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)
 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
 537                 ret = kernel_sendpage(cmd->queue->sock, page, cmd->offset,
 538                                         left, MSG_DONTWAIT | MSG_MORE);
 539                 if (ret <= 0)
 540                         return ret;
 541
 542                 cmd->offset += ret;
 543                 cmd->wbytes_done += ret;
 544
 545                 /* Done with sg?*/
 546                 if (cmd->offset == cmd->cur_sg->length) {
 547                         cmd->cur_sg = sg_next(cmd->cur_sg);
 548                         cmd->offset = 0;
 549                 }
 550         }
 551
 552         if (queue->data_digest) {
 553                 cmd->state = NVMET_TCP_SEND_DDGST;
 554                 cmd->offset = 0;
 555         } else {
 556                 if (queue->nvme_sq.sqhd_disabled) {
 557                         cmd->queue->snd_cmd = NULL;
 558                         nvmet_tcp_put_cmd(cmd);
 559                 } else {
 560                         nvmet_setup_response_pdu(cmd);
 561                 }
 562         }
 563
 564         if (queue->nvme_sq.sqhd_disabled) {
 565                 kfree(cmd->iov);
 566                 sgl_free(cmd->req.sg);
 567         }
 568
 569         return 1;
 570
 571 }
 572
 573 static int nvmet_try_send_response(struct nvmet_tcp_cmd *cmd,
 574                 bool last_in_batch)
 575 {
 576         u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
 577         int left = sizeof(*cmd->rsp_pdu) - cmd->offset + hdgst;
 578         int flags = MSG_DONTWAIT;
 579         int ret;
 580
 581         if (!last_in_batch && cmd->queue->send_list_len)
 582                 flags |= MSG_MORE;
 583         else
 584                 flags |= MSG_EOR;
 585
 586         ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->rsp_pdu),
 587                 offset_in_page(cmd->rsp_pdu) + cmd->offset, left, flags);
 588         if (ret <= 0)
 589                 return ret;
 590         cmd->offset += ret;
 591         left -= ret;
 592
 593         if (left)
 594                 return -EAGAIN;
 595
 596         kfree(cmd->iov);
 597         sgl_free(cmd->req.sg);
 598         cmd->queue->snd_cmd = NULL;
 599         nvmet_tcp_put_cmd(cmd);
 600         return 1;
 601 }
 602
 603 static int nvmet_try_send_r2t(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
 604 {
 605         u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
 606         int left = sizeof(*cmd->r2t_pdu) - cmd->offset + hdgst;
 607         int flags = MSG_DONTWAIT;
 608         int ret;
 609
 610         if (!last_in_batch && cmd->queue->send_list_len)
 611                 flags |= MSG_MORE;
 612         else
 613                 flags |= MSG_EOR;
 614
 615         ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->r2t_pdu),
 616                 offset_in_page(cmd->r2t_pdu) + cmd->offset, left, flags);
 617         if (ret <= 0)
 618                 return ret;
 619         cmd->offset += ret;
 620         left -= ret;
 621
 622         if (left)
 623                 return -EAGAIN;
 624
 625         cmd->queue->snd_cmd = NULL;
 626         return 1;
 627 }
 628
 629 static int nvmet_try_send_ddgst(struct nvmet_tcp_cmd *cmd)
 630 {
 631         struct nvmet_tcp_queue *queue = cmd->queue;
 632         struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
 633         struct kvec iov = {
 634                 .iov_base = &cmd->exp_ddgst + cmd->offset,
 635                 .iov_len = NVME_TCP_DIGEST_LENGTH - cmd->offset
 636         };
 637         int ret;
 638
 639         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
 640         if (unlikely(ret <= 0))
 641                 return ret;
 642
 643         cmd->offset += ret;
 644
 645         if (queue->nvme_sq.sqhd_disabled) {
 646                 cmd->queue->snd_cmd = NULL;
 647                 nvmet_tcp_put_cmd(cmd);
 648         } else {
 649                 nvmet_setup_response_pdu(cmd);
 650         }
 651         return 1;
 652 }
 653
 654 static int nvmet_tcp_try_send_one(struct nvmet_tcp_queue *queue,
 655                 bool last_in_batch)
 656 {
 657         struct nvmet_tcp_cmd *cmd = queue->snd_cmd;
 658         int ret = 0;
 659
 660         if (!cmd || queue->state == NVMET_TCP_Q_DISCONNECTING) {
 661                 cmd = nvmet_tcp_fetch_cmd(queue);
 662                 if (unlikely(!cmd))
 663                         return 0;
 664         }
 665
 666         if (cmd->state == NVMET_TCP_SEND_DATA_PDU) {
 667                 ret = nvmet_try_send_data_pdu(cmd);
 668                 if (ret <= 0)
 669                         goto done_send;
 670         }
 671
 672         if (cmd->state == NVMET_TCP_SEND_DATA) {
 673                 ret = nvmet_try_send_data(cmd);
 674                 if (ret <= 0)
 675                         goto done_send;
 676         }
 677
 678         if (cmd->state == NVMET_TCP_SEND_DDGST) {
 679                 ret = nvmet_try_send_ddgst(cmd);
 680                 if (ret <= 0)
 681                         goto done_send;
 682         }
 683
 684         if (cmd->state == NVMET_TCP_SEND_R2T) {
