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 NVMET_TCP_RECV_BUDGET           8
  23 #define NVMET_TCP_SEND_BUDGET           8
  24 #define NVMET_TCP_IO_WORK_BUDGET        64
  25
  26 enum nvmet_tcp_send_state {
  27         NVMET_TCP_SEND_DATA_PDU,
  28         NVMET_TCP_SEND_DATA,
  29         NVMET_TCP_SEND_R2T,
  30         NVMET_TCP_SEND_DDGST,
  31         NVMET_TCP_SEND_RESPONSE
  32 };
  33
  34 enum nvmet_tcp_recv_state {
  35         NVMET_TCP_RECV_PDU,
  36         NVMET_TCP_RECV_DATA,
  37         NVMET_TCP_RECV_DDGST,
  38         NVMET_TCP_RECV_ERR,
  39 };
  40
  41 enum {
  42         NVMET_TCP_F_INIT_FAILED = (1 << 0),
  43 };
  44
  45 struct nvmet_tcp_cmd {
  46         struct nvmet_tcp_queue          *queue;
  47         struct nvmet_req                req;
  48
  49         struct nvme_tcp_cmd_pdu         *cmd_pdu;
  50         struct nvme_tcp_rsp_pdu         *rsp_pdu;
  51         struct nvme_tcp_data_pdu        *data_pdu;
  52         struct nvme_tcp_r2t_pdu         *r2t_pdu;
  53
  54         u32                             rbytes_done;
  55         u32                             wbytes_done;
  56
  57         u32                             pdu_len;
  58         u32                             pdu_recv;
  59         int                             sg_idx;
  60         int                             nr_mapped;
  61         struct msghdr                   recv_msg;
  62         struct kvec                     *iov;
  63         u32                             flags;
  64
  65         struct list_head                entry;
  66         struct llist_node               lentry;
  67
  68         /* send state */
  69         u32                             offset;
  70         struct scatterlist              *cur_sg;
  71         enum nvmet_tcp_send_state       state;
  72
  73         __le32                          exp_ddgst;
  74         __le32                          recv_ddgst;
  75 };
  76
  77 enum nvmet_tcp_queue_state {
  78         NVMET_TCP_Q_CONNECTING,
  79         NVMET_TCP_Q_LIVE,
  80         NVMET_TCP_Q_DISCONNECTING,
  81 };
  82
  83 struct nvmet_tcp_queue {
  84         struct socket           *sock;
  85         struct nvmet_tcp_port   *port;
  86         struct work_struct      io_work;
  87         int                     cpu;
  88         struct nvmet_cq         nvme_cq;
  89         struct nvmet_sq         nvme_sq;
  90
  91         /* send state */
  92         struct nvmet_tcp_cmd    *cmds;
  93         unsigned int            nr_cmds;
  94         struct list_head        free_list;
  95         struct llist_head       resp_list;
  96         struct list_head        resp_send_list;
  97         int                     send_list_len;
  98         struct nvmet_tcp_cmd    *snd_cmd;
  99
 100         /* recv state */
 101         int                     offset;
 102         int                     left;
 103         enum nvmet_tcp_recv_state rcv_state;
 104         struct nvmet_tcp_cmd    *cmd;
 105         union nvme_tcp_pdu      pdu;
 106
 107         /* digest state */
 108         bool                    hdr_digest;
 109         bool                    data_digest;
 110         struct ahash_request    *snd_hash;
 111         struct ahash_request    *rcv_hash;
 112
 113         spinlock_t              state_lock;
 114         enum nvmet_tcp_queue_state state;
 115
 116         struct sockaddr_storage sockaddr;
 117         struct sockaddr_storage sockaddr_peer;
 118         struct work_struct      release_work;
 119
 120         int                     idx;
 121         struct list_head        queue_list;
 122
 123         struct nvmet_tcp_cmd    connect;
 124
 125         struct page_frag_cache  pf_cache;
 126
 127         void (*data_ready)(struct sock *);
 128         void (*state_change)(struct sock *);
 129         void (*write_space)(struct sock *);
 130 };
 131
 132 struct nvmet_tcp_port {
 133         struct socket           *sock;
 134         struct work_struct      accept_work;
 135         struct nvmet_port       *nport;
 136         struct sockaddr_storage addr;
 137         int                     last_cpu;
 138         void (*data_ready)(struct sock *);
 139 };
 140
 141 static DEFINE_IDA(nvmet_tcp_queue_ida);
 142 static LIST_HEAD(nvmet_tcp_queue_list);
 143 static DEFINE_MUTEX(nvmet_tcp_queue_mutex);
 144
 145 static struct workqueue_struct *nvmet_tcp_wq;
 146 static struct nvmet_fabrics_ops nvmet_tcp_ops;
 147 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c);
 148 static void nvmet_tcp_finish_cmd(struct nvmet_tcp_cmd *cmd);
 149
 150 static inline u16 nvmet_tcp_cmd_tag(struct nvmet_tcp_queue *queue,
 151                 struct nvmet_tcp_cmd *cmd)
 152 {
 153         return cmd - queue->cmds;
 154 }
 155
 156 static inline bool nvmet_tcp_has_data_in(struct nvmet_tcp_cmd *cmd)
 157 {
 158         return nvme_is_write(cmd->req.cmd) &&
 159                 cmd->rbytes_done < cmd->req.transfer_len;
 160 }
 161
 162 static inline bool nvmet_tcp_need_data_in(struct nvmet_tcp_cmd *cmd)
 163 {
 164         return nvmet_tcp_has_data_in(cmd) && !cmd->req.rsp->status;
 165 }
 166
 167 static inline bool nvmet_tcp_need_data_out(struct nvmet_tcp_cmd *cmd)
 168 {
 169         return !nvme_is_write(cmd->req.cmd) &&
 170                 cmd->req.transfer_len > 0 &&
 171                 !cmd->req.rsp->status;
 172 }
 173
 174 static inline bool nvmet_tcp_has_inline_data(struct nvmet_tcp_cmd *cmd)
 175 {
 176         return nvme_is_write(cmd->req.cmd) && cmd->pdu_len &&
 177                 !cmd->rbytes_done;
 178 }
 179
 180 static inline struct nvmet_tcp_cmd *
 181 nvmet_tcp_get_cmd(struct nvmet_tcp_queue *queue)
 182 {
 183         struct nvmet_tcp_cmd *cmd;
 184
 185         cmd = list_first_entry_or_null(&queue->free_list,
 186                                 struct nvmet_tcp_cmd, entry);
 187         if (!cmd)
 188                 return NULL;
 189         list_del_init(&cmd->entry);
 190
 191         cmd->rbytes_done = cmd->wbytes_done = 0;
 192         cmd->pdu_len = 0;
 193         cmd->pdu_recv = 0;
 194         cmd->iov = NULL;
 195         cmd->flags = 0;
 196         return cmd;
 197 }
 198
 199 static inline void nvmet_tcp_put_cmd(struct nvmet_tcp_cmd *cmd)
 200 {
 201         if (unlikely(cmd == &cmd->queue->connect))
 202                 return;
 203
 204         list_add_tail(&cmd->entry, &cmd->queue->free_list);
 205 }
 206
 207 static inline u8 nvmet_tcp_hdgst_len(struct nvmet_tcp_queue *queue)
 208 {
 209         return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 210 }
 211
 212 static inline u8 nvmet_tcp_ddgst_len(struct nvmet_tcp_queue *queue)
 213 {
 214         return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 215 }
 216
 217 static inline void nvmet_tcp_hdgst(struct ahash_request *hash,
 218                 void *pdu, size_t len)
 219 {
 220         struct scatterlist sg;
 221
 222         sg_init_one(&sg, pdu, len);
 223         ahash_request_set_crypt(hash, &sg, pdu + len, len);
 224         crypto_ahash_digest(hash);
 225 }
 226
 227 static int nvmet_tcp_verify_hdgst(struct nvmet_tcp_queue *queue,
 228         void *pdu, size_t len)
 229 {
 230         struct nvme_tcp_hdr *hdr = pdu;
 231         __le32 recv_digest;
 232         __le32 exp_digest;
 233
 234         if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
 235                 pr_err("queue %d: header digest enabled but no header digest\n",
 236                         queue->idx);
 237                 return -EPROTO;
 238         }
 239
 240         recv_digest = *(__le32 *)(pdu + hdr->hlen);
 241         nvmet_tcp_hdgst(queue->rcv_hash, pdu, len);
 242         exp_digest = *(__le32 *)(pdu + hdr->hlen);
 243         if (recv_digest != exp_digest) {
