drivers/nvme/target/tcp.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * NVMe over Fabrics TCP target.
   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/inet.h>
  15 #include <linux/llist.h>
  16 #include <crypto/hash.h>
  17 #include <trace/events/sock.h>
  18
  19 #include "nvmet.h"
  20
  21 #define NVMET_TCP_DEF_INLINE_DATA_SIZE  (4 * PAGE_SIZE)
  22
  23 static int param_store_val(const char *str, int *val, int min, int max)
  24 {
  25         int ret, new_val;
  26
  27         ret = kstrtoint(str, 10, &new_val);
  28         if (ret)
  29                 return -EINVAL;
  30
  31         if (new_val < min || new_val > max)
  32                 return -EINVAL;
  33
  34         *val = new_val;
  35         return 0;
  36 }
  37
  38 static int set_params(const char *str, const struct kernel_param *kp)
  39 {
  40         return param_store_val(str, kp->arg, 0, INT_MAX);
  41 }
  42
  43 static const struct kernel_param_ops set_param_ops = {
  44         .set    = set_params,
  45         .get    = param_get_int,
  46 };
  47
  48 /* Define the socket priority to use for connections were it is desirable
  49  * that the NIC consider performing optimized packet processing or filtering.
  50  * A non-zero value being sufficient to indicate general consideration of any
  51  * possible optimization.  Making it a module param allows for alternative
  52  * values that may be unique for some NIC implementations.
  53  */
  54 static int so_priority;
  55 device_param_cb(so_priority, &set_param_ops, &so_priority, 0644);
  56 MODULE_PARM_DESC(so_priority, "nvmet tcp socket optimize priority: Default 0");
  57
  58 /* Define a time period (in usecs) that io_work() shall sample an activated
  59  * queue before determining it to be idle.  This optional module behavior
  60  * can enable NIC solutions that support socket optimized packet processing
  61  * using advanced interrupt moderation techniques.
  62  */
  63 static int idle_poll_period_usecs;
  64 device_param_cb(idle_poll_period_usecs, &set_param_ops,
  65                 &idle_poll_period_usecs, 0644);
  66 MODULE_PARM_DESC(idle_poll_period_usecs,
  67                 "nvmet tcp io_work poll till idle time period in usecs: Default 0");
  68
  69 #define NVMET_TCP_RECV_BUDGET           8
  70 #define NVMET_TCP_SEND_BUDGET           8
  71 #define NVMET_TCP_IO_WORK_BUDGET        64
  72
  73 enum nvmet_tcp_send_state {
  74         NVMET_TCP_SEND_DATA_PDU,
  75         NVMET_TCP_SEND_DATA,
  76         NVMET_TCP_SEND_R2T,
  77         NVMET_TCP_SEND_DDGST,
  78         NVMET_TCP_SEND_RESPONSE
  79 };
  80
  81 enum nvmet_tcp_recv_state {
  82         NVMET_TCP_RECV_PDU,
  83         NVMET_TCP_RECV_DATA,
  84         NVMET_TCP_RECV_DDGST,
  85         NVMET_TCP_RECV_ERR,
  86 };
  87
  88 enum {
  89         NVMET_TCP_F_INIT_FAILED = (1 << 0),
  90 };
  91
  92 struct nvmet_tcp_cmd {
  93         struct nvmet_tcp_queue          *queue;
  94         struct nvmet_req                req;
  95
  96         struct nvme_tcp_cmd_pdu         *cmd_pdu;
  97         struct nvme_tcp_rsp_pdu         *rsp_pdu;
  98         struct nvme_tcp_data_pdu        *data_pdu;
  99         struct nvme_tcp_r2t_pdu         *r2t_pdu;
 100
 101         u32                             rbytes_done;
 102         u32                             wbytes_done;
 103
 104         u32                             pdu_len;
 105         u32                             pdu_recv;
 106         int                             sg_idx;
 107         struct msghdr                   recv_msg;
 108         struct bio_vec                  *iov;
 109         u32                             flags;
 110
 111         struct list_head                entry;
 112         struct llist_node               lentry;
 113
 114         /* send state */
 115         u32                             offset;
 116         struct scatterlist              *cur_sg;
 117         enum nvmet_tcp_send_state       state;
 118
 119         __le32                          exp_ddgst;
 120         __le32                          recv_ddgst;
 121 };
 122
 123 enum nvmet_tcp_queue_state {
 124         NVMET_TCP_Q_CONNECTING,
 125         NVMET_TCP_Q_LIVE,
 126         NVMET_TCP_Q_DISCONNECTING,
 127 };
 128
 129 struct nvmet_tcp_queue {
 130         struct socket           *sock;
 131         struct nvmet_tcp_port   *port;
 132         struct work_struct      io_work;
 133         struct nvmet_cq         nvme_cq;
 134         struct nvmet_sq         nvme_sq;
 135
 136         /* send state */
 137         struct nvmet_tcp_cmd    *cmds;
 138         unsigned int            nr_cmds;
 139         struct list_head        free_list;
 140         struct llist_head       resp_list;
 141         struct list_head        resp_send_list;
 142         int                     send_list_len;
 143         struct nvmet_tcp_cmd    *snd_cmd;
 144
 145         /* recv state */
 146         int                     offset;
 147         int                     left;
 148         enum nvmet_tcp_recv_state rcv_state;
 149         struct nvmet_tcp_cmd    *cmd;
 150         union nvme_tcp_pdu      pdu;
 151
 152         /* digest state */
 153         bool                    hdr_digest;
 154         bool                    data_digest;
 155         struct ahash_request    *snd_hash;
 156         struct ahash_request    *rcv_hash;
 157
 158         unsigned long           poll_end;
 159
 160         spinlock_t              state_lock;
 161         enum nvmet_tcp_queue_state state;
 162
 163         struct sockaddr_storage sockaddr;
 164         struct sockaddr_storage sockaddr_peer;
 165         struct work_struct      release_work;
 166
 167         int                     idx;
 168         struct list_head        queue_list;
 169
 170         struct nvmet_tcp_cmd    connect;
 171
 172         struct page_frag_cache  pf_cache;
 173
 174         void (*data_ready)(struct sock *);
 175         void (*state_change)(struct sock *);
 176         void (*write_space)(struct sock *);
 177 };
 178
 179 struct nvmet_tcp_port {
 180         struct socket           *sock;
 181         struct work_struct      accept_work;
 182         struct nvmet_port       *nport;
 183         struct sockaddr_storage addr;
 184         void (*data_ready)(struct sock *);
 185 };
 186
 187 static DEFINE_IDA(nvmet_tcp_queue_ida);
 188 static LIST_HEAD(nvmet_tcp_queue_list);
 189 static DEFINE_MUTEX(nvmet_tcp_queue_mutex);
 190
 191 static struct workqueue_struct *nvmet_tcp_wq;
 192 static const struct nvmet_fabrics_ops nvmet_tcp_ops;
 193 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c);
 194 static void nvmet_tcp_free_cmd_buffers(struct nvmet_tcp_cmd *cmd);
 195
 196 static inline u16 nvmet_tcp_cmd_tag(struct nvmet_tcp_queue *queue,
 197                 struct nvmet_tcp_cmd *cmd)
 198 {
 199         if (unlikely(!queue->nr_cmds)) {
 200                 /* We didn't allocate cmds yet, send 0xffff */
 201                 return USHRT_MAX;
 202         }
 203
 204         return cmd - queue->cmds;
 205 }
 206
 207 static inline bool nvmet_tcp_has_data_in(struct nvmet_tcp_cmd *cmd)
 208 {
 209         return nvme_is_write(cmd->req.cmd) &&
 210                 cmd->rbytes_done < cmd->req.transfer_len;
 211 }
 212
 213 static inline bool nvmet_tcp_need_data_in(struct nvmet_tcp_cmd *cmd)
 214 {
 215         return nvmet_tcp_has_data_in(cmd) && !cmd->req.cqe->status;
 216 }
 217
 218 static inline bool nvmet_tcp_need_data_out(struct nvmet_tcp_cmd *cmd)
 219 {
 220         return !nvme_is_write(cmd->req.cmd) &&
 221                 cmd->req.transfer_len > 0 &&