 685                 ret = nvmet_try_send_r2t(cmd, last_in_batch);
 686                 if (ret <= 0)
 687                         goto done_send;
 688         }
 689
 690         if (cmd->state == NVMET_TCP_SEND_RESPONSE)
 691                 ret = nvmet_try_send_response(cmd, last_in_batch);
 692
 693 done_send:
 694         if (ret < 0) {
 695                 if (ret == -EAGAIN)
 696                         return 0;
 697                 return ret;
 698         }
 699
 700         return 1;
 701 }
 702
 703 static int nvmet_tcp_try_send(struct nvmet_tcp_queue *queue,
 704                 int budget, int *sends)
 705 {
 706         int i, ret = 0;
 707
 708         for (i = 0; i < budget; i++) {
 709                 ret = nvmet_tcp_try_send_one(queue, i == budget - 1);
 710                 if (ret <= 0)
 711                         break;
 712                 (*sends)++;
 713         }
 714
 715         return ret;
 716 }
 717
 718 static void nvmet_prepare_receive_pdu(struct nvmet_tcp_queue *queue)
 719 {
 720         queue->offset = 0;
 721         queue->left = sizeof(struct nvme_tcp_hdr);
 722         queue->cmd = NULL;
 723         queue->rcv_state = NVMET_TCP_RECV_PDU;
 724 }
 725
 726 static void nvmet_tcp_free_crypto(struct nvmet_tcp_queue *queue)
 727 {
 728         struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
 729
 730         ahash_request_free(queue->rcv_hash);
 731         ahash_request_free(queue->snd_hash);
 732         crypto_free_ahash(tfm);
 733 }
 734
 735 static int nvmet_tcp_alloc_crypto(struct nvmet_tcp_queue *queue)
 736 {
 737         struct crypto_ahash *tfm;
 738
 739         tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
 740         if (IS_ERR(tfm))
 741                 return PTR_ERR(tfm);
 742
 743         queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
 744         if (!queue->snd_hash)
 745                 goto free_tfm;
 746         ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
 747
 748         queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
 749         if (!queue->rcv_hash)
 750                 goto free_snd_hash;
 751         ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
 752
 753         return 0;
 754 free_snd_hash:
 755         ahash_request_free(queue->snd_hash);
 756 free_tfm:
 757         crypto_free_ahash(tfm);
 758         return -ENOMEM;
 759 }
 760
 761
 762 static int nvmet_tcp_handle_icreq(struct nvmet_tcp_queue *queue)
 763 {
 764         struct nvme_tcp_icreq_pdu *icreq = &queue->pdu.icreq;
 765         struct nvme_tcp_icresp_pdu *icresp = &queue->pdu.icresp;
 766         struct msghdr msg = {};
 767         struct kvec iov;
 768         int ret;
 769
 770         if (le32_to_cpu(icreq->hdr.plen) != sizeof(struct nvme_tcp_icreq_pdu)) {
 771                 pr_err("bad nvme-tcp pdu length (%d)\n",
 772                         le32_to_cpu(icreq->hdr.plen));
 773                 nvmet_tcp_fatal_error(queue);
 774         }
 775
 776         if (icreq->pfv != NVME_TCP_PFV_1_0) {
 777                 pr_err("queue %d: bad pfv %d\n", queue->idx, icreq->pfv);
 778                 return -EPROTO;
 779         }
 780
 781         if (icreq->hpda != 0) {
 782                 pr_err("queue %d: unsupported hpda %d\n", queue->idx,
 783                         icreq->hpda);
 784                 return -EPROTO;
 785         }
 786
 787         queue->hdr_digest = !!(icreq->digest & NVME_TCP_HDR_DIGEST_ENABLE);
 788         queue->data_digest = !!(icreq->digest & NVME_TCP_DATA_DIGEST_ENABLE);
 789         if (queue->hdr_digest || queue->data_digest) {
 790                 ret = nvmet_tcp_alloc_crypto(queue);
 791                 if (ret)
 792                         return ret;
 793         }
 794
 795         memset(icresp, 0, sizeof(*icresp));
 796         icresp->hdr.type = nvme_tcp_icresp;
 797         icresp->hdr.hlen = sizeof(*icresp);
 798         icresp->hdr.pdo = 0;
 799         icresp->hdr.plen = cpu_to_le32(icresp->hdr.hlen);
 800         icresp->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
 801         icresp->maxdata = cpu_to_le32(0xffff); /* FIXME: support r2t */
 802         icresp->cpda = 0;
 803         if (queue->hdr_digest)
 804                 icresp->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
 805         if (queue->data_digest)
 806                 icresp->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
 807
 808         iov.iov_base = icresp;
 809         iov.iov_len = sizeof(*icresp);
 810         ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
 811         if (ret < 0)
 812                 goto free_crypto;
 813
 814         queue->state = NVMET_TCP_Q_LIVE;
 815         nvmet_prepare_receive_pdu(queue);
 816         return 0;
 817 free_crypto:
 818         if (queue->hdr_digest || queue->data_digest)
 819                 nvmet_tcp_free_crypto(queue);
 820         return ret;
 821 }
 822
 823 static void nvmet_tcp_handle_req_failure(struct nvmet_tcp_queue *queue,
 824                 struct nvmet_tcp_cmd *cmd, struct nvmet_req *req)
 825 {
 826         size_t data_len = le32_to_cpu(req->cmd->common.dptr.sgl.length);
 827         int ret;
 828
 829         if (!nvme_is_write(cmd->req.cmd) ||
