_admin_queue(ctrl, shutdown);
2105 }
2106
2107 static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
2108 {
2109         nvme_tcp_teardown_ctrl(ctrl, true);
2110 }
2111
2112 static void nvme_reset_ctrl_work(struct work_struct *work)
2113 {
2114         struct nvme_ctrl *ctrl =
2115                 container_of(work, struct nvme_ctrl, reset_work);
2116
2117         nvme_stop_ctrl(ctrl);
2118         nvme_tcp_teardown_ctrl(ctrl, false);
2119
2120         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2121                 /* state change failure is ok if we started ctrl delete */
2122                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2123                              ctrl->state != NVME_CTRL_DELETING_NOIO);
2124                 return;
2125         }
2126
2127         if (nvme_tcp_setup_ctrl(ctrl, false))
2128                 goto out_fail;
2129
2130         return;
2131
2132 out_fail:
2133         ++ctrl->nr_reconnects;
2134         nvme_tcp_reconnect_or_remove(ctrl);
2135 }
2136
2137 static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
2138 {
2139         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
2140
2141         if (list_empty(&ctrl->list))
2142                 goto free_ctrl;
2143
2144         mutex_lock(&nvme_tcp_ctrl_mutex);
2145         list_del(&ctrl->list);
2146         mutex_unlock(&nvme_tcp_ctrl_mutex);
2147
2148         nvmf_free_options(nctrl->opts);
2149 free_ctrl:
2150         kfree(ctrl->queues);
2151         kfree(ctrl);
2152 }
2153
2154 static void nvme_tcp_set_sg_null(struct nvme_command *c)
2155 {
2156         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2157
2158         sg->addr = 0;
2159         sg->length = 0;
2160         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2161                         NVME_SGL_FMT_TRANSPORT_A;
2162 }
2163
2164 static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
2165                 struct nvme_command *c, u32 data_len)
2166 {
2167         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2168
2169         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
2170         sg->length = cpu_to_le32(data_len);
2171         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
2172 }
2173
2174 static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
2175                 u32 data_len)
2176 {
2177         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2178
2179         sg->addr = 0;
2180         sg->length = cpu_to_le32(data_len);
2181         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2182                         NVME_SGL_FMT_TRANSPORT_A;
2183 }
2184
2185 static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
2186 {
2187         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2188         struct nvme_tcp_queue *queue = &ctrl->queues[0];
2189         struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2190         struct nvme_command *cmd = &pdu->cmd;
2191         u8 hdgst = nvme_tcp_hdgst_len(queue);
2192
2193         memset(pdu, 0, sizeof(*pdu));
2194         pdu->hdr.type = nvme_tcp_cmd;
2195         if (queue->hdr_digest)
2196                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2197         pdu->hdr.hlen = sizeof(*pdu);
2198         pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2199
2200         cmd->common.opcode = nvme_admin_async_event;
2201         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2202         cmd->common.flags |= NVME_CMD_SGL_METABUF;
2203         nvme_tcp_set_sg_null(cmd);
2204
2205         ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2206         ctrl->async_req.offset = 0;
2207         ctrl->async_req.curr_bio = NULL;
2208         ctrl->async_req.data_len = 0;
2209
2210         nvme_tcp_queue_request(&ctrl->async_req, true, true);
2211 }
2212
2213 static void nvme_tcp_complete_timed_out(struct request *rq)
2214 {
2215         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2216         struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2217
2218         nvme_tcp_stop_queue(ctrl, nvme_tcp_queue_id(req->queue));
2219         if (blk_mq_request_started(rq) && !blk_mq_request_completed(rq)) {
2220                 nvme_req(rq)->status = NVME_SC_HOST_ABORTED_CMD;
2221                 blk_mq_complete_request(rq);
2222         }
2223 }
2224
2225 static enum blk_eh_timer_return
2226 nvme_tcp_timeout(struct request *rq, bool reserved)
2227 {
2228         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2229         struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2230         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2231
2232         dev_warn(ctrl->device,
2233                 "queue %d: timeout request %#x type %d\n",
2234                 nvme_tcp_queue_id(req->queue), rq->tag, pdu->hdr.type);
2235
2236         if (ctrl->state != NVME_CTRL_LIVE) {
2237                 /*
2238                  * If we are resetting, connecting or deleting we should
2239                  * complete immediately because we may block controller
2240                  * teardown or setup sequence
2241                  * - ctrl disable/shutdown fabrics requests
2242                  * - connect requests
2243                  * - initialization admin requests
2244                  * - I/O requests that entered after unquiescing and
2245                  *   the controller stopped responding
2246                  *
2247                  * All other requests should be cancelled by the error
2248                  * recovery work, so it's fine that we fail it here.
