drivers/nvme/host/tcp.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * NVMe over Fabrics TCP host.
   4  * Copyright (c) 2018 Lightbits Labs. All rights reserved.
   5  */
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7 #include <linux/module.h>
   8 #include <linux/init.h>
   9 #include <linux/slab.h>
  10 #include <linux/err.h>
  11 #include <linux/nvme-tcp.h>
  12 #include <net/sock.h>
  13 #include <net/tcp.h>
  14 #include <linux/blk-mq.h>
  15 #include <crypto/hash.h>
  16 #include <net/busy_poll.h>
  17
  18 #include "nvme.h"
  19 #include "fabrics.h"
  20
  21 struct nvme_tcp_queue;
  22
  23 /* Define the socket priority to use for connections were it is desirable
  24  * that the NIC consider performing optimized packet processing or filtering.
  25  * A non-zero value being sufficient to indicate general consideration of any
  26  * possible optimization.  Making it a module param allows for alternative
  27  * values that may be unique for some NIC implementations.
  28  */
  29 static int so_priority;
  30 module_param(so_priority, int, 0644);
  31 MODULE_PARM_DESC(so_priority, "nvme tcp socket optimize priority");
  32
  33 enum nvme_tcp_send_state {
  34         NVME_TCP_SEND_CMD_PDU = 0,
  35         NVME_TCP_SEND_H2C_PDU,
  36         NVME_TCP_SEND_DATA,
  37         NVME_TCP_SEND_DDGST,
  38 };
  39
  40 struct nvme_tcp_request {
  41         struct nvme_request     req;
  42         void                    *pdu;
  43         struct nvme_tcp_queue   *queue;
  44         u32                     data_len;
  45         u32                     pdu_len;
  46         u32                     pdu_sent;
  47         u16                     ttag;
  48         struct list_head        entry;
  49         struct llist_node       lentry;
  50         __le32                  ddgst;
  51
  52         struct bio              *curr_bio;
  53         struct iov_iter         iter;
  54
  55         /* send state */
  56         size_t                  offset;
  57         size_t                  data_sent;
  58         enum nvme_tcp_send_state state;
  59 };
  60
  61 enum nvme_tcp_queue_flags {
  62         NVME_TCP_Q_ALLOCATED    = 0,
  63         NVME_TCP_Q_LIVE         = 1,
  64         NVME_TCP_Q_POLLING      = 2,
  65 };
  66
  67 enum nvme_tcp_recv_state {
  68         NVME_TCP_RECV_PDU = 0,
  69         NVME_TCP_RECV_DATA,
  70         NVME_TCP_RECV_DDGST,
  71 };
  72
  73 struct nvme_tcp_ctrl;
  74 struct nvme_tcp_queue {
  75         struct socket           *sock;
  76         struct work_struct      io_work;
  77         int                     io_cpu;
  78
  79         struct mutex            queue_lock;
  80         struct mutex            send_mutex;
  81         struct llist_head       req_list;
  82         struct list_head        send_list;
  83         bool                    more_requests;
  84
  85         /* recv state */
  86         void                    *pdu;
  87         int                     pdu_remaining;
  88         int                     pdu_offset;
  89         size_t                  data_remaining;
  90         size_t                  ddgst_remaining;
  91         unsigned int            nr_cqe;
  92
  93         /* send state */
  94         struct nvme_tcp_request *request;
  95
  96         int                     queue_size;
  97         size_t                  cmnd_capsule_len;
  98         struct nvme_tcp_ctrl    *ctrl;
  99         unsigned long           flags;
 100         bool                    rd_enabled;
 101
 102         bool                    hdr_digest;
 103         bool                    data_digest;
 104         struct ahash_request    *rcv_hash;
 105         struct ahash_request    *snd_hash;
 106         __le32                  exp_ddgst;
 107         __le32                  recv_ddgst;
 108
 109         struct page_frag_cache  pf_cache;
 110
 111         void (*state_change)(struct sock *);
 112         void (*data_ready)(struct sock *);
 113         void (*write_space)(struct sock *);
 114 };
 115
 116 struct nvme_tcp_ctrl {
 117         /* read only in the hot path */
 118         struct nvme_tcp_queue   *queues;
 119         struct blk_mq_tag_set   tag_set;
 120
 121         /* other member variables */
 122         struct list_head        list;
 123         struct blk_mq_tag_set   admin_tag_set;
 124         struct sockaddr_storage addr;
 125         struct sockaddr_storage src_addr;
 126         struct nvme_ctrl        ctrl;
 127
 128         struct work_struct      err_work;
 129         struct delayed_work     connect_work;
 130         struct nvme_tcp_request async_req;
 131         u32                     io_queues[HCTX_MAX_TYPES];
 132 };
 133
 134 static LIST_HEAD(nvme_tcp_ctrl_list);
 135 static DEFINE_MUTEX(nvme_tcp_ctrl_mutex);
 136 static struct workqueue_struct *nvme_tcp_wq;
 137 static const struct blk_mq_ops nvme_tcp_mq_ops;
 138 static const struct blk_mq_ops nvme_tcp_admin_mq_ops;
 139 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue);
 140
 141 static inline struct nvme_tcp_ctrl *to_tcp_ctrl(struct nvme_ctrl *ctrl)
 142 {
 143         return container_of(ctrl, struct nvme_tcp_ctrl, ctrl);
 144 }
 145
 146 static inline int nvme_tcp_queue_id(struct nvme_tcp_queue *queue)
 147 {
 148         return queue - queue->ctrl->queues;
 149 }
 150
 151 static inline struct blk_mq_tags *nvme_tcp_tagset(struct nvme_tcp_queue *queue)
 152 {
 153         u32 queue_idx = nvme_tcp_queue_id(queue);
 154
 155         if (queue_idx == 0)
 156                 return queue->ctrl->admin_tag_set.tags[queue_idx];
 157         return queue->ctrl->tag_set.tags[queue_idx - 1];
 158 }
 159
 160 static inline u8 nvme_tcp_hdgst_len(struct nvme_tcp_queue *queue)
 161 {
 162         return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 163 }
 164
 165 static inline u8 nvme_tcp_ddgst_len(struct nvme_tcp_queue *queue)
 166 {
 167         return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 168 }
 169
 170 static inline size_t nvme_tcp_inline_data_size(struct nvme_tcp_queue *queue)
 171 {
 172         return queue->cmnd_capsule_len - sizeof(struct nvme_command);
 173 }
 174
 175 static inline bool nvme_tcp_async_req(struct nvme_tcp_request *req)
 176 {
 177         return req == &req->queue->ctrl->async_req;
 178 }
 179
 180 static inline bool nvme_tcp_has_inline_data(struct nvme_tcp_request *req)
 181 {
 182         struct request *rq;
 183
 184         if (unlikely(nvme_tcp_async_req(req)))
 185                 return false; /* async events don't have a request */
 186
 187         rq = blk_mq_rq_from_pdu(req);
 188
 189         return rq_data_dir(rq) == WRITE && req->data_len &&
 190                 req->data_len <= nvme_tcp_inline_data_size(req->queue);
 191 }
 192
 193 static inline struct page *nvme_tcp_req_cur_page(struct nvme_tcp_request *req)
 194 {
 195         return req->iter.bvec->bv_page;
 196 }
 197
 198 static inline size_t nvme_tcp_req_cur_offset(struct nvme_tcp_request *req)
 199 {
 200         return req->iter.bvec->bv_offset + req->iter.iov_offset;
 201 }
 202
 203 static inline size_t nvme_tcp_req_cur_length(struct nvme_tcp_request *req)
 204 {
 205         return min_t(size_t, iov_iter_single_seg_count(&req->iter),
 206                         req->pdu_len - req->pdu_sent);
 207 }
 208
 209 static inline size_t nvme_tcp_pdu_data_left(struct nvme_tcp_request *req)
 210 {
 211         return rq_data_dir(blk_mq_rq_from_pdu(req)) == WRITE ?
 212                         req->pdu_len - req->pdu_sent : 0;
 213 }
 214
 215 static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req,
 216                 int len)
 217 {
 218         return nvme_tcp_pdu_data_left(req) <= len;
 219 }
 220
 221 static void nvme_tcp_init_iter(struct nvme_tcp_request *req,
 222                 unsigned int dir)
 223 {
 224         struct request *rq = blk_mq_rq_from_pdu(req);
 225         struct bio_vec *vec;
 226         unsigned int size;
 227         int nr_bvec;
 228         size_t offset;
 229
 230         if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) {
 231                 vec = &rq->special_vec;
 232                 nr_bvec = 1;
 233                 size = blk_rq_payload_bytes(rq);
 234                 offset = 0;
 235         } else {
 236                 struct bio *bio = req->curr_bio;
 237                 struct bvec_iter bi;
 238                 struct bio_vec bv;
 239
 240                 vec = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
 241                 nr_bvec = 0;
 242                 bio_for_each_bvec(bv, bio, bi) {
 243                         nr_bvec++;
 244                 }
 245                 size = bio->bi_iter.bi_size;
 246                 offset = bio->bi_iter.bi_bvec_done;
 247         }
 248
 249         iov_iter_bvec(&req->iter, dir, vec, nr_bvec, size);
 250         req->iter.iov_offset = offset;
 251 }
 252
 253 static inline void nvme_tcp_advance_req(struct nvme_tcp_request *req,
 254                 int len)
 255 {
 256         req->data_sent += len;
 257         req->pdu_sent += len;
 258         iov_iter_advance(&req->iter, len);
 259         if (!iov_iter_count(&req->iter) &&
 260             req->data_sent < req->data_len) {
 261                 req->curr_bio = req->curr_bio->bi_next;
 262                 nvme_tcp_init_iter(req, WRITE);
 263         }
 264 }
 265
 266 static inline void nvme_tcp_send_all(struct nvme_tcp_queue *queue)
 267 {
 268         int ret;
 269
 270         /* drain the send queue as much as we can... */
 271         do {
 272                 ret = nvme_tcp_try_send(queue);
 273         } while (ret > 0);
 274 }
 275
 276 static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req,
 277                 bool sync, bool last)
 278 {
 279         struct nvme_tcp_queue *queue = req->queue;
 280         bool empty;
 281
 282         empty = llist_add(&req->lentry, &queue->req_list) &&
 283                 list_empty(&queue->send_list) && !queue->request;
 284
 285         /*
 286          * if we're the first on the send_list and we can try to send
 287          * directly, otherwise queue io_work. Also, only do that if we
 288          * are on the same cpu, so we don't introduce contention.
