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         __le16                  status;
  49         struct list_head        entry;
  50         struct llist_node       lentry;
  51         __le32                  ddgst;
  52
  53         struct bio              *curr_bio;
  54         struct iov_iter         iter;
  55
  56         /* send state */
  57         size_t                  offset;
  58         size_t                  data_sent;
  59         enum nvme_tcp_send_state state;
  60 };
  61
  62 enum nvme_tcp_queue_flags {
  63         NVME_TCP_Q_ALLOCATED    = 0,
  64         NVME_TCP_Q_LIVE         = 1,
  65         NVME_TCP_Q_POLLING      = 2,
  66 };
  67
  68 enum nvme_tcp_recv_state {
  69         NVME_TCP_RECV_PDU = 0,
  70         NVME_TCP_RECV_DATA,
  71         NVME_TCP_RECV_DDGST,
  72 };
  73
  74 struct nvme_tcp_ctrl;
  75 struct nvme_tcp_queue {
  76         struct socket           *sock;
  77         struct work_struct      io_work;
  78         int                     io_cpu;
  79
  80         struct mutex            queue_lock;
  81         struct mutex            send_mutex;
  82         struct llist_head       req_list;
  83         struct list_head        send_list;
  84         bool                    more_requests;
  85
  86         /* recv state */
  87         void                    *pdu;
  88         int                     pdu_remaining;
  89         int                     pdu_offset;
  90         size_t                  data_remaining;
  91         size_t                  ddgst_remaining;
  92         unsigned int            nr_cqe;
  93
  94         /* send state */
  95         struct nvme_tcp_request *request;
  96
  97         int                     queue_size;
  98         size_t                  cmnd_capsule_len;
  99         struct nvme_tcp_ctrl    *ctrl;
 100         unsigned long           flags;
 101         bool                    rd_enabled;
 102
 103         bool                    hdr_digest;
 104         bool                    data_digest;
 105         struct ahash_request    *rcv_hash;
 106         struct ahash_request    *snd_hash;
 107         __le32                  exp_ddgst;
 108         __le32                  recv_ddgst;
 109
 110         struct page_frag_cache  pf_cache;
 111
 112         void (*state_change)(struct sock *);
 113         void (*data_ready)(struct sock *);
 114         void (*write_space)(struct sock *);
 115 };
 116
 117 struct nvme_tcp_ctrl {
 118         /* read only in the hot path */
 119         struct nvme_tcp_queue   *queues;
 120         struct blk_mq_tag_set   tag_set;
 121
 122         /* other member variables */
 123         struct list_head        list;
 124         struct blk_mq_tag_set   admin_tag_set;
 125         struct sockaddr_storage addr;
 126         struct sockaddr_storage src_addr;
 127         struct nvme_ctrl        ctrl;
 128
 129         struct work_struct      err_work;
 130         struct delayed_work     connect_work;
 131         struct nvme_tcp_request async_req;
 132         u32                     io_queues[HCTX_MAX_TYPES];
 133 };
 134
 135 static LIST_HEAD(nvme_tcp_ctrl_list);
 136 static DEFINE_MUTEX(nvme_tcp_ctrl_mutex);
 137 static struct workqueue_struct *nvme_tcp_wq;
 138 static const struct blk_mq_ops nvme_tcp_mq_ops;
 139 static const struct blk_mq_ops nvme_tcp_admin_mq_ops;
 140 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue);
 141
 142 static inline struct nvme_tcp_ctrl *to_tcp_ctrl(struct nvme_ctrl *ctrl)
 143 {
 144         return container_of(ctrl, struct nvme_tcp_ctrl, ctrl);
 145 }
 146
 147 static inline int nvme_tcp_queue_id(struct nvme_tcp_queue *queue)
 148 {
 149         return queue - queue->ctrl->queues;
 150 }
 151
 152 static inline struct blk_mq_tags *nvme_tcp_tagset(struct nvme_tcp_queue *queue)
 153 {
 154         u32 queue_idx = nvme_tcp_queue_id(queue);
 155
 156         if (queue_idx == 0)
 157                 return queue->ctrl->admin_tag_set.tags[queue_idx];
 158         return queue->ctrl->tag_set.tags[queue_idx - 1];
 159 }
 160
 161 static inline u8 nvme_tcp_hdgst_len(struct nvme_tcp_queue *queue)
 162 {
 163         return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 164 }
 165
 166 static inline u8 nvme_tcp_ddgst_len(struct nvme_tcp_queue *queue)
 167 {
 168         return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 169 }
 170
 171 static inline size_t nvme_tcp_inline_data_size(struct nvme_tcp_queue *queue)
 172 {
 173         return queue->cmnd_capsule_len - sizeof(struct nvme_command);
 174 }
 175
 176 static inline bool nvme_tcp_async_req(struct nvme_tcp_request *req)
 177 {
 178         return req == &req->queue->ctrl->async_req;
 179 }
 180
 181 static inline bool nvme_tcp_has_inline_data(struct nvme_tcp_request *req)
 182 {
 183         struct request *rq;
 184
 185         if (unlikely(nvme_tcp_async_req(req)))
 186                 return false; /* async events don't have a request */
 187
 188         rq = blk_mq_rq_from_pdu(req);
 189
 190         return rq_data_dir(rq) == WRITE && req->data_len &&
 191                 req->data_len <= nvme_tcp_inline_data_size(req->queue);
 192 }
 193
 194 static inline struct page *nvme_tcp_req_cur_page(struct nvme_tcp_request *req)
 195 {
 196         return req->iter.bvec->bv_page;
 197 }
 198
 199 static inline size_t nvme_tcp_req_cur_offset(struct nvme_tcp_request *req)
 200 {
 201         return req->iter.bvec->bv_offset + req->iter.iov_offset;
 202 }
 203
 204 static inline size_t nvme_tcp_req_cur_length(struct nvme_tcp_request *req)
 205 {
 206         return min_t(size_t, iov_iter_single_seg_count(&req->iter),
 207                         req->pdu_len - req->pdu_sent);
 208 }
 209
 210 static inline size_t nvme_tcp_pdu_data_left(struct nvme_tcp_request *req)
 211 {
 212         return rq_data_dir(blk_mq_rq_from_pdu(req)) == WRITE ?
 213                         req->pdu_len - req->pdu_sent : 0;
 214 }
 215
 216 static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req,
 217                 int len)
 218 {
 219         return nvme_tcp_pdu_data_left(req) <= len;
 220 }
 221
 222 static void nvme_tcp_init_iter(struct nvme_tcp_request *req,
 223                 unsigned int dir)
 224 {
 225         struct request *rq = blk_mq_rq_from_pdu(req);
 226         struct bio_vec *vec;
 227         unsigned int size;
 228         int nr_bvec;
 229         size_t offset;
 230
 231         if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) {
 232                 vec = &rq->special_vec;
 233                 nr_bvec = 1;
 234                 size = blk_rq_payload_bytes(rq);
 235                 offset = 0;
 236         } else {
 237                 struct bio *bio = req->curr_bio;
 238                 struct bvec_iter bi;
 239                 struct bio_vec bv;
 240
 241                 vec = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
 242                 nr_bvec = 0;
 243                 bio_for_each_bvec(bv, bio, bi) {
 244                         nr_bvec++;
 245                 }
 246                 size = bio->bi_iter.bi_size;
 247                 offset = bio->bi_iter.bi_bvec_done;
 248         }
 249
 250         iov_iter_bvec(&req->iter, dir, vec, nr_bvec, size);
 251         req->iter.iov_offset = offset;
 252 }
 253
 254 static inline void nvme_tcp_advance_req(struct nvme_tcp_request *req,
 255                 int len)
 256 {
 257         req->data_sent += len;
 258         req->pdu_sent += len;
 259         iov_iter_advance(&req->iter, len);
 260         if (!iov_iter_count(&req->iter) &&
 261             req->data_sent < req->data_len) {
 262                 req->curr_bio = req->curr_bio->bi_next;
 263                 nvme_tcp_init_iter(req, WRITE);
 264         }
 265 }
 266
 267 static inline void nvme_tcp_send_all(struct nvme_tcp_queue *queue)
 268 {
 269         int ret;
 270
 271         /* drain the send queue as much as we can... */
 272         do {
 273                 ret = nvme_tcp_try_send(queue);
 274         } while (ret > 0);
 275 }
 276
 277 static inline bool nvme_tcp_queue_more(struct nvme_tcp_queue *queue)
 278 {
 279         return !list_empty(&queue->send_list) ||
 280                 !llist_empty(&queue->req_list) || queue->more_requests;
 281 }
 282
 283 static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req,
 284                 bool sync, bool last)
 285 {
 286         struct nvme_tcp_queue *queue = req->queue;
 287         bool empty;
 288
 289         empty = llist_add(&req->lentry, &queue->req_list) &&
 290                 list_empty(&queue->send_list) && !queue->request;
 291
 292         /*
 293          * if we're the first on the send_list and we can try to send
 294          * directly, otherwise queue io_work. Also, only do that if we
 295          * are on the same cpu, so we don't introduce contention.
