1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * NVMe over Fabrics TCP target. 4 * Copyright (c) 2018 Lightbits Labs. All rights reserved. 5 */ 6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 7 #include <linux/module.h> 8 #include <linux/init.h> 9 #include <linux/slab.h> 10 #include <linux/err.h> 11 #include <linux/nvme-tcp.h> 12 #include <net/sock.h> 13 #include <net/tcp.h> 14 #include <linux/inet.h> 15 #include <linux/llist.h> 16 #include <crypto/hash.h> 17 18 #include "nvmet.h" 19 20 #define NVMET_TCP_DEF_INLINE_DATA_SIZE (4 * PAGE_SIZE) 21 22 /* Define the socket priority to use for connections were it is desirable 23 * that the NIC consider performing optimized packet processing or filtering. 24 * A non-zero value being sufficient to indicate general consideration of any 25 * possible optimization. Making it a module param allows for alternative 26 * values that may be unique for some NIC implementations. 27 */ 28 static int so_priority; 29 module_param(so_priority, int, 0644); 30 MODULE_PARM_DESC(so_priority, "nvmet tcp socket optimize priority"); 31 32 #define NVMET_TCP_RECV_BUDGET 8 33 #define NVMET_TCP_SEND_BUDGET 8 34 #define NVMET_TCP_IO_WORK_BUDGET 64 35 36 enum nvmet_tcp_send_state { 37 NVMET_TCP_SEND_DATA_PDU, 38 NVMET_TCP_SEND_DATA, 39 NVMET_TCP_SEND_R2T, 40 NVMET_TCP_SEND_DDGST, 41 NVMET_TCP_SEND_RESPONSE 42 }; 43 44 enum nvmet_tcp_recv_state { 45 NVMET_TCP_RECV_PDU, 46 NVMET_TCP_RECV_DATA, 47 NVMET_TCP_RECV_DDGST, 48 NVMET_TCP_RECV_ERR, 49 }; 50 51 enum { 52 NVMET_TCP_F_INIT_FAILED = (1 << 0), 53 }; 54 55 struct nvmet_tcp_cmd { 56 struct nvmet_tcp_queue *queue; 57 struct nvmet_req req; 58 59 struct nvme_tcp_cmd_pdu *cmd_pdu; 60 struct nvme_tcp_rsp_pdu *rsp_pdu; 61 struct nvme_tcp_data_pdu *data_pdu; 62 struct nvme_tcp_r2t_pdu *r2t_pdu; 63 64 u32 rbytes_done; 65 u32 wbytes_done; 66 67 u32 pdu_len; 68 u32 pdu_recv; 69 int sg_idx; 70 int nr_mapped; 71 struct msghdr recv_msg; 72 struct kvec *iov; 73 u32 flags; 74 75 struct list_head entry; 76 struct llist_node lentry; 77 78 /* send state */ 79 u32 offset; 80 struct scatterlist *cur_sg; 81 enum nvmet_tcp_send_state state; 82 83 __le32 exp_ddgst; 84 __le32 recv_ddgst; 85 }; 86 87 enum nvmet_tcp_queue_state { 88 NVMET_TCP_Q_CONNECTING, 89 NVMET_TCP_Q_LIVE, 90 NVMET_TCP_Q_DISCONNECTING, 91 }; 92 93 struct nvmet_tcp_queue { 94 struct socket *sock; 95 struct nvmet_tcp_port *port; 96 struct work_struct io_work; 97 struct nvmet_cq nvme_cq; 98 struct nvmet_sq nvme_sq; 99 100 /* send state */ 101 struct nvmet_tcp_cmd *cmds; 102 unsigned int nr_cmds; 103 struct list_head free_list; 104 struct llist_head resp_list; 105 struct list_head resp_send_list; 106 int send_list_len; 107 struct nvmet_tcp_cmd *snd_cmd; 108 109 /* recv state */ 110 int offset; 111 int left; 112 enum nvmet_tcp_recv_state rcv_state; 113 struct nvmet_tcp_cmd *cmd; 114 union nvme_tcp_pdu pdu; 115 116 /* digest state */ 117 bool hdr_digest; 118 bool data_digest; 119 struct ahash_request *snd_hash; 120 struct ahash_request *rcv_hash; 121 122 spinlock_t state_lock; 123 enum nvmet_tcp_queue_state state; 124 125 struct sockaddr_storage sockaddr; 126 struct sockaddr_storage sockaddr_peer; 127 struct work_struct release_work; 128 129 int idx; 130 struct list_head queue_list; 131 132 struct nvmet_tcp_cmd connect; 133 134 struct page_frag_cache pf_cache; 135 136 void (*data_ready)(struct sock *); 137 void (*state_change)(struct sock *); 138 void (*write_space)(struct sock *); 139 }; 140 141 struct nvmet_tcp_port { 142 struct socket *sock; 143 struct work_struct accept_work; 144 struct nvmet_port *nport; 145 struct sockaddr_storage addr; 146 void (*data_ready)(struct sock *); 147 }; 148 149 static DEFINE_IDA(nvmet_tcp_queue_ida); 150 static LIST_HEAD(nvmet_tcp_queue_list); 151 static DEFINE_MUTEX(nvmet_tcp_queue_mutex); 152 153 static struct workqueue_struct *nvmet_tcp_wq; 154 static const struct nvmet_fabrics_ops nvmet_tcp_ops; 155 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c); 156 static void nvmet_tcp_finish_cmd(struct nvmet_tcp_cmd *cmd); 157 158 static inline u16 nvmet_tcp_cmd_tag(struct nvmet_tcp_queue *queue, 159 struct nvmet_tcp_cmd *cmd) 160 { 161 if (unlikely(!queue->nr_cmds)) { 162 /* We didn't allocate cmds yet, send 0xffff */ 163 return USHRT_MAX; 164 } 165 166 return cmd - queue->cmds; 167 } 168 169 static inline bool nvmet_tcp_has_data_in(struct nvmet_tcp_cmd *cmd) 170 { 171 return nvme_is_write(cmd->req.cmd) && 172 cmd->rbytes_done < cmd->req.transfer_len; 173 } 174 175 static inline bool nvmet_tcp_need_data_in(struct nvmet_tcp_cmd *cmd) 176 { 177 return nvmet_tcp_has_data_in(cmd) && !cmd->req.cqe->status; 178 } 179 180 static inline bool nvmet_tcp_need_data_out(struct nvmet_tcp_cmd *cmd) 181 { 182 return !nvme_is_write(cmd->req.cmd) && 183 cmd->req.transfer_len > 0 && 184 !cmd->req.cqe->status; 185 } 186 187 static inline bool nvmet_tcp_has_inline_data(struct nvmet_tcp_cmd *cmd) 188 { 189 return nvme_is_write(cmd->req.cmd) && cmd->pdu_len && 190 !cmd->rbytes_done; 191 } 192 193 static inline struct nvmet_tcp_cmd * 194 nvmet_tcp_get_cmd(struct nvmet_tcp_queue *queue) 195 { 196 struct nvmet_tcp_cmd *cmd; 197 198 cmd = list_first_entry_or_null(&queue->free_list, 199 struct nvmet_tcp_cmd, entry); 200 if (!cmd) 201 return NULL; 202 list_del_init(&cmd->entry); 203 204 cmd->rbytes_done = cmd->wbytes_done = 0; 205 cmd->pdu_len = 0; 206 cmd->pdu_recv = 0; 207 cmd->iov = NULL; 208 cmd->flags = 0; 209 return cmd; 210 } 211 212 static inline void nvmet_tcp_put_cmd(struct nvmet_tcp_cmd *cmd) 213 { 214 if (unlikely(cmd == &cmd->queue->connect)) 215 return; 216 217 list_add_tail(&cmd->entry, &cmd->queue->free_list); 218 } 219 220 static inline int queue_cpu(struct nvmet_tcp_queue *queue) 221 { 222 return queue->sock->sk->sk_incoming_cpu; 223 } 224 225 static inline u8 nvmet_tcp_hdgst_len(struct nvmet_tcp_queue *queue) 226 { 227 return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0; 228 } 229 230 static inline u8 nvmet_tcp_ddgst_len(struct nvmet_tcp_queue *queue) 231 { 232 return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0; 233 } 234 235 static inline void nvmet_tcp_hdgst(struct ahash_request *hash, 236 void *pdu, size_t len) 237 { 238 struct scatterlist sg; 239 240 sg_init_one(&sg, pdu, len); 241 ahash_request_set_crypt(hash, &sg, pdu + len, len); 242 crypto_ahash_digest(hash); 243 } 244 245 static int nvmet_tcp_verify_hdgst(struct nvmet_tcp_queue *queue, 246 void *pdu, size_t len) 247 { 248 struct nvme_tcp_hdr *hdr = pdu; 249 __le32 recv_digest; 250 __le32 exp_digest; 251 252 if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) { 253 pr_err("queue %d: header digest enabled but no header digest\n", 254 queue->idx); 255 return -EPROTO; 256 } 257 258 recv_digest = *(__le32 *)(pdu + hdr->hlen); 259 nvmet_tcp_hdgst(queue->rcv_hash, pdu, len); 260 exp_digest = *(__le32 *)(pdu + hdr->hlen); 261 if (recv_digest != exp_digest) { 262 pr_err("queue %d: header digest error: recv %#x expected %#x\n", 263 queue->idx, le32_to_cpu(recv_digest), 264 le32_to_cpu(exp_digest)); 265 return -EPROTO; 266 } 267 268 return 0; 269 } 270 271 static int nvmet_tcp_check_ddgst(struct nvmet_tcp_queue *queue, void *pdu) 272 { 273 struct nvme_tcp_hdr *hdr = pdu; 274 u8 digest_len = nvmet_tcp_hdgst_len(queue); 275 u32 len; 276 277 len = le32_to_cpu(hdr->plen) - hdr->hlen - 278 (hdr->flags & NVME_TCP_F_HDGST ? digest_len : 0); 279 280 if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) { 281 pr_err("queue %d: data digest flag is cleared\n", queue->idx); 282 return -EPROTO; 283 } 284 285 return 0; 286 } 287 288 static void nvmet_tcp_unmap_pdu_iovec(struct nvmet_tcp_cmd *cmd) 289 { 290 struct scatterlist *sg; 291 int i; 292 293 sg = &cmd->req.sg[cmd->sg_idx]; 294 295 for (i = 0; i < cmd->nr_mapped; i++) 296 kunmap(sg_page(&sg[i])); 297 } 298 299 static void nvmet_tcp_map_pdu_iovec(struct nvmet_tcp_cmd *cmd) 300 { 301 struct kvec *iov = cmd->iov; 302 struct scatterlist *sg; 303 u32 length, offset, sg_offset; 304 305 length = cmd->pdu_len; 306 cmd->nr_mapped = DIV_ROUND_UP(length, PAGE_SIZE); 307 offset = cmd->rbytes_done; 308 cmd->sg_idx = offset / PAGE_SIZE; 309 sg_offset = offset % PAGE_SIZE; 310 sg = &cmd->req.sg[cmd->sg_idx]; 311 312 while (length) { 313 u32 iov_len = min_t(u32, length, sg->length - sg_offset); 314 315 iov->iov_base = kmap(sg_page(sg)) + sg->offset + sg_offset; 316 iov->iov_len = iov_len; 317 318 length -= iov_len; 319 sg = sg_next(sg); 320 iov++; 321 sg_offset = 0; 322 } 323 324 iov_iter_kvec(&cmd->recv_msg.msg_iter, READ, cmd->iov, 325 cmd->nr_mapped, cmd->pdu_len); 326 } 327 328 static void nvmet_tcp_fatal_error(struct nvmet_tcp_queue *queue) 329 { 330 queue->rcv_state = NVMET_TCP_RECV_ERR; 331 if (queue->nvme_sq.ctrl) 332 nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl); 333 else 334 kernel_sock_shutdown(queue->sock, SHUT_RDWR); 335 } 336 337 static void nvmet_tcp_socket_error(struct nvmet_tcp_queue *queue, int status) 338 { 339 if (status == -EPIPE || status == -ECONNRESET) 340 kernel_sock_shutdown(queue->sock, SHUT_RDWR); 341 else 342 nvmet_tcp_fatal_error(queue); 343 } 344 345 static int nvmet_tcp_map_data(struct nvmet_tcp_cmd *cmd) 346 { 347 struct nvme_sgl_desc *sgl = &cmd->req.cmd->common.dptr.sgl; 348 u32 len = le32_to_cpu(sgl->length); 349 350 if (!len) 351 return 0; 352 353 if (sgl->type == ((NVME_SGL_FMT_DATA_DESC << 4) | 354 NVME_SGL_FMT_OFFSET)) { 355 if (!nvme_is_write(cmd->req.cmd)) 356 return NVME_SC_INVALID_FIELD | NVME_SC_DNR; 357 358 if (len > cmd->req.port->inline_data_size) 359 return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR; 360 cmd->pdu_len = len; 361 } 362 cmd->req.transfer_len += len; 363 364 cmd->req.sg = sgl_alloc(len, GFP_KERNEL, &cmd->req.sg_cnt); 365 if (!cmd->req.sg) 366 return NVME_SC_INTERNAL; 367 cmd->cur_sg = cmd->req.sg; 368 369 if (nvmet_tcp_has_data_in(cmd)) { 370 cmd->iov = kmalloc_array(cmd->req.sg_cnt, 371 sizeof(*cmd->iov), GFP_KERNEL); 372 if (!cmd->iov) 373 goto err; 374 } 375 376 return 0; 377 err: 378 sgl_free(cmd->req.sg); 379 return NVME_SC_INTERNAL; 380 } 381 382 static void nvmet_tcp_ddgst(struct ahash_request *hash, 383 struct nvmet_tcp_cmd *cmd) 384 { 385 ahash_request_set_crypt(hash, cmd->req.sg, 386 (void *)&cmd->exp_ddgst, cmd->req.transfer_len); 387 crypto_ahash_digest(hash); 388 } 389 390 static void nvmet_setup_c2h_data_pdu(struct nvmet_tcp_cmd *cmd) 391 { 392 struct nvme_tcp_data_pdu *pdu = cmd->data_pdu; 393 struct nvmet_tcp_queue *queue = cmd->queue; 394 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue); 395 u8 ddgst = nvmet_tcp_ddgst_len(cmd->queue); 396 397 cmd->offset = 0; 398 cmd->state = NVMET_TCP_SEND_DATA_PDU; 399 400 pdu->hdr.type = nvme_tcp_c2h_data; 401 pdu->hdr.flags = NVME_TCP_F_DATA_LAST | (queue->nvme_sq.sqhd_disabled ? 402 NVME_TCP_F_DATA_SUCCESS : 0); 403 pdu->hdr.hlen = sizeof(*pdu); 404 pdu->hdr.pdo = pdu->hdr.hlen + hdgst; 405 pdu->hdr.plen = 406 cpu_to_le32(pdu->hdr.hlen + hdgst + 407 cmd->req.transfer_len + ddgst); 408 pdu->command_id = cmd->req.cqe->command_id; 409 pdu->data_length = cpu_to_le32(cmd->req.transfer_len); 410 pdu->data_offset = cpu_to_le32(cmd->wbytes_done); 411 412 if (queue->data_digest) { 413 pdu->hdr.flags |= NVME_TCP_F_DDGST; 414 nvmet_tcp_ddgst(queue->snd_hash, cmd); 415 } 416 417 if (cmd->queue->hdr_digest) { 418 pdu->hdr.flags |= NVME_TCP_F_HDGST; 419 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu)); 420 } 421 } 422 423 static void nvmet_setup_r2t_pdu(struct nvmet_tcp_cmd *cmd) 424 { 425 struct nvme_tcp_r2t_pdu *pdu = cmd->r2t_pdu; 426 struct nvmet_tcp_queue *queue = cmd->queue; 427 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue); 428 429 cmd->offset = 0; 430 cmd->state = NVMET_TCP_SEND_R2T; 431 432 pdu->hdr.type = nvme_tcp_r2t; 433 pdu->hdr.flags = 0; 434 pdu->hdr.hlen = sizeof(*pdu); 435 pdu->hdr.pdo = 0; 436 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst); 437 438 pdu->command_id = cmd->req.cmd->common.command_id; 439 pdu->ttag = nvmet_tcp_cmd_tag(cmd->queue, cmd); 440 pdu->r2t_length = cpu_to_le32(cmd->req.transfer_len - cmd->rbytes_done); 441 pdu->r2t_offset = cpu_to_le32(cmd->rbytes_done); 442 if (cmd->queue->hdr_digest) { 443 pdu->hdr.flags |= NVME_TCP_F_HDGST; 444 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu)); 445 } 446 } 447 448 static void nvmet_setup_response_pdu(struct nvmet_tcp_cmd *cmd) 449 { 450 struct nvme_tcp_rsp_pdu *pdu = cmd->rsp_pdu; 451 struct nvmet_tcp_queue *queue = cmd->queue; 452 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue); 453 454 cmd->offset = 0; 455 cmd->state = NVMET_TCP_SEND_RESPONSE; 456 457 pdu->hdr.type = nvme_tcp_rsp; 458 pdu->hdr.flags = 0; 459 pdu->hdr.hlen = sizeof(*pdu); 460 pdu->hdr.pdo = 0; 461 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst); 462 if (cmd->queue->hdr_digest) { 463 pdu->hdr.flags |= NVME_TCP_F_HDGST; 464 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu)); 465 } 466 } 467 468 static void nvmet_tcp_process_resp_list(struct nvmet_tcp_queue *queue) 469 { 470 struct llist_node *node; 471 struct nvmet_tcp_cmd *cmd; 472 473 for (node = llist_del_all(&queue->resp_list); node; node = node->next) { 474 cmd = llist_entry(node, struct nvmet_tcp_cmd, lentry); 475 list_add(&cmd->entry, &queue->resp_send_list); 476 queue->send_list_len++; 477 } 478 } 479 480 static struct nvmet_tcp_cmd *nvmet_tcp_fetch_cmd(struct nvmet_tcp_queue *queue) 481 { 482 queue->snd_cmd = list_first_entry_or_null(&queue->resp_send_list, 483 struct nvmet_tcp_cmd, entry); 484 if (!queue->snd_cmd) { 485 nvmet_tcp_process_resp_list(queue); 486 queue->snd_cmd = 487 list_first_entry_or_null(&queue->resp_send_list, 488 struct nvmet_tcp_cmd, entry); 489 if (unlikely(!queue->snd_cmd)) 490 return NULL; 491 } 492 493 list_del_init(&queue->snd_cmd->entry); 494 queue->send_list_len--; 495 496 if (nvmet_tcp_need_data_out(queue->snd_cmd)) 497 nvmet_setup_c2h_data_pdu(queue->snd_cmd); 498 else if (nvmet_tcp_need_data_in(queue->snd_cmd)) 499 nvmet_setup_r2t_pdu(queue->snd_cmd); 500 else 501 nvmet_setup_response_pdu(queue->snd_cmd); 502 503 return queue->snd_cmd; 504 } 505 506 static void nvmet_tcp_queue_response(struct nvmet_req *req) 507 { 508 struct nvmet_tcp_cmd *cmd = 509 container_of(req, struct nvmet_tcp_cmd, req); 510 struct nvmet_tcp_queue *queue = cmd->queue; 511 512 llist_add(&cmd->lentry, &queue->resp_list); 513 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &cmd->queue->io_work); 514 } 515 516 static int nvmet_try_send_data_pdu(struct nvmet_tcp_cmd *cmd) 517 { 518 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue); 519 int left = sizeof(*cmd->data_pdu) - cmd->offset + hdgst; 520 int ret; 521 522 ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->data_pdu), 523 offset_in_page(cmd->data_pdu) + cmd->offset, 