 830             data_len > cmd->req.port->inline_data_size) {
 831                 nvmet_prepare_receive_pdu(queue);
 832                 return;
 833         }
 834
 835         ret = nvmet_tcp_map_data(cmd);
 836         if (unlikely(ret)) {
 837                 pr_err("queue %d: failed to map data\n", queue->idx);
 838                 nvmet_tcp_fatal_error(queue);
 839                 return;
 840         }
 841
 842         queue->rcv_state = NVMET_TCP_RECV_DATA;
 843         nvmet_tcp_map_pdu_iovec(cmd);
 844         cmd->flags |= NVMET_TCP_F_INIT_FAILED;
 845 }
 846
 847 static int nvmet_tcp_handle_h2c_data_pdu(struct nvmet_tcp_queue *queue)
 848 {
 849         struct nvme_tcp_data_pdu *data = &queue->pdu.data;
 850         struct nvmet_tcp_cmd *cmd;
 851
 852         cmd = &queue->cmds[data->ttag];
 853
 854         if (le32_to_cpu(data->data_offset) != cmd->rbytes_done) {
 855                 pr_err("ttag %u unexpected data offset %u (expected %u)\n",
 856                         data->ttag, le32_to_cpu(data->data_offset),
 857                         cmd->rbytes_done);
 858                 /* FIXME: use path and transport errors */
 859                 nvmet_req_complete(&cmd->req,
 860                         NVME_SC_INVALID_FIELD | NVME_SC_DNR);
 861                 return -EPROTO;
 862         }
 863
 864         cmd->pdu_len = le32_to_cpu(data->data_length);
 865         cmd->pdu_recv = 0;
 866         nvmet_tcp_map_pdu_iovec(cmd);
 867         queue->cmd = cmd;
 868         queue->rcv_state = NVMET_TCP_RECV_DATA;
 869
 870         return 0;
 871 }
 872
 873 static int nvmet_tcp_done_recv_pdu(struct nvmet_tcp_queue *queue)
 874 {
 875         struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
 876         struct nvme_command *nvme_cmd = &queue->pdu.cmd.cmd;
 877         struct nvmet_req *req;
 878         int ret;
 879
 880         if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
 881                 if (hdr->type != nvme_tcp_icreq) {
 882                         pr_err("unexpected pdu type (%d) before icreq\n",
 883                                 hdr->type);
 884                         nvmet_tcp_fatal_error(queue);
 885                         return -EPROTO;
 886                 }
 887                 return nvmet_tcp_handle_icreq(queue);
 888         }
 889
 890         if (hdr->type == nvme_tcp_h2c_data) {
 891                 ret = nvmet_tcp_handle_h2c_data_pdu(queue);
 892                 if (unlikely(ret))
 893                         return ret;
 894                 return 0;
 895         }
 896
 897         queue->cmd = nvmet_tcp_get_cmd(queue);
 898         if (unlikely(!queue->cmd)) {
 899                 /* This should never happen */
 900                 pr_err("queue %d: out of commands (%d) send_list_len: %d, opcode: %d",
 901                         queue->idx, queue->nr_cmds, queue->send_list_len,
 902                         nvme_cmd->common.opcode);
 903                 nvmet_tcp_fatal_error(queue);
 904                 return -ENOMEM;
 905         }
 906
 907         req = &queue->cmd->req;
 908         memcpy(req->cmd, nvme_cmd, sizeof(*nvme_cmd));
 909
 910         if (unlikely(!nvmet_req_init(req, &queue->nvme_cq,
 911                         &queue->nvme_sq, &nvmet_tcp_ops))) {
 912                 pr_err("failed cmd %p id %d opcode %d, data_len: %d\n",
 913                         req->cmd, req->cmd->common.command_id,
 914                         req->cmd->common.opcode,
 915                         le32_to_cpu(req->cmd->common.dptr.sgl.length));
 916
 917                 nvmet_tcp_handle_req_failure(queue, queue->cmd, req);
 918                 return -EAGAIN;
 919         }
 920
 921         ret = nvmet_tcp_map_data(queue->cmd);
 922         if (unlikely(ret)) {
 923                 pr_err("queue %d: failed to map data\n", queue->idx);
 924                 if (nvmet_tcp_has_inline_data(queue->cmd))
 925                         nvmet_tcp_fatal_error(queue);
 926                 else
 927                         nvmet_req_complete(req, ret);
 928                 ret = -EAGAIN;
 929                 goto out;
 930         }
 931
 932         if (nvmet_tcp_need_data_in(queue->cmd)) {
 933                 if (nvmet_tcp_has_inline_data(queue->cmd)) {
 934                         queue->rcv_state = NVMET_TCP_RECV_DATA;
 935                         nvmet_tcp_map_pdu_iovec(queue->cmd);
 936                         return 0;
 937                 }
 938                 /* send back R2T */
 939                 nvmet_tcp_queue_response(&queue->cmd->req);
 940                 goto out;
 941         }
 942
 943         queue->cmd->req.execute(&queue->cmd->req);
 944 out:
 945         nvmet_prepare_receive_pdu(queue);
 946         return ret;
 947 }
 948
 949 static const u8 nvme_tcp_pdu_sizes[] = {
 950         [nvme_tcp_icreq]        = sizeof(struct nvme_tcp_icreq_pdu),
 951         [nvme_tcp_cmd]          = sizeof(struct nvme_tcp_cmd_pdu),
 952         [nvme_tcp_h2c_data]     = sizeof(struct nvme_tcp_data_pdu),
 953 };
 954
 955 static inline u8 nvmet_tcp_pdu_size(u8 type)
 956 {
 957         size_t idx = type;
 958
 959         return (idx < ARRAY_SIZE(nvme_tcp_pdu_sizes) &&
 960                 nvme_tcp_pdu_sizes[idx]) ?