 289          */
 290         if (queue->io_cpu == raw_smp_processor_id() &&
 291             sync && empty && mutex_trylock(&queue->send_mutex)) {
 292                 queue->more_requests = !last;
 293                 nvme_tcp_send_all(queue);
 294                 queue->more_requests = false;
 295                 mutex_unlock(&queue->send_mutex);
 296         } else if (last) {
 297                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 298         }
 299 }
 300
 301 static void nvme_tcp_process_req_list(struct nvme_tcp_queue *queue)
 302 {
 303         struct nvme_tcp_request *req;
 304         struct llist_node *node;
 305
 306         for (node = llist_del_all(&queue->req_list); node; node = node->next) {
 307                 req = llist_entry(node, struct nvme_tcp_request, lentry);
 308                 list_add(&req->entry, &queue->send_list);
 309         }
 310 }
 311
 312 static inline struct nvme_tcp_request *
 313 nvme_tcp_fetch_request(struct nvme_tcp_queue *queue)
 314 {
 315         struct nvme_tcp_request *req;
 316
 317         req = list_first_entry_or_null(&queue->send_list,
 318                         struct nvme_tcp_request, entry);
 319         if (!req) {
 320                 nvme_tcp_process_req_list(queue);
 321                 req = list_first_entry_or_null(&queue->send_list,
 322                                 struct nvme_tcp_request, entry);
 323                 if (unlikely(!req))
 324                         return NULL;
 325         }
 326
 327         list_del(&req->entry);
 328         return req;
 329 }
 330
 331 static inline void nvme_tcp_ddgst_final(struct ahash_request *hash,
 332                 __le32 *dgst)
 333 {
 334         ahash_request_set_crypt(hash, NULL, (u8 *)dgst, 0);
 335         crypto_ahash_final(hash);
 336 }
 337
 338 static inline void nvme_tcp_ddgst_update(struct ahash_request *hash,
 339                 struct page *page, off_t off, size_t len)
 340 {
 341         struct scatterlist sg;
 342
 343         sg_init_marker(&sg, 1);
 344         sg_set_page(&sg, page, len, off);
 345         ahash_request_set_crypt(hash, &sg, NULL, len);
 346         crypto_ahash_update(hash);
 347 }
 348
 349 static inline void nvme_tcp_hdgst(struct ahash_request *hash,
 350                 void *pdu, size_t len)
 351 {
 352         struct scatterlist sg;
 353
 354         sg_init_one(&sg, pdu, len);
 355         ahash_request_set_crypt(hash, &sg, pdu + len, len);
 356         crypto_ahash_digest(hash);
 357 }
 358
 359 static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue,
 360                 void *pdu, size_t pdu_len)
 361 {
 362         struct nvme_tcp_hdr *hdr = pdu;
 363         __le32 recv_digest;
 364         __le32 exp_digest;
 365
 366         if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
 367                 dev_err(queue->ctrl->ctrl.device,
 368                         "queue %d: header digest flag is cleared\n",
 369                         nvme_tcp_queue_id(queue));
 370                 return -EPROTO;
 371         }
 372
 373         recv_digest = *(__le32 *)(pdu + hdr->hlen);
 374         nvme_tcp_hdgst(queue->rcv_hash, pdu, pdu_len);
 375         exp_digest = *(__le32 *)(pdu + hdr->hlen);
 376         if (recv_digest != exp_digest) {
 377                 dev_err(queue->ctrl->ctrl.device,
 378                         "header digest error: recv %#x expected %#x\n",
 379                         le32_to_cpu(recv_digest), le32_to_cpu(exp_digest));
 380                 return -EIO;
 381         }
 382
 383         return 0;
 384 }
 385
 386 static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu)
 387 {
 388         struct nvme_tcp_hdr *hdr = pdu;
 389         u8 digest_len = nvme_tcp_hdgst_len(queue);
 390         u32 len;
 391
 392         len = le32_to_cpu(hdr->plen) - hdr->hlen -
 393                 ((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0);
 394
 395         if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
 396                 dev_err(queue->ctrl->ctrl.device,
 397                         "queue %d: data digest flag is cleared\n",
 398                 nvme_tcp_queue_id(queue));
 399                 return -EPROTO;
 400         }
 401         crypto_ahash_init(queue->rcv_hash);
 402
 403         return 0;
 404 }
 405
 406 static void nvme_tcp_exit_request(struct blk_mq_tag_set *set,
 407                 struct request *rq, unsigned int hctx_idx)
 408 {
 409         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 410
 411         page_frag_free(req->pdu);
 412 }
 413
 414 static int nvme_tcp_init_request(struct blk_mq_tag_set *set,
 415                 struct request *rq, unsigned int hctx_idx,
 416                 unsigned int numa_node)
 417 {
 418         struct nvme_tcp_ctrl *ctrl = set->driver_data;
 419         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 420         struct nvme_tcp_cmd_pdu *pdu;
 421         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 422         struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx];
 423         u8 hdgst = nvme_tcp_hdgst_len(queue);
 424
 425         req->pdu = page_frag_alloc(&queue->pf_cache,
 426                 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
 427                 GFP_KERNEL | __GFP_ZERO);
 428         if (!req->pdu)
 429                 return -ENOMEM;
 430
 431         pdu = req->pdu;
 432         req->queue = queue;
 433         nvme_req(rq)->ctrl = &ctrl->ctrl;
 434         nvme_req(rq)->cmd = &pdu->cmd;
 435
 436         return 0;
 437 }
 438
 439 static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 440                 unsigned int hctx_idx)
 441 {
 442         struct nvme_tcp_ctrl *ctrl = data;
 443         struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1];
 444
 445         hctx->driver_data = queue;
 446         return 0;
 447 }
 448
 449 static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 450                 unsigned int hctx_idx)
 451 {
 452         struct nvme_tcp_ctrl *ctrl = data;
 453         struct nvme_tcp_queue *queue = &ctrl->queues[0];
 454
 455         hctx->driver_data = queue;
 456         return 0;
 457 }
 458
 459 static enum nvme_tcp_recv_state
 460 nvme_tcp_recv_state(struct nvme_tcp_queue *queue)
 461 {
 462         return  (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
 463                 (queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST :
 464                 NVME_TCP_RECV_DATA;
 465 }
 466
 467 static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue)
 468 {
 469         queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) +
 470                                 nvme_tcp_hdgst_len(queue);
 471         queue->pdu_offset = 0;
 472         queue->data_remaining = -1;
 473         queue->ddgst_remaining = 0;
 474 }
 475
 476 static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl)
 477 {
 478         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
 479                 return;
 480
 481         dev_warn(ctrl->device, "starting error recovery\n");
 482         queue_work(nvme_reset_wq, &to_tcp_ctrl(ctrl)->err_work);
 483 }
 484
 485 static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
 486                 struct nvme_completion *cqe)
 487 {
 488         struct request *rq;
 489
 490         rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), cqe->command_id);
 491         if (!rq) {
 492                 dev_err(queue->ctrl->ctrl.device,
 493                         "queue %d tag 0x%x not found\n",
 494                         nvme_tcp_queue_id(queue), cqe->command_id);
 495                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 496                 return -EINVAL;
 497         }
 498
 499         if (!nvme_try_complete_req(rq, cqe->status, cqe->result))
 500                 nvme_complete_rq(rq);
 501         queue->nr_cqe++;
 502
 503         return 0;
 504 }
 505
 506 static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue,
 507                 struct nvme_tcp_data_pdu *pdu)
 508 {
 509         struct request *rq;
 510
 511         rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
 512         if (!rq) {
 513                 dev_err(queue->ctrl->ctrl.device,
 514                         "queue %d tag %#x not found\n",
 515                         nvme_tcp_queue_id(queue), pdu->command_id);
 516                 return -ENOENT;
 517         }
 518
 519         if (!blk_rq_payload_bytes(rq)) {
 520                 dev_err(queue->ctrl->ctrl.device,
 521                         "queue %d tag %#x unexpected data\n",
 522                         nvme_tcp_queue_id(queue), rq->tag);
 523                 return -EIO;
 524         }
 525
 526         queue->data_remaining = le32_to_cpu(pdu->data_length);
 527
 528         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS &&
 529             unlikely(!(pdu->hdr.flags & NVME_TCP_F_DATA_LAST))) {
 530                 dev_err(queue->ctrl->ctrl.device,
 531                         "queue %d tag %#x SUCCESS set but not last PDU\n",
 532                         nvme_tcp_queue_id(queue), rq->tag);
 533                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 534                 return -EPROTO;
 535         }
 536
 537         return 0;
 538 }
 539
 540 static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue,
 541                 struct nvme_tcp_rsp_pdu *pdu)
 542 {
 543         struct nvme_completion *cqe = &pdu->cqe;
 544         int ret = 0;
 545
 546         /*
 547          * AEN requests are special as they don't time out and can
 548          * survive any kind of queue freeze and often don't respond to
 549          * aborts.  We don't even bother to allocate a struct request
 550          * for them but rather special case them here.