 244                 pr_err("queue %d: header digest error: recv %#x expected %#x\n",
 245                         queue->idx, le32_to_cpu(recv_digest),
 246                         le32_to_cpu(exp_digest));
 247                 return -EPROTO;
 248         }
 249
 250         return 0;
 251 }
 252
 253 static int nvmet_tcp_check_ddgst(struct nvmet_tcp_queue *queue, void *pdu)
 254 {
 255         struct nvme_tcp_hdr *hdr = pdu;
 256         u8 digest_len = nvmet_tcp_hdgst_len(queue);
 257         u32 len;
 258
 259         len = le32_to_cpu(hdr->plen) - hdr->hlen -
 260                 (hdr->flags & NVME_TCP_F_HDGST ? digest_len : 0);
 261
 262         if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
 263                 pr_err("queue %d: data digest flag is cleared\n", queue->idx);
 264                 return -EPROTO;
 265         }
 266
 267         return 0;
 268 }
 269
 270 static void nvmet_tcp_unmap_pdu_iovec(struct nvmet_tcp_cmd *cmd)
 271 {
 272         struct scatterlist *sg;
 273         int i;
 274
 275         sg = &cmd->req.sg[cmd->sg_idx];
 276
 277         for (i = 0; i < cmd->nr_mapped; i++)
 278                 kunmap(sg_page(&sg[i]));
 279 }
 280
 281 static void nvmet_tcp_map_pdu_iovec(struct nvmet_tcp_cmd *cmd)
 282 {
 283         struct kvec *iov = cmd->iov;
 284         struct scatterlist *sg;
 285         u32 length, offset, sg_offset;
 286
 287         length = cmd->pdu_len;
 288         cmd->nr_mapped = DIV_ROUND_UP(length, PAGE_SIZE);
 289         offset = cmd->rbytes_done;
 290         cmd->sg_idx = DIV_ROUND_UP(offset, PAGE_SIZE);
 291         sg_offset = offset % PAGE_SIZE;
 292         sg = &cmd->req.sg[cmd->sg_idx];
 293
 294         while (length) {
 295                 u32 iov_len = min_t(u32, length, sg->length - sg_offset);
 296
 297                 iov->iov_base = kmap(sg_page(sg)) + sg->offset + sg_offset;
 298                 iov->iov_len = iov_len;
 299
 300                 length -= iov_len;
 301                 sg = sg_next(sg);
 302                 iov++;
 303         }
 304
 305         iov_iter_kvec(&cmd->recv_msg.msg_iter, READ, cmd->iov,
 306                 cmd->nr_mapped, cmd->pdu_len);
 307 }
 308
 309 static void nvmet_tcp_fatal_error(struct nvmet_tcp_queue *queue)
 310 {
 311         queue->rcv_state = NVMET_TCP_RECV_ERR;
 312         if (queue->nvme_sq.ctrl)
 313                 nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
 314         else
 315                 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
 316 }
 317
 318 static int nvmet_tcp_map_data(struct nvmet_tcp_cmd *cmd)
 319 {
 320         struct nvme_sgl_desc *sgl = &cmd->req.cmd->common.dptr.sgl;
 321         u32 len = le32_to_cpu(sgl->length);
 322
 323         if (!cmd->req.data_len)
 324                 return 0;
 325
 326         if (sgl->type == ((NVME_SGL_FMT_DATA_DESC << 4) |
 327                           NVME_SGL_FMT_OFFSET)) {
 328                 if (!nvme_is_write(cmd->req.cmd))
 329                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 330
 331                 if (len > cmd->req.port->inline_data_size)
 332                         return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
 333                 cmd->pdu_len = len;
 334         }
 335         cmd->req.transfer_len += len;
 336
 337         cmd->req.sg = sgl_alloc(len, GFP_KERNEL, &cmd->req.sg_cnt);
 338         if (!cmd->req.sg)
 339                 return NVME_SC_INTERNAL;
 340         cmd->cur_sg = cmd->req.sg;
 341
 342         if (nvmet_tcp_has_data_in(cmd)) {
 343                 cmd->iov = kmalloc_array(cmd->req.sg_cnt,
 344                                 sizeof(*cmd->iov), GFP_KERNEL);
 345                 if (!cmd->iov)
 346                         goto err;
 347         }
 348
 349         return 0;
 350 err:
 351         sgl_free(cmd->req.sg);
 352         return NVME_SC_INTERNAL;
 353 }
 354
 355 static void nvmet_tcp_ddgst(struct ahash_request *hash,
 356                 struct nvmet_tcp_cmd *cmd)
 357 {
 358         ahash_request_set_crypt(hash, cmd->req.sg,
 359                 (void *)&cmd->exp_ddgst, cmd->req.transfer_len);
 360         crypto_ahash_digest(hash);
 361 }
 362
 363 static void nvmet_setup_c2h_data_pdu(struct nvmet_tcp_cmd *cmd)
 364 {
 365         struct nvme_tcp_data_pdu *pdu = cmd->data_pdu;
 366         struct nvmet_tcp_queue *queue = cmd->queue;
 367         u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
 368         u8 ddgst = nvmet_tcp_ddgst_len(cmd->queue);
 369
 370         cmd->offset = 0;
 371         cmd->state = NVMET_TCP_SEND_DATA_PDU;
 372
 373         pdu->hdr.type = nvme_tcp_c2h_data;
 374         pdu->hdr.flags = NVME_TCP_F_DATA_LAST;
 375         pdu->hdr.hlen = sizeof(*pdu);
 376         pdu->hdr.pdo = pdu->hdr.hlen + hdgst;
 377         pdu->hdr.plen =
 378                 cpu_to_le32(pdu->hdr.hlen + hdgst +
 379                                 cmd->req.transfer_len + ddgst);
 380         pdu->command_id = cmd->req.rsp->command_id;
 381         pdu->data_length = cpu_to_le32(cmd->req.transfer_len);
 382         pdu->data_offset = cpu_to_le32(cmd->wbytes_done);
 383
 384         if (queue->data_digest) {
 385                 pdu->hdr.flags |= NVME_TCP_F_DDGST;
 386                 nvmet_tcp_ddgst(queue->snd_hash, cmd);
 387         }
 388
 389         if (cmd->queue->hdr_digest) {
 390                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
 391                 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
 392         }
 393 }
 394
 395 static void nvmet_setup_r2t_pdu(struct nvmet_tcp_cmd *cmd)
 396 {
 397         struct nvme_tcp_r2t_pdu *pdu = cmd->r2t_pdu;
 398         struct nvmet_tcp_queue *queue = cmd->queue;
 399         u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
 400
 401         cmd->offset = 0;
 402         cmd->state = NVMET_TCP_SEND_R2T;
 403
 404         pdu->hdr.type = nvme_tcp_r2t;
 405         pdu->hdr.flags = 0;
 406         pdu->hdr.hlen = sizeof(*pdu);
 407         pdu->hdr.pdo = 0;
 408         pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
 409
 410         pdu->command_id = cmd->req.cmd->common.command_id;
 411         pdu->ttag = nvmet_tcp_cmd_tag(cmd->queue, cmd);
 412         pdu->r2t_length = cpu_to_le32(cmd->req.transfer_len - cmd->rbytes_done);
 413         pdu->r2t_offset = cpu_to_le32(cmd->rbytes_done);
 414         if (cmd->queue->hdr_digest) {
 415                 pdu->hdr.flags |= NVME_TCP_F_HDGST;
 416                 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
 417         }
 418 }
 419
 420 static void nvmet_setup_response_pdu(struct nvmet_tcp_cmd *cmd)
 421 {
 422         struct nvme_tcp_rsp_pdu *pdu = cmd->rsp_pdu;
 423         struct nvmet_tcp_queue *queue = cmd->queue;
 424         u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
 425
 426         cmd->offset = 0;
 427         cmd->state = NVMET_TCP_SEND_RESPONSE;
 428
 429         pdu->hdr.type = nvme_tcp_rsp;
 430         pdu->hdr.flags = 0;
 431         pdu->hdr.hlen = sizeof(*pdu);
 432         pdu->hdr.pdo = 0;
 433         pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
 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_tcp_process_resp_list(struct nvmet_tcp_queue *queue)
 441 {
 442         struct llist_node *node;
 443
 444         node = llist_del_all(&queue->resp_list);
 445         if (!node)
 446                 return;
 447
 448         while (node) {
 449                 struct nvmet_tcp_cmd *cmd = llist_entry(node,
 450                                         struct nvmet_tcp_cmd, lentry);
 451
 452                 list_add(&cmd->entry, &queue->resp_send_list);
 453                 node = node->next;
 454                 queue->send_list_len++;
