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