 296          */
 297         if (queue->io_cpu == raw_smp_processor_id() &&
 298             sync && empty && mutex_trylock(&queue->send_mutex)) {
 299                 queue->more_requests = !last;
 300                 nvme_tcp_send_all(queue);
 301                 queue->more_requests = false;
 302                 mutex_unlock(&queue->send_mutex);
 303         }
 304
 305         if (last && nvme_tcp_queue_more(queue))
 306                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 307 }
 308
 309 static void nvme_tcp_process_req_list(struct nvme_tcp_queue *queue)
 310 {
 311         struct nvme_tcp_request *req;
 312         struct llist_node *node;
 313
 314         for (node = llist_del_all(&queue->req_list); node; node = node->next) {
 315                 req = llist_entry(node, struct nvme_tcp_request, lentry);
 316                 list_add(&req->entry, &queue->send_list);
 317         }
 318 }
 319
 320 static inline struct nvme_tcp_request *
 321 nvme_tcp_fetch_request(struct nvme_tcp_queue *queue)
 322 {
 323         struct nvme_tcp_request *req;
 324
 325         req = list_first_entry_or_null(&queue->send_list,
 326                         struct nvme_tcp_request, entry);
 327         if (!req) {
 328                 nvme_tcp_process_req_list(queue);
 329                 req = list_first_entry_or_null(&queue->send_list,
 330                                 struct nvme_tcp_request, entry);
 331                 if (unlikely(!req))
 332                         return NULL;
 333         }
 334
 335         list_del(&req->entry);
 336         return req;
 337 }
 338
 339 static inline void nvme_tcp_ddgst_final(struct ahash_request *hash,
 340                 __le32 *dgst)
 341 {
 342         ahash_request_set_crypt(hash, NULL, (u8 *)dgst, 0);
 343         crypto_ahash_final(hash);
 344 }
 345
 346 static inline void nvme_tcp_ddgst_update(struct ahash_request *hash,
 347                 struct page *page, off_t off, size_t len)
 348 {
 349         struct scatterlist sg;
 350
 351         sg_init_marker(&sg, 1);
 352         sg_set_page(&sg, page, len, off);
 353         ahash_request_set_crypt(hash, &sg, NULL, len);
 354         crypto_ahash_update(hash);
 355 }
 356
 357 static inline void nvme_tcp_hdgst(struct ahash_request *hash,
 358                 void *pdu, size_t len)
 359 {
 360         struct scatterlist sg;
 361
 362         sg_init_one(&sg, pdu, len);
 363         ahash_request_set_crypt(hash, &sg, pdu + len, len);
 364         crypto_ahash_digest(hash);
 365 }
 366
 367 static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue,
 368                 void *pdu, size_t pdu_len)
 369 {
 370         struct nvme_tcp_hdr *hdr = pdu;
 371         __le32 recv_digest;
 372         __le32 exp_digest;
 373
 374         if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
 375                 dev_err(queue->ctrl->ctrl.device,
 376                         "queue %d: header digest flag is cleared\n",
 377                         nvme_tcp_queue_id(queue));
 378                 return -EPROTO;
 379         }
 380
 381         recv_digest = *(__le32 *)(pdu + hdr->hlen);
 382         nvme_tcp_hdgst(queue->rcv_hash, pdu, pdu_len);
 383         exp_digest = *(__le32 *)(pdu + hdr->hlen);
 384         if (recv_digest != exp_digest) {
 385                 dev_err(queue->ctrl->ctrl.device,
 386                         "header digest error: recv %#x expected %#x\n",
 387                         le32_to_cpu(recv_digest), le32_to_cpu(exp_digest));
 388                 return -EIO;
 389         }
 390
 391         return 0;
 392 }
 393
 394 static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu)
 395 {
 396         struct nvme_tcp_hdr *hdr = pdu;
 397         u8 digest_len = nvme_tcp_hdgst_len(queue);
 398         u32 len;
 399
 400         len = le32_to_cpu(hdr->plen) - hdr->hlen -
 401                 ((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0);
 402
 403         if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
 404                 dev_err(queue->ctrl->ctrl.device,
 405                         "queue %d: data digest flag is cleared\n",
 406                 nvme_tcp_queue_id(queue));
 407                 return -EPROTO;
 408         }
 409         crypto_ahash_init(queue->rcv_hash);
 410
 411         return 0;
 412 }
 413
 414 static void nvme_tcp_exit_request(struct blk_mq_tag_set *set,
 415                 struct request *rq, unsigned int hctx_idx)
 416 {
 417         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 418
 419         page_frag_free(req->pdu);
 420 }
 421
 422 static int nvme_tcp_init_request(struct blk_mq_tag_set *set,
 423                 struct request *rq, unsigned int hctx_idx,
 424                 unsigned int numa_node)
 425 {
 426         struct nvme_tcp_ctrl *ctrl = set->driver_data;
 427         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 428         struct nvme_tcp_cmd_pdu *pdu;
 429         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 430         struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx];
 431         u8 hdgst = nvme_tcp_hdgst_len(queue);
 432
 433         req->pdu = page_frag_alloc(&queue->pf_cache,
 434                 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
 435                 GFP_KERNEL | __GFP_ZERO);
 436         if (!req->pdu)
 437                 return -ENOMEM;
 438
 439         pdu = req->pdu;
 440         req->queue = queue;
 441         nvme_req(rq)->ctrl = &ctrl->ctrl;
 442         nvme_req(rq)->cmd = &pdu->cmd;
 443
 444         return 0;
 445 }
 446
 447 static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 448                 unsigned int hctx_idx)
 449 {
 450         struct nvme_tcp_ctrl *ctrl = data;
 451         struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1];
 452
 453         hctx->driver_data = queue;
 454         return 0;
 455 }
 456
 457 static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 458                 unsigned int hctx_idx)
 459 {
 460         struct nvme_tcp_ctrl *ctrl = data;
 461         struct nvme_tcp_queue *queue = &ctrl->queues[0];
 462
 463         hctx->driver_data = queue;
 464         return 0;
 465 }
 466
 467 static enum nvme_tcp_recv_state
 468 nvme_tcp_recv_state(struct nvme_tcp_queue *queue)
 469 {
 470         return  (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
 471                 (queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST :
 472                 NVME_TCP_RECV_DATA;
 473 }
 474
 475 static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue)
 476 {
 477         queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) +
 478                                 nvme_tcp_hdgst_len(queue);
 479         queue->pdu_offset = 0;
 480         queue->data_remaining = -1;
 481         queue->ddgst_remaining = 0;
 482 }
 483
 484 static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl)
 485 {
 486         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
 487                 return;
 488
 489         dev_warn(ctrl->device, "starting error recovery\n");
 490         queue_work(nvme_reset_wq, &to_tcp_ctrl(ctrl)->err_work);
 491 }
 492
 493 static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
 494                 struct nvme_completion *cqe)
 495 {
 496         struct nvme_tcp_request *req;
 497         struct request *rq;
 498
 499         rq = nvme_find_rq(nvme_tcp_tagset(queue), cqe->command_id);
 500         if (!rq) {
 501                 dev_err(queue->ctrl->ctrl.device,
 502                         "got bad cqe.command_id %#x on queue %d\n",
 503                         cqe->command_id, nvme_tcp_queue_id(queue));
 504                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 505                 return -EINVAL;
 506         }
 507
 508         req = blk_mq_rq_to_pdu(rq);
 509         if (req->status == cpu_to_le16(NVME_SC_SUCCESS))
 510                 req->status = cqe->status;
 511
 512         if (!nvme_try_complete_req(rq, req->status, cqe->result))
 513                 nvme_complete_rq(rq);
 514         queue->nr_cqe++;
 515
 516         return 0;
 517 }
 518
 519 static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue,
 520                 struct nvme_tcp_data_pdu *pdu)
 521 {
 522         struct request *rq;
 523
 524         rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
 525         if (!rq) {
 526                 dev_err(queue->ctrl->ctrl.device,
 527                         "got bad c2hdata.command_id %#x on queue %d\n",
 528                         pdu->command_id, nvme_tcp_queue_id(queue));
 529                 return -ENOENT;
 530         }
 531
 532         if (!blk_rq_payload_bytes(rq)) {
 533                 dev_err(queue->ctrl->ctrl.device,
 534                         "queue %d tag %#x unexpected data\n",
 535                         nvme_tcp_queue_id(queue), rq->tag);
 536                 return -EIO;
 537         }
 538
 539         queue->data_remaining = le32_to_cpu(pdu->data_length);
 540
 541         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS &&
 542             unlikely(!(pdu->hdr.flags & NVME_TCP_F_DATA_LAST))) {
 543                 dev_err(queue->ctrl->ctrl.device,
 544                         "queue %d tag %#x SUCCESS set but not last PDU\n",
 545                         nvme_tcp_queue_id(queue), rq->tag);
 546                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 547                 return -EPROTO;
 548         }
 549
 550         return 0;
 551 }
 552
 553 static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue,
 554                 struct nvme_tcp_rsp_pdu *pdu)
 555 {
 556         struct nvme_completion *cqe = &pdu->cqe;
 557         int ret = 0;
 558
 559         /*
 560          * AEN requests are special as they don't time out and can
 561          * survive any kind of queue freeze and often don't respond to
 562          * aborts.  We don't even bother to allocate a struct request
 563          * for them but rather special case them here.
 564          */
 565         if (unlikely(nvme_is_aen_req(nvme_tcp_queue_id(queue),
 566                                      cqe->command_id)))
 567                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
 568                                 &cqe->result);
 569         else
 570                 ret = nvme_tcp_process_nvme_cqe(queue, cqe);
 571
 572         return ret;
 573 }
 574
 575 static void nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req,
 576                 struct nvme_tcp_r2t_pdu *pdu)
 577 {
 578         struct nvme_tcp_data_pdu *data = req->pdu;
 579         struct nvme_tcp_queue *queue = req->queue;
 580         struct request *rq = blk_mq_rq_from_pdu(req);
 581         u8 hdgst = nvme_tcp_hdgst_len(queue);
 582         u8 ddgst = nvme_tcp_ddgst_len(queue);
 583
 584         req->state = NVME_TCP_SEND_H2C_PDU;
 585         req->offset = 0;
 586         req->pdu_len = le32_to_cpu(pdu->r2t_length);
 587         req->pdu_sent = 0;
 588
 589         memset(data, 0, sizeof(*data));
 590         data->hdr.type = nvme_tcp_h2c_data;
 591         data->hdr.flags = NVME_TCP_F_DATA_LAST;
 592         if (queue->hdr_digest)
 593                 data->hdr.flags |= NVME_TCP_F_HDGST;
 594         if (queue->data_digest)
 595                 data->hdr.flags |= NVME_TCP_F_DDGST;
 596         data->hdr.hlen = sizeof(*data);
 597         data->hdr.pdo = data->hdr.hlen + hdgst;
 598         data->hdr.plen =
 599                 cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst);
 600         data->ttag = pdu->ttag;
 601         data->command_id = nvme_cid(rq);
 602         data->data_offset = pdu->r2t_offset;
 603         data->data_length = cpu_to_le32(req->pdu_len);
 604 }
 605
 606 static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue,
 607                 struct nvme_tcp_r2t_pdu *pdu)
 608 {
 609         struct nvme_tcp_request *req;
 610         struct request *rq;
 611         u32 r2t_length = le32_to_cpu(pdu->r2t_length);
 612
 613         rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
 614         if (!rq) {
 615                 dev_err(queue->ctrl->ctrl.device,
 616                         "got bad r2t.command_id %#x on queue %d\n",
 617                         pdu->command_id, nvme_tcp_queue_id(queue));
 618                 return -ENOENT;
 619         }
 620         req = blk_mq_rq_to_pdu(rq);
 621
 622         if (unlikely(!r2t_length)) {
 623                 dev_err(queue->ctrl->ctrl.device,
 624                         "req %d r2t len is %u, probably a bug...\n",
 625                         rq->tag, r2t_length);
 626                 return -EPROTO;
 627         }
 628
 629         if (unlikely(req->data_sent + r2t_length > req->data_len)) {
 630                 dev_err(queue->ctrl->ctrl.device,
 631                         "req %d r2t len %u exceeded data len %u (%zu sent)\n",