524 left, MSG_DONTWAIT | MSG_MORE | MSG_SENDPAGE_NOTLAST); 525 if (ret <= 0) 526 return ret; 527 528 cmd->offset += ret; 529 left -= ret; 530 531 if (left) 532 return -EAGAIN; 533 534 cmd->state = NVMET_TCP_SEND_DATA; 535 cmd->offset = 0; 536 return 1; 537 } 538 539 static int nvmet_try_send_data(struct nvmet_tcp_cmd *cmd, bool last_in_batch) 540 { 541 struct nvmet_tcp_queue *queue = cmd->queue; 542 int ret; 543 544 while (cmd->cur_sg) { 545 struct page *page = sg_page(cmd->cur_sg); 546 u32 left = cmd->cur_sg->length - cmd->offset; 547 int flags = MSG_DONTWAIT; 548 549 if ((!last_in_batch && cmd->queue->send_list_len) || 550 cmd->wbytes_done + left < cmd->req.transfer_len || 551 queue->data_digest || !queue->nvme_sq.sqhd_disabled) 552 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST; 553 554 ret = kernel_sendpage(cmd->queue->sock, page, cmd->offset, 555 left, flags); 556 if (ret <= 0) 557 return ret; 558 559 cmd->offset += ret; 560 cmd->wbytes_done += ret; 561 562 /* Done with sg?*/ 563 if (cmd->offset == cmd->cur_sg->length) { 564 cmd->cur_sg = sg_next(cmd->cur_sg); 565 cmd->offset = 0; 566 } 567 } 568 569 if (queue->data_digest) { 570 cmd->state = NVMET_TCP_SEND_DDGST; 571 cmd->offset = 0; 572 } else { 573 if (queue->nvme_sq.sqhd_disabled) { 574 cmd->queue->snd_cmd = NULL; 575 nvmet_tcp_put_cmd(cmd); 576 } else { 577 nvmet_setup_response_pdu(cmd); 578 } 579 } 580 581 if (queue->nvme_sq.sqhd_disabled) { 582 kfree(cmd->iov); 583 sgl_free(cmd->req.sg); 584 } 585 586 return 1; 587 588 } 589 590 static int nvmet_try_send_response(struct nvmet_tcp_cmd *cmd, 591 bool last_in_batch) 592 { 593 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue); 594 int left = sizeof(*cmd->rsp_pdu) - cmd->offset + hdgst; 595 int flags = MSG_DONTWAIT; 596 int ret; 597 598 if (!last_in_batch && cmd->queue->send_list_len) 599 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST; 600 else 601 flags |= MSG_EOR; 602 603 ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->rsp_pdu), 604 offset_in_page(cmd->rsp_pdu) + cmd->offset, left, flags); 605 if (ret <= 0) 606 return ret; 607 cmd->offset += ret; 608 left -= ret; 609 610 if (left) 611 return -EAGAIN; 612 613 kfree(cmd->iov); 614 sgl_free(cmd->req.sg); 615 cmd->queue->snd_cmd = NULL; 616 nvmet_tcp_put_cmd(cmd); 617 return 1; 618 } 619 620 static int nvmet_try_send_r2t(struct nvmet_tcp_cmd *cmd, bool last_in_batch) 621 { 622 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue); 623 int left = sizeof(*cmd->r2t_pdu) - cmd->offset + hdgst; 624 int flags = MSG_DONTWAIT; 625 int ret; 626 627 if (!last_in_batch && cmd->queue->send_list_len) 628 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST; 629 else 630 flags |= MSG_EOR; 631 632 ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->r2t_pdu), 633 offset_in_page(cmd->r2t_pdu) + cmd->offset, left, flags); 634 if (ret <= 0) 635 return ret; 636 cmd->offset += ret; 637 left -= ret; 638 639 if (left) 640 return -EAGAIN; 641 642 cmd->queue->snd_cmd = NULL; 643 return 1; 644 } 645 646 static int nvmet_try_send_ddgst(struct nvmet_tcp_cmd *cmd, bool last_in_batch) 647 { 648 struct nvmet_tcp_queue *queue = cmd->queue; 649 struct msghdr msg = { .msg_flags = MSG_DONTWAIT }; 650 struct kvec iov = { 651 .iov_base = &cmd->exp_ddgst + cmd->offset, 652 .iov_len = NVME_TCP_DIGEST_LENGTH - cmd->offset 653 }; 654 int ret; 655 656 if (!last_in_batch && cmd->queue->send_list_len) 657 msg.msg_flags |= MSG_MORE; 658 else 659 msg.msg_flags |= MSG_EOR; 660 661 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len); 662 if (unlikely(ret <= 0)) 663 return ret; 664 665 cmd->offset += ret; 666 667 if (queue->nvme_sq.sqhd_disabled) { 668 cmd->queue->snd_cmd = NULL; 669 nvmet_tcp_put_cmd(cmd); 670 } else { 671 nvmet_setup_response_pdu(cmd); 672 } 673 return 1; 674 } 675 676 static int nvmet_tcp_try_send_one(struct nvmet_tcp_queue *queue, 677 bool last_in_batch) 678 { 679 struct nvmet_tcp_cmd *cmd = queue->snd_cmd; 680 int ret = 0; 681 682 if (!cmd || queue->state == NVMET_TCP_Q_DISCONNECTING) { 683 cmd = nvmet_tcp_fetch_cmd(queue); 684 if (unlikely(!cmd)) 685 return 0; 686 } 687 688 if (cmd->state == NVMET_TCP_SEND_DATA_PDU) { 689 ret = nvmet_try_send_data_pdu(cmd); 690 if (ret <= 0) 691 goto done_send; 692 } 693 694 if (cmd->state == NVMET_TCP_SEND_DATA) { 695 ret = nvmet_try_send_data(cmd, last_in_batch); 696 if (ret <= 0) 697 goto done_send; 698 } 699 700 if (cmd->state == NVMET_TCP_SEND_DDGST) { 701 ret = nvmet_try_send_ddgst(cmd, last_in_batch); 702 if (ret <= 0) 703 goto done_send; 704 } 705 706 if (cmd->state == NVMET_TCP_SEND_R2T) { 707 ret = nvmet_try_send_r2t(cmd, last_in_batch); 708 if (ret <= 0) 709 goto done_send; 710 } 711 712 if (cmd->state == NVMET_TCP_SEND_RESPONSE) 713 ret = nvmet_try_send_response(cmd, last_in_batch); 714 715 done_send: 716 if (ret < 0) { 717 if (ret == -EAGAIN) 718 return 0; 719 return ret; 720 } 721 722 return 