 551          */
 552         if (unlikely(nvme_is_aen_req(nvme_tcp_queue_id(queue),
 553                                      cqe->command_id)))
 554                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
 555                                 &cqe->result);
 556         else
 557                 ret = nvme_tcp_process_nvme_cqe(queue, cqe);
 558
 559         return ret;
 560 }
 561
 562 static int nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req,
 563                 struct nvme_tcp_r2t_pdu *pdu)
 564 {
 565         struct nvme_tcp_data_pdu *data = req->pdu;
 566         struct nvme_tcp_queue *queue = req->queue;
 567         struct request *rq = blk_mq_rq_from_pdu(req);
 568         u8 hdgst = nvme_tcp_hdgst_len(queue);
 569         u8 ddgst = nvme_tcp_ddgst_len(queue);
 570
 571         req->pdu_len = le32_to_cpu(pdu->r2t_length);
 572         req->pdu_sent = 0;
 573
 574         if (unlikely(!req->pdu_len)) {
 575                 dev_err(queue->ctrl->ctrl.device,
 576                         "req %d r2t len is %u, probably a bug...\n",
 577                         rq->tag, req->pdu_len);
 578                 return -EPROTO;
 579         }
 580
 581         if (unlikely(req->data_sent + req->pdu_len > req->data_len)) {
 582                 dev_err(queue->ctrl->ctrl.device,
 583                         "req %d r2t len %u exceeded data len %u (%zu sent)\n",
 584                         rq->tag, req->pdu_len, req->data_len,
 585                         req->data_sent);
 586                 return -EPROTO;
 587         }
 588
 589         if (unlikely(le32_to_cpu(pdu->r2t_offset) < req->data_sent)) {
 590                 dev_err(queue->ctrl->ctrl.device,
 591                         "req %d unexpected r2t offset %u (expected %zu)\n",
 592                         rq->tag, le32_to_cpu(pdu->r2t_offset),
 593                         req->data_sent);
 594                 return -EPROTO;
 595         }
 596
 597         memset(data, 0, sizeof(*data));
 598         data->hdr.type = nvme_tcp_h2c_data;
 599         data->hdr.flags = NVME_TCP_F_DATA_LAST;
 600         if (queue->hdr_digest)
 601                 data->hdr.flags |= NVME_TCP_F_HDGST;
 602         if (queue->data_digest)
 603                 data->hdr.flags |= NVME_TCP_F_DDGST;
 604         data->hdr.hlen = sizeof(*data);
 605         data->hdr.pdo = data->hdr.hlen + hdgst;
 606         data->hdr.plen =
 607                 cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst);
 608         data->ttag = pdu->ttag;
 609         data->command_id = rq->tag;
 610         data->data_offset = cpu_to_le32(req->data_sent);
 611         data->data_length = cpu_to_le32(req->pdu_len);
 612         return 0;
 613 }
 614
 615 static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue,
 616                 struct nvme_tcp_r2t_pdu *pdu)
 617 {
 618         struct nvme_tcp_request *req;
 619         struct request *rq;
 620         int ret;
 621
 622         rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
 623         if (!rq) {
 624                 dev_err(queue->ctrl->ctrl.device,
 625                         "queue %d tag %#x not found\n",
 626                         nvme_tcp_queue_id(queue), pdu->command_id);
 627                 return -ENOENT;
 628         }
 629         req = blk_mq_rq_to_pdu(rq);
 630
 631         ret = nvme_tcp_setup_h2c_data_pdu(req, pdu);
 632         if (unlikely(ret))
 633                 return ret;
 634
 635         req->state = NVME_TCP_SEND_H2C_PDU;
 636         req->offset = 0;
 637
 638         nvme_tcp_queue_request(req, false, true);
 639
 640         return 0;
 641 }
 642
 643 static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 644                 unsigned int *offset, size_t *len)
 645 {
 646         struct nvme_tcp_hdr *hdr;
 647         char *pdu = queue->pdu;
 648         size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
 649         int ret;
 650
 651         ret = skb_copy_bits(skb, *offset,
 652                 &pdu[queue->pdu_offset], rcv_len);
 653         if (unlikely(ret))
 654                 return ret;
 655
 656         queue->pdu_remaining -= rcv_len;
 657         queue->pdu_offset += rcv_len;
 658         *offset += rcv_len;
 659         *len -= rcv_len;
 660         if (queue->pdu_remaining)
 661                 return 0;
 662
 663         hdr = queue->pdu;
 664         if (queue->hdr_digest) {
 665                 ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen);
 666                 if (unlikely(ret))
 667                         return ret;
 668         }
 669
 670
 671         if (queue->data_digest) {
 672                 ret = nvme_tcp_check_ddgst(queue, queue->pdu);
 673                 if (unlikely(ret))
 674                         return ret;
 675         }
 676
 677         switch (hdr->type) {
 678         case nvme_tcp_c2h_data:
 679                 return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
 680         case nvme_tcp_rsp:
 681                 nvme_tcp_init_recv_ctx(queue);
 682                 return nvme_tcp_handle_comp(queue, (void *)queue->pdu);
 683         case nvme_tcp_r2t:
 684                 nvme_tcp_init_recv_ctx(queue);
 685                 return nvme_tcp_handle_r2t(queue, (void *)queue->pdu);
 686         default:
 687                 dev_err(queue->ctrl->ctrl.device,
 688                         "unsupported pdu type (%d)\n", hdr->type);
 689                 return -EINVAL;
 690         }
 691 }
 692
 693 static inline void nvme_tcp_end_request(struct request *rq, u16 status)
 694 {
 695         union nvme_result res = {};
 696
 697         if (!nvme_try_complete_req(rq, cpu_to_le16(status << 1), res))
 698                 nvme_complete_rq(rq);
 699 }
 700
 701 static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 702                               unsigned int *offset, size_t *len)
 703 {
 704         struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
 705         struct nvme_tcp_request *req;
 706         struct request *rq;
 707
 708         rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
 709         if (!rq) {
 710                 dev_err(queue->ctrl->ctrl.device,
 711                         "queue %d tag %#x not found\n",
 712                         nvme_tcp_queue_id(queue), pdu->command_id);
 713                 return -ENOENT;
 714         }
 715         req = blk_mq_rq_to_pdu(rq);
 716
 717         while (true) {
 718                 int recv_len, ret;
 719
 720                 recv_len = min_t(size_t, *len, queue->data_remaining);
 721                 if (!recv_len)
 722                         break;
 723
 724                 if (!iov_iter_count(&req->iter)) {
 725                         req->curr_bio = req->curr_bio->bi_next;
 726
 727                         /*
 728                          * If we don`t have any bios it means that controller
 729                          * sent more data than we requested, hence error
 730                          */
 731                         if (!req->curr_bio) {
 732                                 dev_err(queue->ctrl->ctrl.device,
 733                                         "queue %d no space in request %#x",
 734                                         nvme_tcp_queue_id(queue), rq->tag);
 735                                 nvme_tcp_init_recv_ctx(queue);
 736                                 return -EIO;
 737                         }
 738                         nvme_tcp_init_iter(req, READ);
 739                 }
 740
 741                 /* we can read only from what is left in this bio */
 742                 recv_len = min_t(size_t, recv_len,
 743                                 iov_iter_count(&req->iter));
 744
 745                 if (queue->data_digest)
 746                         ret = skb_copy_and_hash_datagram_iter(skb, *offset,
 747                                 &req->iter, recv_len, queue->rcv_hash);
 748                 else
 749                         ret = skb_copy_datagram_iter(skb, *offset,
 750                                         &req->iter, recv_len);
 751                 if (ret) {
 752                         dev_err(queue->ctrl->ctrl.device,
 753                                 "queue %d failed to copy request %#x data",
 754                                 nvme_tcp_queue_id(queue), rq->tag);
 755                         return ret;
 756                 }
 757
 758                 *len -= recv_len;
 759                 *offset += recv_len;
 760                 queue->data_remaining -= recv_len;
 761         }
 762
 763         if (!queue->data_remaining) {
 764                 if (queue->data_digest) {
 765                         nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst);
 766                         queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
 767                 } else {
 768                         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
 769                                 nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
 770                                 queue->nr_cqe++;
 771                         }
 772                         nvme_tcp_init_recv_ctx(queue);
 773                 }
 774         }
 775
 776         return 0;
 777 }
 778
 779 static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
 780                 struct sk_buff *skb, unsigned int *offset, size_t *len)
 781 {
 782         struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
 783         char *ddgst = (char *)&queue->recv_ddgst;
 784         size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
 785         off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
 786         int ret;
 787
 788         ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
 789         if (unlikely(ret))
 790                 return ret;
 791
 792         queue->ddgst_remaining -= recv_len;
 793         *offset += recv_len;
 794         *len -= recv_len;
 795         if (queue->ddgst_remaining)
 796                 return 0;
 797
 798         if (queue->recv_ddgst != queue->exp_ddgst) {
 799                 dev_err(queue->ctrl->ctrl.device,
 800                         "data digest error: recv %#x expected %#x\n",
 801                         le32_to_cpu(queue->recv_ddgst),
 802                         le32_to_cpu(queue->exp_ddgst));
 803                 return -EIO;
 804         }
 805
 806         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
 807                 struct request *rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue),
 808                                                 pdu->command_id);
 809
 810                 nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
 811                 queue->nr_cqe++;
 812         }
 813
 814         nvme_tcp_init_recv_ctx(queue);
 815         return 0;
 816 }
 817
 818 static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
 819                              unsigned int offset, size_t len)
 820 {
 821         struct nvme_tcp_queue *queue = desc->arg.data;
 822         size_t consumed = len;
 823         int result;
 824
 825         while (len) {
 826                 switch (nvme_tcp_recv_state(queue)) {
 827                 case NVME_TCP_RECV_PDU:
 828                         result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
 829                         break;
 830                 case NVME_TCP_RECV_DATA:
 831                         result = nvme_tcp_recv_data(queue, skb, &offset, &len);
 832                         break;
 833                 case NVME_TCP_RECV_DDGST:
 834                         result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
 835                         break;
 836                 default:
 837                         result = -EFAULT;
 838                 }
 839                 if (result) {
 840                         dev_err(queue->ctrl->ctrl.device,
 841                                 "receive failed:  %d\n", result);
 842                         queue->rd_enabled = false;
 843                         nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 844                         return result;
 845                 }
 846         }
 847
 848         return consumed;
 849 }
 850
 851 static void nvme_tcp_data_ready(struct sock *sk)
 852 {
 853         struct nvme_tcp_queue *queue;
 854
 855         read_lock_bh(&sk->sk_callback_lock);
 856         queue = sk->sk_user_data;
 857         if (likely(queue && queue->rd_enabled) &&
 858             !test_bit(NVME_TCP_Q_POLLING, &queue->flags))
 859                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 860         read_unlock_bh(&sk->sk_callback_lock);
 861 }
 862
 863 static void nvme_tcp_write_space(struct sock *sk)
 864 {
 865         struct nvme_tcp_queue *queue;
 866
 867         read_lock_bh(&sk->sk_callback_lock);
 868         queue = sk->sk_user_data;
 869         if (likely(queue && sk_stream_is_writeable(sk))) {
 870                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 871                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 872         }
 873         read_unlock_bh(&sk->sk_callback_lock);
 874 }
 875
 876 static void nvme_tcp_state_change(struct sock *sk)
 877 {
 878         struct nvme_tcp_queue *queue;
 879
 880         read_lock_bh(&sk->sk_callback_lock);
 881         queue = sk->sk_user_data;
 882         if (!queue)
 883                 goto done;
 884
 885         switch (sk->sk_state) {
 886         case TCP_CLOSE:
 887         case TCP_CLOSE_WAIT:
 888         case TCP_LAST_ACK:
 889         case TCP_FIN_WAIT1:
 890         case TCP_FIN_WAIT2:
 891                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 892                 break;
 893         default:
 894                 dev_info(queue->ctrl->ctrl.device,
 895                         "queue %d socket state %d\n",
 896                         nvme_tcp_queue_id(queue), sk->sk_state);
 897         }
 898
 899         queue->state_change(sk);
 900 done:
 901         read_unlock_bh(&sk->sk_callback_lock);
 902 }
 903
 904 static inline bool nvme_tcp_queue_more(struct nvme_tcp_queue *queue)
 905 {
 906         return !list_empty(&queue->send_list) ||
 907                 !llist_empty(&queue->req_list) || queue->more_requests;
 908 }
 909
 910 static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
 911 {
 912         queue->request = NULL;
 913 }
 914
 915 static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
 916 {
 917         nvme_tcp_end_request(blk_mq_rq_from_pdu(req), NVME_SC_HOST_PATH_ERROR);
 918 }
 919
 920 static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
 921 {
 922         struct nvme_tcp_queue *queue = req->queue;
 923
 924         while (true) {
 925                 struct page *page = nvme_tcp_req_cur_page(req);
 926                 size_t offset = nvme_tcp_req_cur_offset(req);
 927                 size_t len = nvme_tcp_req_cur_length(req);
 928                 bool last = nvme_tcp_pdu_last_send(req, len);