2249                  */
2250                 nvme_tcp_complete_timed_out(rq);
2251                 return BLK_EH_DONE;
2252         }
2253
2254         /*
2255          * LIVE state should trigger the normal error recovery which will
2256          * handle completing this request.
2257          */
2258         nvme_tcp_error_recovery(ctrl);
2259         return BLK_EH_RESET_TIMER;
2260 }
2261
2262 static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2263                         struct request *rq)
2264 {
2265         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2266         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2267         struct nvme_command *c = &pdu->cmd;
2268
2269         c->common.flags |= NVME_CMD_SGL_METABUF;
2270
2271         if (!blk_rq_nr_phys_segments(rq))
2272                 nvme_tcp_set_sg_null(c);
2273         else if (rq_data_dir(rq) == WRITE &&
2274             req->data_len <= nvme_tcp_inline_data_size(queue))
2275                 nvme_tcp_set_sg_inline(queue, c, req->data_len);
2276         else
2277                 nvme_tcp_set_sg_host_data(c, req->data_len);
2278
2279         return 0;
2280 }
2281
2282 static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2283                 struct request *rq)
2284 {
2285         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2286         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2287         struct nvme_tcp_queue *queue = req->queue;
2288         u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2289         blk_status_t ret;
2290
2291         ret = nvme_setup_cmd(ns, rq);
2292         if (ret)
2293                 return ret;
2294
2295         req->state = NVME_TCP_SEND_CMD_PDU;
2296         req->offset = 0;
2297         req->data_sent = 0;
2298         req->pdu_len = 0;
2299         req->pdu_sent = 0;
2300         req->data_len = blk_rq_nr_phys_segments(rq) ?
2301                                 blk_rq_payload_bytes(rq) : 0;
2302         req->curr_bio = rq->bio;
2303         if (req->curr_bio && req->data_len)
2304                 nvme_tcp_init_iter(req, rq_data_dir(rq));
2305
2306         if (rq_data_dir(rq) == WRITE &&
2307             req->data_len <= nvme_tcp_inline_data_size(queue))
2308                 req->pdu_len = req->data_len;
2309
2310         pdu->hdr.type = nvme_tcp_cmd;
2311         pdu->hdr.flags = 0;
2312         if (queue->hdr_digest)
2313                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2314         if (queue->data_digest && req->pdu_len) {
2315                 pdu->hdr.flags |= NVME_TCP_F_DDGST;
2316                 ddgst = nvme_tcp_ddgst_len(queue);
2317         }
2318         pdu->hdr.hlen = sizeof(*pdu);
2319         pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2320         pdu->hdr.plen =
2321                 cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2322
2323         ret = nvme_tcp_map_data(queue, rq);
2324         if (unlikely(ret)) {
2325                 nvme_cleanup_cmd(rq);
2326                 dev_err(queue->ctrl->ctrl.device,
2327                         "Failed to map data (%d)\n", ret);
2328                 return ret;
2329         }
2330
2331         return 0;
2332 }
2333
2334 static void nvme_tcp_commit_rqs(struct blk_mq_hw_ctx *hctx)
2335 {
2336         struct nvme_tcp_queue *queue = hctx->driver_data;
2337
2338         if (!llist_empty(&queue->req_list))
2339                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
2340 }
2341
2342 static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2343                 const struct blk_mq_queue_data *bd)
2344 {
2345         struct nvme_ns *ns = hctx->queue->queuedata;
2346         struct nvme_tcp_queue *queue = hctx->driver_data;
2347         struct request *rq = bd->rq;
2348         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2349         bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2350         blk_status_t ret;
2351
2352         if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2353                 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
2354
2355         ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2356         if (unlikely(ret))
2357                 return ret;
2358
2359         blk_mq_start_request(rq);
2360
2361         nvme_tcp_queue_request(req, true, bd->last);
2362
2363         return BLK_STS_OK;
2364 }
2365
2366 static int nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2367 {
2368         struct nvme_tcp_ctrl *ctrl = set->driver_data;
2369         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2370
2371         if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2372                 /* separate read/write queues */
2373                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2374                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2375                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2376                 set->map[HCTX_TYPE_READ].nr_queues =
2377                         ctrl->io_queues[HCTX_TYPE_READ];