1; 723 } 724 725 static int nvmet_tcp_try_send(struct nvmet_tcp_queue *queue, 726 int budget, int *sends) 727 { 728 int i, ret = 0; 729 730 for (i = 0; i < budget; i++) { 731 ret = nvmet_tcp_try_send_one(queue, i == budget - 1); 732 if (unlikely(ret < 0)) { 733 nvmet_tcp_socket_error(queue, ret); 734 goto done; 735 } else if (ret == 0) { 736 break; 737 } 738 (*sends)++; 739 } 740 done: 741 return ret; 742 } 743 744 static void nvmet_prepare_receive_pdu(struct nvmet_tcp_queue *queue) 745 { 746 queue->offset = 0; 747 queue->left = sizeof(struct nvme_tcp_hdr); 748 queue->cmd = NULL; 749 queue->rcv_state = NVMET_TCP_RECV_PDU; 750 } 751 752 static void nvmet_tcp_free_crypto(struct nvmet_tcp_queue *queue) 753 { 754 struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash); 755 756 ahash_request_free(queue->rcv_hash); 757 ahash_request_free(queue->snd_hash); 758 crypto_free_ahash(tfm); 759 } 760 761 static int nvmet_tcp_alloc_crypto(struct nvmet_tcp_queue *queue) 762 { 763 struct crypto_ahash *tfm; 764 765 tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC); 766 if (IS_ERR(tfm)) 767 return PTR_ERR(tfm); 768 769 queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL); 770 if (!queue->snd_hash) 771 goto free_tfm; 772 ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL); 773 774 queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL); 775 if (!queue->rcv_hash) 776 goto free_snd_hash; 777 ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL); 778 779 return 0; 780 free_snd_hash: 781 ahash_request_free(queue->snd_hash); 782 free_tfm: 783 crypto_free_ahash(tfm); 784 return -ENOMEM; 785 } 786 787 788 static int nvmet_tcp_handle_icreq(struct nvmet_tcp_queue *queue) 789 { 790 struct nvme_tcp_icreq_pdu *icreq = &queue->pdu.icreq; 791 struct nvme_tcp_icresp_pdu *icresp = &queue->pdu.icresp; 792 struct msghdr msg = {}; 793 struct kvec iov; 794 int ret; 795 796 if (le32_to_cpu(icreq->hdr.plen) != sizeof(struct nvme_tcp_icreq_pdu)) { 797 pr_err("bad nvme-tcp pdu length (%d)\n", 798 le32_to_cpu(icreq->hdr.plen)); 799 nvmet_tcp_fatal_error(queue); 800 } 801 802 if (icreq->pfv != NVME_TCP_PFV_1_0) { 803 pr_err("queue %d: bad pfv %d\n", queue->idx, icreq->pfv); 804 return -EPROTO; 805 } 806 807 if (icreq->hpda != 0) { 808 pr_err("queue %d: unsupported hpda %d\n", queue->idx, 809 icreq->hpda); 810 return -EPROTO; 811 } 812 813 queue->hdr_digest = !!(icreq->digest & NVME_TCP_HDR_DIGEST_ENABLE); 814 queue->data_digest = !!(icreq->digest & NVME_TCP_DATA_DIGEST_ENABLE); 815 if (queue->hdr_digest || queue->data_digest) { 816 ret = nvmet_tcp_alloc_crypto(queue); 817 if (ret) 818 return ret; 819 } 820 821 memset(icresp, 0, sizeof(*icresp)); 822 icresp->hdr.type = nvme_tcp_icresp; 823 icresp->hdr.hlen = sizeof(*icresp); 824 icresp->hdr.pdo = 0; 825 icresp->hdr.plen = cpu_to_le32(icresp->hdr.hlen); 826 icresp->pfv = cpu_to_le16(NVME_TCP_PFV_1_0); 827 icresp->maxdata = cpu_to_le32(0x400000); /* 16M arbitrary limit */ 828 icresp->cpda = 0; 829 if (queue->hdr_digest) 830 icresp->digest |= NVME_TCP_HDR_DIGEST_ENABLE; 831 if (queue->data_digest) 832 icresp->digest |= NVME_TCP_DATA_DIGEST_ENABLE; 833 834 iov.iov_base = icresp; 835 iov.iov_len = sizeof(*icresp); 836 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len); 837 if (ret < 0) 838 goto free_crypto; 839 840 queue->state = NVMET_TCP_Q_LIVE; 841 nvmet_prepare_receive_pdu(queue); 842 return 0; 843 free_crypto: 844 if (queue->hdr_digest || queue->data_digest) 845 nvmet_tcp_free_crypto(queue); 846 return ret; 847 } 848 849 static void nvmet_tcp_handle_req_failure(struct nvmet_tcp_queue *queue, 850 struct nvmet_tcp_cmd *cmd, struct nvmet_req *req) 851 { 852 size_t data_len = le32_to_cpu(req->cmd->common.dptr.sgl.length); 853 int ret; 854 855 if (!nvme_is_write(cmd->req.cmd) || 856 data_len > cmd->req.port->inline_data_size) { 857 nvmet_prepare_receive_pdu(queue); 858 return; 859 } 860 861 ret = nvmet_tcp_map_data(cmd); 862 if (unlikely(ret)) { 863 pr_err("queue %d: failed to map data\n", queue->idx); 864 nvmet_tcp_fatal_error(queue); 865 return; 866 } 867 868 queue->rcv_state = NVMET_TCP_RECV_DATA; 869 nvmet_tcp_map_pdu_iovec(cmd); 870 cmd->flags |= NVMET_TCP_F_INIT_FAILED; 871 } 872 873 static int nvmet_tcp_handle_h2c_data_pdu(struct nvmet_tcp_queue *queue) 874 { 875 struct nvme_tcp_data_pdu *data = &queue->pdu.data; 876 struct nvmet_tcp_cmd *cmd; 877 878 if (likely(queue->nr_cmds)) 879 cmd = &queue->cmds[data->ttag]; 880 else 881 cmd = &queue->connect; 882 883 if (le32_to_cpu(data->data_offset) != cmd->rbytes_done) { 884 pr_err("ttag %u unexpected data offset %u (expected %u)\n", 885 data->ttag, le32_to_cpu(data->data_offset), 886 cmd->rbytes_done); 887 /* FIXME: use path and transport errors */ 888 nvmet_req_complete(&cmd->req, 889 NVME_SC_INVALID_FIELD | NVME_SC_DNR); 890 return -EPROTO; 891 } 892 893 cmd->pdu_len = le32_to_cpu(data->data_length); 894 cmd->pdu_recv = 0; 895 