 929                 int ret, flags = MSG_DONTWAIT;
 930
 931                 if (last && !queue->data_digest && !nvme_tcp_queue_more(queue))
 932                         flags |= MSG_EOR;
 933                 else
 934                         flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
 935
 936                 if (sendpage_ok(page)) {
 937                         ret = kernel_sendpage(queue->sock, page, offset, len,
 938                                         flags);
 939                 } else {
 940                         ret = sock_no_sendpage(queue->sock, page, offset, len,
 941                                         flags);
 942                 }
 943                 if (ret <= 0)
 944                         return ret;
 945
 946                 nvme_tcp_advance_req(req, ret);
 947                 if (queue->data_digest)
 948                         nvme_tcp_ddgst_update(queue->snd_hash, page,
 949                                         offset, ret);
 950
 951                 /* fully successful last write*/
 952                 if (last && ret == len) {
 953                         if (queue->data_digest) {
 954                                 nvme_tcp_ddgst_final(queue->snd_hash,
 955                                         &req->ddgst);
 956                                 req->state = NVME_TCP_SEND_DDGST;
 957                                 req->offset = 0;
 958                         } else {
 959                                 nvme_tcp_done_send_req(queue);
 960                         }
 961                         return 1;
 962                 }
 963         }
 964         return -EAGAIN;
 965 }
 966
 967 static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
 968 {
 969         struct nvme_tcp_queue *queue = req->queue;
 970         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
 971         bool inline_data = nvme_tcp_has_inline_data(req);
 972         u8 hdgst = nvme_tcp_hdgst_len(queue);
 973         int len = sizeof(*pdu) + hdgst - req->offset;
 974         int flags = MSG_DONTWAIT;
 975         int ret;
 976
 977         if (inline_data || nvme_tcp_queue_more(queue))
 978                 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
 979         else
 980                 flags |= MSG_EOR;
 981
 982         if (queue->hdr_digest && !req->offset)
 983                 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
 984
 985         ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
 986                         offset_in_page(pdu) + req->offset, len,  flags);
 987         if (unlikely(ret <= 0))
 988                 return ret;
 989
 990         len -= ret;
 991         if (!len) {
 992                 if (inline_data) {
 993                         req->state = NVME_TCP_SEND_DATA;
 994                         if (queue->data_digest)
 995                                 crypto_ahash_init(queue->snd_hash);
 996                 } else {
 997                         nvme_tcp_done_send_req(queue);
 998                 }
 999                 return 1;
1000         }
1001         req->offset += ret;
1002
1003         return -EAGAIN;
1004 }
1005
1006 static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
1007 {
1008         struct nvme_tcp_queue *queue = req->queue;
1009         struct nvme_tcp_data_pdu *pdu = req->pdu;
1010         u8 hdgst = nvme_tcp_hdgst_len(queue);
1011         int len = sizeof(*pdu) - req->offset + hdgst;
1012         int ret;
1013
1014         if (queue->hdr_digest && !req->offset)
1015                 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
1016
1017         ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
1018                         offset_in_page(pdu) + req->offset, len,
1019                         MSG_DONTWAIT | MSG_MORE | MSG_SENDPAGE_NOTLAST);
1020         if (unlikely(ret <= 0))
1021                 return ret;
1022
1023         len -= ret;
1024         if (!len) {
1025                 req->state = NVME_TCP_SEND_DATA;
1026                 if (queue->data_digest)
1027                         crypto_ahash_init(queue->snd_hash);
1028                 return 1;
1029         }
1030         req->offset += ret;
1031
1032         return -EAGAIN;
1033 }
1034
1035 static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
1036 {
1037         struct nvme_tcp_queue *queue = req->queue;
1038         int ret;
1039         struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1040         struct kvec iov = {
1041                 .iov_base = &req->ddgst + req->offset,
1042                 .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
1043         };
1044
1045         if (nvme_tcp_queue_more(queue))
1046                 msg.msg_flags |= MSG_MORE;
1047         else
1048                 msg.msg_flags |= MSG_EOR;
1049
1050         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1051         if (unlikely(ret <= 0))
1052                 return ret;
1053
1054         if (req->offset + ret == NVME_TCP_DIGEST_LENGTH) {
1055                 nvme_tcp_done_send_req(queue);
1056                 return 1;
1057         }
1058
1059         req->offset += ret;
1060         return -EAGAIN;
1061 }
1062
1063 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
1064 {
1065         struct nvme_tcp_request *req;
1066         int ret = 1;
1067
1068         if (!queue->request) {
1069                 queue->request = nvme_tcp_fetch_request(queue);
1070                 if (!queue->request)
1071                         return 0;
1072         }
1073         req = queue->request;
1074
1075         if (req->state == NVME_TCP_SEND_CMD_PDU) {
1076                 ret = nvme_tcp_try_send_cmd_pdu(req);
1077                 if (ret <= 0)
1078                         goto done;
1079                 if (!nvme_tcp_has_inline_data(req))
1080                         return ret;
1081         }
1082
1083         if (req->state == NVME_TCP_SEND_H2C_PDU) {
1084                 ret = nvme_tcp_try_send_data_pdu(req);
1085                 if (ret <= 0)
1086                         goto done;
1087         }
1088
1089         if (req->state == NVME_TCP_SEND_DATA) {
1090                 ret = nvme_tcp_try_send_data(req);
1091                 if (ret <= 0)
1092                         goto done;
1093         }
1094
1095         if (req->state == NVME_TCP_SEND_DDGST)
1096                 ret = nvme_tcp_try_send_ddgst(req);
1097 done:
1098         if (ret == -EAGAIN) {
1099                 ret = 0;
1100         } else if (ret < 0) {
1101                 dev_err(queue->ctrl->ctrl.device,
1102                         "failed to send request %d\n", ret);
1103                 if (ret != -EPIPE && ret != -ECONNRESET)
1104                         nvme_tcp_fail_request(queue->request);
1105                 nvme_tcp_done_send_req(queue);
1106         }
1107         return ret;
1108 }
1109
1110 static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1111 {
1112         struct socket *sock = queue->sock;
1113         struct sock *sk = sock->sk;
1114         read_descriptor_t rd_desc;
1115         int consumed;
1116
1117         rd_desc.arg.data = queue;
1118         rd_desc.count = 1;
1119         lock_sock(sk);
1120         queue->nr_cqe = 0;
1121         consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1122         release_sock(sk);
1123         return consumed;
1124 }
1125
1126 static void nvme_tcp_io_work(struct work_struct *w)
1127 {
1128         struct nvme_tcp_queue *queue =
1129                 container_of(w, struct nvme_tcp_queue, io_work);
1130         unsigned long deadline = jiffies + msecs_to_jiffies(1);
1131
1132         do {
1133                 bool pending = false;
1134                 int result;
1135
1136                 if (mutex_trylock(&queue->send_mutex)) {
1137                         result = nvme_tcp_try_send(queue);
1138                         mutex_unlock(&queue->send_mutex);
1139                         if (result > 0)
1140                                 pending = true;
1141                         else if (unlikely(result < 0))
1142                                 break;
1143                 }
1144
1145                 result = nvme_tcp_try_recv(queue);
1146                 if (result > 0)
1147                         pending = true;
1148                 else if (unlikely(result < 0))
1149                         return;
1150
1151                 if (!pending)
1152                         return;
1153
1154         } while (!time_after(jiffies, deadline)); /* quota is exhausted */
1155
1156         queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1157 }
1158
1159 static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue)
1160 {
1161         struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
1162
1163         ahash_request_free(queue->rcv_hash);
1164         ahash_request_free(queue->snd_hash);
1165         crypto_free_ahash(tfm);
1166 }
1167
1168 static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue)
1169 {
1170         struct crypto_ahash *tfm;
1171
1172         tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
1173         if (IS_ERR(tfm))
1174                 return PTR_ERR(tfm);
1175
1176         queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1177         if (!queue->snd_hash)
1178                 goto free_tfm;
1179         ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
1180
1181         queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1182         if (!queue->rcv_hash)
1183                 goto free_snd_hash;
1184         ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
1185
1186         return 0;
1187 free_snd_hash:
1188         ahash_request_free(queue->snd_hash);
1189 free_tfm:
1190         crypto_free_ahash(tfm);
1191         return -ENOMEM;
1192 }
1193
1194 static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1195 {
1196         struct nvme_tcp_request *async = &ctrl->async_req;
1197
1198         page_frag_free(async->pdu);
1199 }
1200
1201 static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1202 {
1203         struct nvme_tcp_queue *queue = &ctrl->queues[0];
1204         struct nvme_tcp_request *async = &ctrl->async_req;
1205         u8 hdgst = nvme_tcp_hdgst_len(queue);
1206
1207         async->pdu = page_frag_alloc(&queue->pf_cache,
1208                 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1209                 GFP_KERNEL | __GFP_ZERO);
1210         if (!async->pdu)
1211                 return -ENOMEM;
1212
1213         async->queue = &ctrl->queues[0];
1214         return 0;
1215 }
1216
1217 static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1218 {
1219         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1220         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1221
1222         if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1223                 return;
1224
1225         if (queue->hdr_digest || queue->data_digest)
1226                 nvme_tcp_free_crypto(queue);
1227
1228         sock_release(queue->sock);
1229         kfree(queue->pdu);
1230         mutex_destroy(&queue->queue_lock);
1231 }
1232
1233 static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1234 {
1235         struct nvme_tcp_icreq_pdu *icreq;
1236         struct nvme_tcp_icresp_pdu *icresp;
1237         struct msghdr msg = {};
1238         struct kvec iov;
1239         bool ctrl_hdgst, ctrl_ddgst;
1240         int ret;
1241
1242         icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1243         if (!icreq)
1244                 return -ENOMEM;
1245
1246         icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1247         if (!icresp) {
1248                 ret = -ENOMEM;
1249                 goto free_icreq;
1250         }
1251
1252         icreq->hdr.type = nvme_tcp_icreq;
1253         icreq->hdr.hlen = sizeof(*icreq);
1254         icreq->hdr.pdo = 0;
1255         icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1256         icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1257         icreq->maxr2t = 0; /* single inflight r2t supported */
1258         icreq->hpda = 0; /* no alignment constraint */
1259         if (queue->hdr_digest)
1260                 icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1261         if (queue->data_digest)
1262                 icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1263
1264         iov.iov_base = icreq;
1265         iov.iov_len = sizeof(*icreq);
1266         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1267         if (ret < 0)
1268                 goto free_icresp;
1269
1270         memset(&msg, 0, sizeof(msg));
1271         iov.iov_base = icresp;
1272         iov.iov_len = sizeof(*icresp);
1273         ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1274                         iov.iov_len, msg.msg_flags);
1275         if (ret < 0)
1276                 goto free_icresp;
1277
1278         ret = -EINVAL;
1279         if (icresp->hdr.type != nvme_tcp_icresp) {
1280                 pr_err("queue %d: bad type returned %d\n",
1281                         nvme_tcp_queue_id(queue), icresp->hdr.type);
1282                 goto free_icresp;
1283         }
1284
1285         if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1286                 pr_err("queue %d: bad pdu length returned %d\n",
1287                         nvme_tcp_queue_id(queue), icresp->hdr.plen);
1288                 goto free_icresp;
1289         }
1290
1291         if (icresp->pfv != NVME_TCP_PFV_1_0) {
1292                 pr_err("queue %d: bad pfv returned %d\n",
1293                         nvme_tcp_queue_id(queue), icresp->pfv);
1294                 goto free_icresp;
1295         }
1296
1297         ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1298         if ((queue->data_digest && !ctrl_ddgst) ||
1299             (!queue->data_digest && ctrl_ddgst)) {
1300                 pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1301                         nvme_tcp_queue_id(queue),
1302                         queue->data_digest ? "enabled" : "disabled",
1303                         ctrl_ddgst ? "enabled" : "disabled");
1304                 goto free_icresp;
1305         }
1306
1307         ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1308         if ((queue->hdr_digest && !ctrl_hdgst) ||
1309             (!queue->hdr_digest && ctrl_hdgst)) {
1310                 pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1311                         nvme_tcp_queue_id(queue),
1312                         queue->hdr_digest ? "enabled" : "disabled",
1313                         ctrl_hdgst ? "enabled" : "disabled");
1314                 goto free_icresp;
1315         }
1316
1317         if (icresp->cpda != 0) {
1318                 pr_err("queue %d: unsupported cpda returned %d\n",
1319                         nvme_tcp_queue_id(queue), icresp->cpda);
1320                 goto free_icresp;
1321         }
1322
1323         ret = 0;
1324 free_icresp:
1325         kfree(icresp);