 455         }
 456 }
 457
 458 static struct nvmet_tcp_cmd *nvmet_tcp_fetch_cmd(struct nvmet_tcp_queue *queue)
 459 {
 460         queue->snd_cmd = list_first_entry_or_null(&queue->resp_send_list,
 461                                 struct nvmet_tcp_cmd, entry);
 462         if (!queue->snd_cmd) {
 463                 nvmet_tcp_process_resp_list(queue);
 464                 queue->snd_cmd =
 465                         list_first_entry_or_null(&queue->resp_send_list,
 466                                         struct nvmet_tcp_cmd, entry);
 467                 if (unlikely(!queue->snd_cmd))
 468                         return NULL;
 469         }
 470
 471         list_del_init(&queue->snd_cmd->entry);
 472         queue->send_list_len--;
 473
 474         if (nvmet_tcp_need_data_out(queue->snd_cmd))
 475                 nvmet_setup_c2h_data_pdu(queue->snd_cmd);
 476         else if (nvmet_tcp_need_data_in(queue->snd_cmd))
 477                 nvmet_setup_r2t_pdu(queue->snd_cmd);
 478         else
 479                 nvmet_setup_response_pdu(queue->snd_cmd);
 480
 481         return queue->snd_cmd;
 482 }
 483
 484 static void nvmet_tcp_queue_response(struct nvmet_req *req)
 485 {
 486         struct nvmet_tcp_cmd *cmd =
 487                 container_of(req, struct nvmet_tcp_cmd, req);
 488         struct nvmet_tcp_queue  *queue = cmd->queue;
 489
 490         llist_add(&cmd->lentry, &queue->resp_list);
 491         queue_work_on(cmd->queue->cpu, nvmet_tcp_wq, &cmd->queue->io_work);
 492 }
 493
 494 static int nvmet_try_send_data_pdu(struct nvmet_tcp_cmd *cmd)
 495 {
 496         u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
 497         int left = sizeof(*cmd->data_pdu) - cmd->offset + hdgst;
 498         int ret;
 499
 500         ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->data_pdu),
 501                         offset_in_page(cmd->data_pdu) + cmd->offset,
 502                         left, MSG_DONTWAIT | MSG_MORE);
 503         if (ret <= 0)
 504                 return ret;
 505
 506         cmd->offset += ret;
 507         left -= ret;
 508
 509         if (left)
 510                 return -EAGAIN;
 511
 512         cmd->state = NVMET_TCP_SEND_DATA;
 513         cmd->offset  = 0;
 514         return 1;
 515 }
 516
 517 static int nvmet_try_send_data(struct nvmet_tcp_cmd *cmd)
 518 {
 519         struct nvmet_tcp_queue *queue = cmd->queue;
 520         int ret;
 521
 522         while (cmd->cur_sg) {
 523                 struct page *page = sg_page(cmd->cur_sg);
 524                 u32 left = cmd->cur_sg->length - cmd->offset;
 525
 526                 ret = kernel_sendpage(cmd->queue->sock, page, cmd->offset,
 527                                         left, MSG_DONTWAIT | MSG_MORE);
 528                 if (ret <= 0)
 529                         return ret;
 530
 531                 cmd->offset += ret;
 532                 cmd->wbytes_done += ret;
 533
 534                 /* Done with sg?*/
 535                 if (cmd->offset == cmd->cur_sg->length) {
 536                         cmd->cur_sg = sg_next(cmd->cur_sg);
 537                         cmd->offset = 0;
 538                 }
 539         }
 540
 541         if (queue->data_digest) {
 542                 cmd->state = NVMET_TCP_SEND_DDGST;
 543                 cmd->offset = 0;
 544         } else {
 545                 nvmet_setup_response_pdu(cmd);
 546         }
 547         return 1;
 548
 549 }
 550
 551 static int nvmet_try_send_response(struct nvmet_tcp_cmd *cmd,
 552                 bool last_in_batch)
 553 {
 554         u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
 555         int left = sizeof(*cmd->rsp_pdu) - cmd->offset + hdgst;
 556         int flags = MSG_DONTWAIT;
 557         int ret;
 558
 559         if (!last_in_batch && cmd->queue->send_list_len)
 560                 flags |= MSG_MORE;
 561         else
 562                 flags |= MSG_EOR;
 563
 564         ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->rsp_pdu),
 565                 offset_in_page(cmd->rsp_pdu) + cmd->offset, left, flags);
 566         if (ret <= 0)
 567                 return ret;
 568         cmd->offset += ret;
 569         left -= ret;
 570
 571         if (left)
 572                 return -EAGAIN;
 573
 574         kfree(cmd->iov);
 575         sgl_free(cmd->req.sg);
 576         cmd->queue->snd_cmd = NULL;
 577         nvmet_tcp_put_cmd(cmd);
 578         return 1;
 579 }
 580
 581 static int nvmet_try_send_r2t(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
 582 {
 583         u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
 584         int left = sizeof(*cmd->r2t_pdu) - cmd->offset + hdgst;
 585         int flags = MSG_DONTWAIT;
 586         int ret;
 587
 588         if (!last_in_batch && cmd->queue->send_list_len)
 589                 flags |= MSG_MORE;
 590         else
 591                 flags |= MSG_EOR;
 592
 593         ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->r2t_pdu),
 594                 offset_in_page(cmd->r2t_pdu) + cmd->offset, left, flags);
 595         if (ret <= 0)
 596                 return ret;
 597         cmd->offset += ret;
 598         left -= ret;
 599
 600         if (left)
 601                 return -EAGAIN;
 602
 603         cmd->queue->snd_cmd = NULL;
 604         return 1;
 605 }
 606
 607 static int nvmet_try_send_ddgst(struct nvmet_tcp_cmd *cmd)
 608 {
 609         struct nvmet_tcp_queue *queue = cmd->queue;
 610         struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
 611         struct kvec iov = {
 612                 .iov_base = &cmd->exp_ddgst + cmd->offset,
 613                 .iov_len = NVME_TCP_DIGEST_LENGTH - cmd->offset
 614         };
 615         int ret;
 616
 617         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
 618         if (unlikely(ret <= 0))
 619                 return ret;
 620
 621         cmd->offset += ret;
 622         nvmet_setup_response_pdu(cmd);
 623         return 1;
 624 }
 625
 626 static int nvmet_tcp_try_send_one(struct nvmet_tcp_queue *queue,
 627                 bool last_in_batch)
 628 {
 629         struct nvmet_tcp_cmd *cmd = queue->snd_cmd;
 630         int ret = 0;
 631
 632         if (!cmd || queue->state == NVMET_TCP_Q_DISCONNECTING) {
 633                 cmd = nvmet_tcp_fetch_cmd(queue);
 634                 if (unlikely(!cmd))
 635                         return 0;
 636         }
 637
 638         if (cmd->state == NVMET_TCP_SEND_DATA_PDU) {
 639                 ret = nvmet_try_send_data_pdu(cmd);
 640                 if (ret <= 0)
 641                         goto done_send;
 642         }
 643
 644         if (cmd->state == NVMET_TCP_SEND_DATA) {
 645                 ret = nvmet_try_send_data(cmd);
 646                 if (ret <= 0)
 647                         goto done_send;
 648         }
 649
 650         if (cmd->state == NVMET_TCP_SEND_DDGST) {
 651                 ret = nvmet_try_send_ddgst(cmd);
 652                 if (ret <= 0)
 653                         goto done_send;
 654         }
 655
 656         if (cmd->state == NVMET_TCP_SEND_R2T) {
 657                 ret = nvmet_try_send_r2t(cmd, last_in_batch);
 658                 if (ret <= 0)
 659                         goto done_send;
 660         }
 661
 662         if (cmd->state == NVMET_TCP_SEND_RESPONSE)
 663                 ret = nvmet_try_send_response(cmd, last_in_batch);
 664
 665 done_send:
 666         if (ret < 0) {
 667                 if (ret == -EAGAIN)
 668                         return 0;
 669                 return ret;
 670         }
 671
 672         return 1;
 673 }
 674
 675 static int nvmet_tcp_try_send(struct nvmet_tcp_queue *queue,
 676                 int budget, int *sends)