2378                 set->map[HCTX_TYPE_READ].queue_offset =
2379                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2380         } else {
2381                 /* shared read/write queues */
2382                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2383                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2384                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2385                 set->map[HCTX_TYPE_READ].nr_queues =
2386                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2387                 set->map[HCTX_TYPE_READ].queue_offset = 0;
2388         }
2389         blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2390         blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
2391
2392         if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2393                 /* map dedicated poll queues only if we have queues left */
2394                 set->map[HCTX_TYPE_POLL].nr_queues =
2395                                 ctrl->io_queues[HCTX_TYPE_POLL];
2396                 set->map[HCTX_TYPE_POLL].queue_offset =
2397                         ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2398                         ctrl->io_queues[HCTX_TYPE_READ];
2399                 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2400         }
2401
2402         dev_info(ctrl->ctrl.device,
2403                 "mapped %d/%d/%d default/read/poll queues.\n",
2404                 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2405                 ctrl->io_queues[HCTX_TYPE_READ],
2406                 ctrl->io_queues[HCTX_TYPE_POLL]);
2407
2408         return 0;
2409 }
2410
2411 static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx)
2412 {
2413         struct nvme_tcp_queue *queue = hctx->driver_data;
2414         struct sock *sk = queue->sock->sk;
2415
2416         if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2417                 return 0;
2418
2419         set_bit(NVME_TCP_Q_POLLING, &queue->flags);
2420         if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue))
2421                 sk_busy_loop(sk, true);
2422         nvme_tcp_try_recv(queue);
2423         clear_bit(NVME_TCP_Q_POLLING, &queue->flags);
2424         return queue->nr_cqe;
2425 }
2426
2427 static const struct blk_mq_ops nvme_tcp_mq_ops = {
2428         .queue_rq       = nvme_tcp_queue_rq,
2429         .commit_rqs     = nvme_tcp_commit_rqs,
2430         .complete       = nvme_complete_rq,
2431         .init_request   = nvme_tcp_init_request,
2432         .exit_request   = nvme_tcp_exit_request,
2433         .init_hctx      = nvme_tcp_init_hctx,
2434         .timeout        = nvme_tcp_timeout,
2435         .map_queues     = nvme_tcp_map_queues,
2436         .poll           = nvme_tcp_poll,
2437 };
2438
2439 static const struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2440         .queue_rq       = nvme_tcp_queue_rq,
2441         .complete       = nvme_complete_rq,
2442         .init_request   = nvme_tcp_init_request,
2443         .exit_request   = nvme_tcp_exit_request,
2444         .init_hctx      = nvme_tcp_init_admin_hctx,
2445         .timeout        = nvme_tcp_timeout,
2446 };
2447
2448 static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2449         .name                   = "tcp",
2450         .module                 = THIS_MODULE,
2451         .flags                  = NVME_F_FABRICS,
2452         .reg_read32             = nvmf_reg_read32,
2453         .reg_read64             = nvmf_reg_read64,
2454         .reg_write32            = nvmf_reg_write32,
2455         .free_ctrl              = nvme_tcp_free_ctrl,
2456         .submit_async_event     = nvme_tcp_submit_async_event,
2457         .delete_ctrl            = nvme_tcp_delete_ctrl,
2458         .get_address            = nvmf_get_address,
2459 };
2460
2461 static bool
2462 nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2463 {
2464         struct nvme_tcp_ctrl *ctrl;
2465         bool found = false;
2466
2467         mutex_lock(&nvme_tcp_ctrl_mutex);
2468         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2469                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2470                 if (found)
2471                         break;
2472         }
2473         mutex_unlock(&nvme_tcp_ctrl_mutex);
2474
2475         return found;
2476 }
2477
2478 static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2479                 struct nvmf_ctrl_options *opts)
2480 {
2481         struct nvme_tcp_ctrl *ctrl;
2482         int ret;
2483
2484         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2485         if (!ctrl)
2486                 return ERR_PTR(-ENOMEM);
2487
2488         INIT_LIST_HEAD(&ctrl->list);
2489         ctrl->ctrl.opts = opts;
2490         ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2491                                 opts->nr_poll_queues + 1;
2492         ctrl->ctrl.sqsize = opts->queue_size - 1;
2493         ctrl->ctrl.kato = opts->kato;
2494
2495         INIT_DELAYED_WORK(&ctrl->connect_work,