 961                         nvme_tcp_pdu_sizes[idx] : 0;
 962 }
 963
 964 static inline bool nvmet_tcp_pdu_valid(u8 type)
 965 {
 966         switch (type) {
 967         case nvme_tcp_icreq:
 968         case nvme_tcp_cmd:
 969         case nvme_tcp_h2c_data:
 970                 /* fallthru */
 971                 return true;
 972         }
 973
 974         return false;
 975 }
 976
 977 static int nvmet_tcp_try_recv_pdu(struct nvmet_tcp_queue *queue)
 978 {
 979         struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
 980         int len;
 981         struct kvec iov;
 982         struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
 983
 984 recv:
 985         iov.iov_base = (void *)&queue->pdu + queue->offset;
 986         iov.iov_len = queue->left;
 987         len = kernel_recvmsg(queue->sock, &msg, &iov, 1,
 988                         iov.iov_len, msg.msg_flags);
 989         if (unlikely(len < 0))
 990                 return len;
 991
 992         queue->offset += len;
 993         queue->left -= len;
 994         if (queue->left)
 995                 return -EAGAIN;
 996
 997         if (queue->offset == sizeof(struct nvme_tcp_hdr)) {
 998                 u8 hdgst = nvmet_tcp_hdgst_len(queue);
 999
1000                 if (unlikely(!nvmet_tcp_pdu_valid(hdr->type))) {
1001                         pr_err("unexpected pdu type %d\n", hdr->type);
1002                         nvmet_tcp_fatal_error(queue);
1003                         return -EIO;
1004                 }
1005
1006                 if (unlikely(hdr->hlen != nvmet_tcp_pdu_size(hdr->type))) {
1007                         pr_err("pdu %d bad hlen %d\n", hdr->type, hdr->hlen);
1008                         return -EIO;
1009                 }
1010
1011                 queue->left = hdr->hlen - queue->offset + hdgst;
1012                 goto recv;
1013         }
1014
1015         if (queue->hdr_digest &&
1016             nvmet_tcp_verify_hdgst(queue, &queue->pdu, queue->offset)) {
1017                 nvmet_tcp_fatal_error(queue); /* fatal */
1018                 return -EPROTO;
1019         }
1020
1021         if (queue->data_digest &&
1022             nvmet_tcp_check_ddgst(queue, &queue->pdu)) {
1023                 nvmet_tcp_fatal_error(queue); /* fatal */
1024                 return -EPROTO;
1025         }
1026
1027         return nvmet_tcp_done_recv_pdu(queue);
1028 }
1029
1030 static void nvmet_tcp_prep_recv_ddgst(struct nvmet_tcp_cmd *cmd)
1031 {
1032         struct nvmet_tcp_queue *queue = cmd->queue;
1033
1034         nvmet_tcp_ddgst(queue->rcv_hash, cmd);
1035         queue->offset = 0;
1036         queue->left = NVME_TCP_DIGEST_LENGTH;
1037         queue->rcv_state = NVMET_TCP_RECV_DDGST;
1038 }
1039
1040 static int nvmet_tcp_try_recv_data(struct nvmet_tcp_queue *queue)
1041 {
1042         struct nvmet_tcp_cmd  *cmd = queue->cmd;
1043         int ret;
1044
1045         while (msg_data_left(&cmd->recv_msg)) {
1046                 ret = sock_recvmsg(cmd->queue->sock, &cmd->recv_msg,
1047                         cmd->recv_msg.msg_flags);
1048                 if (ret <= 0)
1049                         return ret;
1050
1051                 cmd->pdu_recv += ret;
1052                 cmd->rbytes_done += ret;
1053         }
1054
1055         nvmet_tcp_unmap_pdu_iovec(cmd);
1056
1057         if (!(cmd->flags & NVMET_TCP_F_INIT_FAILED) &&
1058             cmd->rbytes_done == cmd->req.transfer_len) {
1059                 if (queue->data_digest) {
1060                         nvmet_tcp_prep_recv_ddgst(cmd);
1061                         return 0;
1062                 }
1063                 cmd->req.execute(&cmd->req);
1064         }
1065
1066         nvmet_prepare_receive_pdu(queue);
1067         return 0;
1068 }
1069
1070 static int nvmet_tcp_try_recv_ddgst(struct nvmet_tcp_queue *queue)
1071 {
1072         struct nvmet_tcp_cmd *cmd = queue->cmd;
1073         int ret;
1074         struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1075         struct kvec iov = {
1076                 .iov_base = (void *)&cmd->recv_ddgst + queue->offset,
1077                 .iov_len = queue->left
1078         };
1079
1080         ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1081                         iov.iov_len, msg.msg_flags);
1082         if (unlikely(ret < 0))
1083                 return ret;
1084
1085         queue->offset += ret;
1086         queue->left -= ret;
1087         if (queue->left)
1088                 return -EAGAIN;
1089
1090         if (queue->data_digest && cmd->exp_ddgst != cmd->recv_ddgst) {
1091                 pr_err("queue %d: cmd %d pdu (%d) data digest error: recv %#x expected %#x\n",
1092                         queue->idx, cmd->req.cmd->common.command_id,
1093                         queue->pdu.cmd.hdr.type, le32_to_cpu(cmd->recv_ddgst),
1094                         le32_to_cpu(cmd->exp_ddgst));
1095                 nvmet_tcp_finish_cmd(cmd);
1096                 nvmet_tcp_fatal_error(queue);
1097                 ret = -EPROTO;
1098                 goto out;
1099         }
1100
1101         if (!(cmd->flags & NVMET_TCP_F_INIT_FAILED) &&
1102             cmd->rbytes_done == cmd->req.transfer_len)
1103                 cmd->req.execute(&cmd->req);
1104         ret = 0;
1105 out:
1106         nvmet_prepare_receive_pdu(queue);
1107         return ret;
1108 }
1109
1110 static int nvmet_tcp_try_recv_one(struct nvmet_tcp_queue *queue)
1111 {
1112         int result = 0;
1113
1114         if (unlikely(queue->rcv_state == NVMET_TCP_RECV_ERR))
1115                 return 0;
1116
1117         if (queue->rcv_state == NVMET_TCP_RECV_PDU) {