 677 {
 678         int i, ret = 0;
 679
 680         for (i = 0; i < budget; i++) {
 681                 ret = nvmet_tcp_try_send_one(queue, i == budget - 1);
 682                 if (ret <= 0)
 683                         break;
 684                 (*sends)++;
 685         }
 686
 687         return ret;
 688 }
 689
 690 static void nvmet_prepare_receive_pdu(struct nvmet_tcp_queue *queue)
 691 {
 692         queue->offset = 0;
 693         queue->left = sizeof(struct nvme_tcp_hdr);
 694         queue->cmd = NULL;
 695         queue->rcv_state = NVMET_TCP_RECV_PDU;
 696 }
 697
 698 static void nvmet_tcp_free_crypto(struct nvmet_tcp_queue *queue)
 699 {
 700         struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
 701
 702         ahash_request_free(queue->rcv_hash);
 703         ahash_request_free(queue->snd_hash);
 704         crypto_free_ahash(tfm);
 705 }
 706
 707 static int nvmet_tcp_alloc_crypto(struct nvmet_tcp_queue *queue)
 708 {
 709         struct crypto_ahash *tfm;
 710
 711         tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
 712         if (IS_ERR(tfm))
 713                 return PTR_ERR(tfm);
 714
 715         queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
 716         if (!queue->snd_hash)
 717                 goto free_tfm;
 718         ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
 719
 720         queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
 721         if (!queue->rcv_hash)
 722                 goto free_snd_hash;
 723         ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
 724
 725         return 0;
 726 free_snd_hash:
 727         ahash_request_free(queue->snd_hash);
 728 free_tfm:
 729         crypto_free_ahash(tfm);
 730         return -ENOMEM;
 731 }
 732
 733
 734 static int nvmet_tcp_handle_icreq(struct nvmet_tcp_queue *queue)
 735 {
 736         struct nvme_tcp_icreq_pdu *icreq = &queue->pdu.icreq;
 737         struct nvme_tcp_icresp_pdu *icresp = &queue->pdu.icresp;
 738         struct msghdr msg = {};
 739         struct kvec iov;
 740         int ret;
 741
 742         if (le32_to_cpu(icreq->hdr.plen) != sizeof(struct nvme_tcp_icreq_pdu)) {
 743                 pr_err("bad nvme-tcp pdu length (%d)\n",
 744                         le32_to_cpu(icreq->hdr.plen));
 745                 nvmet_tcp_fatal_error(queue);
 746         }
 747
 748         if (icreq->pfv != NVME_TCP_PFV_1_0) {
 749                 pr_err("queue %d: bad pfv %d\n", queue->idx, icreq->pfv);
 750                 return -EPROTO;
 751         }
 752
 753         if (icreq->hpda != 0) {
 754                 pr_err("queue %d: unsupported hpda %d\n", queue->idx,
 755                         icreq->hpda);
 756                 return -EPROTO;
 757         }
 758
 759         if (icreq->maxr2t != 0) {
 760                 pr_err("queue %d: unsupported maxr2t %d\n", queue->idx,
 761                         le32_to_cpu(icreq->maxr2t) + 1);
 762                 return -EPROTO;
 763         }
 764
 765         queue->hdr_digest = !!(icreq->digest & NVME_TCP_HDR_DIGEST_ENABLE);
 766         queue->data_digest = !!(icreq->digest & NVME_TCP_DATA_DIGEST_ENABLE);
 767         if (queue->hdr_digest || queue->data_digest) {
 768                 ret = nvmet_tcp_alloc_crypto(queue);
 769                 if (ret)
 770                         return ret;
 771         }
 772
 773         memset(icresp, 0, sizeof(*icresp));
 774         icresp->hdr.type = nvme_tcp_icresp;
 775         icresp->hdr.hlen = sizeof(*icresp);
 776         icresp->hdr.pdo = 0;
 777         icresp->hdr.plen = cpu_to_le32(icresp->hdr.hlen);
 778         icresp->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
 779         icresp->maxdata = cpu_to_le32(0xffff); /* FIXME: support r2t */
 780         icresp->cpda = 0;
 781         if (queue->hdr_digest)
 782                 icresp->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
 783         if (queue->data_digest)
 784                 icresp->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
 785
 786         iov.iov_base = icresp;
 787         iov.iov_len = sizeof(*icresp);
 788         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
 789         if (ret < 0)
 790                 goto free_crypto;
 791
 792         queue->state = NVMET_TCP_Q_LIVE;
 793         nvmet_prepare_receive_pdu(queue);
 794         return 0;
 795 free_crypto:
 796         if (queue->hdr_digest || queue->data_digest)
 797                 nvmet_tcp_free_crypto(queue);
 798         return ret;
 799 }
 800
 801 static void nvmet_tcp_handle_req_failure(struct nvmet_tcp_queue *queue,
 802                 struct nvmet_tcp_cmd *cmd, struct nvmet_req *req)
 803 {
 804         int ret;
 805
 806         /* recover the expected data transfer length */
 807         req->data_len = le32_to_cpu(req->cmd->common.dptr.sgl.length);
 808
 809         if (!nvme_is_write(cmd->req.cmd) ||
 810             req->data_len > cmd->req.port->inline_data_size) {
 811                 nvmet_prepare_receive_pdu(queue);
 812                 return;
 813         }
 814
 815         ret = nvmet_tcp_map_data(cmd);
 816         if (unlikely(ret)) {
 817                 pr_err("queue %d: failed to map data\n", queue->idx);
 818                 nvmet_tcp_fatal_error(queue);
 819                 return;
 820         }
 821
 822         queue->rcv_state = NVMET_TCP_RECV_DATA;
 823         nvmet_tcp_map_pdu_iovec(cmd);
 824         cmd->flags |= NVMET_TCP_F_INIT_FAILED;
 825 }
 826
 827 static int nvmet_tcp_handle_h2c_data_pdu(struct nvmet_tcp_queue *queue)
 828 {
 829         struct nvme_tcp_data_pdu *data = &queue->pdu.data;
 830         struct nvmet_tcp_cmd *cmd;
 831
 832         cmd = &queue->cmds[data->ttag];
 833
 834         if (le32_to_cpu(data->data_offset) != cmd->rbytes_done) {
 835                 pr_err("ttag %u unexpected data offset %u (expected %u)\n",
 836                         data->ttag, le32_to_cpu(data->data_offset),
 837                         cmd->rbytes_done);
 838                 /* FIXME: use path and transport errors */
 839                 nvmet_req_complete(&cmd->req,
 840                         NVME_SC_INVALID_FIELD | NVME_SC_DNR);
 841                 return -EPROTO;
 842         }
 843
 844         cmd->pdu_len = le32_to_cpu(data->data_length);
 845         cmd->pdu_recv = 0;
 846         nvmet_tcp_map_pdu_iovec(cmd);
 847         queue->cmd = cmd;
 848         queue->rcv_state = NVMET_TCP_RECV_DATA;
 849
 850         return 0;
 851 }
 852
 853 static int nvmet_tcp_done_recv_pdu(struct nvmet_tcp_queue *queue)
 854 {
 855         struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
 856         struct nvme_command *nvme_cmd = &queue->pdu.cmd.cmd;
 857         struct nvmet_req *req;
 858         int ret;
 859
 860         if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
 861                 if (hdr->type != nvme_tcp_icreq) {
 862                         pr_err("unexpected pdu type (%d) before icreq\n",
 863                                 hdr->type);
 864                         nvmet_tcp_fatal_error(queue);
 865                         return -EPROTO;
 866                 }
 867                 return nvmet_tcp_handle_icreq(queue);
 868         }
 869
 870         if (hdr->type == nvme_tcp_h2c_data) {
 871                 ret = nvmet_tcp_handle_h2c_data_pdu(queue);
 872                 if (unlikely(ret))
 873                         return ret;
 874                 return 0;
 875         }
 876
 877         queue->cmd = nvmet_tcp_get_cmd(queue);
 878         if (unlikely(!queue->cmd)) {
 879                 /* This should never happen */
 880                 pr_err("queue %d: out of commands (%d) send_list_len: %d, opcode: %d",