 222                 !cmd->req.cqe->status;
 223 }
 224
 225 static inline bool nvmet_tcp_has_inline_data(struct nvmet_tcp_cmd *cmd)
 226 {
 227         return nvme_is_write(cmd->req.cmd) && cmd->pdu_len &&
 228                 !cmd->rbytes_done;
 229 }
 230
 231 static inline struct nvmet_tcp_cmd *
 232 nvmet_tcp_get_cmd(struct nvmet_tcp_queue *queue)
 233 {
 234         struct nvmet_tcp_cmd *cmd;
 235
 236         cmd = list_first_entry_or_null(&queue->free_list,
 237                                 struct nvmet_tcp_cmd, entry);
 238         if (!cmd)
 239                 return NULL;
 240         list_del_init(&cmd->entry);
 241
 242         cmd->rbytes_done = cmd->wbytes_done = 0;
 243         cmd->pdu_len = 0;
 244         cmd->pdu_recv = 0;
 245         cmd->iov = NULL;
 246         cmd->flags = 0;
 247         return cmd;
 248 }
 249
 250 static inline void nvmet_tcp_put_cmd(struct nvmet_tcp_cmd *cmd)
 251 {
 252         if (unlikely(cmd == &cmd->queue->connect))
 253                 return;
 254
 255         list_add_tail(&cmd->entry, &cmd->queue->free_list);
 256 }
 257
 258 static inline int queue_cpu(struct nvmet_tcp_queue *queue)
 259 {
 260         return queue->sock->sk->sk_incoming_cpu;
 261 }
 262
 263 static inline u8 nvmet_tcp_hdgst_len(struct nvmet_tcp_queue *queue)
 264 {
 265         return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 266 }
 267
 268 static inline u8 nvmet_tcp_ddgst_len(struct nvmet_tcp_queue *queue)
 269 {
 270         return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 271 }
 272
 273 static inline void nvmet_tcp_hdgst(struct ahash_request *hash,
 274                 void *pdu, size_t len)
 275 {
 276         struct scatterlist sg;
 277
 278         sg_init_one(&sg, pdu, len);
 279         ahash_request_set_crypt(hash, &sg, pdu + len, len);
 280         crypto_ahash_digest(hash);
 281 }
 282
 283 static int nvmet_tcp_verify_hdgst(struct nvmet_tcp_queue *queue,
 284         void *pdu, size_t len)
 285 {
 286         struct nvme_tcp_hdr *hdr = pdu;
 287         __le32 recv_digest;
 288         __le32 exp_digest;
 289
 290         if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
 291                 pr_err("queue %d: header digest enabled but no header digest\n",
 292                         queue->idx);
 293                 return -EPROTO;
 294         }
 295
 296         recv_digest = *(__le32 *)(pdu + hdr->hlen);
 297         nvmet_tcp_hdgst(queue->rcv_hash, pdu, len);
 298         exp_digest = *(__le32 *)(pdu + hdr->hlen);
 299         if (recv_digest != exp_digest) {
 300                 pr_err("queue %d: header digest error: recv %#x expected %#x\n",
 301                         queue->idx, le32_to_cpu(recv_digest),
 302                         le32_to_cpu(exp_digest));
 303                 return -EPROTO;
 304         }
 305
 306         return 0;
 307 }
 308
 309 static int nvmet_tcp_check_ddgst(struct nvmet_tcp_queue *queue, void *pdu)
 310 {
 311         struct nvme_tcp_hdr *hdr = pdu;
 312         u8 digest_len = nvmet_tcp_hdgst_len(queue);
 313         u32 len;
 314
 315         len = le32_to_cpu(hdr->plen) - hdr->hlen -
 316                 (hdr->flags & NVME_TCP_F_HDGST ? digest_len : 0);
 317
 318         if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
 319                 pr_err("queue %d: data digest flag is cleared\n", queue->idx);
 320                 return -EPROTO;
 321         }
 322
 323         return 0;
 324 }
 325
 326 static void nvmet_tcp_free_cmd_buffers(struct nvmet_tcp_cmd *cmd)
 327 {
 328         kfree(cmd->iov);
 329         sgl_free(cmd->req.sg);
 330         cmd->iov = NULL;
 331         cmd->req.sg = NULL;
 332 }
 333
 334 static void nvmet_tcp_build_pdu_iovec(struct nvmet_tcp_cmd *cmd)
 335 {
 336         struct bio_vec *iov = cmd->iov;
 337         struct scatterlist *sg;
 338         u32 length, offset, sg_offset;
 339         int nr_pages;
 340
 341         length = cmd->pdu_len;
 342         nr_pages = DIV_ROUND_UP(length, PAGE_SIZE);
 343         offset = cmd->rbytes_done;
 344         cmd->sg_idx = offset / PAGE_SIZE;
 345         sg_offset = offset % PAGE_SIZE;
 346         sg = &cmd->req.sg[cmd->sg_idx];
 347
 348         while (length) {
 349                 u32 iov_len = min_t(u32, length, sg->length - sg_offset);
 350
 351                 bvec_set_page(iov, sg_page(sg), sg->length,
 352                                 sg->offset + sg_offset);
 353
 354                 length -= iov_len;
 355                 sg = sg_next(sg);
 356                 iov++;
 357                 sg_offset = 0;
 358         }
 359
 360         iov_iter_bvec(&cmd->recv_msg.msg_iter, ITER_DEST, cmd->iov,
 361                       nr_pages, cmd->pdu_len);
 362 }
 363
 364 static void nvmet_tcp_fatal_error(struct nvmet_tcp_queue *queue)
 365 {
 366         queue->rcv_state = NVMET_TCP_RECV_ERR;
 367         if (queue->nvme_sq.ctrl)
 368                 nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
 369         else
 370                 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
 371 }
 372
 373 static void nvmet_tcp_socket_error(struct nvmet_tcp_queue *queue, int status)
 374 {
 375         if (status == -EPIPE || status == -ECONNRESET)
 376                 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
 377         else
 378                 nvmet_tcp_fatal_error(queue);
 379 }
 380
 381 static int nvmet_tcp_map_data(struct nvmet_tcp_cmd *cmd)
 382 {
 383         struct nvme_sgl_desc *sgl = &cmd->req.cmd->common.dptr.sgl;
 384         u32 len = le32_to_cpu(sgl->length);
 385
 386         if (!len)
 387                 return 0;
 388
 389         if (sgl->type == ((NVME_SGL_FMT_DATA_DESC << 4) |
 390                           NVME_SGL_FMT_OFFSET)) {
 391                 if (!nvme_is_write(cmd->req.cmd))
 392                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 393
 394                 if (len > cmd->req.port->inline_data_size)
 395                         return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
 396                 cmd->pdu_len = len;
 397         }
 398         cmd->req.transfer_len += len;
 399
 400         cmd->req.sg = sgl_alloc(len, GFP_KERNEL, &cmd->req.sg_cnt);
 401         if (!cmd->req.sg)
 402                 return NVME_SC_INTERNAL;
 403         cmd->cur_sg = cmd->req.sg;
 404
 405         if (nvmet_tcp_has_data_in(cmd)) {
 406                 cmd->iov = kmalloc_array(cmd->req.sg_cnt,
 407                                 sizeof(*cmd->iov), GFP_KERNEL);
 408                 if (!cmd->iov)
 409                         goto err;
 410         }
 411
 412         return 0;
 413 err:
 414         nvmet_tcp_free_cmd_buffers(cmd);
 415         return NVME_SC_INTERNAL;
 416 }
 417
 418 static void nvmet_tcp_calc_ddgst(struct ahash_request *hash,
 419                 struct nvmet_tcp_cmd *cmd)
 420 {
 421         ahash_request_set_crypt(hash, cmd->req.sg,
 422                 (void *)&cmd->exp_ddgst, cmd->req.transfer_len);
 423         crypto_ahash_digest(hash);
 424 }
 425
 426 static void nvmet_setup_c2h_data_pdu(struct nvmet_tcp_cmd *cmd)
 427 {
 428         struct nvme_tcp_data_pdu *pdu = cmd->data_pdu;
 429         struct nvmet_tcp_queue *queue = cmd->queue;
 430         u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
 431         u8 ddgst = nvmet_tcp_ddgst_len(cmd->queue);
 432
 433         cmd->offset = 0;
 434         cmd->state = NVMET_TCP_SEND_DATA_PDU;
 435
 436         pdu->hdr.type = nvme_tcp_c2h_data;
 437         pdu->hdr.flags = NVME_TCP_F_DATA_LAST | (queue->nvme_sq.sqhd_disabled ?