nvmet_tcp_map_pdu_iovec(cmd); 896 queue->cmd = cmd; 897 queue->rcv_state = NVMET_TCP_RECV_DATA; 898 899 return 0; 900 } 901 902 static int nvmet_tcp_done_recv_pdu(struct nvmet_tcp_queue *queue) 903 { 904 struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr; 905 struct nvme_command *nvme_cmd = &queue->pdu.cmd.cmd; 906 struct nvmet_req *req; 907 int ret; 908 909 if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) { 910 if (hdr->type != nvme_tcp_icreq) { 911 pr_err("unexpected pdu type (%d) before icreq\n", 912 hdr->type); 913 nvmet_tcp_fatal_error(queue); 914 return -EPROTO; 915 } 916 return nvmet_tcp_handle_icreq(queue); 917 } 918 919 if (hdr->type == nvme_tcp_h2c_data) { 920 ret = nvmet_tcp_handle_h2c_data_pdu(queue); 921 if (unlikely(ret)) 922 return ret; 923 return 0; 924 } 925 926 queue->cmd = nvmet_tcp_get_cmd(queue); 927 if (unlikely(!queue->cmd)) { 928 /* This should never happen */ 929 pr_err("queue %d: out of commands (%d) send_list_len: %d, opcode: %d", 930 queue->idx, queue->nr_cmds, queue->send_list_len, 931 nvme_cmd->common.opcode); 932 nvmet_tcp_fatal_error(queue); 933 return -ENOMEM; 934 } 935 936 req = &queue->cmd->req; 937 memcpy(req->cmd, nvme_cmd, sizeof(*nvme_cmd)); 938 939 if (unlikely(!nvmet_req_init(req, &queue->nvme_cq, 940 &queue->nvme_sq, &nvmet_tcp_ops))) { 941 pr_err("failed cmd %p id %d opcode %d, data_len: %d\n", 942 req->cmd, req->cmd->common.command_id, 943 req->cmd->common.opcode, 944 le32_to_cpu(req->cmd->common.dptr.sgl.length)); 945 946 nvmet_tcp_handle_req_failure(queue, queue->cmd, req); 947 return -EAGAIN; 948 } 949 950 ret = nvmet_tcp_map_data(queue->cmd); 951 if (unlikely(ret)) { 952 pr_err("queue %d: failed to map data\n", queue->idx); 953 if (nvmet_tcp_has_inline_data(queue->cmd)) 954 nvmet_tcp_fatal_error(queue); 955 else 956 nvmet_req_complete(req, ret); 957 ret = -EAGAIN; 958 goto out; 959 } 960 961 if (nvmet_tcp_need_data_in(queue->cmd)) { 962 if (nvmet_tcp_has_inline_data(queue->cmd)) { 963 queue->rcv_state = NVMET_TCP_RECV_DATA; 964 nvmet_tcp_map_pdu_iovec(queue->cmd); 965 return 0; 966 } 967 /* send back R2T */ 968 nvmet_tcp_queue_response(&queue->cmd->req); 969 goto out; 970 } 971 972 queue->cmd->req.execute(&queue->cmd->req); 973 out: 974 nvmet_prepare_receive_pdu(queue); 975 return ret; 976 } 977 978 static const u8 nvme_tcp_pdu_sizes[] = { 979 [nvme_tcp_icreq] = sizeof(struct nvme_tcp_icreq_pdu), 980 [nvme_tcp_cmd] = sizeof(struct nvme_tcp_cmd_pdu), 981 [nvme_tcp_h2c_data] = sizeof(struct nvme_tcp_data_pdu), 982 }; 983 984 static inline u8 nvmet_tcp_pdu_size(u8 type) 985 { 986 size_t idx = type; 987 988 return (idx < ARRAY_SIZE(nvme_tcp_pdu_sizes) && 989 nvme_tcp_pdu_sizes[idx]) ? 990 nvme_tcp_pdu_sizes[idx] : 0; 991 } 992 993 static inline bool nvmet_tcp_pdu_valid(u8 type) 994 { 995 switch (type) { 996 case nvme_tcp_icreq: 997 case nvme_tcp_cmd: 998 case nvme_tcp_h2c_data: 999 /* fallthru */ 1000 return true; 1001 } 1002 1003 return false; 1004 } 1005 1006 static int nvmet_tcp_try_recv_pdu(struct nvmet_tcp_queue *queue) 1007 { 1008 struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr; 1009 int len; 1010 struct kvec iov; 1011 struct msghdr msg = { .msg_flags = MSG_DONTWAIT }; 1012 1013 recv: 1014 iov.iov_base = (void *)&queue->pdu + queue->offset; 1015 iov.iov_len = queue->left; 1016 len = kernel_recvmsg(queue->sock, &msg, &iov, 1, 1017 iov.iov_len, msg.msg_flags); 1018 if (unlikely(len < 0)) 1019 return len; 1020 1021 queue->offset += len; 1022 queue->left -= len; 1023 if (queue->left) 1024 return -EAGAIN; 1025 1026 if (queue->offset == sizeof(struct nvme_tcp_hdr)) { 1027 u8 hdgst = nvmet_tcp_hdgst_len(queue); 1028 1029 if (unlikely(!nvmet_tcp_pdu_valid(hdr->type))) { 1030 pr_err("unexpected pdu type %d\n", hdr->type); 1031 nvmet_tcp_fatal_error(queue); 1032 return -EIO; 1033 } 1034 1035 if (unlikely(hdr->hlen != nvmet_tcp_pdu_size(hdr->type))) { 1036 pr_err("pdu %d bad hlen %d\n", hdr->type, hdr->hlen); 1037 return -EIO; 1038 } 1039 1040 queue->left = hdr->hlen - queue->offset + hdgst; 1041 goto recv; 1042 } 1043 1044 if (queue->hdr_digest && 1045 nvmet_tcp_verify_hdgst(queue, &queue->pdu, queue->offset)) { 1046 nvmet_tcp_fatal_error(queue); /* fatal */ 1047 return -EPROTO; 1048 } 1049 1050 if (queue->data_digest && 1051 nvmet_tcp_check_ddgst(queue, &queue->pdu)) { 1052 nvmet_tcp_fatal_error(queue); /* fatal */ 1053 return -EPROTO; 1054 } 1055 1056 return nvmet_tcp_done_recv_pdu(queue); 1057 } 1058 1059 static void nvmet_tcp_prep_recv_ddgst(struct nvmet_tcp_cmd *cmd) 1060 { 1061 struct nvmet_tcp_queue *queue = cmd->queue; 1062 1063 nvmet_tcp_ddgst(queue->rcv_hash, cmd); 1064 queue->offset = 0; 1065 queue->left = NVME_TCP_DIGEST_LENGTH; 1066 queue->rcv_state = NVMET_TCP_RECV_DDGST; 1067 } 1068 1069 static int nvmet_tcp_try_recv_data(struct nvmet_tcp_queue *queue) 1070 { 1071 struct nvmet_tcp_cmd *cmd = queue->cmd; 1072 int ret; 1073 1074 while (msg_data_left(&cmd->recv_msg)) { 1075 ret = sock_recvmsg(cmd->queue->sock, &cmd->recv_msg, 1076 cmd->recv_msg.msg_flags); 1077 if (ret <= 0) 1078 return ret; 1079 1080 cmd->pdu_recv += ret; 1081 cmd->rbytes_done += ret; 1082 } 1083 1084 nvmet_tcp_unmap_pdu_iovec(cmd); 1085 1086 if (!