 881                         queue->idx, queue->nr_cmds, queue->send_list_len,
 882                         nvme_cmd->common.opcode);
 883                 nvmet_tcp_fatal_error(queue);
 884                 return -ENOMEM;
 885         }
 886
 887         req = &queue->cmd->req;
 888         memcpy(req->cmd, nvme_cmd, sizeof(*nvme_cmd));
 889
 890         if (unlikely(!nvmet_req_init(req, &queue->nvme_cq,
 891                         &queue->nvme_sq, &nvmet_tcp_ops))) {
 892                 pr_err("failed cmd %p id %d opcode %d, data_len: %d\n",
 893                         req->cmd, req->cmd->common.command_id,
 894                         req->cmd->common.opcode,
 895                         le32_to_cpu(req->cmd->common.dptr.sgl.length));
 896
 897                 nvmet_tcp_handle_req_failure(queue, queue->cmd, req);
 898                 return -EAGAIN;
 899         }
 900
 901         ret = nvmet_tcp_map_data(queue->cmd);
 902         if (unlikely(ret)) {
 903                 pr_err("queue %d: failed to map data\n", queue->idx);
 904                 if (nvmet_tcp_has_inline_data(queue->cmd))
 905                         nvmet_tcp_fatal_error(queue);
 906                 else
 907                         nvmet_req_complete(req, ret);
 908                 ret = -EAGAIN;
 909                 goto out;
 910         }
 911
 912         if (nvmet_tcp_need_data_in(queue->cmd)) {
 913                 if (nvmet_tcp_has_inline_data(queue->cmd)) {
 914                         queue->rcv_state = NVMET_TCP_RECV_DATA;
 915                         nvmet_tcp_map_pdu_iovec(queue->cmd);
 916                         return 0;
 917                 }
 918                 /* send back R2T */
 919                 nvmet_tcp_queue_response(&queue->cmd->req);
 920                 goto out;
 921         }
 922
 923         nvmet_req_execute(&queue->cmd->req);
 924 out:
 925         nvmet_prepare_receive_pdu(queue);
 926         return ret;
 927 }
 928
 929 static const u8 nvme_tcp_pdu_sizes[] = {
 930         [nvme_tcp_icreq]        = sizeof(struct nvme_tcp_icreq_pdu),
 931         [nvme_tcp_cmd]          = sizeof(struct nvme_tcp_cmd_pdu),
 932         [nvme_tcp_h2c_data]     = sizeof(struct nvme_tcp_data_pdu),
 933 };
 934
 935 static inline u8 nvmet_tcp_pdu_size(u8 type)
 936 {
 937         size_t idx = type;
 938
 939         return (idx < ARRAY_SIZE(nvme_tcp_pdu_sizes) &&
 940                 nvme_tcp_pdu_sizes[idx]) ?
 941                         nvme_tcp_pdu_sizes[idx] : 0;
 942 }
 943
 944 static inline bool nvmet_tcp_pdu_valid(u8 type)
 945 {
 946         switch (type) {
 947         case nvme_tcp_icreq:
 948         case nvme_tcp_cmd:
 949         case nvme_tcp_h2c_data:
 950                 /* fallthru */
 951                 return true;
 952         }
 953
 954         return false;
 955 }
 956
 957 static int nvmet_tcp_try_recv_pdu(struct nvmet_tcp_queue *queue)
 958 {
 959         struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
 960         int len;
 961         struct kvec iov;
 962         struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
 963
 964 recv:
 965         iov.iov_base = (void *)&queue->pdu + queue->offset;
 966         iov.iov_len = queue->left;
 967         len = kernel_recvmsg(queue->sock, &msg, &iov, 1,
 968                         iov.iov_len, msg.msg_flags);
 969         if (unlikely(len < 0))
 970                 return len;
 971
 972         queue->offset += len;
 973         queue->left -= len;
 974         if (queue->left)
 975                 return -EAGAIN;
 976
 977         if (queue->offset == sizeof(struct nvme_tcp_hdr)) {
 978                 u8 hdgst = nvmet_tcp_hdgst_len(queue);
 979
 980                 if (unlikely(!nvmet_tcp_pdu_valid(hdr->type))) {
 981                         pr_err("unexpected pdu type %d\n", hdr->type);
 982                         nvmet_tcp_fatal_error(queue);
 983                         return -EIO;
 984                 }
 985
 986                 if (unlikely(hdr->hlen != nvmet_tcp_pdu_size(hdr->type))) {
 987                         pr_err("pdu %d bad hlen %d\n", hdr->type, hdr->hlen);
 988                         return -EIO;
 989                 }
 990
 991                 queue->left = hdr->hlen - queue->offset + hdgst;
 992                 goto recv;
 993         }
 994
 995         if (queue->hdr_digest &&
 996             nvmet_tcp_verify_hdgst(queue, &queue->pdu, queue->offset)) {
 997                 nvmet_tcp_fatal_error(queue); /* fatal */
 998                 return -EPROTO;
 999         }
1000
1001         if (queue->data_digest &&
1002             nvmet_tcp_check_ddgst(queue, &queue->pdu)) {
1003                 nvmet_tcp_fatal_error(queue); /* fatal */
1004                 return -EPROTO;
1005         }
1006
1007         return nvmet_tcp_done_recv_pdu(queue);
1008 }
1009
1010 static void nvmet_tcp_prep_recv_ddgst(struct nvmet_tcp_cmd *cmd)
1011 {
1012         struct nvmet_tcp_queue *queue = cmd->queue;
1013
1014         nvmet_tcp_ddgst(queue->rcv_hash, cmd);
1015         queue->offset = 0;
1016         queue->left = NVME_TCP_DIGEST_LENGTH;
1017         queue->rcv_state = NVMET_TCP_RECV_DDGST;
1018 }
1019
1020 static int nvmet_tcp_try_recv_data(struct nvmet_tcp_queue *queue)
1021 {
1022         struct nvmet_tcp_cmd  *cmd = queue->cmd;
1023         int ret;
1024
1025         while (msg_data_left(&cmd->recv_msg)) {
1026                 ret = sock_recvmsg(cmd->queue->sock, &cmd->recv_msg,
1027                         cmd->recv_msg.msg_flags);
1028                 if (ret <= 0)
1029                         return ret;
1030
1031                 cmd->pdu_recv += ret;
1032                 cmd->rbytes_done += ret;
1033         }
1034
1035         nvmet_tcp_unmap_pdu_iovec(cmd);
1036
1037         if (!(cmd->flags & NVMET_TCP_F_INIT_FAILED) &&
1038             cmd->rbytes_done == cmd->req.transfer_len) {
1039                 if (queue->data_digest) {
1040                         nvmet_tcp_prep_recv_ddgst(cmd);
1041                         return 0;
1042                 }
1043                 nvmet_req_execute(&cmd->req);
1044         }
1045
1046         nvmet_prepare_receive_pdu(queue);
1047         return 0;
1048 }
1049
1050 static int nvmet_tcp_try_recv_ddgst(struct nvmet_tcp_queue *queue)
1051 {
1052         struct nvmet_tcp_cmd *cmd = queue->cmd;
1053         int ret;
1054         struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1055         struct kvec iov = {
1056                 .iov_base = (void *)&cmd->recv_ddgst + queue->offset,
1057                 .iov_len = queue->left
1058         };
1059
1060         ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1061                         iov.iov_len, msg.msg_flags);
1062         if (unlikely(ret < 0))
1063                 return ret;
1064
1065         queue->offset += ret;
1066         queue->left -= ret;
1067         if (queue->left)
1068                 return -EAGAIN;
1069
1070         if (queue->data_digest && cmd->exp_ddgst != cmd->recv_ddgst) {
1071                 pr_err("queue %d: cmd %d pdu (%d) data digest error: recv %#x expected %#x\n",
1072                         queue->idx, cmd->req.cmd->common.command_id,
1073                         queue->pdu.cmd.hdr.type, le32_to_cpu(cmd->recv_ddgst),
1074                         le32_to_cpu(cmd->exp_ddgst));
1075                 nvmet_tcp_finish_cmd(cmd);
1076                 nvmet_tcp_fatal_error(queue);
1077                 ret = -EPROTO;
1078                 goto out;
1079         }
1080
1081         if (!(cmd->flags & NVMET_TCP_F_INIT_FAILED) &&
1082             cmd->rbytes_done == cmd->req.transfer_len)
1083                 nvmet_req_execute(&cmd->req);
1084         ret = 0;
1085 out:
1086         nvmet_prepare_receive_pdu(queue);