 438                                                 NVME_TCP_F_DATA_SUCCESS : 0);
 439         pdu->hdr.hlen = sizeof(*pdu);
 440         pdu->hdr.pdo = pdu->hdr.hlen + hdgst;
 441         pdu->hdr.plen =
 442                 cpu_to_le32(pdu->hdr.hlen + hdgst +
 443                                 cmd->req.transfer_len + ddgst);
 444         pdu->command_id = cmd->req.cqe->command_id;
 445         pdu->data_length = cpu_to_le32(cmd->req.transfer_len);
 446         pdu->data_offset = cpu_to_le32(cmd->wbytes_done);
 447
 448         if (queue->data_digest) {
 449                 pdu->hdr.flags |= NVME_TCP_F_DDGST;
 450                 nvmet_tcp_calc_ddgst(queue->snd_hash, cmd);
 451         }
 452
 453         if (cmd->queue->hdr_digest) {
 454                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
 455                 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
 456         }
 457 }
 458
 459 static void nvmet_setup_r2t_pdu(struct nvmet_tcp_cmd *cmd)
 460 {
 461         struct nvme_tcp_r2t_pdu *pdu = cmd->r2t_pdu;
 462         struct nvmet_tcp_queue *queue = cmd->queue;
 463         u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
 464
 465         cmd->offset = 0;
 466         cmd->state = NVMET_TCP_SEND_R2T;
 467
 468         pdu->hdr.type = nvme_tcp_r2t;
 469         pdu->hdr.flags = 0;
 470         pdu->hdr.hlen = sizeof(*pdu);
 471         pdu->hdr.pdo = 0;
 472         pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
 473
 474         pdu->command_id = cmd->req.cmd->common.command_id;
 475         pdu->ttag = nvmet_tcp_cmd_tag(cmd->queue, cmd);
 476         pdu->r2t_length = cpu_to_le32(cmd->req.transfer_len - cmd->rbytes_done);
 477         pdu->r2t_offset = cpu_to_le32(cmd->rbytes_done);
 478         if (cmd->queue->hdr_digest) {
 479                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
 480                 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
 481         }
 482 }
 483
 484 static void nvmet_setup_response_pdu(struct nvmet_tcp_cmd *cmd)
 485 {
 486         struct nvme_tcp_rsp_pdu *pdu = cmd->rsp_pdu;
 487         struct nvmet_tcp_queue *queue = cmd->queue;
 488         u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
 489
 490         cmd->offset = 0;
 491         cmd->state = NVMET_TCP_SEND_RESPONSE;
 492
 493         pdu->hdr.type = nvme_tcp_rsp;
 494         pdu->hdr.flags = 0;
 495         pdu->hdr.hlen = sizeof(*pdu);
 496         pdu->hdr.pdo = 0;
 497         pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
 498         if (cmd->queue->hdr_digest) {
 499                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
 500                 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
 501         }
 502 }
 503
 504 static void nvmet_tcp_process_resp_list(struct nvmet_tcp_queue *queue)
 505 {
 506         struct llist_node *node;
 507         struct nvmet_tcp_cmd *cmd;
 508
 509         for (node = llist_del_all(&queue->resp_list); node; node = node->next) {
 510                 cmd = llist_entry(node, struct nvmet_tcp_cmd, lentry);
 511                 list_add(&cmd->entry, &queue->resp_send_list);
 512                 queue->send_list_len++;
 513         }
 514 }
 515
 516 static struct nvmet_tcp_cmd *nvmet_tcp_fetch_cmd(struct nvmet_tcp_queue *queue)
 517 {
 518         queue->snd_cmd = list_first_entry_or_null(&queue->resp_send_list,
 519                                 struct nvmet_tcp_cmd, entry);
 520         if (!queue->snd_cmd) {
 521                 nvmet_tcp_process_resp_list(queue);
 522                 queue->snd_cmd =
 523                         list_first_entry_or_null(&queue->resp_send_list,
 524                                         struct nvmet_tcp_cmd, entry);
 525                 if (unlikely(!queue->snd_cmd))
 526                         return NULL;
 527         }
 528
 529         list_del_init(&queue->snd_cmd->entry);
 530         queue->send_list_len--;
 531
 532         if (nvmet_tcp_need_data_out(queue->snd_cmd))
 533                 nvmet_setup_c2h_data_pdu(queue->snd_cmd);
 534         else if (nvmet_tcp_need_data_in(queue->snd_cmd))
 535                 nvmet_setup_r2t_pdu(queue->snd_cmd);
 536         else
 537                 nvmet_setup_response_pdu(queue->snd_cmd);
 538
 539         return queue->snd_cmd;
 540 }
 541
 542 static void nvmet_tcp_queue_response(struct nvmet_req *req)
 543 {
 544         struct nvmet_tcp_cmd *cmd =
 545                 container_of(req, struct nvmet_tcp_cmd, req);
 546         struct nvmet_tcp_queue  *queue = cmd->queue;
 547         struct nvme_sgl_desc *sgl;
 548         u32 len;
 549
 550         if (unlikely(cmd == queue->cmd)) {
 551                 sgl = &cmd->req.cmd->common.dptr.sgl;
 552                 len = le32_to_cpu(sgl->length);
 553
 554                 /*
 555                  * Wait for inline data before processing the response.
 556                  * Avoid using helpers, this might happen before
 557                  * nvmet_req_init is completed.
 558                  */
 559                 if (queue->rcv_state == NVMET_TCP_RECV_PDU &&
 560                     len && len <= cmd->req.port->inline_data_size &&
 561                     nvme_is_write(cmd->req.cmd))
 562                         return;
 563         }
 564
 565         llist_add(&cmd->lentry, &queue->resp_list);
 566         queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &cmd->queue->io_work);
 567 }
 568
 569 static void nvmet_tcp_execute_request(struct nvmet_tcp_cmd *cmd)
 570 {
 571         if (unlikely(cmd->flags & NVMET_TCP_F_INIT_FAILED))
 572                 nvmet_tcp_queue_response(&cmd->req);
 573         else
 574                 cmd->req.execute(&cmd->req);
 575 }
 576
 577 static int nvmet_try_send_data_pdu(struct nvmet_tcp_cmd *cmd)
 578 {
 579         u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
 580         int left = sizeof(*cmd->data_pdu) - cmd->offset + hdgst;
 581         int ret;
 582
 583         ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->data_pdu),
 584                         offset_in_page(cmd->data_pdu) + cmd->offset,
 585                         left, MSG_DONTWAIT | MSG_MORE | MSG_SENDPAGE_NOTLAST);
 586         if (ret <= 0)
 587                 return ret;
 588
 589         cmd->offset += ret;
 590         left -= ret;
 591
 592         if (left)
 593                 return -EAGAIN;
 594
 595         cmd->state = NVMET_TCP_SEND_DATA;
 596         cmd->offset  = 0;
 597         return 1;
 598 }
 599
 600 static int nvmet_try_send_data(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
 601 {
 602         struct nvmet_tcp_queue *queue = cmd->queue;
 603         int ret;
 604
 605         while (cmd->cur_sg) {
 606                 struct page *page = sg_page(cmd->cur_sg);
 607                 u32 left = cmd->cur_sg->length - cmd->offset;
 608                 int flags = MSG_DONTWAIT;
 609
 610                 if ((!last_in_batch && cmd->queue->send_list_len) ||
 611                     cmd->wbytes_done + left < cmd->req.transfer_len ||
 612                     queue->data_digest || !queue->nvme_sq.sqhd_disabled)
 613                         flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
 614
 615                 ret = kernel_sendpage(cmd->queue->sock, page, cmd->offset,
 616                                         left, flags);
 617                 if (ret <= 0)
 618                         return ret;
 619
 620                 cmd->offset += ret;
 621                 cmd->wbytes_done += ret;
 622
 623                 /* Done with sg?*/
 624                 if (cmd->offset == cmd->cur_sg->length) {
 625                         cmd->cur_sg = sg_next(cmd->cur_sg);
 626                         cmd->offset = 0;
 627                 }
 628         }
 629
 630         if (queue->data_digest) {
 631                 cmd->state = NVMET_TCP_SEND_DDGST;
 632                 cmd->offset = 0;
 633         } else {
 634                 if (queue->nvme_sq.sqhd_disabled) {
 635                         cmd->queue->snd_cmd = NULL;
 636                         nvmet_tcp_put_cmd(cmd);
 637                 } else {
 638                         nvmet_setup_response_pdu(cmd);
 639                 }
 640         }
 641
 642         if (queue->nvme_sq.sqhd_disabled)
 643                 nvmet_tcp_free_cmd_buffers(cmd);
 644
 645         return 1;
 646
 647 }
 648
 649 static int nvmet_try_send_response(struct nvmet_tcp_cmd *cmd,
 650                 bool last_in_batch)
 651 {
 652         u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
 653         int left = sizeof(*cmd->rsp_pdu) - cmd->offset + hdgst;
 654         int flags = MSG_DONTWAIT;
 655         int ret;
 656
 657         if (!last_in_batch && cmd->queue->send_list_len)
 658                 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
 659         else
 660                 flags |= MSG_EOR;
 661
 662         ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->rsp_pdu),
 663                 offset_in_page(cmd->rsp_pdu) + cmd->offset, left, flags);
 664         if (ret <= 0)
 665                 return ret;
 666         cmd->offset += ret;
 667         left -= ret;
 668