2496                         nvme_tcp_reconnect_ctrl_work);
2497         INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2498         INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2499
2500         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2501                 opts->trsvcid =
2502                         kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2503                 if (!opts->trsvcid) {
2504                         ret = -ENOMEM;
2505                         goto out_free_ctrl;
2506                 }
2507                 opts->mask |= NVMF_OPT_TRSVCID;
2508         }
2509
2510         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2511                         opts->traddr, opts->trsvcid, &ctrl->addr);
2512         if (ret) {
2513                 pr_err("malformed address passed: %s:%s\n",
2514                         opts->traddr, opts->trsvcid);
2515                 goto out_free_ctrl;
2516         }
2517
2518         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2519                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2520                         opts->host_traddr, NULL, &ctrl->src_addr);
2521                 if (ret) {
2522                         pr_err("malformed src address passed: %s\n",
2523                                opts->host_traddr);
2524                         goto out_free_ctrl;
2525                 }
2526         }
2527
2528         if (opts->mask & NVMF_OPT_HOST_IFACE) {
2529                 if (!__dev_get_by_name(&init_net, opts->host_iface)) {
2530                         pr_err("invalid interface passed: %s\n",
2531                                opts->host_iface);
2532                         ret = -ENODEV;
2533                         goto out_free_ctrl;
2534                 }
2535         }
2536
2537         if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2538                 ret = -EALREADY;
2539                 goto out_free_ctrl;
2540         }
2541
2542         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2543                                 GFP_KERNEL);
2544         if (!ctrl->queues) {
2545                 ret = -ENOMEM;
2546                 goto out_free_ctrl;
2547         }
2548
2549         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2550         if (ret)
2551                 goto out_kfree_queues;
2552
2553         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2554                 WARN_ON_ONCE(1);
2555                 ret = -EINTR;
2556                 goto out_uninit_ctrl;
2557         }
2558
2559         ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2560         if (ret)
2561                 goto out_uninit_ctrl;
2562
2563         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
2564                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2565
2566         mutex_lock(&nvme_tcp_ctrl_mutex);
2567         list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2568         mutex_unlock(&nvme_tcp_ctrl_mutex);
2569
2570         return &ctrl->ctrl;
2571
2572 out_uninit_ctrl:
2573         nvme_uninit_ctrl(&ctrl->ctrl);
2574         nvme_put_ctrl(&ctrl->ctrl);
2575         if (ret > 0)
2576                 ret = -EIO;
2577         return ERR_PTR(ret);
2578 out_kfree_queues:
2579         kfree(ctrl->queues);
2580 out_free_ctrl:
2581         kfree(ctrl);
2582         return ERR_PTR(ret);
2583 }
2584
2585 static struct nvmf_transport_ops nvme_tcp_transport = {
2586         .name           = "tcp",
2587         .module         = THIS_MODULE,
2588         .required_opts  = NVMF_OPT_TRADDR,
2589         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2590                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2591                           NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
2592                           NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2593                           NVMF_OPT_TOS | NVMF_OPT_HOST_IFACE,
2594         .create_ctrl    = nvme_tcp_create_ctrl,
2595 };
2596
2597 static int __init nvme_tcp_init_module(void)
2598 {
2599         nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
2600                         WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2601         if (!nvme_tcp_wq)
2602                 return -ENOMEM;
2603
2604         nvmf_register_transport(&nvme_tcp_transport);
2605         return 0;
2606 }
2607
2608 static void __exit nvme_tcp_cleanup_module(void)
2609 {
2610         struct nvme_tcp_ctrl *ctrl;
2611
2612         nvmf_unregister_transport(&nvme_tcp_transport);
2613
2614         mutex_lock(&nvme_tcp_ctrl_mutex);
2615         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
2616                 nvme_delete_ctrl(&ctrl->ctrl);
2617         mutex_unlock(&nvme_tcp_ctrl_mutex);
2618         flush_workqueue(nvme_delete_wq);
2619
2620         destroy_workqueue(nvme_tcp_wq);
2621 }
2622
2623 module_init(nvme_tcp_init_module);
2624 module_exit(nvme_tcp_cleanup_module);
2625
2626 MODULE_LICENSE("GPL v2");