 632                         rq->tag, r2t_length, req->data_len, req->data_sent);
 633                 return -EPROTO;
 634         }
 635
 636         if (unlikely(le32_to_cpu(pdu->r2t_offset) < req->data_sent)) {
 637                 dev_err(queue->ctrl->ctrl.device,
 638                         "req %d unexpected r2t offset %u (expected %zu)\n",
 639                         rq->tag, le32_to_cpu(pdu->r2t_offset), req->data_sent);
 640                 return -EPROTO;
 641         }
 642
 643         nvme_tcp_setup_h2c_data_pdu(req, pdu);
 644         nvme_tcp_queue_request(req, false, true);
 645
 646         return 0;
 647 }
 648
 649 static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 650                 unsigned int *offset, size_t *len)
 651 {
 652         struct nvme_tcp_hdr *hdr;
 653         char *pdu = queue->pdu;
 654         size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
 655         int ret;
 656
 657         ret = skb_copy_bits(skb, *offset,
 658                 &pdu[queue->pdu_offset], rcv_len);
 659         if (unlikely(ret))
 660                 return ret;
 661
 662         queue->pdu_remaining -= rcv_len;
 663         queue->pdu_offset += rcv_len;
 664         *offset += rcv_len;
 665         *len -= rcv_len;
 666         if (queue->pdu_remaining)
 667                 return 0;
 668
 669         hdr = queue->pdu;
 670         if (queue->hdr_digest) {
 671                 ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen);
 672                 if (unlikely(ret))
 673                         return ret;
 674         }
 675
 676
 677         if (queue->data_digest) {
 678                 ret = nvme_tcp_check_ddgst(queue, queue->pdu);
 679                 if (unlikely(ret))
 680                         return ret;
 681         }
 682
 683         switch (hdr->type) {
 684         case nvme_tcp_c2h_data:
 685                 return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
 686         case nvme_tcp_rsp:
 687                 nvme_tcp_init_recv_ctx(queue);
 688                 return nvme_tcp_handle_comp(queue, (void *)queue->pdu);
 689         case nvme_tcp_r2t:
 690                 nvme_tcp_init_recv_ctx(queue);
 691                 return nvme_tcp_handle_r2t(queue, (void *)queue->pdu);
 692         default:
 693                 dev_err(queue->ctrl->ctrl.device,
 694                         "unsupported pdu type (%d)\n", hdr->type);
 695                 return -EINVAL;
 696         }
 697 }
 698
 699 static inline void nvme_tcp_end_request(struct request *rq, u16 status)
 700 {
 701         union nvme_result res = {};
 702
 703         if (!nvme_try_complete_req(rq, cpu_to_le16(status << 1), res))
 704                 nvme_complete_rq(rq);
 705 }
 706
 707 static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 708                               unsigned int *offset, size_t *len)
 709 {
 710         struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
 711         struct request *rq =
 712                 nvme_cid_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
 713         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 714
 715         while (true) {
 716                 int recv_len, ret;
 717
 718                 recv_len = min_t(size_t, *len, queue->data_remaining);
 719                 if (!recv_len)
 720                         break;
 721
 722                 if (!iov_iter_count(&req->iter)) {
 723                         req->curr_bio = req->curr_bio->bi_next;
 724
 725                         /*
 726                          * If we don`t have any bios it means that controller
 727                          * sent more data than we requested, hence error
 728                          */
 729                         if (!req->curr_bio) {
 730                                 dev_err(queue->ctrl->ctrl.device,
 731                                         "queue %d no space in request %#x",
 732                                         nvme_tcp_queue_id(queue), rq->tag);
 733                                 nvme_tcp_init_recv_ctx(queue);
 734                                 return -EIO;
 735                         }
 736                         nvme_tcp_init_iter(req, READ);
 737                 }
 738
 739                 /* we can read only from what is left in this bio */
 740                 recv_len = min_t(size_t, recv_len,
 741                                 iov_iter_count(&req->iter));
 742
 743                 if (queue->data_digest)
 744                         ret = skb_copy_and_hash_datagram_iter(skb, *offset,
 745                                 &req->iter, recv_len, queue->rcv_hash);
 746                 else
 747                         ret = skb_copy_datagram_iter(skb, *offset,
 748                                         &req->iter, recv_len);
 749                 if (ret) {
 750                         dev_err(queue->ctrl->ctrl.device,
 751                                 "queue %d failed to copy request %#x data",
 752                                 nvme_tcp_queue_id(queue), rq->tag);
 753                         return ret;
 754                 }
 755
 756                 *len -= recv_len;
 757                 *offset += recv_len;
 758                 queue->data_remaining -= recv_len;
 759         }
 760
 761         if (!queue->data_remaining) {
 762                 if (queue->data_digest) {
 763                         nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst);
 764                         queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
 765                 } else {
 766                         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
 767                                 nvme_tcp_end_request(rq,
 768                                                 le16_to_cpu(req->status));
 769                                 queue->nr_cqe++;
 770                         }
 771                         nvme_tcp_init_recv_ctx(queue);
 772                 }
 773         }
 774
 775         return 0;
 776 }
 777
 778 static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
 779                 struct sk_buff *skb, unsigned int *offset, size_t *len)
 780 {
 781         struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
 782         char *ddgst = (char *)&queue->recv_ddgst;
 783         size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
 784         off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
 785         int ret;
 786
 787         ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
 788         if (unlikely(ret))
 789                 return ret;
 790
 791         queue->ddgst_remaining -= recv_len;
 792         *offset += recv_len;
 793         *len -= recv_len;
 794         if (queue->ddgst_remaining)
 795                 return 0;
 796
 797         if (queue->recv_ddgst != queue->exp_ddgst) {
 798                 struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
 799                                         pdu->command_id);
 800                 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 801
 802                 req->status = cpu_to_le16(NVME_SC_DATA_XFER_ERROR);
 803
 804                 dev_err(queue->ctrl->ctrl.device,
 805                         "data digest error: recv %#x expected %#x\n",
 806                         le32_to_cpu(queue->recv_ddgst),
 807                         le32_to_cpu(queue->exp_ddgst));
 808         }
 809
 810         if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
 811                 struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
 812                                         pdu->command_id);
 813                 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 814
 815                 nvme_tcp_end_request(rq, le16_to_cpu(req->status));
 816                 queue->nr_cqe++;
 817         }
 818
 819         nvme_tcp_init_recv_ctx(queue);
 820         return 0;
 821 }
 822
 823 static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
 824                              unsigned int offset, size_t len)
 825 {
 826         struct nvme_tcp_queue *queue = desc->arg.data;
 827         size_t consumed = len;
 828         int result;
 829
 830         while (len) {
 831                 switch (nvme_tcp_recv_state(queue)) {
 832                 case NVME_TCP_RECV_PDU:
 833                         result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
 834                         break;
 835                 case NVME_TCP_RECV_DATA:
 836                         result = nvme_tcp_recv_data(queue, skb, &offset, &len);
 837                         break;
 838                 case NVME_TCP_RECV_DDGST:
 839                         result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
 840                         break;
 841                 default:
 842                         result = -EFAULT;
 843                 }
 844                 if (result) {
 845                         dev_err(queue->ctrl->ctrl.device,
 846                                 "receive failed:  %d\n", result);
 847                         queue->rd_enabled = false;
 848                         nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 849                         return result;
 850                 }
 851         }
 852
 853         return consumed;
 854 }
 855
 856 static void nvme_tcp_data_ready(struct sock *sk)
 857 {
 858         struct nvme_tcp_queue *queue;
 859
 860         read_lock_bh(&sk->sk_callback_lock);
 861         queue = sk->sk_user_data;
 862         if (likely(queue && queue->rd_enabled) &&
 863             !test_bit(NVME_TCP_Q_POLLING, &queue->flags))
 864                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 865         read_unlock_bh(&sk->sk_callback_lock);
 866 }
 867
 868 static void nvme_tcp_write_space(struct sock *sk)
 869 {
 870         struct nvme_tcp_queue *queue;
 871
 872         read_lock_bh(&sk->sk_callback_lock);
 873         queue = sk->sk_user_data;
 874         if (likely(queue && sk_stream_is_writeable(sk))) {
 875                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 876                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 877         }
 878         read_unlock_bh(&sk->sk_callback_lock);
 879 }
 880
 881 static void nvme_tcp_state_change(struct sock *sk)
 882 {
 883         struct nvme_tcp_queue *queue;
 884
 885         read_lock_bh(&sk->sk_callback_lock);
 886         queue = sk->sk_user_data;
 887         if (!queue)
 888                 goto done;
 889
 890         switch (sk->sk_state) {
 891         case TCP_CLOSE:
 892         case TCP_CLOSE_WAIT:
 893         case TCP_LAST_ACK:
 894         case TCP_FIN_WAIT1:
 895         case TCP_FIN_WAIT2:
 896                 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 897                 break;
 898         default:
 899                 dev_info(queue->ctrl->ctrl.device,
 900                         "queue %d socket state %d\n",
 901                         nvme_tcp_queue_id(queue), sk->sk_state);
 902         }
 903
 904         queue->state_change(sk);
 905 done:
 906         read_unlock_bh(&sk->sk_callback_lock);
 907 }
 908
 909 static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
 910 {
 911         queue->request = NULL;
 912 }
 913
 914 static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
 915 {
 916         nvme_tcp_end_request(blk_mq_rq_from_pdu(req), NVME_SC_HOST_PATH_ERROR);
 917 }
 918
 919 static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
 920 {
 921         struct nvme_tcp_queue *queue = req->queue;
 922         int req_data_len = req->data_len;
 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 req_data_sent = req->data_sent;
 930                 int ret, flags = MSG_DONTWAIT;
 931
 932                 if (last && !queue->data_digest && !nvme_tcp_queue_more(queue))
 933                         flags |= MSG_EOR;
 934                 else
 935                         flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
 936
 937                 if (sendpage_ok(page)) {
 938                         ret = kernel_sendpage(queue->sock, page, offset, len,
 939                                         flags);
 940                 } else {
 941                         ret = sock_no_sendpage(queue->sock, page, offset, len,
 942                                         flags);
 943                 }
 944                 if (ret <= 0)
 945                         return ret;
 946
 947                 if (queue->data_digest)
 948                         nvme_tcp_ddgst_update(queue->snd_hash, page,
 949                                         offset, ret);
 950
 951                 /*
 952                  * update the request iterator except for the last payload send
 953                  * in the request where we don't want to modify it as we may
 954                  * compete with the RX path completing the request.