1326 free_icreq:
1327         kfree(icreq);
1328         return ret;
1329 }
1330
1331 static bool nvme_tcp_admin_queue(struct nvme_tcp_queue *queue)
1332 {
1333         return nvme_tcp_queue_id(queue) == 0;
1334 }
1335
1336 static bool nvme_tcp_default_queue(struct nvme_tcp_queue *queue)
1337 {
1338         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1339         int qid = nvme_tcp_queue_id(queue);
1340
1341         return !nvme_tcp_admin_queue(queue) &&
1342                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT];
1343 }
1344
1345 static bool nvme_tcp_read_queue(struct nvme_tcp_queue *queue)
1346 {
1347         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1348         int qid = nvme_tcp_queue_id(queue);
1349
1350         return !nvme_tcp_admin_queue(queue) &&
1351                 !nvme_tcp_default_queue(queue) &&
1352                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1353                           ctrl->io_queues[HCTX_TYPE_READ];
1354 }
1355
1356 static bool nvme_tcp_poll_queue(struct nvme_tcp_queue *queue)
1357 {
1358         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1359         int qid = nvme_tcp_queue_id(queue);
1360
1361         return !nvme_tcp_admin_queue(queue) &&
1362                 !nvme_tcp_default_queue(queue) &&
1363                 !nvme_tcp_read_queue(queue) &&
1364                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1365                           ctrl->io_queues[HCTX_TYPE_READ] +
1366                           ctrl->io_queues[HCTX_TYPE_POLL];
1367 }
1368
1369 static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
1370 {
1371         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1372         int qid = nvme_tcp_queue_id(queue);
1373         int n = 0;
1374
1375         if (nvme_tcp_default_queue(queue))
1376                 n = qid - 1;
1377         else if (nvme_tcp_read_queue(queue))
1378                 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] - 1;
1379         else if (nvme_tcp_poll_queue(queue))
1380                 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] -
1381                                 ctrl->io_queues[HCTX_TYPE_READ] - 1;
1382         queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
1383 }
1384
1385 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl,
1386                 int qid, size_t queue_size)
1387 {
1388         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1389         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1390         int ret, rcv_pdu_size;
1391
1392         mutex_init(&queue->queue_lock);
1393         queue->ctrl = ctrl;
1394         init_llist_head(&queue->req_list);
1395         INIT_LIST_HEAD(&queue->send_list);
1396         mutex_init(&queue->send_mutex);
1397         INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1398         queue->queue_size = queue_size;
1399
1400         if (qid > 0)
1401                 queue->cmnd_capsule_len = nctrl->ioccsz * 16;
1402         else
1403                 queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1404                                                 NVME_TCP_ADMIN_CCSZ;
1405
1406         ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM,
1407                         IPPROTO_TCP, &queue->sock);
1408         if (ret) {
1409                 dev_err(nctrl->device,
1410                         "failed to create socket: %d\n", ret);
1411                 goto err_destroy_mutex;
1412         }
1413
1414         /* Single syn retry */
1415         tcp_sock_set_syncnt(queue->sock->sk, 1);
1416
1417         /* Set TCP no delay */
1418         tcp_sock_set_nodelay(queue->sock->sk);
1419
1420         /*
1421          * Cleanup whatever is sitting in the TCP transmit queue on socket
1422          * close. This is done to prevent stale data from being sent should
1423          * the network connection be restored before TCP times out.
1424          */
1425         sock_no_linger(queue->sock->sk);
1426
1427         if (so_priority > 0)
1428                 sock_set_priority(queue->sock->sk, so_priority);
1429
1430         /* Set socket type of service */
1431         if (nctrl->opts->tos >= 0)
1432                 ip_sock_set_tos(queue->sock->sk, nctrl->opts->tos);
1433
1434         /* Set 10 seconds timeout for icresp recvmsg */
1435         queue->sock->sk->sk_rcvtimeo = 10 * HZ;
1436
1437         queue->sock->sk->sk_allocation = GFP_ATOMIC;
1438         nvme_tcp_set_queue_io_cpu(queue);
1439         queue->request = NULL;
1440         queue->data_remaining = 0;
1441         queue->ddgst_remaining = 0;
1442         queue->pdu_remaining = 0;
1443         queue->pdu_offset = 0;
1444         sk_set_memalloc(queue->sock->sk);
1445
1446         if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
1447                 ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1448                         sizeof(ctrl->src_addr));
1449                 if (ret) {
1450                         dev_err(nctrl->device,
1451                                 "failed to bind queue %d socket %d\n",
1452                                 qid, ret);
1453                         goto err_sock;
1454                 }
1455         }
1456
1457         queue->hdr_digest = nctrl->opts->hdr_digest;
1458         queue->data_digest = nctrl->opts->data_digest;
1459         if (queue->hdr_digest || queue->data_digest) {
1460                 ret = nvme_tcp_alloc_crypto(queue);
1461                 if (ret) {
1462                         dev_err(nctrl->device,
1463                                 "failed to allocate queue %d crypto\n", qid);
1464                         goto err_sock;
1465                 }
1466         }
1467
1468         rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1469                         nvme_tcp_hdgst_len(queue);
1470         queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1471         if (!queue->pdu) {
1472                 ret = -ENOMEM;
1473                 goto err_crypto;
1474         }
1475
1476         dev_dbg(nctrl->device, "connecting queue %d\n",
1477                         nvme_tcp_queue_id(queue));
1478
1479         ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1480                 sizeof(ctrl->addr), 0);
1481         if (ret) {
1482                 dev_err(nctrl->device,
1483                         "failed to connect socket: %d\n", ret);
1484                 goto err_rcv_pdu;
1485         }
1486
1487         ret = nvme_tcp_init_connection(queue);
1488         if (ret)
1489                 goto err_init_connect;
1490
1491         queue->rd_enabled = true;
1492         set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1493         nvme_tcp_init_recv_ctx(queue);
1494
1495         write_lock_bh(&queue->sock->sk->sk_callback_lock);
1496         queue->sock->sk->sk_user_data = queue;
1497         queue->state_change = queue->sock->sk->sk_state_change;
1498         queue->data_ready = queue->sock->sk->sk_data_ready;
1499         queue->write_space = queue->sock->sk->sk_write_space;
1500         queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1501         queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1502         queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1503 #ifdef CONFIG_NET_RX_BUSY_POLL
1504         queue->sock->sk->sk_ll_usec = 1;
1505 #endif
1506         write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1507
1508         return 0;
1509
1510 err_init_connect:
1511         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1512 err_rcv_pdu:
1513         kfree(queue->pdu);
1514 err_crypto:
1515         if (queue->hdr_digest || queue->data_digest)
1516                 nvme_tcp_free_crypto(queue);
1517 err_sock:
1518         sock_release(queue->sock);
1519         queue->sock = NULL;
1520 err_destroy_mutex:
1521         mutex_destroy(&queue->queue_lock);
1522         return ret;
1523 }
1524
1525 static void nvme_tcp_restore_sock_calls(struct nvme_tcp_queue *queue)
1526 {
1527         struct socket *sock = queue->sock;
1528
1529         write_lock_bh(&sock->sk->sk_callback_lock);
1530         sock->sk->sk_user_data  = NULL;
1531         sock->sk->sk_data_ready = queue->data_ready;
1532         sock->sk->sk_state_change = queue->state_change;
1533         sock->sk->sk_write_space  = queue->write_space;
1534         write_unlock_bh(&sock->sk->sk_callback_lock);
1535 }
1536
1537 static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1538 {
1539         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1540         nvme_tcp_restore_sock_calls(queue);
1541         cancel_work_sync(&queue->io_work);
1542 }
1543
1544 static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1545 {
1546         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1547         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1548
1549         mutex_lock(&queue->queue_lock);
1550         if (test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1551                 __nvme_tcp_stop_queue(queue);
1552         mutex_unlock(&queue->queue_lock);
1553 }
1554
1555 static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1556 {
1557         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1558         int ret;
1559
1560         if (idx)
1561                 ret = nvmf_connect_io_queue(nctrl, idx, false);
1562         else
1563                 ret = nvmf_connect_admin_queue(nctrl);
1564
1565         if (!ret) {
1566                 set_bit(NVME_TCP_Q_LIVE, &ctrl->queues[idx].flags);
1567         } else {
1568                 if (test_bit(NVME_TCP_Q_ALLOCATED, &ctrl->queues[idx].flags))
1569                         __nvme_tcp_stop_queue(&ctrl->queues[idx]);
1570                 dev_err(nctrl->device,
1571                         "failed to connect queue: %d ret=%d\n", idx, ret);
1572         }
1573         return ret;
1574 }
1575
1576 static struct blk_mq_tag_set *nvme_tcp_alloc_tagset(struct nvme_ctrl *nctrl,
1577                 bool admin)
1578 {
1579         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1580         struct blk_mq_tag_set *set;
1581         int ret;
1582
1583         if (admin) {
1584                 set = &ctrl->admin_tag_set;
1585                 memset(set, 0, sizeof(*set));
1586                 set->ops = &nvme_tcp_admin_mq_ops;
1587                 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1588                 set->reserved_tags = NVMF_RESERVED_TAGS;
1589                 set->numa_node = nctrl->numa_node;
1590                 set->flags = BLK_MQ_F_BLOCKING;
1591                 set->cmd_size = sizeof(struct nvme_tcp_request);
1592                 set->driver_data = ctrl;
1593                 set->nr_hw_queues = 1;
1594                 set->timeout = NVME_ADMIN_TIMEOUT;
1595         } else {
1596                 set = &ctrl->tag_set;
1597                 memset(set, 0, sizeof(*set));
1598                 set->ops = &nvme_tcp_mq_ops;
1599                 set->queue_depth = nctrl->sqsize + 1;
1600                 set->reserved_tags = NVMF_RESERVED_TAGS;
1601                 set->numa_node = nctrl->numa_node;
1602                 set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_BLOCKING;
1603                 set->cmd_size = sizeof(struct nvme_tcp_request);
1604                 set->driver_data = ctrl;
1605                 set->nr_hw_queues = nctrl->queue_count - 1;
1606                 set->timeout = NVME_IO_TIMEOUT;
1607                 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
1608         }
1609
1610         ret = blk_mq_alloc_tag_set(set);
1611         if (ret)
1612                 return ERR_PTR(ret);
1613
1614         return set;
1615 }
1616
1617 static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
1618 {
1619         if (to_tcp_ctrl(ctrl)->async_req.pdu) {
1620                 cancel_work_sync(&ctrl->async_event_work);
1621                 nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
1622                 to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
1623         }
1624
1625         nvme_tcp_free_queue(ctrl, 0);
1626 }
1627
1628 static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
1629 {
1630         int i;
1631
1632         for (i = 1; i < ctrl->queue_count; i++)
1633                 nvme_tcp_free_queue(ctrl, i);
1634 }
1635
1636 static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
1637 {
1638         int i;
1639
1640         for (i = 1; i < ctrl->queue_count; i++)
1641                 nvme_tcp_stop_queue(ctrl, i);
1642 }
1643
1644 static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl)
1645 {
1646         int i, ret = 0;
1647
1648         for (i = 1; i < ctrl->queue_count; i++) {
1649                 ret = nvme_tcp_start_queue(ctrl, i);
1650                 if (ret)
1651                         goto out_stop_queues;
1652         }
1653
1654         return 0;
1655
1656 out_stop_queues:
1657         for (i--; i >= 1; i--)
1658                 nvme_tcp_stop_queue(ctrl, i);
1659         return ret;
1660 }
1661
1662 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
1663 {
1664         int ret;
1665
1666         ret = nvme_tcp_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
1667         if (ret)
1668                 return ret;
1669
1670         ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
1671         if (ret)
1672                 goto out_free_queue;
1673
1674         return 0;
1675
1676 out_free_queue:
1677         nvme_tcp_free_queue(ctrl, 0);
1678         return ret;
1679 }
1680
1681 static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1682 {
1683         int i, ret;
1684
1685         for (i = 1; i < ctrl->queue_count; i++) {
1686                 ret = nvme_tcp_alloc_queue(ctrl, i,
1687                                 ctrl->sqsize + 1);
1688                 if (ret)
1689                         goto out_free_queues;
1690         }
1691
1692         return 0;
1693
1694 out_free_queues:
1695         for (i--; i >= 1; i--)
1696                 nvme_tcp_free_queue(ctrl, i);
1697
1698         return ret;
1699 }
1700
1701 static unsigned int nvme_tcp_nr_io_queues(struct nvme_ctrl *ctrl)
1702 {
1703         unsigned int nr_io_queues;
1704
1705         nr_io_queues = min(ctrl->opts->nr_io_queues, num_online_cpus());
1706         nr_io_queues += min(ctrl->opts->nr_write_queues, num_online_cpus());
1707         nr_io_queues += min(ctrl->opts->nr_poll_queues, num_online_cpus());
1708
1709         return nr_io_queues;
1710 }
1711
1712 static void nvme_tcp_set_io_queues(struct nvme_ctrl *nctrl,
1713                 unsigned int nr_io_queues)
1714 {
1715         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1716         struct nvmf_ctrl_options *opts = nctrl->opts;
1717
1718         if (opts->nr_write_queues && opts->nr_io_queues < nr_io_queues) {
1719                 /*
1720                  * separate read/write queues
1721                  * hand out dedicated default queues only after we have
1722                  * sufficient read queues.