(cmd->flags & NVMET_TCP_F_INIT_FAILED) && 1087 cmd->rbytes_done == cmd->req.transfer_len) { 1088 if (queue->data_digest) { 1089 nvmet_tcp_prep_recv_ddgst(cmd); 1090 return 0; 1091 } 1092 cmd->req.execute(&cmd->req); 1093 } 1094 1095 nvmet_prepare_receive_pdu(queue); 1096 return 0; 1097 } 1098 1099 static int nvmet_tcp_try_recv_ddgst(struct nvmet_tcp_queue *queue) 1100 { 1101 struct nvmet_tcp_cmd *cmd = queue->cmd; 1102 int ret; 1103 struct msghdr msg = { .msg_flags = MSG_DONTWAIT }; 1104 struct kvec iov = { 1105 .iov_base = (void *)&cmd->recv_ddgst + queue->offset, 1106 .iov_len = queue->left 1107 }; 1108 1109 ret = kernel_recvmsg(queue->sock, &msg, &iov, 1, 1110 iov.iov_len, msg.msg_flags); 1111 if (unlikely(ret < 0)) 1112 return ret; 1113 1114 queue->offset += ret; 1115 queue->left -= ret; 1116 if (queue->left) 1117 return -EAGAIN; 1118 1119 if (queue->data_digest && cmd->exp_ddgst != cmd->recv_ddgst) { 1120 pr_err("queue %d: cmd %d pdu (%d) data digest error: recv %#x expected %#x\n", 1121 queue->idx, cmd->req.cmd->common.command_id, 1122 queue->pdu.cmd.hdr.type, le32_to_cpu(cmd->recv_ddgst), 1123 le32_to_cpu(cmd->exp_ddgst)); 1124 nvmet_tcp_finish_cmd(cmd); 1125 nvmet_tcp_fatal_error(queue); 1126 ret = -EPROTO; 1127 goto out; 1128 } 1129 1130 if (!(cmd->flags & NVMET_TCP_F_INIT_FAILED) && 1131 cmd->rbytes_done == cmd->req.transfer_len) 1132 cmd->req.execute(&cmd->req); 1133 ret = 0; 1134 out: 1135 nvmet_prepare_receive_pdu(queue); 1136 return ret; 1137 } 1138 1139 static int nvmet_tcp_try_recv_one(struct nvmet_tcp_queue *queue) 1140 { 1141 int result = 0; 1142 1143 if (unlikely(queue->rcv_state == NVMET_TCP_RECV_ERR)) 1144 return 0; 1145 1146 if (queue->rcv_state == NVMET_TCP_RECV_PDU) { 1147 result = nvmet_tcp_try_recv_pdu(queue); 1148 if (result != 0) 1149 goto done_recv; 1150 } 1151 1152 if (queue->rcv_state == NVMET_TCP_RECV_DATA) { 1153 result = nvmet_tcp_try_recv_data(queue); 1154 if (result != 0) 1155 goto done_recv; 1156 } 1157 1158 if (queue->rcv_state == NVMET_TCP_RECV_DDGST) { 1159 result = nvmet_tcp_try_recv_ddgst(queue); 1160 if (result != 0) 1161 goto done_recv; 1162 } 1163 1164 done_recv: 1165 if (result < 0) { 1166 if (result == -EAGAIN) 1167 return 0; 1168 return result; 1169 } 1170 return 1; 1171 } 1172 1173 static int nvmet_tcp_try_recv(struct nvmet_tcp_queue *queue, 1174 int budget, int *recvs) 1175 { 1176 int i, ret = 0; 1177 1178 for (i = 0; i < budget; i++) { 1179 ret = nvmet_tcp_try_recv_one(queue); 1180 if (unlikely(ret < 0)) { 1181 nvmet_tcp_socket_error(queue, ret); 1182 goto done; 1183 } else if (ret == 0) { 1184 break; 1185 } 1186 (*recvs)++; 1187 } 1188 done: 1189 return ret; 1190 } 1191 1192 static void nvmet_tcp_schedule_release_queue(struct nvmet_tcp_queue *queue) 1193 { 1194 spin_lock(&queue->state_lock); 1195 if (queue->state != NVMET_TCP_Q_DISCONNECTING) { 1196 queue->state = NVMET_TCP_Q_DISCONNECTING; 1197 schedule_work(&queue->release_work); 1198 } 1199 spin_unlock(&queue->state_lock); 1200 } 1201 1202 static void nvmet_tcp_io_work(struct work_struct *w) 1203 { 1204 struct nvmet_tcp_queue *queue = 1205 container_of(w, struct nvmet_tcp_queue, io_work); 1206 bool pending; 1207 int ret, ops = 0; 1208 1209 do { 1210 pending = false; 1211 1212 ret = nvmet_tcp_try_recv(queue, NVMET_TCP_RECV_BUDGET, &ops); 1213 if (ret > 0) 1214 pending = true; 1215 else if (ret < 0) 1216 return; 1217 1218 ret = nvmet_tcp_try_send(queue, NVMET_TCP_SEND_BUDGET, &ops); 1219 if (ret > 0) 1220 pending = true; 1221 else if (ret < 0) 1222 return; 1223 1224 } while (pending && ops < NVMET_TCP_IO_WORK_BUDGET); 1225 1226 /* 1227 * We exahusted our budget, requeue our selves 1228 */ 1229 if (pending) 1230 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work); 1231 } 1232 1233 static int nvmet_tcp_alloc_cmd(struct nvmet_tcp_queue *queue, 1234 struct nvmet_tcp_cmd *c) 1235 { 1236 u8 hdgst = nvmet_tcp_hdgst_len(queue); 1237 1238 c->queue = queue; 1239 c->req.port = queue->port->nport; 1240 1241 c->cmd_pdu = page_frag_alloc(&queue->pf_cache, 1242 sizeof(*c->cmd_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO); 1243 if (!c->cmd_pdu) 1244 return -ENOMEM; 1245 c->req.cmd = &c->cmd_pdu->cmd; 1246 1247 c->rsp_pdu = page_frag_alloc(&queue->pf_cache, 1248 sizeof(*c->rsp_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO); 1249 if (!c->rsp_pdu) 1250 goto out_free_cmd; 1251 c->req.cqe = &c->rsp_pdu->cqe; 1252 1253 c->data_pdu = page_frag_alloc(&queue->pf_cache, 1254 sizeof(*c->data_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO); 1255 if (!c->data_pdu) 1256 goto out_free_rsp; 1257 1258 c->r2t_pdu = page_frag_alloc(&queue->pf_cache, 1259 sizeof(*c->r2t_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO); 1260 if (!c->r2t_pdu) 1261 goto out_free_data; 1262 1263 c->recv_msg.