 955                  */
 956                 if (req_data_sent + ret < req_data_len)
 957                         nvme_tcp_advance_req(req, ret);
 958
 959                 /* fully successful last send in current PDU */
 960                 if (last && ret == len) {
 961                         if (queue->data_digest) {
 962                                 nvme_tcp_ddgst_final(queue->snd_hash,
 963                                         &req->ddgst);
 964                                 req->state = NVME_TCP_SEND_DDGST;
 965                                 req->offset = 0;
 966                         } else {
 967                                 nvme_tcp_done_send_req(queue);
 968                         }
 969                         return 1;
 970                 }
 971         }
 972         return -EAGAIN;
 973 }
 974
 975 static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
 976 {
 977         struct nvme_tcp_queue *queue = req->queue;
 978         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
 979         bool inline_data = nvme_tcp_has_inline_data(req);
 980         u8 hdgst = nvme_tcp_hdgst_len(queue);
 981         int len = sizeof(*pdu) + hdgst - req->offset;
 982         int flags = MSG_DONTWAIT;
 983         int ret;
 984
 985         if (inline_data || nvme_tcp_queue_more(queue))
 986                 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
 987         else
 988                 flags |= MSG_EOR;
 989
 990         if (queue->hdr_digest && !req->offset)
 991                 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
 992
 993         ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
 994                         offset_in_page(pdu) + req->offset, len,  flags);
 995         if (unlikely(ret <= 0))
 996                 return ret;
 997
 998         len -= ret;
 999         if (!len) {
1000                 if (inline_data) {
1001                         req->state = NVME_TCP_SEND_DATA;
1002                         if (queue->data_digest)
1003                                 crypto_ahash_init(queue->snd_hash);
1004                 } else {
1005                         nvme_tcp_done_send_req(queue);
1006                 }
1007                 return 1;
1008         }
1009         req->offset += ret;
1010
1011         return -EAGAIN;
1012 }
1013
1014 static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
1015 {
1016         struct nvme_tcp_queue *queue = req->queue;
1017         struct nvme_tcp_data_pdu *pdu = req->pdu;
1018         u8 hdgst = nvme_tcp_hdgst_len(queue);
1019         int len = sizeof(*pdu) - req->offset + hdgst;
1020         int ret;
1021
1022         if (queue->hdr_digest && !req->offset)
1023                 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
1024
1025         ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
1026                         offset_in_page(pdu) + req->offset, len,
1027                         MSG_DONTWAIT | MSG_MORE | MSG_SENDPAGE_NOTLAST);
1028         if (unlikely(ret <= 0))
1029                 return ret;
1030
1031         len -= ret;
1032         if (!len) {
1033                 req->state = NVME_TCP_SEND_DATA;
1034                 if (queue->data_digest)
1035                         crypto_ahash_init(queue->snd_hash);
1036                 return 1;
1037         }
1038         req->offset += ret;
1039
1040         return -EAGAIN;
1041 }
1042
1043 static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
1044 {
1045         struct nvme_tcp_queue *queue = req->queue;
1046         size_t offset = req->offset;
1047         int ret;
1048         struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1049         struct kvec iov = {
1050                 .iov_base = (u8 *)&req->ddgst + req->offset,
1051                 .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
1052         };
1053
1054         if (nvme_tcp_queue_more(queue))
1055                 msg.msg_flags |= MSG_MORE;
1056         else
1057                 msg.msg_flags |= MSG_EOR;
1058
1059         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1060         if (unlikely(ret <= 0))
1061                 return ret;
1062
1063         if (offset + ret == NVME_TCP_DIGEST_LENGTH) {
1064                 nvme_tcp_done_send_req(queue);
1065                 return 1;
1066         }
1067
1068         req->offset += ret;
1069         return -EAGAIN;
1070 }
1071
1072 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
1073 {
1074         struct nvme_tcp_request *req;
1075         int ret = 1;
1076
1077         if (!queue->request) {
1078                 queue->request = nvme_tcp_fetch_request(queue);
1079                 if (!queue->request)
1080                         return 0;
1081         }
1082         req = queue->request;
1083
1084         if (req->state == NVME_TCP_SEND_CMD_PDU) {
1085                 ret = nvme_tcp_try_send_cmd_pdu(req);
1086                 if (ret <= 0)
1087                         goto done;
1088                 if (!nvme_tcp_has_inline_data(req))
1089                         return ret;
1090         }
1091
1092         if (req->state == NVME_TCP_SEND_H2C_PDU) {
1093                 ret = nvme_tcp_try_send_data_pdu(req);
1094                 if (ret <= 0)
1095                         goto done;
1096         }
1097
1098         if (req->state == NVME_TCP_SEND_DATA) {
1099                 ret = nvme_tcp_try_send_data(req);
1100                 if (ret <= 0)
1101                         goto done;
1102         }
1103
1104         if (req->state == NVME_TCP_SEND_DDGST)
1105                 ret = nvme_tcp_try_send_ddgst(req);
1106 done:
1107         if (ret == -EAGAIN) {
1108                 ret = 0;
1109         } else if (ret < 0) {
1110                 dev_err(queue->ctrl->ctrl.device,
1111                         "failed to send request %d\n", ret);
1112                 if (ret != -EPIPE && ret != -ECONNRESET)
1113                         nvme_tcp_fail_request(queue->request);
1114                 nvme_tcp_done_send_req(queue);
1115         }
1116         return ret;
1117 }
1118
1119 static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1120 {
1121         struct socket *sock = queue->sock;
1122         struct sock *sk = sock->sk;
1123         read_descriptor_t rd_desc;
1124         int consumed;
1125
1126         rd_desc.arg.data = queue;
1127         rd_desc.count = 1;
1128         lock_sock(sk);
1129         queue->nr_cqe = 0;
1130         consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1131         release_sock(sk);
1132         return consumed;
1133 }
1134
1135 static void nvme_tcp_io_work(struct work_struct *w)
1136 {
1137         struct nvme_tcp_queue *queue =
1138                 container_of(w, struct nvme_tcp_queue, io_work);
1139         unsigned long deadline = jiffies + msecs_to_jiffies(1);
1140
1141         do {
1142                 bool pending = false;
1143                 int result;
1144
1145                 if (mutex_trylock(&queue->send_mutex)) {
1146                         result = nvme_tcp_try_send(queue);
1147                         mutex_unlock(&queue->send_mutex);
1148                         if (result > 0)
1149                                 pending = true;
1150                         else if (unlikely(result < 0))
1151                                 break;
1152                 }
1153
1154                 result = nvme_tcp_try_recv(queue);
1155                 if (result > 0)
1156                         pending = true;
1157                 else if (unlikely(result < 0))
1158                         return;
1159
1160                 if (!pending)
1161                         return;
1162
1163         } while (!time_after(jiffies, deadline)); /* quota is exhausted */
1164
1165         queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1166 }
1167
1168 static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue)
1169 {
1170         struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
1171
1172         ahash_request_free(queue->rcv_hash);
1173         ahash_request_free(queue->snd_hash);
1174         crypto_free_ahash(tfm);
1175 }
1176
1177 static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue)
1178 {
1179         struct crypto_ahash *tfm;
1180
1181         tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
1182         if (IS_ERR(tfm))
1183                 return PTR_ERR(tfm);
1184
1185         queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1186         if (!queue->snd_hash)
1187                 goto free_tfm;
1188         ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
1189
1190         queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1191         if (!queue->rcv_hash)
1192                 goto free_snd_hash;
1193         ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
1194
1195         return 0;
1196 free_snd_hash:
1197         ahash_request_free(queue->snd_hash);
1198 free_tfm:
1199         crypto_free_ahash(tfm);
1200         return -ENOMEM;
1201 }
1202
1203 static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1204 {
1205         struct nvme_tcp_request *async = &ctrl->async_req;
1206
1207         page_frag_free(async->pdu);
1208 }
1209
1210 static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1211 {
1212         struct nvme_tcp_queue *queue = &ctrl->queues[0];
1213         struct nvme_tcp_request *async = &ctrl->async_req;
1214         u8 hdgst = nvme_tcp_hdgst_len(queue);
1215
1216         async->pdu = page_frag_alloc(&queue->pf_cache,
1217                 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1218                 GFP_KERNEL | __GFP_ZERO);
1219         if (!async->pdu)
1220                 return -ENOMEM;
1221
1222         async->queue = &ctrl->queues[0];
1223         return 0;
1224 }
1225
1226 static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1227 {
1228         struct page *page;
1229         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1230         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1231
1232         if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1233                 return;
1234
1235         if (queue->hdr_digest || queue->data_digest)
1236                 nvme_tcp_free_crypto(queue);
1237
1238         if (queue->pf_cache.va) {
1239                 page = virt_to_head_page(queue->pf_cache.va);
1240                 __page_frag_cache_drain(page, queue->pf_cache.pagecnt_bias);
1241                 queue->pf_cache.va = NULL;
1242         }
1243         sock_release(queue->sock);
1244         kfree(queue->pdu);
1245         mutex_destroy(&queue->send_mutex);
1246         mutex_destroy(&queue->queue_lock);
1247 }
1248
1249 static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1250 {
1251         struct nvme_tcp_icreq_pdu *icreq;
1252         struct nvme_tcp_icresp_pdu *icresp;
1253         struct msghdr msg = {};
1254         struct kvec iov;
1255         bool ctrl_hdgst, ctrl_ddgst;
1256         int ret;
1257
1258         icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1259         if (!icreq)
1260                 return -ENOMEM;
1261
1262         icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1263         if (!icresp) {
1264                 ret = -ENOMEM;
1265                 goto free_icreq;
1266         }
1267
1268         icreq->hdr.type = nvme_tcp_icreq;
1269         icreq->hdr.hlen = sizeof(*icreq);
1270         icreq->hdr.pdo = 0;
1271         icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1272         icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1273         icreq->maxr2t = 0; /* single inflight r2t supported */
1274         icreq->hpda = 0; /* no alignment constraint */
1275         if (queue->hdr_digest)
1276                 icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1277         if (queue->data_digest)
1278                 icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1279
1280         iov.iov_base = icreq;
1281         iov.iov_len = sizeof(*icreq);
1282         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1283         if (ret < 0)
1284                 goto free_icresp;
1285
1286         memset(&msg, 0, sizeof(msg));
1287         iov.iov_base = icresp;
1288         iov.iov_len = sizeof(*icresp);
1289         ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1290                         iov.iov_len, msg.msg_flags);
1291         if (ret < 0)
1292                 goto free_icresp;
1293
1294         ret = -EINVAL;
1295         if (icresp->hdr.type != nvme_tcp_icresp) {
1296                 pr_err("queue %d: bad type returned %d\n",
1297                         nvme_tcp_queue_id(queue), icresp->hdr.type);
1298                 goto free_icresp;
1299         }
1300
1301         if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1302                 pr_err("queue %d: bad pdu length returned %d\n",
1303                         nvme_tcp_queue_id(queue), icresp->hdr.plen);
1304                 goto free_icresp;
1305         }
1306
1307         if (icresp->pfv != NVME_TCP_PFV_1_0) {
1308                 pr_err("queue %d: bad pfv returned %d\n",
1309                         nvme_tcp_queue_id(queue), icresp->pfv);
1310                 goto free_icresp;
1311         }
1312
1313         ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1314         if ((queue->data_digest && !ctrl_ddgst) ||
1315             (!queue->data_digest && ctrl_ddgst)) {
1316                 pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1317                         nvme_tcp_queue_id(queue),
1318                         queue->data_digest ? "enabled" : "disabled",
1319                         ctrl_ddgst ? "enabled" : "disabled");
1320                 goto free_icresp;
1321         }
1322
1323         ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1324         if ((queue->hdr_digest && !ctrl_hdgst) ||
1325             (!queue->hdr_digest && ctrl_hdgst)) {
1326                 pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1327                         nvme_tcp_queue_id(queue),
1328                         queue->hdr_digest ? "enabled" : "disabled",
1329                         ctrl_hdgst ? "enabled" : "disabled");
1330                 goto free_icresp;
1331         }
1332
1333         if (icresp->cpda != 0) {
1334                 pr_err("queue %d: unsupported cpda returned %d\n",
1335                         nvme_tcp_queue_id(queue), icresp->cpda);
1336                 goto free_icresp;
1337         }
1338
1339         ret = 0;
1340 free_icresp:
1341         kfree(icresp);
1342 free_icreq:
1343         kfree(icreq);
1344         return ret;
1345 }
1346
1347 static bool nvme_tcp_admin_queue(struct nvme_tcp_queue *queue)
1348 {
1349         return nvme_tcp_queue_id(queue) == 0;
1350 }
1351
1352 static bool nvme_tcp_default_queue(struct nvme_tcp_queue *queue)
1353 {
1354         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1355         int qid = nvme_tcp_queue_id(queue);
1356
1357         return !nvme_tcp_admin_queue(queue) &&
1358                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT];
1359 }
1360
1361 static bool nvme_tcp_read_queue(struct nvme_tcp_queue *queue)
1362 {
1363         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1364         int qid = nvme_tcp_queue_id(queue);
1365
1366         return !nvme_tcp_admin_queue(queue) &&
1367                 !nvme_tcp_default_queue(queue) &&
1368                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1369                           ctrl->io_queues[HCTX_TYPE_READ];
1370 }
1371
1372 static bool nvme_tcp_poll_queue(struct nvme_tcp_queue *queue)
1373 {
1374         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1375         int qid = nvme_tcp_queue_id(queue);
1376
1377         return !nvme_tcp_admin_queue(queue) &&
1378                 !nvme_tcp_default_queue(queue) &&
1379                 !nvme_tcp_read_queue(queue) &&
1380                 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1381                           ctrl->io_queues[HCTX_TYPE_READ] +
1382                           ctrl->io_queues[HCTX_TYPE_POLL];
1383 }
1384
1385 static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
1386 {
1387         struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1388         int qid = nvme_tcp_queue_id(queue);
1389         int n = 0;
1390
1391         if (nvme_tcp_default_queue(queue))
1392                 n = qid - 1;
1393         else if (nvme_tcp_read_queue(queue))
1394                 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] - 1;
1395         else if (nvme_tcp_poll_queue(queue))
1396                 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] -
1397                                 ctrl->io_queues[HCTX_TYPE_READ] - 1;
1398         queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
1399 }
1400
1401 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl,
1402                 int qid, size_t queue_size)
1403 {
1404         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1405         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1406         int ret, rcv_pdu_size;
1407
1408         mutex_init(&queue->queue_lock);
1409         queue->ctrl = ctrl;
1410         init_llist_head(&queue->req_list);
1411         INIT_LIST_HEAD(&queue->send_list);
1412         mutex_init(&queue->send_mutex);
1413         INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1414         queue->queue_size = queue_size;
1415
1416         if (qid > 0)
1417                 queue->cmnd_capsule_len = nctrl->ioccsz * 16;
1418         else
1419                 queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1420                                                 NVME_TCP_ADMIN_CCSZ;
1421
1422         ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM,
1423                         IPPROTO_TCP, &queue->sock);
1424         if (ret) {
1425                 dev_err(nctrl->device,
1426                         "failed to create socket: %d\n", ret);
1427                 goto err_destroy_mutex;
1428         }
1429
1430         /* Single syn retry */
1431         tcp_sock_set_syncnt(queue->sock->sk, 1);
1432
1433         /* Set TCP no delay */
1434         tcp_sock_set_nodelay(queue->sock->sk);
1435
1436         /*
1437          * Cleanup whatever is sitting in the TCP transmit queue on socket
1438          * close. This is done to prevent stale data from being sent should
1439          * the network connection be restored before TCP times out.