1723                  */
1724                 ctrl->io_queues[HCTX_TYPE_READ] = opts->nr_io_queues;
1725                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
1726                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1727                         min(opts->nr_write_queues, nr_io_queues);
1728                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1729         } else {
1730                 /*
1731                  * shared read/write queues
1732                  * either no write queues were requested, or we don't have
1733                  * sufficient queue count to have dedicated default queues.
1734                  */
1735                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1736                         min(opts->nr_io_queues, nr_io_queues);
1737                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1738         }
1739
1740         if (opts->nr_poll_queues && nr_io_queues) {
1741                 /* map dedicated poll queues only if we have queues left */
1742                 ctrl->io_queues[HCTX_TYPE_POLL] =
1743                         min(opts->nr_poll_queues, nr_io_queues);
1744         }
1745 }
1746
1747 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1748 {
1749         unsigned int nr_io_queues;
1750         int ret;
1751
1752         nr_io_queues = nvme_tcp_nr_io_queues(ctrl);
1753         ret = nvme_set_queue_count(ctrl, &nr_io_queues);
1754         if (ret)
1755                 return ret;
1756
1757         ctrl->queue_count = nr_io_queues + 1;
1758         if (ctrl->queue_count < 2) {
1759                 dev_err(ctrl->device,
1760                         "unable to set any I/O queues\n");
1761                 return -ENOMEM;
1762         }
1763
1764         dev_info(ctrl->device,
1765                 "creating %d I/O queues.\n", nr_io_queues);
1766
1767         nvme_tcp_set_io_queues(ctrl, nr_io_queues);
1768
1769         return __nvme_tcp_alloc_io_queues(ctrl);
1770 }
1771
1772 static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove)
1773 {
1774         nvme_tcp_stop_io_queues(ctrl);
1775         if (remove) {
1776                 blk_cleanup_queue(ctrl->connect_q);
1777                 blk_mq_free_tag_set(ctrl->tagset);
1778         }
1779         nvme_tcp_free_io_queues(ctrl);
1780 }
1781
1782 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
1783 {
1784         int ret;
1785
1786         ret = nvme_tcp_alloc_io_queues(ctrl);
1787         if (ret)
1788                 return ret;
1789
1790         if (new) {
1791                 ctrl->tagset = nvme_tcp_alloc_tagset(ctrl, false);
1792                 if (IS_ERR(ctrl->tagset)) {
1793                         ret = PTR_ERR(ctrl->tagset);
1794                         goto out_free_io_queues;
1795                 }
1796
1797                 ctrl->connect_q = blk_mq_init_queue(ctrl->tagset);
1798                 if (IS_ERR(ctrl->connect_q)) {
1799                         ret = PTR_ERR(ctrl->connect_q);
1800                         goto out_free_tag_set;
1801                 }
1802         }
1803
1804         ret = nvme_tcp_start_io_queues(ctrl);
1805         if (ret)
1806                 goto out_cleanup_connect_q;
1807
1808         if (!new) {
1809                 nvme_start_queues(ctrl);
1810                 if (!nvme_wait_freeze_timeout(ctrl, NVME_IO_TIMEOUT)) {
1811                         /*
1812                          * If we timed out waiting for freeze we are likely to
1813                          * be stuck.  Fail the controller initialization just
1814                          * to be safe.
1815                          */
1816                         ret = -ENODEV;
1817                         goto out_wait_freeze_timed_out;
1818                 }
1819                 blk_mq_update_nr_hw_queues(ctrl->tagset,
1820                         ctrl->queue_count - 1);
1821                 nvme_unfreeze(ctrl);
1822         }
1823
1824         return 0;
1825
1826 out_wait_freeze_timed_out:
1827         nvme_stop_queues(ctrl);
1828         nvme_sync_io_queues(ctrl);
1829         nvme_tcp_stop_io_queues(ctrl);
1830 out_cleanup_connect_q:
1831         nvme_cancel_tagset(ctrl);
1832         if (new)
1833                 blk_cleanup_queue(ctrl->connect_q);
1834 out_free_tag_set:
1835         if (new)
1836                 blk_mq_free_tag_set(ctrl->tagset);
1837 out_free_io_queues:
1838         nvme_tcp_free_io_queues(ctrl);
1839         return ret;
1840 }
1841
1842 static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove)
1843 {
1844         nvme_tcp_stop_queue(ctrl, 0);
1845         if (remove) {
1846                 blk_cleanup_queue(ctrl->admin_q);
1847                 blk_cleanup_queue(ctrl->fabrics_q);
1848                 blk_mq_free_tag_set(ctrl->admin_tagset);
1849         }
1850         nvme_tcp_free_admin_queue(ctrl);
1851 }
1852
1853 static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
1854 {
1855         int error;
1856
1857         error = nvme_tcp_alloc_admin_queue(ctrl);
1858         if (error)
1859                 return error;
1860
1861         if (new) {
1862                 ctrl->admin_tagset = nvme_tcp_alloc_tagset(ctrl, true);
1863                 if (IS_ERR(ctrl->admin_tagset)) {
1864                         error = PTR_ERR(ctrl->admin_tagset);
1865                         goto out_free_queue;
1866                 }
1867
1868                 ctrl->fabrics_q = blk_mq_init_queue(ctrl->admin_tagset);
1869                 if (IS_ERR(ctrl->fabrics_q)) {
1870                         error = PTR_ERR(ctrl->fabrics_q);
1871                         goto out_free_tagset;
1872                 }
1873
1874                 ctrl->admin_q = blk_mq_init_queue(ctrl->admin_tagset);
1875                 if (IS_ERR(ctrl->admin_q)) {
1876                         error = PTR_ERR(ctrl->admin_q);
1877                         goto out_cleanup_fabrics_q;
1878                 }
1879         }
1880
1881         error = nvme_tcp_start_queue(ctrl, 0);
1882         if (error)
1883                 goto out_cleanup_queue;
1884
1885         error = nvme_enable_ctrl(ctrl);
1886         if (error)
1887                 goto out_stop_queue;
1888
1889         blk_mq_unquiesce_queue(ctrl->admin_q);
1890
1891         error = nvme_init_ctrl_finish(ctrl);
1892         if (error)
1893                 goto out_quiesce_queue;
1894
1895         return 0;
1896
1897 out_quiesce_queue:
1898         blk_mq_quiesce_queue(ctrl->admin_q);
1899         blk_sync_queue(ctrl->admin_q);
1900 out_stop_queue:
1901         nvme_tcp_stop_queue(ctrl, 0);
1902         nvme_cancel_admin_tagset(ctrl);
1903 out_cleanup_queue:
1904         if (new)
1905                 blk_cleanup_queue(ctrl->admin_q);
1906 out_cleanup_fabrics_q:
1907         if (new)
1908                 blk_cleanup_queue(ctrl->fabrics_q);
1909 out_free_tagset:
1910         if (new)
1911                 blk_mq_free_tag_set(ctrl->admin_tagset);
1912 out_free_queue:
1913         nvme_tcp_free_admin_queue(ctrl);
1914         return error;
1915 }
1916
1917 static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
1918                 bool remove)
1919 {
1920         blk_mq_quiesce_queue(ctrl->admin_q);
1921         blk_sync_queue(ctrl->admin_q);
1922         nvme_tcp_stop_queue(ctrl, 0);
1923         nvme_cancel_admin_tagset(ctrl);
1924         if (remove)
1925                 blk_mq_unquiesce_queue(ctrl->admin_q);
1926         nvme_tcp_destroy_admin_queue(ctrl, remove);
1927 }
1928
1929 static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
1930                 bool remove)
1931 {
1932         if (ctrl->queue_count <= 1)
1933                 return;
1934         blk_mq_quiesce_queue(ctrl->admin_q);
1935         nvme_start_freeze(ctrl);
1936         nvme_stop_queues(ctrl);
1937         nvme_sync_io_queues(ctrl);
1938         nvme_tcp_stop_io_queues(ctrl);
1939         nvme_cancel_tagset(ctrl);
1940         if (remove)
1941                 nvme_start_queues(ctrl);
1942         nvme_tcp_destroy_io_queues(ctrl, remove);
1943 }
1944
1945 static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
1946 {
1947         /* If we are resetting/deleting then do nothing */
1948         if (ctrl->state != NVME_CTRL_CONNECTING) {
1949                 WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW ||
1950                         ctrl->state == NVME_CTRL_LIVE);
1951                 return;
1952         }
1953
1954         if (nvmf_should_reconnect(ctrl)) {
1955                 dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
1956                         ctrl->opts->reconnect_delay);
1957                 queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
1958                                 ctrl->opts->reconnect_delay * HZ);
1959         } else {
1960                 dev_info(ctrl->device, "Removing controller...\n");
1961                 nvme_delete_ctrl(ctrl);
1962         }
1963 }
1964
1965 static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
1966 {
1967         struct nvmf_ctrl_options *opts = ctrl->opts;
1968         int ret;
1969
1970         ret = nvme_tcp_configure_admin_queue(ctrl, new);
1971         if (ret)
1972                 return ret;
1973
1974         if (ctrl->icdoff) {
1975                 dev_err(ctrl->device, "icdoff is not supported!\n");
1976                 goto destroy_admin;
1977         }
1978
1979         if (!(ctrl->sgls & ((1 << 0) | (1 << 1)))) {
1980                 dev_err(ctrl->device, "Mandatory sgls are not supported!\n");
1981                 goto destroy_admin;
1982         }
1983
1984         if (opts->queue_size > ctrl->sqsize + 1)
1985                 dev_warn(ctrl->device,
1986                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
1987                         opts->queue_size, ctrl->sqsize + 1);
1988
1989         if (ctrl->sqsize + 1 > ctrl->maxcmd) {
1990                 dev_warn(ctrl->device,
1991                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
1992                         ctrl->sqsize + 1, ctrl->maxcmd);
1993                 ctrl->sqsize = ctrl->maxcmd - 1;
1994         }
1995
1996         if (ctrl->queue_count > 1) {
1997                 ret = nvme_tcp_configure_io_queues(ctrl, new);
1998                 if (ret)
1999                         goto destroy_admin;
2000         }
2001
2002         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
2003                 /*
2004                  * state change failure is ok if we started ctrl delete,
2005                  * unless we're during creation of a new controller to
2006                  * avoid races with teardown flow.