1118                 result = nvmet_tcp_try_recv_pdu(queue);
1119                 if (result != 0)
1120                         goto done_recv;
1121         }
1122
1123         if (queue->rcv_state == NVMET_TCP_RECV_DATA) {
1124                 result = nvmet_tcp_try_recv_data(queue);
1125                 if (result != 0)
1126                         goto done_recv;
1127         }
1128
1129         if (queue->rcv_state == NVMET_TCP_RECV_DDGST) {
1130                 result = nvmet_tcp_try_recv_ddgst(queue);
1131                 if (result != 0)
1132                         goto done_recv;
1133         }
1134
1135 done_recv:
1136         if (result < 0) {
1137                 if (result == -EAGAIN)
1138                         return 0;
1139                 return result;
1140         }
1141         return 1;
1142 }
1143
1144 static int nvmet_tcp_try_recv(struct nvmet_tcp_queue *queue,
1145                 int budget, int *recvs)
1146 {
1147         int i, ret = 0;
1148
1149         for (i = 0; i < budget; i++) {
1150                 ret = nvmet_tcp_try_recv_one(queue);
1151                 if (ret <= 0)
1152                         break;
1153                 (*recvs)++;
1154         }
1155
1156         return ret;
1157 }
1158
1159 static void nvmet_tcp_schedule_release_queue(struct nvmet_tcp_queue *queue)
1160 {
1161         spin_lock(&queue->state_lock);
1162         if (queue->state != NVMET_TCP_Q_DISCONNECTING) {
1163                 queue->state = NVMET_TCP_Q_DISCONNECTING;
1164                 schedule_work(&queue->release_work);
1165         }
1166         spin_unlock(&queue->state_lock);
1167 }
1168
1169 static void nvmet_tcp_io_work(struct work_struct *w)
1170 {
1171         struct nvmet_tcp_queue *queue =
1172                 container_of(w, struct nvmet_tcp_queue, io_work);
1173         bool pending;
1174         int ret, ops = 0;
1175
1176         do {
1177                 pending = false;
1178
1179                 ret = nvmet_tcp_try_recv(queue, NVMET_TCP_RECV_BUDGET, &ops);
1180                 if (ret > 0) {
1181                         pending = true;
1182                 } else if (ret < 0) {
1183                         if (ret == -EPIPE || ret == -ECONNRESET)
1184                                 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1185                         else
1186                                 nvmet_tcp_fatal_error(queue);
1187                         return;
1188                 }
1189
1190                 ret = nvmet_tcp_try_send(queue, NVMET_TCP_SEND_BUDGET, &ops);
1191                 if (ret > 0) {
1192                         /* transmitted message/data */
1193                         pending = true;
1194                 } else if (ret < 0) {
1195                         if (ret == -EPIPE || ret == -ECONNRESET)
1196                                 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1197                         else
1198                                 nvmet_tcp_fatal_error(queue);
1199                         return;
1200                 }
1201
1202         } while (pending && ops < NVMET_TCP_IO_WORK_BUDGET);
1203
1204         /*
1205          * We exahusted our budget, requeue our selves
1206          */
1207         if (pending)
1208                 queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1209 }
1210
1211 static int nvmet_tcp_alloc_cmd(struct nvmet_tcp_queue *queue,
1212                 struct nvmet_tcp_cmd *c)
1213 {
1214         u8 hdgst = nvmet_tcp_hdgst_len(queue);
1215
1216         c->queue = queue;
1217         c->req.port = queue->port->nport;
1218
1219         c->cmd_pdu = page_frag_alloc(&queue->pf_cache,
1220                         sizeof(*c->cmd_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1221         if (!c->cmd_pdu)
1222                 return -ENOMEM;
1223         c->req.cmd = &c->cmd_pdu->cmd;
1224
1225         c->rsp_pdu = page_frag_alloc(&queue->pf_cache,
1226                         sizeof(*c->rsp_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1227         if (!c->rsp_pdu)
1228                 goto out_free_cmd;
1229         c->req.cqe = &c->rsp_pdu->cqe;
1230
1231         c->data_pdu = page_frag_alloc(&queue->pf_cache,
1232                         sizeof(*c->data_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1233         if (!c->data_pdu)
1234                 goto out_free_rsp;
1235
1236         c->r2t_pdu = page_frag_alloc(&queue->pf_cache,
1237                         sizeof(*c->r2t_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1238         if (!c->r2t_pdu)
1239                 goto out_free_data;
1240
1241         c->recv_msg.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1242
1243         list_add_tail(&c->entry, &queue->free_list);
1244
1245         return 0;
1246 out_free_data:
1247         page_frag_free(c->data_pdu);
1248 out_free_rsp:
1249         page_frag_free(c->rsp_pdu);
1250 out_free_cmd:
1251         page_frag_free(c->cmd_pdu);
1252         return -ENOMEM;
1253 }
1254
1255 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c)
1256 {
1257         page_frag_free(c->r2t_pdu);
1258         page_frag_free(c->data_pdu);
1259         page_frag_free(c->rsp_pdu);
1260         page_frag_free(c->cmd_pdu);
1261 }
1262
1263 static int nvmet_tcp_alloc_cmds(struct nvmet_tcp_queue *queue)
1264 {
1265         struct nvmet_tcp_cmd *cmds;
1266         int i, ret = -EINVAL, nr_cmds = queue->nr_cmds;
1267
1268         cmds = kcalloc(nr_cmds, sizeof(struct nvmet_tcp_cmd), GFP_KERNEL);
1269         if (!cmds)
1270                 goto out;
1271
1272         for (i = 0; i < nr_cmds; i++) {
1273                 ret = nvmet_tcp_alloc_cmd(queue, cmds + i);