1087         return ret;
1088 }
1089
1090 static int nvmet_tcp_try_recv_one(struct nvmet_tcp_queue *queue)
1091 {
1092         int result;
1093
1094         if (unlikely(queue->rcv_state == NVMET_TCP_RECV_ERR))
1095                 return 0;
1096
1097         if (queue->rcv_state == NVMET_TCP_RECV_PDU) {
1098                 result = nvmet_tcp_try_recv_pdu(queue);
1099                 if (result != 0)
1100                         goto done_recv;
1101         }
1102
1103         if (queue->rcv_state == NVMET_TCP_RECV_DATA) {
1104                 result = nvmet_tcp_try_recv_data(queue);
1105                 if (result != 0)
1106                         goto done_recv;
1107         }
1108
1109         if (queue->rcv_state == NVMET_TCP_RECV_DDGST) {
1110                 result = nvmet_tcp_try_recv_ddgst(queue);
1111                 if (result != 0)
1112                         goto done_recv;
1113         }
1114
1115 done_recv:
1116         if (result < 0) {
1117                 if (result == -EAGAIN)
1118                         return 0;
1119                 return result;
1120         }
1121         return 1;
1122 }
1123
1124 static int nvmet_tcp_try_recv(struct nvmet_tcp_queue *queue,
1125                 int budget, int *recvs)
1126 {
1127         int i, ret = 0;
1128
1129         for (i = 0; i < budget; i++) {
1130                 ret = nvmet_tcp_try_recv_one(queue);
1131                 if (ret <= 0)
1132                         break;
1133                 (*recvs)++;
1134         }
1135
1136         return ret;
1137 }
1138
1139 static void nvmet_tcp_schedule_release_queue(struct nvmet_tcp_queue *queue)
1140 {
1141         spin_lock(&queue->state_lock);
1142         if (queue->state != NVMET_TCP_Q_DISCONNECTING) {
1143                 queue->state = NVMET_TCP_Q_DISCONNECTING;
1144                 schedule_work(&queue->release_work);
1145         }
1146         spin_unlock(&queue->state_lock);
1147 }
1148
1149 static void nvmet_tcp_io_work(struct work_struct *w)
1150 {
1151         struct nvmet_tcp_queue *queue =
1152                 container_of(w, struct nvmet_tcp_queue, io_work);
1153         bool pending;
1154         int ret, ops = 0;
1155
1156         do {
1157                 pending = false;
1158
1159                 ret = nvmet_tcp_try_recv(queue, NVMET_TCP_RECV_BUDGET, &ops);
1160                 if (ret > 0) {
1161                         pending = true;
1162                 } else if (ret < 0) {
1163                         if (ret == -EPIPE || ret == -ECONNRESET)
1164                                 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1165                         else
1166                                 nvmet_tcp_fatal_error(queue);
1167                         return;
1168                 }
1169
1170                 ret = nvmet_tcp_try_send(queue, NVMET_TCP_SEND_BUDGET, &ops);
1171                 if (ret > 0) {
1172                         /* transmitted message/data */
1173                         pending = true;
1174                 } else if (ret < 0) {
1175                         if (ret == -EPIPE || ret == -ECONNRESET)
1176                                 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1177                         else
1178                                 nvmet_tcp_fatal_error(queue);
1179                         return;
1180                 }
1181
1182         } while (pending && ops < NVMET_TCP_IO_WORK_BUDGET);
1183
1184         /*
1185          * We exahusted our budget, requeue our selves
1186          */
1187         if (pending)
1188                 queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1189 }
1190
1191 static int nvmet_tcp_alloc_cmd(struct nvmet_tcp_queue *queue,
1192                 struct nvmet_tcp_cmd *c)
1193 {
1194         u8 hdgst = nvmet_tcp_hdgst_len(queue);
1195
1196         c->queue = queue;
1197         c->req.port = queue->port->nport;
1198
1199         c->cmd_pdu = page_frag_alloc(&queue->pf_cache,
1200                         sizeof(*c->cmd_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1201         if (!c->cmd_pdu)
1202                 return -ENOMEM;
1203         c->req.cmd = &c->cmd_pdu->cmd;
1204
1205         c->rsp_pdu = page_frag_alloc(&queue->pf_cache,
1206                         sizeof(*c->rsp_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1207         if (!c->rsp_pdu)
1208                 goto out_free_cmd;
1209         c->req.rsp = &c->rsp_pdu->cqe;
1210
1211         c->data_pdu = page_frag_alloc(&queue->pf_cache,
1212                         sizeof(*c->data_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1213         if (!c->data_pdu)
1214                 goto out_free_rsp;
1215
1216         c->r2t_pdu = page_frag_alloc(&queue->pf_cache,
1217                         sizeof(*c->r2t_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1218         if (!c->r2t_pdu)
1219                 goto out_free_data;
1220
1221         c->recv_msg.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1222
1223         list_add_tail(&c->entry, &queue->free_list);
1224
1225         return 0;
1226 out_free_data:
1227         page_frag_free(c->data_pdu);
1228 out_free_rsp:
1229         page_frag_free(c->rsp_pdu);
1230 out_free_cmd:
1231         page_frag_free(c->cmd_pdu);
1232         return -ENOMEM;
1233 }
1234
1235 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c)
1236 {
1237         page_frag_free(c->r2t_pdu);
1238         page_frag_free(c->data_pdu);
1239         page_frag_free(c->rsp_pdu);
1240         page_frag_free(c->cmd_pdu);
1241 }
1242
1243 static int nvmet_tcp_alloc_cmds(struct nvmet_tcp_queue *queue)
1244 {
1245         struct nvmet_tcp_cmd *cmds;
1246         int i, ret = -EINVAL, nr_cmds = queue->nr_cmds;
1247
1248         cmds = kcalloc(nr_cmds, sizeof(struct nvmet_tcp_cmd), GFP_KERNEL);
1249         if (!cmds)
1250                 goto out;
1251
1252         for (i = 0; i < nr_cmds; i++) {
1253                 ret = nvmet_tcp_alloc_cmd(queue, cmds + i);
1254                 if (ret)
1255                         goto out_free;
1256         }
1257
1258         queue->cmds = cmds;
1259
1260         return 0;
1261 out_free:
1262         while (--i >= 0)
1263                 nvmet_tcp_free_cmd(cmds + i);
1264         kfree(cmds);
1265 out:
1266         return ret;
1267 }
1268
1269 static void nvmet_tcp_free_cmds(struct nvmet_tcp_queue *queue)
1270 {
1271         struct nvmet_tcp_cmd *cmds = queue->cmds;
1272         int i;
1273
1274         for (i = 0; i < queue->nr_cmds; i++)
1275                 nvmet_tcp_free_cmd(cmds + i);
1276
1277         nvmet_tcp_free_cmd(&queue->connect);
1278         kfree(cmds);
1279 }
1280
1281 static void nvmet_tcp_restore_socket_callbacks(struct nvmet_tcp_queue *queue)
1282 {
1283         struct socket *sock = queue->sock;
1284
1285         write_lock_bh(&sock->sk->sk_callback_lock);
1286         sock->sk->sk_data_ready =  queue->data_ready;
1287         sock->sk->sk_state_change = queue->state_change;
1288         sock->sk->sk_write_space = queue->write_space;
1289         sock->sk->sk_user_data = NULL;
1290         write_unlock_bh(&sock->sk->sk_callback_lock);
1291 }
1292
1293 static void nvmet_tcp_finish_cmd(struct nvmet_tcp_cmd *cmd)
1294 {
1295         nvmet_req_uninit(&cmd->req);
1296         nvmet_tcp_unmap_pdu_iovec(cmd);
1297         sgl_free(cmd->req.sg);
1298 }
1299
1300 static void nvmet_tcp_uninit_data_in_cmds(struct nvmet_tcp_queue *queue)
1301 {
1302         struct nvmet_tcp_cmd *cmd = queue->cmds;
1303         int i;
1304
1305         for (i = 0; i < queue->nr_cmds; i++, cmd++) {
1306                 if (nvmet_tcp_need_data_in(cmd))
1307                         nvmet_tcp_finish_cmd(cmd);
1308         }
1309