1440          */
1441         sock_no_linger(queue->sock->sk);
1442
1443         if (so_priority > 0)
1444                 sock_set_priority(queue->sock->sk, so_priority);
1445
1446         /* Set socket type of service */
1447         if (nctrl->opts->tos >= 0)
1448                 ip_sock_set_tos(queue->sock->sk, nctrl->opts->tos);
1449
1450         /* Set 10 seconds timeout for icresp recvmsg */
1451         queue->sock->sk->sk_rcvtimeo = 10 * HZ;
1452
1453         queue->sock->sk->sk_allocation = GFP_ATOMIC;
1454         nvme_tcp_set_queue_io_cpu(queue);
1455         queue->request = NULL;
1456         queue->data_remaining = 0;
1457         queue->ddgst_remaining = 0;
1458         queue->pdu_remaining = 0;
1459         queue->pdu_offset = 0;
1460         sk_set_memalloc(queue->sock->sk);
1461
1462         if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
1463                 ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1464                         sizeof(ctrl->src_addr));
1465                 if (ret) {
1466                         dev_err(nctrl->device,
1467                                 "failed to bind queue %d socket %d\n",
1468                                 qid, ret);
1469                         goto err_sock;
1470                 }
1471         }
1472
1473         if (nctrl->opts->mask & NVMF_OPT_HOST_IFACE) {
1474                 char *iface = nctrl->opts->host_iface;
1475                 sockptr_t optval = KERNEL_SOCKPTR(iface);
1476
1477                 ret = sock_setsockopt(queue->sock, SOL_SOCKET, SO_BINDTODEVICE,
1478                                       optval, strlen(iface));
1479                 if (ret) {
1480                         dev_err(nctrl->device,
1481                           "failed to bind to interface %s queue %d err %d\n",
1482                           iface, qid, ret);
1483                         goto err_sock;
1484                 }
1485         }
1486
1487         queue->hdr_digest = nctrl->opts->hdr_digest;
1488         queue->data_digest = nctrl->opts->data_digest;
1489         if (queue->hdr_digest || queue->data_digest) {
1490                 ret = nvme_tcp_alloc_crypto(queue);
1491                 if (ret) {
1492                         dev_err(nctrl->device,
1493                                 "failed to allocate queue %d crypto\n", qid);
1494                         goto err_sock;
1495                 }
1496         }
1497
1498         rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1499                         nvme_tcp_hdgst_len(queue);
1500         queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1501         if (!queue->pdu) {
1502                 ret = -ENOMEM;
1503                 goto err_crypto;
1504         }
1505
1506         dev_dbg(nctrl->device, "connecting queue %d\n",
1507                         nvme_tcp_queue_id(queue));
1508
1509         ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1510                 sizeof(ctrl->addr), 0);
1511         if (ret) {
1512                 dev_err(nctrl->device,
1513                         "failed to connect socket: %d\n", ret);
1514                 goto err_rcv_pdu;
1515         }
1516
1517         ret = nvme_tcp_init_connection(queue);
1518         if (ret)
1519                 goto err_init_connect;
1520
1521         queue->rd_enabled = true;
1522         set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1523         nvme_tcp_init_recv_ctx(queue);
1524
1525         write_lock_bh(&queue->sock->sk->sk_callback_lock);
1526         queue->sock->sk->sk_user_data = queue;
1527         queue->state_change = queue->sock->sk->sk_state_change;
1528         queue->data_ready = queue->sock->sk->sk_data_ready;
1529         queue->write_space = queue->sock->sk->sk_write_space;
1530         queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1531         queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1532         queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1533 #ifdef CONFIG_NET_RX_BUSY_POLL
1534         queue->sock->sk->sk_ll_usec = 1;
1535 #endif
1536         write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1537
1538         return 0;
1539
1540 err_init_connect:
1541         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1542 err_rcv_pdu:
1543         kfree(queue->pdu);
1544 err_crypto:
1545         if (queue->hdr_digest || queue->data_digest)
1546                 nvme_tcp_free_crypto(queue);
1547 err_sock:
1548         sock_release(queue->sock);
1549         queue->sock = NULL;
1550 err_destroy_mutex:
1551         mutex_destroy(&queue->send_mutex);
1552         mutex_destroy(&queue->queue_lock);
1553         return ret;
1554 }
1555
1556 static void nvme_tcp_restore_sock_calls(struct nvme_tcp_queue *queue)
1557 {
1558         struct socket *sock = queue->sock;
1559
1560         write_lock_bh(&sock->sk->sk_callback_lock);
1561         sock->sk->sk_user_data  = NULL;
1562         sock->sk->sk_data_ready = queue->data_ready;
1563         sock->sk->sk_state_change = queue->state_change;
1564         sock->sk->sk_write_space  = queue->write_space;
1565         write_unlock_bh(&sock->sk->sk_callback_lock);
1566 }
1567
1568 static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1569 {
1570         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1571         nvme_tcp_restore_sock_calls(queue);
1572         cancel_work_sync(&queue->io_work);
1573 }
1574
1575 static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1576 {
1577         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1578         struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1579
1580         mutex_lock(&queue->queue_lock);
1581         if (test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1582                 __nvme_tcp_stop_queue(queue);
1583         mutex_unlock(&queue->queue_lock);
1584 }
1585
1586 static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1587 {
1588         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1589         int ret;
1590
1591         if (idx)
1592                 ret = nvmf_connect_io_queue(nctrl, idx);
1593         else
1594                 ret = nvmf_connect_admin_queue(nctrl);
1595
1596         if (!ret) {
1597                 set_bit(NVME_TCP_Q_LIVE, &ctrl->queues[idx].flags);
1598         } else {
1599                 if (test_bit(NVME_TCP_Q_ALLOCATED, &ctrl->queues[idx].flags))
1600                         __nvme_tcp_stop_queue(&ctrl->queues[idx]);
1601                 dev_err(nctrl->device,
1602                         "failed to connect queue: %d ret=%d\n", idx, ret);
1603         }
1604         return ret;
1605 }
1606
1607 static struct blk_mq_tag_set *nvme_tcp_alloc_tagset(struct nvme_ctrl *nctrl,
1608                 bool admin)
1609 {
1610         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1611         struct blk_mq_tag_set *set;
1612         int ret;
1613
1614         if (admin) {
1615                 set = &ctrl->admin_tag_set;
1616                 memset(set, 0, sizeof(*set));
1617                 set->ops = &nvme_tcp_admin_mq_ops;
1618                 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1619                 set->reserved_tags = NVMF_RESERVED_TAGS;
1620                 set->numa_node = nctrl->numa_node;
1621                 set->flags = BLK_MQ_F_BLOCKING;
1622                 set->cmd_size = sizeof(struct nvme_tcp_request);
1623                 set->driver_data = ctrl;
1624                 set->nr_hw_queues = 1;
1625                 set->timeout = NVME_ADMIN_TIMEOUT;
1626         } else {
1627                 set = &ctrl->tag_set;
1628                 memset(set, 0, sizeof(*set));
1629                 set->ops = &nvme_tcp_mq_ops;
1630                 set->queue_depth = nctrl->sqsize + 1;
1631                 set->reserved_tags = NVMF_RESERVED_TAGS;
1632                 set->numa_node = nctrl->numa_node;
1633                 set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_BLOCKING;
1634                 set->cmd_size = sizeof(struct nvme_tcp_request);
1635                 set->driver_data = ctrl;
1636                 set->nr_hw_queues = nctrl->queue_count - 1;
1637                 set->timeout = NVME_IO_TIMEOUT;
1638                 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
1639         }
1640
1641         ret = blk_mq_alloc_tag_set(set);
1642         if (ret)
1643                 return ERR_PTR(ret);
1644
1645         return set;
1646 }
1647
1648 static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
1649 {
1650         if (to_tcp_ctrl(ctrl)->async_req.pdu) {
1651                 cancel_work_sync(&ctrl->async_event_work);
1652                 nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
1653                 to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
1654         }
1655
1656         nvme_tcp_free_queue(ctrl, 0);
1657 }
1658
1659 static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
1660 {
1661         int i;
1662
1663         for (i = 1; i < ctrl->queue_count; i++)
1664                 nvme_tcp_free_queue(ctrl, i);
1665 }
1666
1667 static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
1668 {
1669         int i;
1670
1671         for (i = 1; i < ctrl->queue_count; i++)
1672                 nvme_tcp_stop_queue(ctrl, i);
1673 }
1674
1675 static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl)
1676 {
1677         int i, ret = 0;
1678
1679         for (i = 1; i < ctrl->queue_count; i++) {
1680                 ret = nvme_tcp_start_queue(ctrl, i);
1681                 if (ret)
1682                         goto out_stop_queues;
1683         }
1684
1685         return 0;
1686
1687 out_stop_queues:
1688         for (i--; i >= 1; i--)
1689                 nvme_tcp_stop_queue(ctrl, i);
1690         return ret;
1691 }
1692
1693 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
1694 {
1695         int ret;
1696
1697         ret = nvme_tcp_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
1698         if (ret)
1699                 return ret;
1700
1701         ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
1702         if (ret)
1703                 goto out_free_queue;
1704
1705         return 0;
1706
1707 out_free_queue:
1708         nvme_tcp_free_queue(ctrl, 0);
1709         return ret;
1710 }
1711
1712 static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1713 {
1714         int i, ret;
1715
1716         for (i = 1; i < ctrl->queue_count; i++) {
1717                 ret = nvme_tcp_alloc_queue(ctrl, i,
1718                                 ctrl->sqsize + 1);
1719                 if (ret)
1720                         goto out_free_queues;
1721         }
1722
1723         return 0;
1724
1725 out_free_queues:
1726         for (i--; i >= 1; i--)
1727                 nvme_tcp_free_queue(ctrl, i);
1728
1729         return ret;
1730 }
1731
1732 static unsigned int nvme_tcp_nr_io_queues(struct nvme_ctrl *ctrl)
1733 {
1734         unsigned int nr_io_queues;
1735
1736         nr_io_queues = min(ctrl->opts->nr_io_queues, num_online_cpus());
1737         nr_io_queues += min(ctrl->opts->nr_write_queues, num_online_cpus());
1738         nr_io_queues += min(ctrl->opts->nr_poll_queues, num_online_cpus());
1739
1740         return nr_io_queues;
1741 }
1742
1743 static void nvme_tcp_set_io_queues(struct nvme_ctrl *nctrl,
1744                 unsigned int nr_io_queues)
1745 {
1746         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1747         struct nvmf_ctrl_options *opts = nctrl->opts;
1748
1749         if (opts->nr_write_queues && opts->nr_io_queues < nr_io_queues) {
1750                 /*
1751                  * separate read/write queues
1752                  * hand out dedicated default queues only after we have
1753                  * sufficient read queues.