1274                 if (ret)
1275                         goto out_free;
1276         }
1277
1278         queue->cmds = cmds;
1279
1280         return 0;
1281 out_free:
1282         while (--i >= 0)
1283                 nvmet_tcp_free_cmd(cmds + i);
1284         kfree(cmds);
1285 out:
1286         return ret;
1287 }
1288
1289 static void nvmet_tcp_free_cmds(struct nvmet_tcp_queue *queue)
1290 {
1291         struct nvmet_tcp_cmd *cmds = queue->cmds;
1292         int i;
1293
1294         for (i = 0; i < queue->nr_cmds; i++)
1295                 nvmet_tcp_free_cmd(cmds + i);
1296
1297         nvmet_tcp_free_cmd(&queue->connect);
1298         kfree(cmds);
1299 }
1300
1301 static void nvmet_tcp_restore_socket_callbacks(struct nvmet_tcp_queue *queue)
1302 {
1303         struct socket *sock = queue->sock;
1304
1305         write_lock_bh(&sock->sk->sk_callback_lock);
1306         sock->sk->sk_data_ready =  queue->data_ready;
1307         sock->sk->sk_state_change = queue->state_change;
1308         sock->sk->sk_write_space = queue->write_space;
1309         sock->sk->sk_user_data = NULL;
1310         write_unlock_bh(&sock->sk->sk_callback_lock);
1311 }
1312
1313 static void nvmet_tcp_finish_cmd(struct nvmet_tcp_cmd *cmd)
1314 {
1315         nvmet_req_uninit(&cmd->req);
1316         nvmet_tcp_unmap_pdu_iovec(cmd);
1317         kfree(cmd->iov);
1318         sgl_free(cmd->req.sg);
1319 }
1320
1321 static void nvmet_tcp_uninit_data_in_cmds(struct nvmet_tcp_queue *queue)
1322 {
1323         struct nvmet_tcp_cmd *cmd = queue->cmds;
1324         int i;
1325
1326         for (i = 0; i < queue->nr_cmds; i++, cmd++) {
1327                 if (nvmet_tcp_need_data_in(cmd))
1328                         nvmet_tcp_finish_cmd(cmd);
1329         }
1330
1331         if (!queue->nr_cmds && nvmet_tcp_need_data_in(&queue->connect)) {
1332                 /* failed in connect */
1333                 nvmet_tcp_finish_cmd(&queue->connect);
1334         }
1335 }
1336
1337 static void nvmet_tcp_release_queue_work(struct work_struct *w)
1338 {
1339         struct nvmet_tcp_queue *queue =
1340                 container_of(w, struct nvmet_tcp_queue, release_work);
1341
1342         mutex_lock(&nvmet_tcp_queue_mutex);
1343         list_del_init(&queue->queue_list);
1344         mutex_unlock(&nvmet_tcp_queue_mutex);
1345
1346         nvmet_tcp_restore_socket_callbacks(queue);
1347         flush_work(&queue->io_work);
1348
1349         nvmet_tcp_uninit_data_in_cmds(queue);
1350         nvmet_sq_destroy(&queue->nvme_sq);
1351         cancel_work_sync(&queue->io_work);
1352         sock_release(queue->sock);
1353         nvmet_tcp_free_cmds(queue);
1354         if (queue->hdr_digest || queue->data_digest)
1355                 nvmet_tcp_free_crypto(queue);
1356         ida_simple_remove(&nvmet_tcp_queue_ida, queue->idx);
1357
1358         kfree(queue);
1359 }
1360
1361 static void nvmet_tcp_data_ready(struct sock *sk)
1362 {
1363         struct nvmet_tcp_queue *queue;
1364
1365         read_lock_bh(&sk->sk_callback_lock);
1366         queue = sk->sk_user_data;
1367         if (likely(queue))
1368                 queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1369         read_unlock_bh(&sk->sk_callback_lock);
1370 }
1371
1372 static void nvmet_tcp_write_space(struct sock *sk)
1373 {
1374         struct nvmet_tcp_queue *queue;
1375
1376         read_lock_bh(&sk->sk_callback_lock);
1377         queue = sk->sk_user_data;
1378         if (unlikely(!queue))
1379                 goto out;
1380
1381         if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
1382                 queue->write_space(sk);
1383                 goto out;
1384         }
1385
1386         if (sk_stream_is_writeable(sk)) {
1387                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1388                 queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1389         }
1390 out:
1391         read_unlock_bh(&sk->sk_callback_lock);
1392 }
1393
1394 static void nvmet_tcp_state_change(struct sock *sk)
1395 {
1396         struct nvmet_tcp_queue *queue;
1397
1398         write_lock_bh(&sk->sk_callback_lock);
1399         queue = sk->sk_user_data;
1400         if (!queue)
1401                 goto done;
1402
1403         switch (sk->sk_state) {
1404         case TCP_FIN_WAIT1:
1405         case TCP_CLOSE_WAIT:
1406         case TCP_CLOSE:
1407                 /* FALLTHRU */
1408                 sk->sk_user_data = NULL;
1409                 nvmet_tcp_schedule_release_queue(queue);
1410                 break;
1411         default:
1412                 pr_warn("queue %d unhandled state %d\n",
1413                         queue->idx, sk->sk_state);
1414         }
1415 done:
1416         write_unlock_bh(&sk->sk_callback_lock);
1417 }
1418
1419 static int nvmet_tcp_set_queue_sock(struct nvmet_tcp_queue *queue)
1420 {
1421         struct socket *sock = queue->sock;
1422         struct inet_sock *inet = inet_sk(sock->sk);
1423         struct linger sol = { .l_onoff = 1, .l_linger = 0 };
1424         int ret;
1425
1426         ret = kernel_getsockname(sock,
1427                 (struct sockaddr *)&queue->sockaddr);
1428         if (ret < 0)
1429                 return ret;
1430
1431         ret = kernel_getpeername(sock,
1432                 (struct sockaddr *)&queue->sockaddr_peer);
1433         if (ret < 0)
1434                 return ret;
1435
1436         /*
1437          * Cleanup whatever is sitting in the TCP transmit queue on socket
1438          * close. This is done to prevent stale data from being sent should
1439          * the network connection be restored before TCP times out.
