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