2007                  */
2008                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2009                              ctrl->state != NVME_CTRL_DELETING_NOIO);
2010                 WARN_ON_ONCE(new);
2011                 ret = -EINVAL;
2012                 goto destroy_io;
2013         }
2014
2015         nvme_start_ctrl(ctrl);
2016         return 0;
2017
2018 destroy_io:
2019         if (ctrl->queue_count > 1) {
2020                 nvme_stop_queues(ctrl);
2021                 nvme_sync_io_queues(ctrl);
2022                 nvme_tcp_stop_io_queues(ctrl);
2023                 nvme_cancel_tagset(ctrl);
2024                 nvme_tcp_destroy_io_queues(ctrl, new);
2025         }
2026 destroy_admin:
2027         blk_mq_quiesce_queue(ctrl->admin_q);
2028         blk_sync_queue(ctrl->admin_q);
2029         nvme_tcp_stop_queue(ctrl, 0);
2030         nvme_cancel_admin_tagset(ctrl);
2031         nvme_tcp_destroy_admin_queue(ctrl, new);
2032         return ret;
2033 }
2034
2035 static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
2036 {
2037         struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
2038                         struct nvme_tcp_ctrl, connect_work);
2039         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2040
2041         ++ctrl->nr_reconnects;
2042
2043         if (nvme_tcp_setup_ctrl(ctrl, false))
2044                 goto requeue;
2045
2046         dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n",
2047                         ctrl->nr_reconnects);
2048
2049         ctrl->nr_reconnects = 0;
2050
2051         return;
2052
2053 requeue:
2054         dev_info(ctrl->device, "Failed reconnect attempt %d\n",
2055                         ctrl->nr_reconnects);
2056         nvme_tcp_reconnect_or_remove(ctrl);
2057 }
2058
2059 static void nvme_tcp_error_recovery_work(struct work_struct *work)
2060 {
2061         struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
2062                                 struct nvme_tcp_ctrl, err_work);
2063         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2064
2065         nvme_stop_keep_alive(ctrl);
2066         nvme_tcp_teardown_io_queues(ctrl, false);
2067         /* unquiesce to fail fast pending requests */
2068         nvme_start_queues(ctrl);
2069         nvme_tcp_teardown_admin_queue(ctrl, false);
2070         blk_mq_unquiesce_queue(ctrl->admin_q);
2071
2072         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2073                 /* state change failure is ok if we started ctrl delete */
2074                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2075                              ctrl->state != NVME_CTRL_DELETING_NOIO);
2076                 return;
2077         }
2078
2079         nvme_tcp_reconnect_or_remove(ctrl);
2080 }
2081
2082 static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
2083 {
2084         cancel_work_sync(&to_tcp_ctrl(ctrl)->err_work);
2085         cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
2086
2087         nvme_tcp_teardown_io_queues(ctrl, shutdown);
2088         blk_mq_quiesce_queue(ctrl->admin_q);
2089         if (shutdown)
2090                 nvme_shutdown_ctrl(ctrl);
2091         else
2092                 nvme_disable_ctrl(ctrl);
2093         nvme_tcp_teardown_admin_queue(ctrl, shutdown);
2094 }
2095
2096 static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
2097 {
2098         nvme_tcp_teardown_ctrl(ctrl, true);
2099 }
2100
2101 static void nvme_reset_ctrl_work(struct work_struct *work)
2102 {
2103         struct nvme_ctrl *ctrl =
2104                 container_of(work, struct nvme_ctrl, reset_work);
2105
2106         nvme_stop_ctrl(ctrl);
2107         nvme_tcp_teardown_ctrl(ctrl, false);
2108
2109         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2110                 /* state change failure is ok if we started ctrl delete */
2111                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2112                              ctrl->state != NVME_CTRL_DELETING_NOIO);
2113                 return;
2114         }
2115
2116         if (nvme_tcp_setup_ctrl(ctrl, false))
2117                 goto out_fail;
2118
2119         return;
2120
2121 out_fail:
2122         ++ctrl->nr_reconnects;
2123         nvme_tcp_reconnect_or_remove(ctrl);
2124 }
2125
2126 static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
2127 {
2128         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
2129
2130         if (list_empty(&ctrl->list))
2131                 goto free_ctrl;
2132
2133         mutex_lock(&nvme_tcp_ctrl_mutex);
2134         list_del(&ctrl->list);
2135         mutex_unlock(&nvme_tcp_ctrl_mutex);
2136
2137         nvmf_free_options(nctrl->opts);
2138 free_ctrl:
2139         kfree(ctrl->queues);
2140         kfree(ctrl);
2141 }
2142
2143 static void nvme_tcp_set_sg_null(struct nvme_command *c)
2144 {
2145         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2146
2147         sg->addr = 0;
2148         sg->length = 0;
2149         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2150                         NVME_SGL_FMT_TRANSPORT_A;
2151 }
2152
2153 static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
2154                 struct nvme_command *c, u32 data_len)
2155 {
2156         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2157
2158         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
2159         sg->length = cpu_to_le32(data_len);
2160         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
2161 }
2162
2163 static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
2164                 u32 data_len)
2165 {
2166         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2167
2168         sg->addr = 0;
2169         sg->length = cpu_to_le32(data_len);
2170         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2171                         NVME_SGL_FMT_TRANSPORT_A;
2172 }
2173
2174 static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
2175 {
2176         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2177         struct nvme_tcp_queue *queue = &ctrl->queues[0];
2178         struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2179         struct nvme_command *cmd = &pdu->cmd;
2180         u8 hdgst = nvme_tcp_hdgst_len(queue);
2181
2182         memset(pdu, 0, sizeof(*pdu));
2183         pdu->hdr.type = nvme_tcp_cmd;
2184         if (queue->hdr_digest)
2185                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2186         pdu->hdr.hlen = sizeof(*pdu);
2187         pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2188
2189         cmd->common.opcode = nvme_admin_async_event;
2190         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2191         cmd->common.flags |= NVME_CMD_SGL_METABUF;
2192         nvme_tcp_set_sg_null(cmd);
2193
2194         ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2195         ctrl->async_req.offset = 0;
2196         ctrl->async_req.curr_bio = NULL;
2197         ctrl->async_req.data_len = 0;
2198
2199         nvme_tcp_queue_request(&ctrl->async_req, true, true);
2200 }
2201
2202 static void nvme_tcp_complete_timed_out(struct request *rq)
2203 {
2204         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2205         struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2206
2207         nvme_tcp_stop_queue(ctrl, nvme_tcp_queue_id(req->queue));
2208         if (blk_mq_request_started(rq) && !blk_mq_request_completed(rq)) {
2209                 nvme_req(rq)->status = NVME_SC_HOST_ABORTED_CMD;
2210                 blk_mq_complete_request(rq);
2211         }
2212 }
2213
2214 static enum blk_eh_timer_return
2215 nvme_tcp_timeout(struct request *rq, bool reserved)
2216 {
2217         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2218         struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2219         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2220
2221         dev_warn(ctrl->device,
2222                 "queue %d: timeout request %#x type %d\n",
2223                 nvme_tcp_queue_id(req->queue), rq->tag, pdu->hdr.type);
2224
2225         if (ctrl->state != NVME_CTRL_LIVE) {
2226                 /*
2227                  * If we are resetting, connecting or deleting we should
2228                  * complete immediately because we may block controller
2229                  * teardown or setup sequence
2230                  * - ctrl disable/shutdown fabrics requests
2231                  * - connect requests
2232                  * - initialization admin requests
2233                  * - I/O requests that entered after unquiescing and
2234                  *   the controller stopped responding
2235                  *
2236                  * All other requests should be cancelled by the error
2237                  * recovery work, so it's fine that we fail it here.
2238                  */
2239                 nvme_tcp_complete_timed_out(rq);
2240                 return BLK_EH_DONE;
2241         }
2242
2243         /*
2244          * LIVE state should trigger the normal error recovery which will
2245          * handle completing this request.