1440          */
1441         ret = kernel_setsockopt(sock, SOL_SOCKET, SO_LINGER,
1442                         (char *)&sol, sizeof(sol));
1443         if (ret)
1444                 return ret;
1445
1446         if (so_priority > 0) {
1447                 ret = kernel_setsockopt(sock, SOL_SOCKET, SO_PRIORITY,
1448                                 (char *)&so_priority, sizeof(so_priority));
1449                 if (ret)
1450                         return ret;
1451         }
1452
1453         /* Set socket type of service */
1454         if (inet->rcv_tos > 0) {
1455                 int tos = inet->rcv_tos;
1456
1457                 ret = kernel_setsockopt(sock, SOL_IP, IP_TOS,
1458                                 (char *)&tos, sizeof(tos));
1459                 if (ret)
1460                         return ret;
1461         }
1462
1463         write_lock_bh(&sock->sk->sk_callback_lock);
1464         sock->sk->sk_user_data = queue;
1465         queue->data_ready = sock->sk->sk_data_ready;
1466         sock->sk->sk_data_ready = nvmet_tcp_data_ready;
1467         queue->state_change = sock->sk->sk_state_change;
1468         sock->sk->sk_state_change = nvmet_tcp_state_change;
1469         queue->write_space = sock->sk->sk_write_space;
1470         sock->sk->sk_write_space = nvmet_tcp_write_space;
1471         write_unlock_bh(&sock->sk->sk_callback_lock);
1472
1473         return 0;
1474 }
1475
1476 static int nvmet_tcp_alloc_queue(struct nvmet_tcp_port *port,
1477                 struct socket *newsock)
1478 {
1479         struct nvmet_tcp_queue *queue;
1480         int ret;
1481
1482         queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1483         if (!queue)
1484                 return -ENOMEM;
1485
1486         INIT_WORK(&queue->release_work, nvmet_tcp_release_queue_work);
1487         INIT_WORK(&queue->io_work, nvmet_tcp_io_work);
1488         queue->sock = newsock;
1489         queue->port = port;
1490         queue->nr_cmds = 0;
1491         spin_lock_init(&queue->state_lock);
1492         queue->state = NVMET_TCP_Q_CONNECTING;
1493         INIT_LIST_HEAD(&queue->free_list);
1494         init_llist_head(&queue->resp_list);
1495         INIT_LIST_HEAD(&queue->resp_send_list);
1496
1497         queue->idx = ida_simple_get(&nvmet_tcp_queue_ida, 0, 0, GFP_KERNEL);
1498         if (queue->idx < 0) {
1499                 ret = queue->idx;
1500                 goto out_free_queue;
1501         }
1502
1503         ret = nvmet_tcp_alloc_cmd(queue, &queue->connect);
1504         if (ret)
1505                 goto out_ida_remove;
1506
1507         ret = nvmet_sq_init(&queue->nvme_sq);
1508         if (ret)
1509                 goto out_free_connect;
1510
1511         port->last_cpu = cpumask_next_wrap(port->last_cpu,
1512                                 cpu_online_mask, -1, false);
1513         queue->cpu = port->last_cpu;
1514         nvmet_prepare_receive_pdu(queue);
1515
1516         mutex_lock(&nvmet_tcp_queue_mutex);
1517         list_add_tail(&queue->queue_list, &nvmet_tcp_queue_list);
1518         mutex_unlock(&nvmet_tcp_queue_mutex);
1519
1520         ret = nvmet_tcp_set_queue_sock(queue);
1521         if (ret)
1522                 goto out_destroy_sq;
1523
1524         queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1525
1526         return 0;
1527 out_destroy_sq:
1528         mutex_lock(&nvmet_tcp_queue_mutex);
1529         list_del_init(&queue->queue_list);
1530         mutex_unlock(&nvmet_tcp_queue_mutex);
1531         nvmet_sq_destroy(&queue->nvme_sq);
1532 out_free_connect:
1533         nvmet_tcp_free_cmd(&queue->connect);
1534 out_ida_remove:
1535         ida_simple_remove(&nvmet_tcp_queue_ida, queue->idx);
1536 out_free_queue:
1537         kfree(queue);
1538         return ret;
1539 }
1540
1541 static void nvmet_tcp_accept_work(struct work_struct *w)
1542 {
1543         struct nvmet_tcp_port *port =
1544                 container_of(w, struct nvmet_tcp_port, accept_work);
1545         struct socket *newsock;
1546         int ret;
1547
1548         while (true) {
1549                 ret = kernel_accept(port->sock, &newsock, O_NONBLOCK);
1550                 if (ret < 0) {
1551                         if (ret != -EAGAIN)
1552                                 pr_warn("failed to accept err=%d\n", ret);
1553                         return;
1554                 }
1555                 ret = nvmet_tcp_alloc_queue(port, newsock);
1556                 if (ret) {
1557                         pr_err("failed to allocate queue\n");
1558                         sock_release(newsock);
1559                 }
1560         }
1561 }
1562
1563 static void nvmet_tcp_listen_data_ready(struct sock *sk)
1564 {
1565         struct nvmet_tcp_port *port;
1566
1567         read_lock_bh(&sk->sk_callback_lock);
1568         port = sk->sk_user_data;
1569         if (!port)
1570                 goto out;
1571
1572         if (sk->sk_state == TCP_LISTEN)
1573                 schedule_work(&port->accept_work);
1574 out:
1575         read_unlock_bh(&sk->sk_callback_lock);
1576 }
1577
1578 static int nvmet_tcp_add_port(struct nvmet_port *nport)
1579 {
1580         struct nvmet_tcp_port *port;
1581         __kernel_sa_family_t af;
1582         int opt, ret;
1583
1584         port = kzalloc(sizeof(*port), GFP_KERNEL);
1585         if (!port)
1586                 return -ENOMEM;
1587
1588         switch (nport->disc_addr.adrfam) {
1589         case NVMF_ADDR_FAMILY_IP4:
1590                 af = AF_INET;
1591                 break;
1592         case NVMF_ADDR_FAMILY_IP6:
1593                 af = AF_INET6;
1594                 break;
1595         default:
1596                 pr_err("address family %d not supported\n",
1597                                 nport->disc_addr.adrfam);
1598                 ret = -EINVAL;
1599                 goto err_port;
1600         }
1601
1602         ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
1603                         nport->disc_addr.trsvcid, &port->addr);
1604         if (ret) {
1605                 pr_err("malformed ip/port passed: %s:%s\n",
1606                         nport->disc_addr.traddr, nport->disc_addr.trsvcid);
1607                 goto err_port;
1608         }
1609
1610         port->nport = nport;
1611         port->last_cpu = -1;
1612         INIT_WORK(&port->accept_work, nvmet_tcp_accept_work);