1310         if (!queue->nr_cmds && nvmet_tcp_need_data_in(&queue->connect)) {
1311                 /* failed in connect */
1312                 nvmet_tcp_finish_cmd(&queue->connect);
1313         }
1314 }
1315
1316 static void nvmet_tcp_release_queue_work(struct work_struct *w)
1317 {
1318         struct nvmet_tcp_queue *queue =
1319                 container_of(w, struct nvmet_tcp_queue, release_work);
1320
1321         mutex_lock(&nvmet_tcp_queue_mutex);
1322         list_del_init(&queue->queue_list);
1323         mutex_unlock(&nvmet_tcp_queue_mutex);
1324
1325         nvmet_tcp_restore_socket_callbacks(queue);
1326         flush_work(&queue->io_work);
1327
1328         nvmet_tcp_uninit_data_in_cmds(queue);
1329         nvmet_sq_destroy(&queue->nvme_sq);
1330         cancel_work_sync(&queue->io_work);
1331         sock_release(queue->sock);
1332         nvmet_tcp_free_cmds(queue);
1333         if (queue->hdr_digest || queue->data_digest)
1334                 nvmet_tcp_free_crypto(queue);
1335         ida_simple_remove(&nvmet_tcp_queue_ida, queue->idx);
1336
1337         kfree(queue);
1338 }
1339
1340 static void nvmet_tcp_data_ready(struct sock *sk)
1341 {
1342         struct nvmet_tcp_queue *queue;
1343
1344         read_lock_bh(&sk->sk_callback_lock);
1345         queue = sk->sk_user_data;
1346         if (likely(queue))
1347                 queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1348         read_unlock_bh(&sk->sk_callback_lock);
1349 }
1350
1351 static void nvmet_tcp_write_space(struct sock *sk)
1352 {
1353         struct nvmet_tcp_queue *queue;
1354
1355         read_lock_bh(&sk->sk_callback_lock);
1356         queue = sk->sk_user_data;
1357         if (unlikely(!queue))
1358                 goto out;
1359
1360         if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
1361                 queue->write_space(sk);
1362                 goto out;
1363         }
1364
1365         if (sk_stream_is_writeable(sk)) {
1366                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1367                 queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1368         }
1369 out:
1370         read_unlock_bh(&sk->sk_callback_lock);
1371 }
1372
1373 static void nvmet_tcp_state_change(struct sock *sk)
1374 {
1375         struct nvmet_tcp_queue *queue;
1376
1377         write_lock_bh(&sk->sk_callback_lock);
1378         queue = sk->sk_user_data;
1379         if (!queue)
1380                 goto done;
1381
1382         switch (sk->sk_state) {
1383         case TCP_FIN_WAIT1:
1384         case TCP_CLOSE_WAIT:
1385         case TCP_CLOSE:
1386                 /* FALLTHRU */
1387                 sk->sk_user_data = NULL;
1388                 nvmet_tcp_schedule_release_queue(queue);
1389                 break;
1390         default:
1391                 pr_warn("queue %d unhandled state %d\n",
1392                         queue->idx, sk->sk_state);
1393         }
1394 done:
1395         write_unlock_bh(&sk->sk_callback_lock);
1396 }
1397
1398 static int nvmet_tcp_set_queue_sock(struct nvmet_tcp_queue *queue)
1399 {
1400         struct socket *sock = queue->sock;
1401         struct linger sol = { .l_onoff = 1, .l_linger = 0 };
1402         int ret;
1403
1404         ret = kernel_getsockname(sock,
1405                 (struct sockaddr *)&queue->sockaddr);
1406         if (ret < 0)
1407                 return ret;
1408
1409         ret = kernel_getpeername(sock,
1410                 (struct sockaddr *)&queue->sockaddr_peer);
1411         if (ret < 0)
1412                 return ret;
1413
1414         /*
1415          * Cleanup whatever is sitting in the TCP transmit queue on socket
1416          * close. This is done to prevent stale data from being sent should
1417          * the network connection be restored before TCP times out.
1418          */
1419         ret = kernel_setsockopt(sock, SOL_SOCKET, SO_LINGER,
1420                         (char *)&sol, sizeof(sol));
1421         if (ret)
1422                 return ret;
1423
1424         write_lock_bh(&sock->sk->sk_callback_lock);
1425         sock->sk->sk_user_data = queue;
1426         queue->data_ready = sock->sk->sk_data_ready;
1427         sock->sk->sk_data_ready = nvmet_tcp_data_ready;
1428         queue->state_change = sock->sk->sk_state_change;
1429         sock->sk->sk_state_change = nvmet_tcp_state_change;
1430         queue->write_space = sock->sk->sk_write_space;
1431         sock->sk->sk_write_space = nvmet_tcp_write_space;
1432         write_unlock_bh(&sock->sk->sk_callback_lock);
1433
1434         return 0;
1435 }
1436
1437 static int nvmet_tcp_alloc_queue(struct nvmet_tcp_port *port,
1438                 struct socket *newsock)
1439 {
1440         struct nvmet_tcp_queue *queue;
1441         int ret;
1442
1443         queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1444         if (!queue)
1445                 return -ENOMEM;
1446
1447         INIT_WORK(&queue->release_work, nvmet_tcp_release_queue_work);
1448         INIT_WORK(&queue->io_work, nvmet_tcp_io_work);
1449         queue->sock = newsock;
1450         queue->port = port;
1451         queue->nr_cmds = 0;
1452         spin_lock_init(&queue->state_lock);
1453         queue->state = NVMET_TCP_Q_CONNECTING;
1454         INIT_LIST_HEAD(&queue->free_list);
1455         init_llist_head(&queue->resp_list);
1456         INIT_LIST_HEAD(&queue->resp_send_list);
1457
1458         queue->idx = ida_simple_get(&nvmet_tcp_queue_ida, 0, 0, GFP_KERNEL);
1459         if (queue->idx < 0) {
1460                 ret = queue->idx;
1461                 goto out_free_queue;
1462         }
1463
1464         ret = nvmet_tcp_alloc_cmd(queue, &queue->connect);
1465         if (ret)
1466                 goto out_ida_remove;
1467
1468         ret = nvmet_sq_init(&queue->nvme_sq);
1469         if (ret)
1470                 goto out_free_connect;
1471
1472         port->last_cpu = cpumask_next_wrap(port->last_cpu,
1473                                 cpu_online_mask, -1, false);
1474         queue->cpu = port->last_cpu;
1475         nvmet_prepare_receive_pdu(queue);
1476
1477         mutex_lock(&nvmet_tcp_queue_mutex);
1478         list_add_tail(&queue->queue_list, &nvmet_tcp_queue_list);
1479         mutex_unlock(&nvmet_tcp_queue_mutex);
1480
1481         ret = nvmet_tcp_set_queue_sock(queue);
1482         if (ret)
1483                 goto out_destroy_sq;
1484
1485         queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1486
1487         return 0;
1488 out_destroy_sq:
1489         mutex_lock(&nvmet_tcp_queue_mutex);
1490         list_del_init(&queue->queue_list);
1491         mutex_unlock(&nvmet_tcp_queue_mutex);
1492         nvmet_sq_destroy(&queue->nvme_sq);
1493 out_free_connect:
1494         nvmet_tcp_free_cmd(&queue->connect);
1495 out_ida_remove:
1496         ida_simple_remove(&nvmet_tcp_queue_ida, queue->idx);
1497 out_free_queue:
1498         kfree(queue);
1499         return ret;
1500 }
1501
1502 static void nvmet_tcp_accept_work(struct work_struct *w)
1503 {
1504         struct nvmet_tcp_port *port =
1505                 container_of(w, struct nvmet_tcp_port, accept_work);
1506         struct socket *newsock;
1507         int ret;
1508
1509         while (true) {
1510                 ret = kernel_accept(port->sock, &newsock, O_NONBLOCK);
1511                 if (ret < 0) {
1512                         if (ret != -EAGAIN)
1513                                 pr_warn("failed to accept err=%d\n", ret);
1514                         return;
1515                 }
1516                 ret = nvmet_tcp_alloc_queue(port, newsock);
1517                 if (ret) {
1518                         pr_err("failed to allocate queue\n");
1519                         sock_release(newsock);