 669         if (left)
 670                 return -EAGAIN;
 671
 672         nvmet_tcp_free_cmd_buffers(cmd);
 673         cmd->queue->snd_cmd = NULL;
 674         nvmet_tcp_put_cmd(cmd);
 675         return 1;
 676 }
 677
 678 static int nvmet_try_send_r2t(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
 679 {
 680         u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
 681         int left = sizeof(*cmd->r2t_pdu) - cmd->offset + hdgst;
 682         int flags = MSG_DONTWAIT;
 683         int ret;
 684
 685         if (!last_in_batch && cmd->queue->send_list_len)
 686                 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
 687         else
 688                 flags |= MSG_EOR;
 689
 690         ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->r2t_pdu),
 691                 offset_in_page(cmd->r2t_pdu) + cmd->offset, left, flags);
 692         if (ret <= 0)
 693                 return ret;
 694         cmd->offset += ret;
 695         left -= ret;
 696
 697         if (left)
 698                 return -EAGAIN;
 699
 700         cmd->queue->snd_cmd = NULL;
 701         return 1;
 702 }
 703
 704 static int nvmet_try_send_ddgst(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
 705 {
 706         struct nvmet_tcp_queue *queue = cmd->queue;
 707         int left = NVME_TCP_DIGEST_LENGTH - cmd->offset;
 708         struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
 709         struct kvec iov = {
 710                 .iov_base = (u8 *)&cmd->exp_ddgst + cmd->offset,
 711                 .iov_len = left
 712         };
 713         int ret;
 714
 715         if (!last_in_batch && cmd->queue->send_list_len)
 716                 msg.msg_flags |= MSG_MORE;
 717         else
 718                 msg.msg_flags |= MSG_EOR;
 719
 720         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
 721         if (unlikely(ret <= 0))
 722                 return ret;
 723
 724         cmd->offset += ret;
 725         left -= ret;
 726
 727         if (left)
 728                 return -EAGAIN;
 729
 730         if (queue->nvme_sq.sqhd_disabled) {
 731                 cmd->queue->snd_cmd = NULL;
 732                 nvmet_tcp_put_cmd(cmd);
 733         } else {
 734                 nvmet_setup_response_pdu(cmd);
 735         }
 736         return 1;
 737 }
 738
 739 static int nvmet_tcp_try_send_one(struct nvmet_tcp_queue *queue,
 740                 bool last_in_batch)
 741 {
 742         struct nvmet_tcp_cmd *cmd = queue->snd_cmd;
 743         int ret = 0;
 744
 745         if (!cmd || queue->state == NVMET_TCP_Q_DISCONNECTING) {
 746                 cmd = nvmet_tcp_fetch_cmd(queue);
 747                 if (unlikely(!cmd))
 748                         return 0;
 749         }
 750
 751         if (cmd->state == NVMET_TCP_SEND_DATA_PDU) {
 752                 ret = nvmet_try_send_data_pdu(cmd);
 753                 if (ret <= 0)
 754                         goto done_send;
 755         }
 756
 757         if (cmd->state == NVMET_TCP_SEND_DATA) {
 758                 ret = nvmet_try_send_data(cmd, last_in_batch);
 759                 if (ret <= 0)
 760                         goto done_send;
 761         }
 762
 763         if (cmd->state == NVMET_TCP_SEND_DDGST) {
 764                 ret = nvmet_try_send_ddgst(cmd, last_in_batch);
 765                 if (ret <= 0)
 766                         goto done_send;
 767         }
 768
 769         if (cmd->state == NVMET_TCP_SEND_R2T) {
 770                 ret = nvmet_try_send_r2t(cmd, last_in_batch);
 771                 if (ret <= 0)
 772                         goto done_send;
 773         }
 774
 775         if (cmd->state == NVMET_TCP_SEND_RESPONSE)
 776                 ret = nvmet_try_send_response(cmd, last_in_batch);
 777
 778 done_send:
 779         if (ret < 0) {
 780                 if (ret == -EAGAIN)
 781                         return 0;
 782                 return ret;
 783         }
 784
 785         return 1;
 786 }
 787
 788 static int nvmet_tcp_try_send(struct nvmet_tcp_queue *queue,
 789                 int budget, int *sends)
 790 {
 791         int i, ret = 0;
 792
 793         for (i = 0; i < budget; i++) {
 794                 ret = nvmet_tcp_try_send_one(queue, i == budget - 1);
 795                 if (unlikely(ret < 0)) {
 796                         nvmet_tcp_socket_error(queue, ret);
 797                         goto done;
 798                 } else if (ret == 0) {
 799                         break;
 800                 }
 801                 (*sends)++;
 802         }
 803 done:
 804         return ret;
 805 }
 806
 807 static void nvmet_prepare_receive_pdu(struct nvmet_tcp_queue *queue)
 808 {
 809         queue->offset = 0;
 810         queue->left = sizeof(struct nvme_tcp_hdr);
 811         queue->cmd = NULL;
 812         queue->rcv_state = NVMET_TCP_RECV_PDU;
 813 }
 814
 815 static void nvmet_tcp_free_crypto(struct nvmet_tcp_queue *queue)
 816 {
 817         struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
 818
 819         ahash_request_free(queue->rcv_hash);
 820         ahash_request_free(queue->snd_hash);
 821         crypto_free_ahash(tfm);
 822 }
 823
 824 static int nvmet_tcp_alloc_crypto(struct nvmet_tcp_queue *queue)
 825 {
 826         struct crypto_ahash *tfm;
 827
 828         tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
 829         if (IS_ERR(tfm))
 830                 return PTR_ERR(tfm);
 831
 832         queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
 833         if (!queue->snd_hash)
 834                 goto free_tfm;
 835         ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
 836
 837         queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
 838         if (!queue->rcv_hash)
 839                 goto free_snd_hash;
 840         ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
 841
 842         return 0;
 843 free_snd_hash:
 844         ahash_request_free(queue->snd_hash);
 845 free_tfm:
 846         crypto_free_ahash(tfm);
 847         return -ENOMEM;
 848 }
 849
 850
 851 static int nvmet_tcp_handle_icreq(struct nvmet_tcp_queue *queue)
 852 {
 853         struct nvme_tcp_icreq_pdu *icreq = &queue->pdu.icreq;
 854         struct nvme_tcp_icresp_pdu *icresp = &queue->pdu.icresp;
 855         struct msghdr msg = {};
 856         struct kvec iov;
 857         int ret;
 858
 859         if (le32_to_cpu(icreq->hdr.plen) != sizeof(struct nvme_tcp_icreq_pdu)) {
 860                 pr_err("bad nvme-tcp pdu length (%d)\n",
 861                         le32_to_cpu(icreq->hdr.plen));
 862                 nvmet_tcp_fatal_error(queue);
 863         }
 864
 865         if (icreq->pfv != NVME_TCP_PFV_1_0) {
 866                 pr_err("queue %d: bad pfv %d\n", queue->idx, icreq->pfv);
 867                 return -EPROTO;
 868         }
 869
 870         if (icreq->hpda != 0) {
 871                 pr_err("queue %d: unsupported hpda %d\n", queue->idx,
 872                         icreq->hpda);
 873                 return -EPROTO;
 874         }
 875
 876         queue->hdr_digest = !!(icreq->digest & NVME_TCP_HDR_DIGEST_ENABLE);
 877         queue->data_digest = !!(icreq->digest & NVME_TCP_DATA_DIGEST_ENABLE);
 878         if (queue->hdr_digest || queue->data_digest) {
 879                 ret = nvmet_tcp_alloc_crypto(queue);
 880                 if (ret)
 881                         return ret;
 882         }
 883
 884         memset(icresp, 0, sizeof(*icresp));
 885         icresp->hdr.type = nvme_tcp_icresp;
 886         icresp->hdr.hlen = sizeof(*icresp);
 887         icresp->hdr.pdo = 0;
 888         icresp->hdr.plen = cpu_to_le32(icresp->hdr.hlen);
 889         icresp->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
 890         icresp->maxdata = cpu_to_le32(0x400000); /* 16M arbitrary limit */
 891         icresp->cpda = 0;
 892         if (queue->hdr_digest)
 893                 icresp->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
 894         if (queue->data_digest)
 895                 icresp->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
 896
 897         iov.iov_base = icresp;
 898         iov.iov_len = sizeof(*icresp);
 899         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
 900         if (ret < 0)
 901                 goto free_crypto;
 902
 903         queue->state = NVMET_TCP_Q_LIVE;
 904         nvmet_prepare_receive_pdu(queue);
 905         return 0;
 906 free_crypto:
 907         if (queue->hdr_digest || queue->data_digest)
 908                 nvmet_tcp_free_crypto(queue);
 909         return ret;
 910 }
 911
 912 static void nvmet_tcp_handle_req_failure(struct nvmet_tcp_queue *queue,
 913                 struct nvmet_tcp_cmd *cmd, struct nvmet_req *req)
 914 {
 915         size_t data_len = le32_to_cpu(req->cmd->common.dptr.sgl.length);
 916         int ret;
 917
 918         /*
 919          * This command has not been processed yet, hence we are trying to
 920          * figure out if there is still pending data left to receive. If
 921          * we don't, we can simply prepare for the next pdu and bail out,
 922          * otherwise we will need to prepare a buffer and receive the
 923          * stale data before continuing forward.