1754                  */
1755                 ctrl->io_queues[HCTX_TYPE_READ] = opts->nr_io_queues;
1756                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
1757                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1758                         min(opts->nr_write_queues, nr_io_queues);
1759                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1760         } else {
1761                 /*
1762                  * shared read/write queues
1763                  * either no write queues were requested, or we don't have
1764                  * sufficient queue count to have dedicated default queues.
1765                  */
1766                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1767                         min(opts->nr_io_queues, nr_io_queues);
1768                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1769         }
1770
1771         if (opts->nr_poll_queues && nr_io_queues) {
1772                 /* map dedicated poll queues only if we have queues left */
1773                 ctrl->io_queues[HCTX_TYPE_POLL] =
1774                         min(opts->nr_poll_queues, nr_io_queues);
1775         }
1776 }
1777
1778 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1779 {
1780         unsigned int nr_io_queues;
1781         int ret;
1782
1783         nr_io_queues = nvme_tcp_nr_io_queues(ctrl);
1784         ret = nvme_set_queue_count(ctrl, &nr_io_queues);
1785         if (ret)
1786                 return ret;
1787
1788         if (nr_io_queues == 0) {
1789                 dev_err(ctrl->device,
1790                         "unable to set any I/O queues\n");
1791                 return -ENOMEM;
1792         }
1793
1794         ctrl->queue_count = nr_io_queues + 1;
1795         dev_info(ctrl->device,
1796                 "creating %d I/O queues.\n", nr_io_queues);
1797
1798         nvme_tcp_set_io_queues(ctrl, nr_io_queues);
1799
1800         return __nvme_tcp_alloc_io_queues(ctrl);
1801 }
1802
1803 static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove)
1804 {
1805         nvme_tcp_stop_io_queues(ctrl);
1806         if (remove) {
1807                 blk_cleanup_queue(ctrl->connect_q);
1808                 blk_mq_free_tag_set(ctrl->tagset);
1809         }
1810         nvme_tcp_free_io_queues(ctrl);
1811 }
1812
1813 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
1814 {
1815         int ret;
1816
1817         ret = nvme_tcp_alloc_io_queues(ctrl);
1818         if (ret)
1819                 return ret;
1820
1821         if (new) {
1822                 ctrl->tagset = nvme_tcp_alloc_tagset(ctrl, false);
1823                 if (IS_ERR(ctrl->tagset)) {
1824                         ret = PTR_ERR(ctrl->tagset);
1825                         goto out_free_io_queues;
1826                 }
1827
1828                 ctrl->connect_q = blk_mq_init_queue(ctrl->tagset);
1829                 if (IS_ERR(ctrl->connect_q)) {
1830                         ret = PTR_ERR(ctrl->connect_q);
1831                         goto out_free_tag_set;
1832                 }
1833         }
1834
1835         ret = nvme_tcp_start_io_queues(ctrl);
1836         if (ret)
1837                 goto out_cleanup_connect_q;
1838
1839         if (!new) {
1840                 nvme_start_queues(ctrl);
1841                 if (!nvme_wait_freeze_timeout(ctrl, NVME_IO_TIMEOUT)) {
1842                         /*
1843                          * If we timed out waiting for freeze we are likely to
1844                          * be stuck.  Fail the controller initialization just
1845                          * to be safe.
1846                          */
1847                         ret = -ENODEV;
1848                         goto out_wait_freeze_timed_out;
1849                 }
1850                 blk_mq_update_nr_hw_queues(ctrl->tagset,
1851                         ctrl->queue_count - 1);
1852                 nvme_unfreeze(ctrl);
1853         }
1854
1855         return 0;
1856
1857 out_wait_freeze_timed_out:
1858         nvme_stop_queues(ctrl);
1859         nvme_sync_io_queues(ctrl);
1860         nvme_tcp_stop_io_queues(ctrl);
1861 out_cleanup_connect_q:
1862         nvme_cancel_tagset(ctrl);
1863         if (new)
1864                 blk_cleanup_queue(ctrl->connect_q);
1865 out_free_tag_set:
1866         if (new)
1867                 blk_mq_free_tag_set(ctrl->tagset);
1868 out_free_io_queues:
1869         nvme_tcp_free_io_queues(ctrl);
1870         return ret;
1871 }
1872
1873 static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove)
1874 {
1875         nvme_tcp_stop_queue(ctrl, 0);
1876         if (remove) {
1877                 blk_cleanup_queue(ctrl->admin_q);
1878                 blk_cleanup_queue(ctrl->fabrics_q);
1879                 blk_mq_free_tag_set(ctrl->admin_tagset);
1880         }
1881         nvme_tcp_free_admin_queue(ctrl);
1882 }
1883
1884 static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
1885 {
1886         int error;
1887
1888         error = nvme_tcp_alloc_admin_queue(ctrl);
1889         if (error)
1890                 return error;
1891
1892         if (new) {
1893                 ctrl->admin_tagset = nvme_tcp_alloc_tagset(ctrl, true);
1894                 if (IS_ERR(ctrl->admin_tagset)) {
1895                         error = PTR_ERR(ctrl->admin_tagset);
1896                         goto out_free_queue;
1897                 }
1898
1899                 ctrl->fabrics_q = blk_mq_init_queue(ctrl->admin_tagset);
1900                 if (IS_ERR(ctrl->fabrics_q)) {
1901                         error = PTR_ERR(ctrl->fabrics_q);
1902                         goto out_free_tagset;
1903                 }
1904
1905                 ctrl->admin_q = blk_mq_init_queue(ctrl->admin_tagset);
1906                 if (IS_ERR(ctrl->admin_q)) {
1907                         error = PTR_ERR(ctrl->admin_q);
1908                         goto out_cleanup_fabrics_q;
1909                 }
1910         }
1911
1912         error = nvme_tcp_start_queue(ctrl, 0);
1913         if (error)
1914                 goto out_cleanup_queue;
1915
1916         error = nvme_enable_ctrl(ctrl);
1917         if (error)
1918                 goto out_stop_queue;
1919
1920         nvme_start_admin_queue(ctrl);
1921
1922         error = nvme_init_ctrl_finish(ctrl);
1923         if (error)
1924                 goto out_quiesce_queue;
1925
1926         return 0;
1927
1928 out_quiesce_queue:
1929         nvme_stop_admin_queue(ctrl);
1930         blk_sync_queue(ctrl->admin_q);
1931 out_stop_queue:
1932         nvme_tcp_stop_queue(ctrl, 0);
1933         nvme_cancel_admin_tagset(ctrl);
1934 out_cleanup_queue:
1935         if (new)
1936                 blk_cleanup_queue(ctrl->admin_q);
1937 out_cleanup_fabrics_q:
1938         if (new)
1939                 blk_cleanup_queue(ctrl->fabrics_q);
1940 out_free_tagset:
1941         if (new)
1942                 blk_mq_free_tag_set(ctrl->admin_tagset);
1943 out_free_queue:
1944         nvme_tcp_free_admin_queue(ctrl);
1945         return error;
1946 }
1947
1948 static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
1949                 bool remove)
1950 {
1951         nvme_stop_admin_queue(ctrl);
1952         blk_sync_queue(ctrl->admin_q);
1953         nvme_tcp_stop_queue(ctrl, 0);
1954         nvme_cancel_admin_tagset(ctrl);
1955         if (remove)
1956                 nvme_start_admin_queue(ctrl);
1957         nvme_tcp_destroy_admin_queue(ctrl, remove);
1958 }
1959
1960 static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
1961                 bool remove)
1962 {
1963         if (ctrl->queue_count <= 1)
1964                 return;
1965         nvme_stop_admin_queue(ctrl);
1966         nvme_start_freeze(ctrl);
1967         nvme_stop_queues(ctrl);
1968         nvme_sync_io_queues(ctrl);
1969         nvme_tcp_stop_io_queues(ctrl);
1970         nvme_cancel_tagset(ctrl);
1971         if (remove)
1972                 nvme_start_queues(ctrl);
1973         nvme_tcp_destroy_io_queues(ctrl, remove);
1974 }
1975
1976 static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
1977 {
1978         /* If we are resetting/deleting then do nothing */
1979         if (ctrl->state != NVME_CTRL_CONNECTING) {
1980                 WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW ||
1981                         ctrl->state == NVME_CTRL_LIVE);
1982                 return;
1983         }
1984
1985         if (nvmf_should_reconnect(ctrl)) {
1986                 dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
1987                         ctrl->opts->reconnect_delay);
1988                 queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
1989                                 ctrl->opts->reconnect_delay * HZ);
1990         } else {
1991                 dev_info(ctrl->device, "Removing controller...\n");
1992                 nvme_delete_ctrl(ctrl);
1993         }
1994 }
1995
1996 static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
1997 {
1998         struct nvmf_ctrl_options *opts = ctrl->opts;
1999         int ret;
2000
2001         ret = nvme_tcp_configure_admin_queue(ctrl, new);
2002         if (ret)
2003                 return ret;
2004
2005         if (ctrl->icdoff) {
2006                 ret = -EOPNOTSUPP;
2007                 dev_err(ctrl->device, "icdoff is not supported!\n");
2008                 goto destroy_admin;
2009         }
2010
2011         if (!nvme_ctrl_sgl_supported(ctrl)) {
2012                 ret = -EOPNOTSUPP;
2013                 dev_err(ctrl->device, "Mandatory sgls are not supported!\n");
2014                 goto destroy_admin;
2015         }
2016
2017         if (opts->queue_size > ctrl->sqsize + 1)
2018                 dev_warn(ctrl->device,
2019                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
2020                         opts->queue_size, ctrl->sqsize + 1);
2021
2022         if (ctrl->sqsize + 1 > ctrl->maxcmd) {
2023                 dev_warn(ctrl->device,
2024                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
2025                         ctrl->sqsize + 1, ctrl->maxcmd);
2026                 ctrl->sqsize = ctrl->maxcmd - 1;
2027         }
2028
2029         if (ctrl->queue_count > 1) {
2030                 ret = nvme_tcp_configure_io_queues(ctrl, new);
2031                 if (ret)
2032                         goto destroy_admin;
2033         }
2034
2035         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
2036                 /*
2037                  * state change failure is ok if we started ctrl delete,
2038                  * unless we're during creation of a new controller to
2039                  * avoid races with teardown flow.
