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