 924          */
 925         if (!nvme_is_write(cmd->req.cmd) || !data_len ||
 926             data_len > cmd->req.port->inline_data_size) {
 927                 nvmet_prepare_receive_pdu(queue);
 928                 return;
 929         }
 930
 931         ret = nvmet_tcp_map_data(cmd);
 932         if (unlikely(ret)) {
 933                 pr_err("queue %d: failed to map data\n", queue->idx);
 934                 nvmet_tcp_fatal_error(queue);
 935                 return;
 936         }
 937
 938         queue->rcv_state = NVMET_TCP_RECV_DATA;
 939         nvmet_tcp_build_pdu_iovec(cmd);
 940         cmd->flags |= NVMET_TCP_F_INIT_FAILED;
 941 }
 942
 943 static int nvmet_tcp_handle_h2c_data_pdu(struct nvmet_tcp_queue *queue)
 944 {
 945         struct nvme_tcp_data_pdu *data = &queue->pdu.data;
 946         struct nvmet_tcp_cmd *cmd;
 947
 948         if (likely(queue->nr_cmds)) {
 949                 if (unlikely(data->ttag >= queue->nr_cmds)) {
 950                         pr_err("queue %d: received out of bound ttag %u, nr_cmds %u\n",
 951                                 queue->idx, data->ttag, queue->nr_cmds);
 952                         nvmet_tcp_fatal_error(queue);
 953                         return -EPROTO;
 954                 }
 955                 cmd = &queue->cmds[data->ttag];
 956         } else {
 957                 cmd = &queue->connect;
 958         }
 959
 960         if (le32_to_cpu(data->data_offset) != cmd->rbytes_done) {
 961                 pr_err("ttag %u unexpected data offset %u (expected %u)\n",
 962                         data->ttag, le32_to_cpu(data->data_offset),
 963                         cmd->rbytes_done);
 964                 /* FIXME: use path and transport errors */
 965                 nvmet_req_complete(&cmd->req,
 966                         NVME_SC_INVALID_FIELD | NVME_SC_DNR);
 967                 return -EPROTO;
 968         }
 969
 970         cmd->pdu_len = le32_to_cpu(data->data_length);
 971         cmd->pdu_recv = 0;
 972         nvmet_tcp_build_pdu_iovec(cmd);
 973         queue->cmd = cmd;
 974         queue->rcv_state = NVMET_TCP_RECV_DATA;
 975
 976         return 0;
 977 }
 978
 979 static int nvmet_tcp_done_recv_pdu(struct nvmet_tcp_queue *queue)
 980 {
 981         struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
 982         struct nvme_command *nvme_cmd = &queue->pdu.cmd.cmd;
 983         struct nvmet_req *req;
 984         int ret;
 985
 986         if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
 987                 if (hdr->type != nvme_tcp_icreq) {
 988                         pr_err("unexpected pdu type (%d) before icreq\n",
 989                                 hdr->type);
 990                         nvmet_tcp_fatal_error(queue);
 991                         return -EPROTO;
 992                 }
 993                 return nvmet_tcp_handle_icreq(queue);
 994         }
 995
 996         if (unlikely(hdr->type == nvme_tcp_icreq)) {
 997                 pr_err("queue %d: received icreq pdu in state %d\n",
 998                         queue->idx, queue->state);
 999                 nvmet_tcp_fatal_error(queue);
1000                 return -EPROTO;
1001         }
1002
1003         if (hdr->type == nvme_tcp_h2c_data) {
1004                 ret = nvmet_tcp_handle_h2c_data_pdu(queue);
1005                 if (unlikely(ret))
1006                         return ret;
1007                 return 0;
1008         }
1009
1010         queue->cmd = nvmet_tcp_get_cmd(queue);
1011         if (unlikely(!queue->cmd)) {
1012                 /* This should never happen */
1013                 pr_err("queue %d: out of commands (%d) send_list_len: %d, opcode: %d",
1014                         queue->idx, queue->nr_cmds, queue->send_list_len,
1015                         nvme_cmd->common.opcode);
1016                 nvmet_tcp_fatal_error(queue);
1017                 return -ENOMEM;
1018         }
1019
1020         req = &queue->cmd->req;
1021         memcpy(req->cmd, nvme_cmd, sizeof(*nvme_cmd));
1022
1023         if (unlikely(!nvmet_req_init(req, &queue->nvme_cq,
1024                         &queue->nvme_sq, &nvmet_tcp_ops))) {
1025                 pr_err("failed cmd %p id %d opcode %d, data_len: %d\n",
1026                         req->cmd, req->cmd->common.command_id,
1027                         req->cmd->common.opcode,
1028                         le32_to_cpu(req->cmd->common.dptr.sgl.length));
1029
1030                 nvmet_tcp_handle_req_failure(queue, queue->cmd, req);
1031                 return 0;
1032         }
1033
1034         ret = nvmet_tcp_map_data(queue->cmd);
1035         if (unlikely(ret)) {
1036                 pr_err("queue %d: failed to map data\n", queue->idx);
1037                 if (nvmet_tcp_has_inline_data(queue->cmd))
1038                         nvmet_tcp_fatal_error(queue);
1039                 else
1040                         nvmet_req_complete(req, ret);
1041                 ret = -EAGAIN;
1042                 goto out;
1043         }
1044
1045         if (nvmet_tcp_need_data_in(queue->cmd)) {
1046                 if (nvmet_tcp_has_inline_data(queue->cmd)) {
1047                         queue->rcv_state = NVMET_TCP_RECV_DATA;
1048                         nvmet_tcp_build_pdu_iovec(queue->cmd);
1049                         return 0;
1050                 }
1051                 /* send back R2T */
1052                 nvmet_tcp_queue_response(&queue->cmd->req);
1053                 goto out;
1054         }
1055
1056         queue->cmd->req.execute(&queue->cmd->req);
1057 out:
1058         nvmet_prepare_receive_pdu(queue);
1059         return ret;
1060 }
1061
1062 static const u8 nvme_tcp_pdu_sizes[] = {
1063         [nvme_tcp_icreq]        = sizeof(struct nvme_tcp_icreq_pdu),
1064         [nvme_tcp_cmd]          = sizeof(struct nvme_tcp_cmd_pdu),
1065         [nvme_tcp_h2c_data]     = sizeof(struct nvme_tcp_data_pdu),
1066 };
1067
1068 static inline u8 nvmet_tcp_pdu_size(u8 type)
1069 {
1070         size_t idx = type;
1071
1072         return (idx < ARRAY_SIZE(nvme_tcp_pdu_sizes) &&
1073                 nvme_tcp_pdu_sizes[idx]) ?
1074                         nvme_tcp_pdu_sizes[idx] : 0;
1075 }
1076
1077 static inline bool nvmet_tcp_pdu_valid(u8 type)
1078 {
1079         switch (type) {
1080         case nvme_tcp_icreq:
1081         case nvme_tcp_cmd:
1082         case nvme_tcp_h2c_data:
1083                 /* fallthru */
1084                 return true;
1085         }
1086
1087         return false;
1088 }
1089
1090 static int nvmet_tcp_try_recv_pdu(struct nvmet_tcp_queue *queue)
1091 {
1092         struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
1093         int len;
1094         struct kvec iov;
1095         struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1096
1097 recv:
1098         iov.iov_base = (void *)&queue->pdu + queue->offset;
1099         iov.iov_len = queue->left;
1100         len = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1101                         iov.iov_len, msg.msg_flags);
1102         if (unlikely(len < 0))
1103                 return len;
1104
1105         queue->offset += len;
1106         queue->left -= len;
1107         if (queue->left)
1108                 return -EAGAIN;
1109
1110         if (queue->offset == sizeof(struct nvme_tcp_hdr)) {
1111                 u8 hdgst = nvmet_tcp_hdgst_len(queue);
1112
1113                 if (unlikely(!nvmet_tcp_pdu_valid(hdr->type))) {
1114                         pr_err("unexpected pdu type %d\n", hdr->type);
1115                         nvmet_tcp_fatal_error(queue);
1116                         return -EIO;
1117                 }
1118
1119                 if (unlikely(hdr->hlen != nvmet_tcp_pdu_size(hdr->type))) {
1120                         pr_err("pdu %d bad hlen %d\n", hdr->type, hdr->hlen);
1121                         return -EIO;
1122                 }
1123
1124                 queue->left = hdr->hlen - queue->offset + hdgst;
1125                 goto recv;
1126         }
1127
1128         if (queue->hdr_digest &&
1129             nvmet_tcp_verify_hdgst(queue, &queue->pdu, hdr->hlen)) {
1130                 nvmet_tcp_fatal_error(queue); /* fatal */
1131                 return -EPROTO;
1132         }
1133
1134         if (queue->data_digest &&
1135             nvmet_tcp_check_ddgst(queue, &queue->pdu)) {
1136                 nvmet_tcp_fatal_error(queue); /* fatal */
1137                 return -EPROTO;
1138         }
1139
1140         return nvmet_tcp_done_recv_pdu(queue);
1141 }
1142
1143 static void nvmet_tcp_prep_recv_ddgst(struct nvmet_tcp_cmd *cmd)
1144 {
1145         struct nvmet_tcp_queue *queue = cmd->queue;
1146
1147         nvmet_tcp_calc_ddgst(queue->rcv_hash, cmd);
1148         queue->offset = 0;
1149         queue->left = NVME_TCP_DIGEST_LENGTH;
1150         queue->rcv_state = NVMET_TCP_RECV_DDGST;
1151 }
1152
1153 static int nvmet_tcp_try_recv_data(struct nvmet_tcp_queue *queue)
1154 {
1155         struct nvmet_tcp_cmd  *cmd = queue->cmd;
1156         int ret;
1157