2040                  */
2041                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2042                              ctrl->state != NVME_CTRL_DELETING_NOIO);
2043                 WARN_ON_ONCE(new);
2044                 ret = -EINVAL;
2045                 goto destroy_io;
2046         }
2047
2048         nvme_start_ctrl(ctrl);
2049         return 0;
2050
2051 destroy_io:
2052         if (ctrl->queue_count > 1) {
2053                 nvme_stop_queues(ctrl);
2054                 nvme_sync_io_queues(ctrl);
2055                 nvme_tcp_stop_io_queues(ctrl);
2056                 nvme_cancel_tagset(ctrl);
2057                 nvme_tcp_destroy_io_queues(ctrl, new);
2058         }
2059 destroy_admin:
2060         nvme_stop_admin_queue(ctrl);
2061         blk_sync_queue(ctrl->admin_q);
2062         nvme_tcp_stop_queue(ctrl, 0);
2063         nvme_cancel_admin_tagset(ctrl);
2064         nvme_tcp_destroy_admin_queue(ctrl, new);
2065         return ret;
2066 }
2067
2068 static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
2069 {
2070         struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
2071                         struct nvme_tcp_ctrl, connect_work);
2072         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2073
2074         ++ctrl->nr_reconnects;
2075
2076         if (nvme_tcp_setup_ctrl(ctrl, false))
2077                 goto requeue;
2078
2079         dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n",
2080                         ctrl->nr_reconnects);
2081
2082         ctrl->nr_reconnects = 0;
2083
2084         return;
2085
2086 requeue:
2087         dev_info(ctrl->device, "Failed reconnect attempt %d\n",
2088                         ctrl->nr_reconnects);
2089         nvme_tcp_reconnect_or_remove(ctrl);
2090 }
2091
2092 static void nvme_tcp_error_recovery_work(struct work_struct *work)
2093 {
2094         struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
2095                                 struct nvme_tcp_ctrl, err_work);
2096         struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2097
2098         nvme_stop_keep_alive(ctrl);
2099         nvme_tcp_teardown_io_queues(ctrl, false);
2100         /* unquiesce to fail fast pending requests */
2101         nvme_start_queues(ctrl);
2102         nvme_tcp_teardown_admin_queue(ctrl, false);
2103         nvme_start_admin_queue(ctrl);
2104
2105         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2106                 /* state change failure is ok if we started ctrl delete */
2107                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2108                              ctrl->state != NVME_CTRL_DELETING_NOIO);
2109                 return;
2110         }
2111
2112         nvme_tcp_reconnect_or_remove(ctrl);
2113 }
2114
2115 static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
2116 {
2117         cancel_work_sync(&to_tcp_ctrl(ctrl)->err_work);
2118         cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
2119
2120         nvme_tcp_teardown_io_queues(ctrl, shutdown);
2121         nvme_stop_admin_queue(ctrl);
2122         if (shutdown)
2123                 nvme_shutdown_ctrl(ctrl);
2124         else
2125                 nvme_disable_ctrl(ctrl);
2126         nvme_tcp_teardown_admin_queue(ctrl, shutdown);
2127 }
2128
2129 static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
2130 {
2131         nvme_tcp_teardown_ctrl(ctrl, true);
2132 }
2133
2134 static void nvme_reset_ctrl_work(struct work_struct *work)
2135 {
2136         struct nvme_ctrl *ctrl =
2137                 container_of(work, struct nvme_ctrl, reset_work);
2138
2139         nvme_stop_ctrl(ctrl);
2140         nvme_tcp_teardown_ctrl(ctrl, false);
2141
2142         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2143                 /* state change failure is ok if we started ctrl delete */
2144                 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2145                              ctrl->state != NVME_CTRL_DELETING_NOIO);
2146                 return;
2147         }
2148
2149         if (nvme_tcp_setup_ctrl(ctrl, false))
2150                 goto out_fail;
2151
2152         return;
2153
2154 out_fail:
2155         ++ctrl->nr_reconnects;
2156         nvme_tcp_reconnect_or_remove(ctrl);
2157 }
2158
2159 static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
2160 {
2161         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
2162
2163         if (list_empty(&ctrl->list))
2164                 goto free_ctrl;
2165
2166         mutex_lock(&nvme_tcp_ctrl_mutex);
2167         list_del(&ctrl->list);
2168         mutex_unlock(&nvme_tcp_ctrl_mutex);
2169
2170         nvmf_free_options(nctrl->opts);
2171 free_ctrl:
2172         kfree(ctrl->queues);
2173         kfree(ctrl);
2174 }
2175
2176 static void nvme_tcp_set_sg_null(struct nvme_command *c)
2177 {
2178         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2179
2180         sg->addr = 0;
2181         sg->length = 0;
2182         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2183                         NVME_SGL_FMT_TRANSPORT_A;
2184 }
2185
2186 static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
2187                 struct nvme_command *c, u32 data_len)
2188 {
2189         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2190
2191         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
2192         sg->length = cpu_to_le32(data_len);
2193         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
2194 }
2195
2196 static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
2197                 u32 data_len)
2198 {
2199         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2200
2201         sg->addr = 0;
2202         sg->length = cpu_to_le32(data_len);
2203         sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2204                         NVME_SGL_FMT_TRANSPORT_A;
2205 }
2206
2207 static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
2208 {
2209         struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2210         struct nvme_tcp_queue *queue = &ctrl->queues[0];
2211         struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2212         struct nvme_command *cmd = &pdu->cmd;
2213         u8 hdgst = nvme_tcp_hdgst_len(queue);
2214
2215         memset(pdu, 0, sizeof(*pdu));
2216         pdu->hdr.type = nvme_tcp_cmd;
2217         if (queue->hdr_digest)
2218                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2219         pdu->hdr.hlen = sizeof(*pdu);
2220         pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2221
2222         cmd->common.opcode = nvme_admin_async_event;
2223         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2224         cmd->common.flags |= NVME_CMD_SGL_METABUF;
2225         nvme_tcp_set_sg_null(cmd);
2226
2227         ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2228         ctrl->async_req.offset = 0;
2229         ctrl->async_req.curr_bio = NULL;
2230         ctrl->async_req.data_len = 0;
2231
2232         nvme_tcp_queue_request(&ctrl->async_req, true, true);
2233 }
2234
2235 static void nvme_tcp_complete_timed_out(struct request *rq)
2236 {
2237         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2238         struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2239
2240         nvme_tcp_stop_queue(ctrl, nvme_tcp_queue_id(req->queue));
2241         if (blk_mq_request_started(rq) && !blk_mq_request_completed(rq)) {
2242                 nvme_req(rq)->status = NVME_SC_HOST_ABORTED_CMD;
2243                 blk_mq_complete_request(rq);
2244         }
2245 }
2246
2247 static enum blk_eh_timer_return
2248 nvme_tcp_timeout(struct request *rq, bool reserved)
2249 {
2250         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2251         struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2252         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2253
2254         dev_warn(ctrl->device,
2255                 "queue %d: timeout request %#x type %d\n",
2256                 nvme_tcp_queue_id(req->queue), rq->tag, pdu->hdr.type);
2257
2258         if (ctrl->state != NVME_CTRL_LIVE) {
2259                 /*
2260                  * If we are resetting, connecting or deleting we should
2261                  * complete immediately because we may block controller
2262                  * teardown or setup sequence
2263                  * - ctrl disable/shutdown fabrics requests
2264                  * - connect requests
2265                  * - initialization admin requests
2266                  * - I/O requests that entered after unquiescing and
2267                  *   the controller stopped responding
2268                  *
2269                  * All other requests should be cancelled by the error
2270                  * recovery work, so it's fine that we fail it here.
2271                  */
2272                 nvme_tcp_complete_timed_out(rq);
2273                 return BLK_EH_DONE;
2274         }
2275
2276         /*
2277          * LIVE state should trigger the normal error recovery which will
2278          * handle completing this request.
2279          */
2280         nvme_tcp_error_recovery(ctrl);
2281         return BLK_EH_RESET_TIMER;
2282 }
2283
2284 static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2285                         struct request *rq)
2286 {
2287         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2288         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2289         struct nvme_command *c = &pdu->cmd;
2290
2291         c->common.flags |= NVME_CMD_SGL_METABUF;
2292
2293         if (!blk_rq_nr_phys_segments(rq))
2294                 nvme_tcp_set_sg_null(c);
2295         else if (rq_data_dir(rq) == WRITE &&
2296             req->data_len <= nvme_tcp_inline_data_size(queue))
2297                 nvme_tcp_set_sg_inline(queue, c, req->data_len);
2298         else
2299                 nvme_tcp_set_sg_host_data(c, req->data_len);
2300
2301         return 0;
2302 }
2303
2304 static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2305                 struct request *rq)
2306 {
2307         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2308         struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2309         struct nvme_tcp_queue *queue = req->queue;
2310         u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2311         blk_status_t ret;
2312
2313         ret = nvme_setup_cmd(ns, rq);
2314         if (ret)
2315                 return ret;
2316
2317         req->state = NVME_TCP_SEND_CMD_PDU;
2318         req->status = cpu_to_le16(NVME_SC_SUCCESS);
2319         req->offset = 0;
2320         req->data_sent = 0;
2321         req->pdu_len = 0;
2322         req->pdu_sent = 0;
2323         req->data_len = blk_rq_nr_phys_segments(rq) ?