1613         if (port->nport->inline_data_size < 0)
1614                 port->nport->inline_data_size = NVMET_TCP_DEF_INLINE_DATA_SIZE;
1615
1616         ret = sock_create(port->addr.ss_family, SOCK_STREAM,
1617                                 IPPROTO_TCP, &port->sock);
1618         if (ret) {
1619                 pr_err("failed to create a socket\n");
1620                 goto err_port;
1621         }
1622
1623         port->sock->sk->sk_user_data = port;
1624         port->data_ready = port->sock->sk->sk_data_ready;
1625         port->sock->sk->sk_data_ready = nvmet_tcp_listen_data_ready;
1626
1627         opt = 1;
1628         ret = kernel_setsockopt(port->sock, IPPROTO_TCP,
1629                         TCP_NODELAY, (char *)&opt, sizeof(opt));
1630         if (ret) {
1631                 pr_err("failed to set TCP_NODELAY sock opt %d\n", ret);
1632                 goto err_sock;
1633         }
1634
1635         ret = kernel_setsockopt(port->sock, SOL_SOCKET, SO_REUSEADDR,
1636                         (char *)&opt, sizeof(opt));
1637         if (ret) {
1638                 pr_err("failed to set SO_REUSEADDR sock opt %d\n", ret);
1639                 goto err_sock;
1640         }
1641
1642         if (so_priority > 0) {
1643                 ret = kernel_setsockopt(port->sock, SOL_SOCKET, SO_PRIORITY,
1644                                 (char *)&so_priority, sizeof(so_priority));
1645                 if (ret) {
1646                         pr_err("failed to set SO_PRIORITY sock opt %d\n", ret);
1647                         goto err_sock;
1648                 }
1649         }
1650
1651         ret = kernel_bind(port->sock, (struct sockaddr *)&port->addr,
1652                         sizeof(port->addr));
1653         if (ret) {
1654                 pr_err("failed to bind port socket %d\n", ret);
1655                 goto err_sock;
1656         }
1657
1658         ret = kernel_listen(port->sock, 128);
1659         if (ret) {
1660                 pr_err("failed to listen %d on port sock\n", ret);
1661                 goto err_sock;
1662         }
1663
1664         nport->priv = port;
1665         pr_info("enabling port %d (%pISpc)\n",
1666                 le16_to_cpu(nport->disc_addr.portid), &port->addr);
1667
1668         return 0;
1669
1670 err_sock:
1671         sock_release(port->sock);
1672 err_port:
1673         kfree(port);
1674         return ret;
1675 }
1676
1677 static void nvmet_tcp_remove_port(struct nvmet_port *nport)
1678 {
1679         struct nvmet_tcp_port *port = nport->priv;
1680
1681         write_lock_bh(&port->sock->sk->sk_callback_lock);
1682         port->sock->sk->sk_data_ready = port->data_ready;
1683         port->sock->sk->sk_user_data = NULL;
1684         write_unlock_bh(&port->sock->sk->sk_callback_lock);
1685         cancel_work_sync(&port->accept_work);
1686
1687         sock_release(port->sock);
1688         kfree(port);
1689 }
1690
1691 static void nvmet_tcp_delete_ctrl(struct nvmet_ctrl *ctrl)
1692 {
1693         struct nvmet_tcp_queue *queue;
1694
1695         mutex_lock(&nvmet_tcp_queue_mutex);
1696         list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1697                 if (queue->nvme_sq.ctrl == ctrl)
1698                         kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1699         mutex_unlock(&nvmet_tcp_queue_mutex);
1700 }
1701
1702 static u16 nvmet_tcp_install_queue(struct nvmet_sq *sq)
1703 {
1704         struct nvmet_tcp_queue *queue =
1705                 container_of(sq, struct nvmet_tcp_queue, nvme_sq);
1706
1707         if (sq->qid == 0) {
1708                 /* Let inflight controller teardown complete */
1709                 flush_scheduled_work();
1710         }
1711
1712         queue->nr_cmds = sq->size * 2;
1713         if (nvmet_tcp_alloc_cmds(queue))
1714                 return NVME_SC_INTERNAL;
1715         return 0;
1716 }
1717
1718 static void nvmet_tcp_disc_port_addr(struct nvmet_req *req,
1719                 struct nvmet_port *nport, char *traddr)
1720 {
1721         struct nvmet_tcp_port *port = nport->priv;
1722
1723         if (inet_addr_is_any((struct sockaddr *)&port->addr)) {
1724                 struct nvmet_tcp_cmd *cmd =
1725                         container_of(req, struct nvmet_tcp_cmd, req);
1726                 struct nvmet_tcp_queue *queue = cmd->queue;
1727
1728                 sprintf(traddr, "%pISc", (struct sockaddr *)&queue->sockaddr);
1729         } else {
1730                 memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
1731         }
1732 }
1733
1734 static struct nvmet_fabrics_ops nvmet_tcp_ops = {
1735         .owner                  = THIS_MODULE,
1736         .type                   = NVMF_TRTYPE_TCP,
1737         .msdbd                  = 1,
1738         .has_keyed_sgls         = 0,
1739         .add_port               = nvmet_tcp_add_port,
1740         .remove_port            = nvmet_tcp_remove_port,
1741         .queue_response         = nvmet_tcp_queue_response,
1742         .delete_ctrl            = nvmet_tcp_delete_ctrl,
1743         .install_queue          = nvmet_tcp_install_queue,
1744         .disc_traddr            = nvmet_tcp_disc_port_addr,
1745 };
1746
1747 static int __init nvmet_tcp_init(void)
1748 {
1749         int ret;
1750
1751         nvmet_tcp_wq = alloc_workqueue("nvmet_tcp_wq", WQ_HIGHPRI, 0);
1752         if (!nvmet_tcp_wq)
1753                 return -ENOMEM;
1754
1755         ret = nvmet_register_transport(&nvmet_tcp_ops);
1756         if (ret)
1757                 goto err;
1758
1759         return 0;
1760 err:
1761         destroy_workqueue(nvmet_tcp_wq);
1762         return ret;
1763 }
1764
1765 static void __exit nvmet_tcp_exit(void)
1766 {
1767         struct nvmet_tcp_queue *queue;
1768
1769         nvmet_unregister_transport(&nvmet_tcp_ops);
1770
1771         flush_scheduled_work();
1772         mutex_lock(&nvmet_tcp_queue_mutex);
1773         list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1774                 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1775         mutex_unlock(&nvmet_tcp_queue_mutex);
1776         flush_scheduled_work();
1777
1778         destroy_workqueue(nvmet_tcp_wq);
1779 }
1780
1781 module_init(nvmet_tcp_init);
1782 module_exit(nvmet_tcp_exit);
1783
1784 MODULE_LICENSE("GPL v2");
1785 MODULE_ALIAS("nvmet-transport-3"); /* 3 == NVMF_TRTYPE_TCP */