1520                 }
1521         }
1522 }
1523
1524 static void nvmet_tcp_listen_data_ready(struct sock *sk)
1525 {
1526         struct nvmet_tcp_port *port;
1527
1528         read_lock_bh(&sk->sk_callback_lock);
1529         port = sk->sk_user_data;
1530         if (!port)
1531                 goto out;
1532
1533         if (sk->sk_state == TCP_LISTEN)
1534                 schedule_work(&port->accept_work);
1535 out:
1536         read_unlock_bh(&sk->sk_callback_lock);
1537 }
1538
1539 static int nvmet_tcp_add_port(struct nvmet_port *nport)
1540 {
1541         struct nvmet_tcp_port *port;
1542         __kernel_sa_family_t af;
1543         int opt, ret;
1544
1545         port = kzalloc(sizeof(*port), GFP_KERNEL);
1546         if (!port)
1547                 return -ENOMEM;
1548
1549         switch (nport->disc_addr.adrfam) {
1550         case NVMF_ADDR_FAMILY_IP4:
1551                 af = AF_INET;
1552                 break;
1553         case NVMF_ADDR_FAMILY_IP6:
1554                 af = AF_INET6;
1555                 break;
1556         default:
1557                 pr_err("address family %d not supported\n",
1558                                 nport->disc_addr.adrfam);
1559                 ret = -EINVAL;
1560                 goto err_port;
1561         }
1562
1563         ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
1564                         nport->disc_addr.trsvcid, &port->addr);
1565         if (ret) {
1566                 pr_err("malformed ip/port passed: %s:%s\n",
1567                         nport->disc_addr.traddr, nport->disc_addr.trsvcid);
1568                 goto err_port;
1569         }
1570
1571         port->nport = nport;
1572         port->last_cpu = -1;
1573         INIT_WORK(&port->accept_work, nvmet_tcp_accept_work);
1574         if (port->nport->inline_data_size < 0)
1575                 port->nport->inline_data_size = NVMET_TCP_DEF_INLINE_DATA_SIZE;
1576
1577         ret = sock_create(port->addr.ss_family, SOCK_STREAM,
1578                                 IPPROTO_TCP, &port->sock);
1579         if (ret) {
1580                 pr_err("failed to create a socket\n");
1581                 goto err_port;
1582         }
1583
1584         port->sock->sk->sk_user_data = port;
1585         port->data_ready = port->sock->sk->sk_data_ready;
1586         port->sock->sk->sk_data_ready = nvmet_tcp_listen_data_ready;
1587
1588         opt = 1;
1589         ret = kernel_setsockopt(port->sock, IPPROTO_TCP,
1590                         TCP_NODELAY, (char *)&opt, sizeof(opt));
1591         if (ret) {
1592                 pr_err("failed to set TCP_NODELAY sock opt %d\n", ret);
1593                 goto err_sock;
1594         }
1595
1596         ret = kernel_setsockopt(port->sock, SOL_SOCKET, SO_REUSEADDR,
1597                         (char *)&opt, sizeof(opt));
1598         if (ret) {
1599                 pr_err("failed to set SO_REUSEADDR sock opt %d\n", ret);
1600                 goto err_sock;
1601         }
1602
1603         ret = kernel_bind(port->sock, (struct sockaddr *)&port->addr,
1604                         sizeof(port->addr));
1605         if (ret) {
1606                 pr_err("failed to bind port socket %d\n", ret);
1607                 goto err_sock;
1608         }
1609
1610         ret = kernel_listen(port->sock, 128);
1611         if (ret) {
1612                 pr_err("failed to listen %d on port sock\n", ret);
1613                 goto err_sock;
1614         }
1615
1616         nport->priv = port;
1617         pr_info("enabling port %d (%pISpc)\n",
1618                 le16_to_cpu(nport->disc_addr.portid), &port->addr);
1619
1620         return 0;
1621
1622 err_sock:
1623         sock_release(port->sock);
1624 err_port:
1625         kfree(port);
1626         return ret;
1627 }
1628
1629 static void nvmet_tcp_remove_port(struct nvmet_port *nport)
1630 {
1631         struct nvmet_tcp_port *port = nport->priv;
1632
1633         write_lock_bh(&port->sock->sk->sk_callback_lock);
1634         port->sock->sk->sk_data_ready = port->data_ready;
1635         port->sock->sk->sk_user_data = NULL;
1636         write_unlock_bh(&port->sock->sk->sk_callback_lock);
1637         cancel_work_sync(&port->accept_work);
1638
1639         sock_release(port->sock);
1640         kfree(port);
1641 }
1642
1643 static void nvmet_tcp_delete_ctrl(struct nvmet_ctrl *ctrl)
1644 {
1645         struct nvmet_tcp_queue *queue;
1646
1647         mutex_lock(&nvmet_tcp_queue_mutex);
1648         list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1649                 if (queue->nvme_sq.ctrl == ctrl)
1650                         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1651         mutex_unlock(&nvmet_tcp_queue_mutex);
1652 }
1653
1654 static u16 nvmet_tcp_install_queue(struct nvmet_sq *sq)
1655 {
1656         struct nvmet_tcp_queue *queue =
1657                 container_of(sq, struct nvmet_tcp_queue, nvme_sq);
1658
1659         if (sq->qid == 0) {
1660                 /* Let inflight controller teardown complete */
1661                 flush_scheduled_work();
1662         }
1663
1664         queue->nr_cmds = sq->size * 2;
1665         if (nvmet_tcp_alloc_cmds(queue))
1666                 return NVME_SC_INTERNAL;
1667         return 0;
1668 }
1669
1670 static void nvmet_tcp_disc_port_addr(struct nvmet_req *req,
1671                 struct nvmet_port *nport, char *traddr)
1672 {
1673         struct nvmet_tcp_port *port = nport->priv;
1674
1675         if (inet_addr_is_any((struct sockaddr *)&port->addr)) {
1676                 struct nvmet_tcp_cmd *cmd =
1677                         container_of(req, struct nvmet_tcp_cmd, req);
1678                 struct nvmet_tcp_queue *queue = cmd->queue;
1679
1680                 sprintf(traddr, "%pISc", (struct sockaddr *)&queue->sockaddr);
1681         } else {
1682                 memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
1683         }
1684 }
1685
1686 static struct nvmet_fabrics_ops nvmet_tcp_ops = {
1687         .owner                  = THIS_MODULE,
1688         .type                   = NVMF_TRTYPE_TCP,
1689         .msdbd                  = 1,
1690         .has_keyed_sgls         = 0,
1691         .add_port               = nvmet_tcp_add_port,
1692         .remove_port            = nvmet_tcp_remove_port,
1693         .queue_response         = nvmet_tcp_queue_response,
1694         .delete_ctrl            = nvmet_tcp_delete_ctrl,
1695         .install_queue          = nvmet_tcp_install_queue,
1696         .disc_traddr            = nvmet_tcp_disc_port_addr,
1697 };
1698
1699 static int __init nvmet_tcp_init(void)
1700 {
1701         int ret;
1702
1703         nvmet_tcp_wq = alloc_workqueue("nvmet_tcp_wq", WQ_HIGHPRI, 0);
1704         if (!nvmet_tcp_wq)
1705                 return -ENOMEM;
1706
1707         ret = nvmet_register_transport(&nvmet_tcp_ops);
1708         if (ret)
1709                 goto err;
1710
1711         return 0;
1712 err:
1713         destroy_workqueue(nvmet_tcp_wq);
1714         return ret;
1715 }
1716
1717 static void __exit nvmet_tcp_exit(void)
1718 {
1719         struct nvmet_tcp_queue *queue;
1720
1721         nvmet_unregister_transport(&nvmet_tcp_ops);
1722
1723         flush_scheduled_work();
1724         mutex_lock(&nvmet_tcp_queue_mutex);
1725         list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1726                 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1727         mutex_unlock(&nvmet_tcp_queue_mutex);
1728         flush_scheduled_work();
1729
1730         destroy_workqueue(nvmet_tcp_wq);
1731 }
1732
1733 module_init(nvmet_tcp_init);
1734 module_exit(nvmet_tcp_exit);
1735
1736 MODULE_LICENSE("GPL v2");
1737 MODULE_ALIAS("nvmet-transport-3"); /* 3 == NVMF_TRTYPE_TCP */