2324                                 blk_rq_payload_bytes(rq) : 0;
2325         req->curr_bio = rq->bio;
2326         if (req->curr_bio && req->data_len)
2327                 nvme_tcp_init_iter(req, rq_data_dir(rq));
2328
2329         if (rq_data_dir(rq) == WRITE &&
2330             req->data_len <= nvme_tcp_inline_data_size(queue))
2331                 req->pdu_len = req->data_len;
2332
2333         pdu->hdr.type = nvme_tcp_cmd;
2334         pdu->hdr.flags = 0;
2335         if (queue->hdr_digest)
2336                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2337         if (queue->data_digest && req->pdu_len) {
2338                 pdu->hdr.flags |= NVME_TCP_F_DDGST;
2339                 ddgst = nvme_tcp_ddgst_len(queue);
2340         }
2341         pdu->hdr.hlen = sizeof(*pdu);
2342         pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2343         pdu->hdr.plen =
2344                 cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2345
2346         ret = nvme_tcp_map_data(queue, rq);
2347         if (unlikely(ret)) {
2348                 nvme_cleanup_cmd(rq);
2349                 dev_err(queue->ctrl->ctrl.device,
2350                         "Failed to map data (%d)\n", ret);
2351                 return ret;
2352         }
2353
2354         return 0;
2355 }
2356
2357 static void nvme_tcp_commit_rqs(struct blk_mq_hw_ctx *hctx)
2358 {
2359         struct nvme_tcp_queue *queue = hctx->driver_data;
2360
2361         if (!llist_empty(&queue->req_list))
2362                 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
2363 }
2364
2365 static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2366                 const struct blk_mq_queue_data *bd)
2367 {
2368         struct nvme_ns *ns = hctx->queue->queuedata;
2369         struct nvme_tcp_queue *queue = hctx->driver_data;
2370         struct request *rq = bd->rq;
2371         struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2372         bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2373         blk_status_t ret;
2374
2375         if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2376                 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
2377
2378         ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2379         if (unlikely(ret))
2380                 return ret;
2381
2382         blk_mq_start_request(rq);
2383
2384         nvme_tcp_queue_request(req, true, bd->last);
2385
2386         return BLK_STS_OK;
2387 }
2388
2389 static int nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2390 {
2391         struct nvme_tcp_ctrl *ctrl = set->driver_data;
2392         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2393
2394         if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2395                 /* separate read/write queues */
2396                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2397                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2398                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2399                 set->map[HCTX_TYPE_READ].nr_queues =
2400                         ctrl->io_queues[HCTX_TYPE_READ];
2401                 set->map[HCTX_TYPE_READ].queue_offset =
2402                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2403         } else {
2404                 /* shared read/write queues */
2405                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2406                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2407                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2408                 set->map[HCTX_TYPE_READ].nr_queues =
2409                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2410                 set->map[HCTX_TYPE_READ].queue_offset = 0;
2411         }
2412         blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2413         blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
2414
2415         if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2416                 /* map dedicated poll queues only if we have queues left */
2417                 set->map[HCTX_TYPE_POLL].nr_queues =
2418                                 ctrl->io_queues[HCTX_TYPE_POLL];
2419                 set->map[HCTX_TYPE_POLL].queue_offset =
2420                         ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2421                         ctrl->io_queues[HCTX_TYPE_READ];
2422                 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2423         }
2424
2425         dev_info(ctrl->ctrl.device,
2426                 "mapped %d/%d/%d default/read/poll queues.\n",
2427                 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2428                 ctrl->io_queues[HCTX_TYPE_READ],
2429                 ctrl->io_queues[HCTX_TYPE_POLL]);
2430
2431         return 0;
2432 }
2433
2434 static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
2435 {
2436         struct nvme_tcp_queue *queue = hctx->driver_data;
2437         struct sock *sk = queue->sock->sk;
2438
2439         if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2440                 return 0;
2441
2442         set_bit(NVME_TCP_Q_POLLING, &queue->flags);
2443         if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue))
2444                 sk_busy_loop(sk, true);
2445         nvme_tcp_try_recv(queue);
2446         clear_bit(NVME_TCP_Q_POLLING, &queue->flags);
2447         return queue->nr_cqe;
2448 }
2449
2450 static const struct blk_mq_ops nvme_tcp_mq_ops = {
2451         .queue_rq       = nvme_tcp_queue_rq,
2452         .commit_rqs     = nvme_tcp_commit_rqs,
2453         .complete       = nvme_complete_rq,
2454         .init_request   = nvme_tcp_init_request,
2455         .exit_request   = nvme_tcp_exit_request,
2456         .init_hctx      = nvme_tcp_init_hctx,
2457         .timeout        = nvme_tcp_timeout,
2458         .map_queues     = nvme_tcp_map_queues,
2459         .poll           = nvme_tcp_poll,
2460 };
2461
2462 static const struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2463         .queue_rq       = nvme_tcp_queue_rq,
2464         .complete       = nvme_complete_rq,
2465         .init_request   = nvme_tcp_init_request,
2466         .exit_request   = nvme_tcp_exit_request,
2467         .init_hctx      = nvme_tcp_init_admin_hctx,
2468         .timeout        = nvme_tcp_timeout,
2469 };
2470
2471 static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2472         .name                   = "tcp",
2473         .module                 = THIS_MODULE,
2474         .flags                  = NVME_F_FABRICS,
2475         .reg_read32             = nvmf_reg_read32,
2476         .reg_read64             = nvmf_reg_read64,
2477         .reg_write32            = nvmf_reg_write32,
2478         .free_ctrl              = nvme_tcp_free_ctrl,
2479         .submit_async_event     = nvme_tcp_submit_async_event,
2480         .delete_ctrl            = nvme_tcp_delete_ctrl,
2481         .get_address            = nvmf_get_address,
2482 };
2483
2484 static bool
2485 nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2486 {
2487         struct nvme_tcp_ctrl *ctrl;
2488         bool found = false;
2489
2490         mutex_lock(&nvme_tcp_ctrl_mutex);
2491         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2492                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2493                 if (found)
2494                         break;
2495         }
2496         mutex_unlock(&nvme_tcp_ctrl_mutex);
2497
2498         return found;
2499 }
2500
2501 static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2502                 struct nvmf_ctrl_options *opts)
2503 {
2504         struct nvme_tcp_ctrl *ctrl;
2505         int ret;
2506
2507         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2508         if (!ctrl)
2509                 return ERR_PTR(-ENOMEM);
2510
2511         INIT_LIST_HEAD(&ctrl->list);
2512         ctrl->ctrl.opts = opts;
2513         ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2514                                 opts->nr_poll_queues + 1;
2515         ctrl->ctrl.sqsize = opts->queue_size - 1;
2516         ctrl->ctrl.kato = opts->kato;
2517
2518         INIT_DELAYED_WORK(&ctrl->connect_work,
2519                         nvme_tcp_reconnect_ctrl_work);
2520         INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2521         INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2522
2523         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2524                 opts->trsvcid =
2525                         kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2526                 if (!opts->trsvcid) {
2527                         ret = -ENOMEM;
2528                         goto out_free_ctrl;
2529                 }
2530                 opts->mask |= NVMF_OPT_TRSVCID;
2531         }
2532
2533         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2534                         opts->traddr, opts->trsvcid, &ctrl->addr);
2535         if (ret) {
2536                 pr_err("malformed address passed: %s:%s\n",
2537                         opts->traddr, opts->trsvcid);
2538                 goto out_free_ctrl;
2539         }
2540
2541         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2542                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2543                         opts->host_traddr, NULL, &ctrl->src_addr);
2544                 if (ret) {
2545                         pr_err("malformed src address passed: %s\n",
2546                                opts->host_traddr);
2547                         goto out_free_ctrl;
2548                 }
2549         }
2550
2551         if (opts->mask & NVMF_OPT_HOST_IFACE) {
2552                 if (!__dev_get_by_name(&init_net, opts->host_iface)) {
2553                         pr_err("invalid interface passed: %s\n",
2554                                opts->host_iface);
2555                         ret = -ENODEV;
2556                         goto out_free_ctrl;
2557                 }
2558         }
2559
2560         if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2561                 ret = -EALREADY;
2562                 goto out_free_ctrl;
2563         }
2564
2565         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2566                                 GFP_KERNEL);
2567         if (!ctrl->queues) {
2568                 ret = -ENOMEM;
2569                 goto out_free_ctrl;
2570         }
2571
2572         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2573         if (ret)
2574                 goto out_kfree_queues;
2575
2576         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2577                 WARN_ON_ONCE(1);
2578                 ret = -EINTR;
2579                 goto out_uninit_ctrl;
2580         }
2581
2582         ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2583         if (ret)
2584                 goto out_uninit_ctrl;
2585
2586         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
2587                 nvmf_ctrl_subsysnqn(&ctrl->ctrl), &ctrl->addr);
2588
2589         mutex_lock(&nvme_tcp_ctrl_mutex);
2590         list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2591         mutex_unlock(&nvme_tcp_ctrl_mutex);
2592
2593         return &ctrl->ctrl;
2594
2595 out_uninit_ctrl:
2596         nvme_uninit_ctrl(&ctrl->ctrl);
2597         nvme_put_ctrl(&ctrl->ctrl);
2598         if (ret > 0)
2599                 ret = -EIO;
2600         return ERR_PTR(ret);
2601 out_kfree_queues:
2602         kfree(ctrl->queues);
2603 out_free_ctrl:
2604         kfree(ctrl);
2605         return ERR_PTR(ret);
2606 }
2607
2608 static struct nvmf_transport_ops nvme_tcp_transport = {
2609         .name           = "tcp",
2610         .module         = THIS_MODULE,
2611         .required_opts  = NVMF_OPT_TRADDR,
2612         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2613                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2614                           NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
2615                           NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2616                           NVMF_OPT_TOS | NVMF_OPT_HOST_IFACE,
2617         .create_ctrl    = nvme_tcp_create_ctrl,
2618 };
2619
2620 static int __init nvme_tcp_init_module(void)
2621 {
2622         nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
2623                         WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2624         if (!nvme_tcp_wq)
2625                 return -ENOMEM;
2626
2627         nvmf_register_transport(&nvme_tcp_transport);
2628         return 0;
2629 }
2630
2631 static void __exit nvme_tcp_cleanup_module(void)
2632 {
2633         struct nvme_tcp_ctrl *ctrl;
2634
2635         nvmf_unregister_transport(&nvme_tcp_transport);
2636
2637         mutex_lock(&nvme_tcp_ctrl_mutex);
2638         list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
2639                 nvme_delete_ctrl(&ctrl->ctrl);
2640         mutex_unlock(&nvme_tcp_ctrl_mutex);
2641         flush_workqueue(nvme_delete_wq);
2642
2643         destroy_workqueue(nvme_tcp_wq);
2644 }
2645
2646 module_init(nvme_tcp_init_module);
2647 module_exit(nvme_tcp_cleanup_module);
2648
2649 MODULE_LICENSE("GPL v2");