1158         while (msg_data_left(&cmd->recv_msg)) {
1159                 ret = sock_recvmsg(cmd->queue->sock, &cmd->recv_msg,
1160                         cmd->recv_msg.msg_flags);
1161                 if (ret <= 0)
1162                         return ret;
1163
1164                 cmd->pdu_recv += ret;
1165                 cmd->rbytes_done += ret;
1166         }
1167
1168         if (queue->data_digest) {
1169                 nvmet_tcp_prep_recv_ddgst(cmd);
1170                 return 0;
1171         }
1172
1173         if (cmd->rbytes_done == cmd->req.transfer_len)
1174                 nvmet_tcp_execute_request(cmd);
1175
1176         nvmet_prepare_receive_pdu(queue);
1177         return 0;
1178 }
1179
1180 static int nvmet_tcp_try_recv_ddgst(struct nvmet_tcp_queue *queue)
1181 {
1182         struct nvmet_tcp_cmd *cmd = queue->cmd;
1183         int ret;
1184         struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1185         struct kvec iov = {
1186                 .iov_base = (void *)&cmd->recv_ddgst + queue->offset,
1187                 .iov_len = queue->left
1188         };
1189
1190         ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1191                         iov.iov_len, msg.msg_flags);
1192         if (unlikely(ret < 0))
1193                 return ret;
1194
1195         queue->offset += ret;
1196         queue->left -= ret;
1197         if (queue->left)
1198                 return -EAGAIN;
1199
1200         if (queue->data_digest && cmd->exp_ddgst != cmd->recv_ddgst) {
1201                 pr_err("queue %d: cmd %d pdu (%d) data digest error: recv %#x expected %#x\n",
1202                         queue->idx, cmd->req.cmd->common.command_id,
1203                         queue->pdu.cmd.hdr.type, le32_to_cpu(cmd->recv_ddgst),
1204                         le32_to_cpu(cmd->exp_ddgst));
1205                 nvmet_req_uninit(&cmd->req);
1206                 nvmet_tcp_free_cmd_buffers(cmd);
1207                 nvmet_tcp_fatal_error(queue);
1208                 ret = -EPROTO;
1209                 goto out;
1210         }
1211
1212         if (cmd->rbytes_done == cmd->req.transfer_len)
1213                 nvmet_tcp_execute_request(cmd);
1214
1215         ret = 0;
1216 out:
1217         nvmet_prepare_receive_pdu(queue);
1218         return ret;
1219 }
1220
1221 static int nvmet_tcp_try_recv_one(struct nvmet_tcp_queue *queue)
1222 {
1223         int result = 0;
1224
1225         if (unlikely(queue->rcv_state == NVMET_TCP_RECV_ERR))
1226                 return 0;
1227
1228         if (queue->rcv_state == NVMET_TCP_RECV_PDU) {
1229                 result = nvmet_tcp_try_recv_pdu(queue);
1230                 if (result != 0)
1231                         goto done_recv;
1232         }
1233
1234         if (queue->rcv_state == NVMET_TCP_RECV_DATA) {
1235                 result = nvmet_tcp_try_recv_data(queue);
1236                 if (result != 0)
1237                         goto done_recv;
1238         }
1239
1240         if (queue->rcv_state == NVMET_TCP_RECV_DDGST) {
1241                 result = nvmet_tcp_try_recv_ddgst(queue);
1242                 if (result != 0)
1243                         goto done_recv;
1244         }
1245
1246 done_recv:
1247         if (result < 0) {
1248                 if (result == -EAGAIN)
1249                         return 0;
1250                 return result;
1251         }
1252         return 1;
1253 }
1254
1255 static int nvmet_tcp_try_recv(struct nvmet_tcp_queue *queue,
1256                 int budget, int *recvs)
1257 {
1258         int i, ret = 0;
1259
1260         for (i = 0; i < budget; i++) {
1261                 ret = nvmet_tcp_try_recv_one(queue);
1262                 if (unlikely(ret < 0)) {
1263                         nvmet_tcp_socket_error(queue, ret);
1264                         goto done;
1265                 } else if (ret == 0) {
1266                         break;
1267                 }
1268                 (*recvs)++;
1269         }
1270 done:
1271         return ret;
1272 }
1273
1274 static void nvmet_tcp_schedule_release_queue(struct nvmet_tcp_queue *queue)
1275 {
1276         spin_lock(&queue->state_lock);
1277         if (queue->state != NVMET_TCP_Q_DISCONNECTING) {
1278                 queue->state = NVMET_TCP_Q_DISCONNECTING;
1279                 queue_work(nvmet_wq, &queue->release_work);
1280         }
1281         spin_unlock(&queue->state_lock);
1282 }
1283
1284 static inline void nvmet_tcp_arm_queue_deadline(struct nvmet_tcp_queue *queue)
1285 {
1286         queue->poll_end = jiffies + usecs_to_jiffies(idle_poll_period_usecs);
1287 }
1288
1289 static bool nvmet_tcp_check_queue_deadline(struct nvmet_tcp_queue *queue,
1290                 int ops)
1291 {
1292         if (!idle_poll_period_usecs)
1293                 return false;
1294
1295         if (ops)
1296                 nvmet_tcp_arm_queue_deadline(queue);
1297
1298         return !time_after(jiffies, queue->poll_end);
1299 }
1300
1301 static void nvmet_tcp_io_work(struct work_struct *w)
1302 {
1303         struct nvmet_tcp_queue *queue =
1304                 container_of(w, struct nvmet_tcp_queue, io_work);
1305         bool pending;
1306         int ret, ops = 0;
1307
1308         do {
1309                 pending = false;
1310
1311                 ret = nvmet_tcp_try_recv(queue, NVMET_TCP_RECV_BUDGET, &ops);
1312                 if (ret > 0)
1313                         pending = true;
1314                 else if (ret < 0)
1315                         return;
1316
1317                 ret = nvmet_tcp_try_send(queue, NVMET_TCP_SEND_BUDGET, &ops);
1318                 if (ret > 0)
1319                         pending = true;
1320                 else if (ret < 0)
1321                         return;
1322
1323         } while (pending && ops < NVMET_TCP_IO_WORK_BUDGET);
1324
1325         /*
1326          * Requeue the worker if idle deadline period is in progress or any
1327          * ops activity was recorded during the do-while loop above.
1328          */
1329         if (nvmet_tcp_check_queue_deadline(queue, ops) || pending)
1330                 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work);
1331 }
1332
1333 static int nvmet_tcp_alloc_cmd(struct nvmet_tcp_queue *queue,
1334                 struct nvmet_tcp_cmd *c)
1335 {
1336         u8 hdgst = nvmet_tcp_hdgst_len(queue);
1337
1338         c->queue = queue;
1339         c->req.port = queue->port->nport;
1340
1341         c->cmd_pdu = page_frag_alloc(&queue->pf_cache,
1342                         sizeof(*c->cmd_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1343         if (!c->cmd_pdu)
1344                 return -ENOMEM;
1345         c->req.cmd = &c->cmd_pdu->cmd;
1346
1347         c->rsp_pdu = page_frag_alloc(&queue->pf_cache,
1348                         sizeof(*c->rsp_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1349         if (!c->rsp_pdu)
1350                 goto out_free_cmd;
1351         c->req.cqe = &c->rsp_pdu->cqe;
1352
1353         c->data_pdu = page_frag_alloc(&queue->pf_cache,
1354                         sizeof(*c->data_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1355         if (!c->data_pdu)
1356                 goto out_free_rsp;
1357
1358         c->r2t_pdu = page_frag_alloc(&queue->pf_cache,
1359                         sizeof(*c->r2t_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1360         if (!c->r2t_pdu)
1361                 goto out_free_data;
1362
1363         c->recv_msg.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1364
1365         list_add_tail(&c->entry, &queue->free_list);
1366
1367         return 0;
1368 out_free_data:
1369         page_frag_free(c->data_pdu);
1370 out_free_rsp:
1371         page_frag_free(c->rsp_pdu);
1372 out_free_cmd:
1373         page_frag_free(c->cmd_pdu);
1374         return -ENOMEM;
1375 }
1376
1377 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c)
1378 {
1379         page_frag_free(c->r2t_pdu);
1380         page_frag_free(c->data_pdu);
1381         page_frag_free(c->rsp_pdu);
1382         page_frag_free(c->cmd_pdu);
1383 }
1384
1385 static int nvmet_tcp_alloc_cmds(struct nvmet_tcp_queue *queue)
1386 {
1387         struct nvmet_tcp_cmd *cmds;
1388         int i, ret = -EINVAL, nr_cmds = queue->nr_cmds;
1389
1390         cmds = kcalloc(nr_cmds, sizeof(struct nvmet_tcp_cmd), GFP_KERNEL);
1391         if (!cmds)
1392                 goto out;
1393
1394         for (i = 0; i < nr_cmds; i++) {
1395                 ret = nvmet_tcp_alloc_cmd(queue, cmds + i);
1396                 if (ret)
1397                         goto out_free;
1398         }
1399
1400         queue->cmds = cmds;
1401
1402         return 0;
1403 out_free:
1404         while (--i >= 0)
1405                 nvmet_tcp_free_cmd(cmds + i);
1406         kfree(cmds);
1407 out:
1408         return ret;
1409 }
1410
1411 static void nvmet_tcp_free_cmds(struct nvmet_tcp_queue *queue)
1412 {
1413         struct nvmet_tcp_cmd *cmds = queue->cmds;
1414         int i;
1415
1416         for (i = 0; i < queue->nr_cmds; i++)