2246          */
2247         nvme_tcp_error_recovery(ctrl);
2248         return BLK_EH_RESET_TIMER;
2249 }
2250
2251 static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2252                         struct request *rq)
2253 {
2254         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2255         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2256         struct nvme_command *c = &pdu->cmd;
2257
2258         c->common.flags |= NVME_CMD_SGL_METABUF;
2259
2260         if (!blk_rq_nr_phys_segments(rq))
2261                 nvme_tcp_set_sg_null(c);
2262         else if (rq_data_dir(rq) == WRITE &&
2263             req->data_len <= nvme_tcp_inline_data_size(queue))
2264                 nvme_tcp_set_sg_inline(queue, c, req->data_len);
2265         else
2266                 nvme_tcp_set_sg_host_data(c, req->data_len);
2267
2268         return 0;
2269 }
2270
2271 static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2272                 struct request *rq)
2273 {
2274         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2275         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2276         struct nvme_tcp_queue *queue = req->queue;
2277         u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2278         blk_status_t ret;
2279
2280         ret = nvme_setup_cmd(ns, rq);
2281         if (ret)
2282                 return ret;
2283
2284         req->state = NVME_TCP_SEND_CMD_PDU;
2285         req->offset = 0;
2286         req->data_sent = 0;
2287         req->pdu_len = 0;
2288         req->pdu_sent = 0;
2289         req->data_len = blk_rq_nr_phys_segments(rq) ?
2290                                 blk_rq_payload_bytes(rq) : 0;
2291         req->curr_bio = rq->bio;
2292         if (req->curr_bio && req->data_len)
2293                 nvme_tcp_init_iter(req, rq_data_dir(rq));
2294
2295         if (rq_data_dir(rq) == WRITE &&
2296             req->data_len <= nvme_tcp_inline_data_size(queue))
2297                 req->pdu_len = req->data_len;
2298
2299         pdu->hdr.type = nvme_tcp_cmd;
2300         pdu->hdr.flags = 0;
2301         if (queue->hdr_digest)
2302                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2303         if (queue->data_digest && req->pdu_len) {
2304                 pdu->hdr.flags |= NVME_TCP_F_DDGST;
2305                 ddgst = nvme_tcp_ddgst_len(queue);
2306         }
2307         pdu->hdr.hlen = sizeof(*pdu);
2308         pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2309         pdu->hdr.plen =
2310                 cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2311
2312         ret = nvme_tcp_map_data(queue, rq);
2313         if (unlikely(ret)) {
2314                 nvme_cleanup_cmd(rq);
2315                 dev_err(queue->ctrl->ctrl.device,
2316                         "Failed to map data (%d)\n", ret);
2317                 return ret;
2318         }
2319
2320         return 0;
2321 }
2322
2323 static void nvme_tcp_commit_rqs(struct blk_mq_hw_ctx *hctx)
2324 {
2325         struct nvme_tcp_queue *queue = hctx->driver_data;
2326
2327         if (!llist_empty(&queue->req_list))
2328                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
2329 }
2330
2331 static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2332                 const struct blk_mq_queue_data *bd)
2333 {
2334         struct nvme_ns *ns = hctx->queue->queuedata;
2335         struct nvme_tcp_queue *queue = hctx->driver_data;
2336         struct request *rq = bd->rq;
2337         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2338         bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2339         blk_status_t ret;
2340
2341         if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2342                 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
2343
2344         ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2345         if (unlikely(ret))
2346                 return ret;
2347
2348         blk_mq_start_request(rq);
2349
2350         nvme_tcp_queue_request(req, true, bd->last);
2351
2352         return BLK_STS_OK;
2353 }
2354
2355 static int nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2356 {
2357         struct nvme_tcp_ctrl *ctrl = set->driver_data;
2358         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2359
2360         if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2361                 /* separate read/write queues */
2362                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2363                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2364                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2365                 set->map[HCTX_TYPE_READ].nr_queues =
2366                         ctrl->io_queues[HCTX_TYPE_READ];
2367                 set->map[HCTX_TYPE_READ].queue_offset =
2368                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2369         } else {
2370                 /* shared read/write queues */
2371                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2372                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2373                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2374                 set->map[HCTX_TYPE_READ].nr_queues =
2375                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2376                 set->map[HCTX_TYPE_READ].queue_offset = 0;
2377         }
2378         blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2379         blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
2380
2381         if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2382                 /* map dedicated poll queues only if we have queues left */
2383                 set->map[HCTX_TYPE_POLL].nr_queues =
2384                                 ctrl->io_queues[HCTX_TYPE_POLL];
2385                 set->map[HCTX_TYPE_POLL].queue_offset =
2386                         ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2387                         ctrl->io_queues[HCTX_TYPE_READ];
2388                 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2389         }
2390
2391         dev_info(ctrl->ctrl.device,
2392                 "mapped %d/%d/%d default/read/poll queues.\n",
2393                 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2394                 ctrl->io_queues[HCTX_TYPE_READ],
2395                 ctrl->io_queues[HCTX_TYPE_POLL]);
2396
2397         return 0;
2398 }
2399
2400 static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx)
2401 {
2402         struct nvme_tcp_queue *queue = hctx->driver_data;
2403         struct sock *sk = queue->sock->sk;
2404
2405         if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2406                 return 0;
2407
2408         set_bit(NVME_TCP_Q_POLLING, &queue->flags);
2409         if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue))
2410                 sk_busy_loop(sk, true);
2411         nvme_tcp_try_recv(queue);
2412         clear_bit(NVME_TCP_Q_POLLING, &queue->flags);
2413         return queue->nr_cqe;
2414 }
2415
2416 static const struct blk_mq_ops nvme_tcp_mq_ops = {
2417         .queue_rq       = nvme_tcp_queue_rq,
2418         .commit_rqs     = nvme_tcp_commit_rqs,
2419         .complete       = nvme_complete_rq,
2420         .init_request   = nvme_tcp_init_request,
2421         .exit_request   = nvme_tcp_exit_request,
2422         .init_hctx      = nvme_tcp_init_hctx,
2423         .timeout        = nvme_tcp_timeout,
2424         .map_queues     = nvme_tcp_map_queues,
2425         .poll           = nvme_tcp_poll,
2426 };
2427
2428 static const struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2429         .queue_rq       = nvme_tcp_queue_rq,
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_admin_hctx,
2434         .timeout        = nvme_tcp_timeout,
2435 };
2436
2437 static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2438         .name                   = "tcp",
2439         .module                 = THIS_MODULE,
2440         .flags                  = NVME_F_FABRICS,
2441         .reg_read32             = nvmf_reg_read32,
2442         .reg_read64             = nvmf_reg_read64,
2443         .reg_write32            = nvmf_reg_write32,
2444         .free_ctrl              = nvme_tcp_free_ctrl,
2445         .submit_async_event     = nvme_tcp_submit_async_event,
2446         .delete_ctrl            = nvme_tcp_delete_ctrl,
2447         .get_address            = nvmf_get_address,
2448 };
2449
2450 static bool
2451 nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2452 {
2453         struct nvme_tcp_ctrl *ctrl;
2454         bool found = false;
2455
2456         mutex_lock(&nvme_tcp_ctrl_mutex);
2457         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2458                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2459                 if (found)
2460                         break;
2461         }
2462         mutex_unlock(&nvme_tcp_ctrl_mutex);
2463
2464         return found;
2465 }
2466
2467 static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2468                 struct nvmf_ctrl_options *opts)
2469 {
2470         struct nvme_tcp_ctrl *ctrl;
2471         int ret;
2472
2473         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2474         if (!ctrl)
2475                 return ERR_PTR(-ENOMEM);
2476
2477         INIT_LIST_HEAD(&ctrl->list);
2478         ctrl->ctrl.opts = opts;
2479         ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2480                                 opts->nr_poll_queues + 1;
2481         ctrl->ctrl.sqsize = opts->queue_size - 1;
2482         ctrl->ctrl.kato = opts->kato;
2483
2484         INIT_DELAYED_WORK(&ctrl->connect_work,
2485                         nvme_tcp_reconnect_ctrl_work);
2486         INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2487         INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2488
2489         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2490                 opts->trsvcid =
2491                         kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2492                 if (!opts->trsvcid) {
2493                         ret = -ENOMEM;
2494                         goto out_free_ctrl;
2495                 }
2496                 opts->mask |= NVMF_OPT_TRSVCID;
2497         }
2498
2499         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2500                         opts->traddr, opts->trsvcid, &ctrl->addr);
2501         if (ret) {
2502                 pr_err("malformed address passed: %s:%s\n",
2503                         opts->traddr, opts->trsvcid);
2504                 goto out_free_ctrl;
2505         }
2506
2507         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2508                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2509                         opts->host_traddr, NULL, &ctrl->src_addr);
2510                 if (ret) {
2511                         pr_err("malformed src address passed: %s\n",
2512                                opts->host_traddr);
2513                         goto out_free_ctrl;
2514                 }
2515         }
2516
2517         if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2518                 ret = -EALREADY;
2519                 goto out_free_ctrl;
2520         }
2521
2522         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2523                                 GFP_KERNEL);
2524         if (!ctrl->queues) {
2525                 ret = -ENOMEM;
2526                 goto out_free_ctrl;
2527         }
2528
2529         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2530         if (ret)
2531                 goto out_kfree_queues;
2532
2533         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2534                 WARN_ON_ONCE(1);
2535                 ret = -EINTR;
2536                 goto out_uninit_ctrl;
2537         }
2538
2539         ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2540         if (ret)
2541                 goto out_uninit_ctrl;
2542
2543         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
2544                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2545
2546         mutex_lock(&nvme_tcp_ctrl_mutex);
2547         list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2548         mutex_unlock(&nvme_tcp_ctrl_mutex);
2549
2550         return &ctrl->ctrl;
2551
2552 out_uninit_ctrl:
2553         nvme_uninit_ctrl(&ctrl->ctrl);
2554         nvme_put_ctrl(&ctrl->ctrl);
2555         if (ret > 0)
2556                 ret = -EIO;
2557         return ERR_PTR(ret);
2558 out_kfree_queues:
2559         kfree(ctrl->queues);
2560 out_free_ctrl:
2561         kfree(ctrl);
2562         return ERR_PTR(ret);
2563 }
2564
2565 static struct nvmf_transport_ops nvme_tcp_transport = {
2566         .name           = "tcp",
2567         .module         = THIS_MODULE,
2568         .required_opts  = NVMF_OPT_TRADDR,
2569         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2570                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2571                           NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
2572                           NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2573                           NVMF_OPT_TOS,
2574         .create_ctrl    = nvme_tcp_create_ctrl,
2575 };
2576
2577 static int __init nvme_tcp_init_module(void)
2578 {
2579         nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
2580                         WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2581         if (!nvme_tcp_wq)
2582                 return -ENOMEM;
2583
2584         nvmf_register_transport(&nvme_tcp_transport);
2585         return 0;
2586 }
2587
2588 static void __exit nvme_tcp_cleanup_module(void)
2589 {
2590         struct nvme_tcp_ctrl *ctrl;
2591
2592         nvmf_unregister_transport(&nvme_tcp_transport);
2593
2594         mutex_lock(&nvme_tcp_ctrl_mutex);
2595         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
2596                 nvme_delete_ctrl(&ctrl->ctrl);
2597         mutex_unlock(&nvme_tcp_ctrl_mutex);
2598         flush_workqueue(nvme_delete_wq);
2599
2600         destroy_workqueue(nvme_tcp_wq);
2601 }
2602
2603 module_init(nvme_tcp_init_module);
2604 module_exit(nvme_tcp_cleanup_module);
2605
2606 MODULE_LICENSE("GPL v2");