1417                 nvmet_tcp_free_cmd(cmds + i);
1418
1419         nvmet_tcp_free_cmd(&queue->connect);
1420         kfree(cmds);
1421 }
1422
1423 static void nvmet_tcp_restore_socket_callbacks(struct nvmet_tcp_queue *queue)
1424 {
1425         struct socket *sock = queue->sock;
1426
1427         write_lock_bh(&sock->sk->sk_callback_lock);
1428         sock->sk->sk_data_ready =  queue->data_ready;
1429         sock->sk->sk_state_change = queue->state_change;
1430         sock->sk->sk_write_space = queue->write_space;
1431         sock->sk->sk_user_data = NULL;
1432         write_unlock_bh(&sock->sk->sk_callback_lock);
1433 }
1434
1435 static void nvmet_tcp_uninit_data_in_cmds(struct nvmet_tcp_queue *queue)
1436 {
1437         struct nvmet_tcp_cmd *cmd = queue->cmds;
1438         int i;
1439
1440         for (i = 0; i < queue->nr_cmds; i++, cmd++) {
1441                 if (nvmet_tcp_need_data_in(cmd))
1442                         nvmet_req_uninit(&cmd->req);
1443         }
1444
1445         if (!queue->nr_cmds && nvmet_tcp_need_data_in(&queue->connect)) {
1446                 /* failed in connect */
1447                 nvmet_req_uninit(&queue->connect.req);
1448         }
1449 }
1450
1451 static void nvmet_tcp_free_cmd_data_in_buffers(struct nvmet_tcp_queue *queue)
1452 {
1453         struct nvmet_tcp_cmd *cmd = queue->cmds;
1454         int i;
1455
1456         for (i = 0; i < queue->nr_cmds; i++, cmd++) {
1457                 if (nvmet_tcp_need_data_in(cmd))
1458                         nvmet_tcp_free_cmd_buffers(cmd);
1459         }
1460
1461         if (!queue->nr_cmds && nvmet_tcp_need_data_in(&queue->connect))
1462                 nvmet_tcp_free_cmd_buffers(&queue->connect);
1463 }
1464
1465 static void nvmet_tcp_release_queue_work(struct work_struct *w)
1466 {
1467         struct page *page;
1468         struct nvmet_tcp_queue *queue =
1469                 container_of(w, struct nvmet_tcp_queue, release_work);
1470
1471         mutex_lock(&nvmet_tcp_queue_mutex);
1472         list_del_init(&queue->queue_list);
1473         mutex_unlock(&nvmet_tcp_queue_mutex);
1474
1475         nvmet_tcp_restore_socket_callbacks(queue);
1476         cancel_work_sync(&queue->io_work);
1477         /* stop accepting incoming data */
1478         queue->rcv_state = NVMET_TCP_RECV_ERR;
1479
1480         nvmet_tcp_uninit_data_in_cmds(queue);
1481         nvmet_sq_destroy(&queue->nvme_sq);
1482         cancel_work_sync(&queue->io_work);
1483         nvmet_tcp_free_cmd_data_in_buffers(queue);
1484         sock_release(queue->sock);
1485         nvmet_tcp_free_cmds(queue);
1486         if (queue->hdr_digest || queue->data_digest)
1487                 nvmet_tcp_free_crypto(queue);
1488         ida_free(&nvmet_tcp_queue_ida, queue->idx);
1489
1490         page = virt_to_head_page(queue->pf_cache.va);
1491         __page_frag_cache_drain(page, queue->pf_cache.pagecnt_bias);
1492         kfree(queue);
1493 }
1494
1495 static void nvmet_tcp_data_ready(struct sock *sk)
1496 {
1497         struct nvmet_tcp_queue *queue;
1498
1499         trace_sk_data_ready(sk);
1500
1501         read_lock_bh(&sk->sk_callback_lock);
1502         queue = sk->sk_user_data;
1503         if (likely(queue))
1504                 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work);
1505         read_unlock_bh(&sk->sk_callback_lock);
1506 }
1507
1508 static void nvmet_tcp_write_space(struct sock *sk)
1509 {
1510         struct nvmet_tcp_queue *queue;
1511
1512         read_lock_bh(&sk->sk_callback_lock);
1513         queue = sk->sk_user_data;
1514         if (unlikely(!queue))
1515                 goto out;
1516
1517         if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
1518                 queue->write_space(sk);
1519                 goto out;
1520         }
1521
1522         if (sk_stream_is_writeable(sk)) {
1523                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1524                 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work);
1525         }
1526 out:
1527         read_unlock_bh(&sk->sk_callback_lock);
1528 }
1529
1530 static void nvmet_tcp_state_change(struct sock *sk)
1531 {
1532         struct nvmet_tcp_queue *queue;
1533
1534         read_lock_bh(&sk->sk_callback_lock);
1535         queue = sk->sk_user_data;
1536         if (!queue)
1537                 goto done;
1538
1539         switch (sk->sk_state) {
1540         case TCP_FIN_WAIT2:
1541         case TCP_LAST_ACK:
1542                 break;
1543         case TCP_FIN_WAIT1:
1544         case TCP_CLOSE_WAIT:
1545         case TCP_CLOSE:
1546                 /* FALLTHRU */
1547                 nvmet_tcp_schedule_release_queue(queue);
1548                 break;
1549         default:
1550                 pr_warn("queue %d unhandled state %d\n",
1551                         queue->idx, sk->sk_state);
1552         }
1553 done:
1554         read_unlock_bh(&sk->sk_callback_lock);
1555 }
1556
1557 static int nvmet_tcp_set_queue_sock(struct nvmet_tcp_queue *queue)
1558 {
1559         struct socket *sock = queue->sock;
1560         struct inet_sock *inet = inet_sk(sock->sk);
1561         int ret;
1562
1563         ret = kernel_getsockname(sock,
1564                 (struct sockaddr *)&queue->sockaddr);
1565         if (ret < 0)
1566                 return ret;
1567
1568         ret = kernel_getpeername(sock,
1569                 (struct sockaddr *)&queue->sockaddr_peer);
1570         if (ret < 0)
1571                 return ret;
1572
1573         /*
1574          * Cleanup whatever is sitting in the TCP transmit queue on socket
1575          * close. This is done to prevent stale data from being sent should
1576          * the network connection be restored before TCP times out.
1577          */
1578         sock_no_linger(sock->sk);
1579
1580         if (so_priority > 0)
1581                 sock_set_priority(sock->sk, so_priority);
1582
1583         /* Set socket type of service */
1584         if (inet->rcv_tos > 0)
1585                 ip_sock_set_tos(sock->sk, inet->rcv_tos);
1586
1587         ret = 0;
1588         write_lock_bh(&sock->sk->sk_callback_lock);
1589         if (sock->sk->sk_state != TCP_ESTABLISHED) {
1590                 /*
1591                  * If the socket is already closing, don't even start
1592                  * consuming it
1593                  */
1594                 ret = -ENOTCONN;
1595         } else {
1596                 sock->sk->sk_user_data = queue;
1597                 queue->data_ready = sock->sk->sk_data_ready;
1598                 sock->sk->sk_data_ready = nvmet_tcp_data_ready;
1599                 queue->state_change = sock->sk->sk_state_change;
1600                 sock->sk->sk_state_change = nvmet_tcp_state_change;
1601                 queue->write_space = sock->sk->sk_write_space;
1602                 sock->sk->sk_write_space = nvmet_tcp_write_space;
1603                 if (idle_poll_period_usecs)
1604                         nvmet_tcp_arm_queue_deadline(queue);
1605                 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work);
1606         }
1607         write_unlock_bh(&sock->sk->sk_callback_lock);
1608
1609         return ret;
1610 }
1611
1612 static int nvmet_tcp_alloc_queue(struct nvmet_tcp_port *port,
1613                 struct socket *newsock)
1614 {
1615         struct nvmet_tcp_queue *queue;
1616         int ret;
1617
1618         queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1619         if (!queue)
1620                 return -ENOMEM;
1621
1622         INIT_WORK(&queue->release_work, nvmet_tcp_release_queue_work);
1623         INIT_WORK(&queue->io_work, nvmet_tcp_io_work);
1624         queue->sock = newsock;
1625         queue->port = port;
1626         queue->nr_cmds = 0;
1627         spin_lock_init(&queue->state_lock);
1628         queue->state = NVMET_TCP_Q_CONNECTING;
1629         INIT_LIST_HEAD(&queue->free_list);
1630         init_llist_head(&queue->resp_list);
1631         INIT_LIST_HEAD(&queue->resp_send_list);
1632
1633         queue->idx = ida_alloc(&nvmet_tcp_queue_ida, GFP_KERNEL);
1634         if (queue->idx < 0) {
1635                 ret = queue->idx;
1636                 goto out_free_queue;
1637         }
1638
1639         ret = nvmet_tcp_alloc_cmd(queue, &queue->connect);
1640         if (ret)
1641                 goto out_ida_remove;
1642
1643         ret = nvmet_sq_init(&queue->nvme_sq);
1644         if (ret)
1645                 goto out_free_connect;
1646
1647         nvmet_prepare_receive_pdu(queue);
1648
1649         mutex_lock(&nvmet_tcp_queue_mutex);
1650         list_add_tail(&queue->queue_list, &nvmet_tcp_queue_list);
1651         mutex_unlock(&nvmet_tcp_queue_mutex);
1652
1653         ret = nvmet_tcp_set_queue_sock(queue);
1654         if (ret)
1655                 goto out_destroy_sq;
1656
1657         return 0;
1658 out_destroy_sq:
1659         mutex_lock(&nvmet_tcp_queue_mutex);
1660         list_del_init(&queue->queue_list);
1661         mutex_unlock(&nvmet_tcp_queue_mutex);
1662         nvmet_sq_destroy(&queue->nvme_sq);
1663 out_free_connect:
1664         nvmet_tcp_free_cmd(&queue->connect);
1665 out_ida_remove:
1666         ida_free(&nvmet_tcp_queue_ida, queue->idx);
1667 out_free_queue:
1668         kfree(queue);
1669         return ret;
1670 }
1671
1672 static void nvmet_tcp_accept_work(struct work_struct *w)
1673 {
1674         struct nvmet_tcp_port *port =
1675                 container_of(w, struct nvmet_tcp_port, accept_work);
1676         struct socket *newsock;
1677         int ret;
1678
1679         while (true) {
1680                 ret = kernel_accept(port->sock, &newsock, O_NONBLOCK);
1681                 if (ret < 0) {
1682                         if (ret != -EAGAIN)
1683                                 pr_warn("failed to accept err=%d\n", ret);
1684                         return;
1685                 }
1686                 ret = nvmet_tcp_alloc_queue(port, newsock);
1687                 if (ret) {
1688                         pr_err("failed to allocate queue\n");
1689                         sock_release(newsock);
1690                 }
1691         }
1692 }
1693
1694 static void nvmet_tcp_listen_data_ready(struct sock *sk)
1695 {
1696         struct nvmet_tcp_port *port;
1697
1698         trace_sk_data_ready(sk);
1699
1700         read_lock_bh(&sk->sk_callback_lock);
1701         port = sk->sk_user_data;
1702         if (!port)
1703                 goto out;
1704
1705         if (sk->sk_state == TCP_LISTEN)
1706                 queue_work(nvmet_wq, &port->accept_work);
1707 out:
1708         read_unlock_bh(&sk->sk_callback_lock);
1709 }
1710
1711 static int nvmet_tcp_add_port(struct nvmet_port *nport)
1712 {
1713         struct nvmet_tcp_port *port;
1714         __kernel_sa_family_t af;
1715         int ret;
1716
1717         port = kzalloc(sizeof(*port), GFP_KERNEL);
1718         if (!port)
1719                 return -ENOMEM;
1720
1721         switch (nport->disc_addr.adrfam) {
1722         case NVMF_ADDR_FAMILY_IP4:
1723                 af = AF_INET;
1724                 break;
1725         case NVMF_ADDR_FAMILY_IP6:
1726                 af = AF_INET6;
1727                 break;
1728         default:
1729                 pr_err("address family %d not supported\n",
1730                                 nport->disc_addr.adrfam);
1731                 ret = -EINVAL;
1732                 goto err_port;
1733         }
1734
1735         ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
1736                         nport->disc_addr.trsvcid, &port->addr);
1737         if (ret) {
1738                 pr_err("malformed ip/port passed: %s:%s\n",
1739                         nport->disc_addr.traddr, nport->disc_addr.trsvcid);
1740                 goto err_port;
1741         }
1742
1743         port->nport = nport;
1744         INIT_WORK(&port->accept_work, nvmet_tcp_accept_work);
1745         if (port->nport->inline_data_size < 0)
1746                 port->nport->inline_data_size = NVMET_TCP_DEF_INLINE_DATA_SIZE;
1747
1748         ret = sock_create(port->addr.ss_family, SOCK_STREAM,
1749                                 IPPROTO_TCP, &port->sock);
1750         if (ret) {
1751                 pr_err("failed to create a socket\n");
1752                 goto err_port;
1753         }
1754
1755         port->sock->sk->sk_user_data = port;
1756         port->data_ready = port->sock->sk->sk_data_ready;
1757         port->sock->sk->sk_data_ready = nvmet_tcp_listen_data_ready;
1758         sock_set_reuseaddr(port->sock->sk);
1759         tcp_sock_set_nodelay(port->sock->sk);
1760         if (so_priority > 0)
1761                 sock_set_priority(port->sock->sk, so_priority);
1762
1763         ret = kernel_bind(port->sock, (struct sockaddr *)&port->addr,
1764                         sizeof(port->addr));
1765         if (ret) {
1766                 pr_err("failed to bind port socket %d\n", ret);
1767                 goto err_sock;
1768         }
1769
1770         ret = kernel_listen(port->sock, 128);
1771         if (ret) {
1772                 pr_err("failed to listen %d on port sock\n", ret);
1773                 goto err_sock;
1774         }
1775
1776         nport->priv = port;
1777         pr_info("enabling port %d (%pISpc)\n",
1778                 le16_to_cpu(nport->disc_addr.portid), &port->addr);
1779
1780         return 0;
1781
1782 err_sock:
1783         sock_release(port->sock);
1784 err_port:
1785         kfree(port);
1786         return ret;
1787 }
1788
1789 static void nvmet_tcp_destroy_port_queues(struct nvmet_tcp_port *port)
1790 {
1791         struct nvmet_tcp_queue *queue;
1792
1793         mutex_lock(&nvmet_tcp_queue_mutex);
1794         list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1795                 if (queue->port == port)
1796                         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1797         mutex_unlock(&nvmet_tcp_queue_mutex);
1798 }
1799
1800 static void nvmet_tcp_remove_port(struct nvmet_port *nport)
1801 {
1802         struct nvmet_tcp_port *port = nport->priv;
1803
1804         write_lock_bh(&port->sock->sk->sk_callback_lock);
1805         port->sock->sk->sk_data_ready = port->data_ready;
1806         port->sock->sk->sk_user_data = NULL;
1807         write_unlock_bh(&port->sock->sk->sk_callback_lock);
1808         cancel_work_sync(&port->accept_work);
1809         /*
1810          * Destroy the remaining queues, which are not belong to any
1811          * controller yet.
1812          */
1813         nvmet_tcp_destroy_port_queues(port);
1814
1815         sock_release(port->sock);
1816         kfree(port);
1817 }
1818
1819 static void nvmet_tcp_delete_ctrl(struct nvmet_ctrl *ctrl)
1820 {
1821         struct nvmet_tcp_queue *queue;
1822
1823         mutex_lock(&nvmet_tcp_queue_mutex);
1824         list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1825                 if (queue->nvme_sq.ctrl == ctrl)
1826                         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1827         mutex_unlock(&nvmet_tcp_queue_mutex);
1828 }
1829
1830 static u16 nvmet_tcp_install_queue(struct nvmet_sq *sq)
1831 {
1832         struct nvmet_tcp_queue *queue =
1833                 container_of(sq, struct nvmet_tcp_queue, nvme_sq);
1834
1835         if (sq->qid == 0) {
1836                 /* Let inflight controller teardown complete */
1837                 flush_workqueue(nvmet_wq);
1838         }
1839
1840         queue->nr_cmds = sq->size * 2;
1841         if (nvmet_tcp_alloc_cmds(queue))
1842                 return NVME_SC_INTERNAL;
1843         return 0;
1844 }
1845
1846 static void nvmet_tcp_disc_port_addr(struct nvmet_req *req,
1847                 struct nvmet_port *nport, char *traddr)
1848 {
1849         struct nvmet_tcp_port *port = nport->priv;
1850
1851         if (inet_addr_is_any((struct sockaddr *)&port->addr)) {
1852                 struct nvmet_tcp_cmd *cmd =
1853                         container_of(req, struct nvmet_tcp_cmd, req);
1854                 struct nvmet_tcp_queue *queue = cmd->queue;
1855
1856                 sprintf(traddr, "%pISc", (struct sockaddr *)&queue->sockaddr);
1857         } else {
1858                 memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
1859         }
1860 }
1861
1862 static const struct nvmet_fabrics_ops nvmet_tcp_ops = {
1863         .owner                  = THIS_MODULE,
1864         .type                   = NVMF_TRTYPE_TCP,
1865         .msdbd                  = 1,
1866         .add_port               = nvmet_tcp_add_port,
1867         .remove_port            = nvmet_tcp_remove_port,
1868         .queue_response         = nvmet_tcp_queue_response,
1869         .delete_ctrl            = nvmet_tcp_delete_ctrl,
1870         .install_queue          = nvmet_tcp_install_queue,
1871         .disc_traddr            = nvmet_tcp_disc_port_addr,
1872 };
1873
1874 static int __init nvmet_tcp_init(void)
1875 {
1876         int ret;
1877
1878         nvmet_tcp_wq = alloc_workqueue("nvmet_tcp_wq",
1879                                 WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
1880         if (!nvmet_tcp_wq)
1881                 return -ENOMEM;
1882
1883         ret = nvmet_register_transport(&nvmet_tcp_ops);
1884         if (ret)
1885                 goto err;
1886
1887         return 0;
1888 err:
1889         destroy_workqueue(nvmet_tcp_wq);
1890         return ret;
1891 }
1892
1893 static void __exit nvmet_tcp_exit(void)
1894 {
1895         struct nvmet_tcp_queue *queue;
1896
1897         nvmet_unregister_transport(&nvmet_tcp_ops);
1898
1899         flush_workqueue(nvmet_wq);
1900         mutex_lock(&nvmet_tcp_queue_mutex);
1901         list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1902                 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1903         mutex_unlock(&nvmet_tcp_queue_mutex);
1904         flush_workqueue(nvmet_wq);
1905
1906         destroy_workqueue(nvmet_tcp_wq);
1907 }
1908
1909 module_init(nvmet_tcp_init);
1910 module_exit(nvmet_tcp_exit);
1911
1912 MODULE_LICENSE("GPL v2");
1913 MODULE_ALIAS("nvmet-transport-3"); /* 3 == NVMF_TRTYPE_TCP */