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL; 1264 1265 list_add_tail(&c->entry, &queue->free_list); 1266 1267 return 0; 1268 out_free_data: 1269 page_frag_free(c->data_pdu); 1270 out_free_rsp: 1271 page_frag_free(c->rsp_pdu); 1272 out_free_cmd: 1273 page_frag_free(c->cmd_pdu); 1274 return -ENOMEM; 1275 } 1276 1277 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c) 1278 { 1279 page_frag_free(c->r2t_pdu); 1280 page_frag_free(c->data_pdu); 1281 page_frag_free(c->rsp_pdu); 1282 page_frag_free(c->cmd_pdu); 1283 } 1284 1285 static int nvmet_tcp_alloc_cmds(struct nvmet_tcp_queue *queue) 1286 { 1287 struct nvmet_tcp_cmd *cmds; 1288 int i, ret = -EINVAL, nr_cmds = queue->nr_cmds; 1289 1290 cmds = kcalloc(nr_cmds, sizeof(struct nvmet_tcp_cmd), GFP_KERNEL); 1291 if (!cmds) 1292 goto out; 1293 1294 for (i = 0; i < nr_cmds; i++) { 1295 ret = nvmet_tcp_alloc_cmd(queue, cmds + i); 1296 if (ret) 1297 goto out_free; 1298 } 1299 1300 queue->cmds = cmds; 1301 1302 return 0; 1303 out_free: 1304 while (--i >= 0) 1305 nvmet_tcp_free_cmd(cmds + i); 1306 kfree(cmds); 1307 out: 1308 return ret; 1309 } 1310 1311 static void nvmet_tcp_free_cmds(struct nvmet_tcp_queue *queue) 1312 { 1313 struct nvmet_tcp_cmd *cmds = queue->cmds; 1314 int i; 1315 1316 for (i = 0; i < queue->nr_cmds; i++) 1317 nvmet_tcp_free_cmd(cmds + i); 1318 1319 nvmet_tcp_free_cmd(&queue->connect); 1320 kfree(cmds); 1321 } 1322 1323 static void nvmet_tcp_restore_socket_callbacks(struct nvmet_tcp_queue *queue) 1324 { 1325 struct socket *sock = queue->sock; 1326 1327 write_lock_bh(&sock->sk->sk_callback_lock); 1328 sock->sk->sk_data_ready = queue->data_ready; 1329 sock->sk->sk_state_change = queue->state_change; 1330 sock->sk->sk_write_space = queue->write_space; 1331 sock->sk->sk_user_data = NULL; 1332 write_unlock_bh(&sock->sk->sk_callback_lock); 1333 } 1334 1335 static void nvmet_tcp_finish_cmd(struct nvmet_tcp_cmd *cmd) 1336 { 1337 nvmet_req_uninit(&cmd->req); 1338 nvmet_tcp_unmap_pdu_iovec(cmd); 1339 kfree(cmd->iov); 1340 sgl_free(cmd->req.sg); 1341 } 1342 1343 static void nvmet_tcp_uninit_data_in_cmds(struct nvmet_tcp_queue *queue) 1344 { 1345 struct nvmet_tcp_cmd *cmd = queue->cmds; 1346 int i; 1347 1348 for (i = 0; i < queue->nr_cmds; i++, cmd++) { 1349 if (nvmet_tcp_need_data_in(cmd)) 1350 nvmet_tcp_finish_cmd(cmd); 1351 } 1352 1353 if (!queue->nr_cmds && nvmet_tcp_need_data_in(&queue->connect)) { 1354 /* failed in connect */ 1355 nvmet_tcp_finish_cmd(&queue->connect); 1356 } 1357 } 1358 1359 static void nvmet_tcp_release_queue_work(struct work_struct *w) 1360 { 1361 struct nvmet_tcp_queue *queue = 1362 container_of(w, struct nvmet_tcp_queue, release_work); 1363 1364 mutex_lock(&nvmet_tcp_queue_mutex); 1365 list_del_init(&queue->queue_list); 1366 mutex_unlock(&nvmet_tcp_queue_mutex); 1367 1368 nvmet_tcp_restore_socket_callbacks(queue); 1369 flush_work(&queue->io_work); 1370 1371 nvmet_tcp_uninit_data_in_cmds(queue); 1372 nvmet_sq_destroy(&queue->nvme_sq); 1373 cancel_work_sync(&queue->io_work); 1374 sock_release(queue->sock); 1375 nvmet_tcp_free_cmds(queue); 1376 if (queue->hdr_digest || queue->data_digest) 1377 nvmet_tcp_free_crypto(queue); 1378 ida_simple_remove(&nvmet_tcp_queue_ida, queue->idx); 1379 1380 kfree(queue); 1381 } 1382 1383 static void nvmet_tcp_data_ready(struct sock *sk) 1384 { 1385 struct nvmet_tcp_queue *queue; 1386 1387 read_lock_bh(&sk->sk_callback_lock); 1388 queue = sk->sk_user_data; 1389 if (likely(queue)) 1390 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work); 1391 read_unlock_bh(&sk->sk_callback_lock); 1392 } 1393 1394 static void nvmet_tcp_write_space(struct sock *sk) 1395 { 1396 struct nvmet_tcp_queue *queue; 1397 1398 read_lock_bh(&sk->sk_callback_lock); 1399 queue = sk->sk_user_data; 1400 if (unlikely(!queue)) 1401 goto out; 1402 1403 if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) { 1404 queue->write_space(sk); 1405 goto out; 1406 } 1407 1408 if (sk_stream_is_writeable(sk)) { 1409 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1410 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work); 1411 } 1412 out: 1413 read_unlock_bh(&sk->sk_callback_lock); 1414 } 1415 1416 static void nvmet_tcp_state_change(struct sock *sk) 1417 { 1418 struct nvmet_tcp_queue *queue; 1419 1420 write_lock_bh(&sk->sk_callback_lock); 1421 queue = sk->sk_user_data; 1422 if (!queue) 1423 goto done; 1424 1425 switch (sk->sk_state) { 1426 case TCP_FIN_WAIT1: 1427 case TCP_CLOSE_WAIT: 1428 case TCP_CLOSE: 1429 /* FALLTHRU */ 1430 sk->sk_user_data = NULL; 1431 nvmet_tcp_schedule_release_queue(queue); 1432 break; 1433 default: 1434 pr_warn("queue %d unhandled state %d\n", 1435 queue->idx, sk->sk_state); 1436 } 1437 done: 1438 write_unlock_bh(&sk->sk_callback_lock); 1439 } 1440 1441 static int nvmet_tcp_set_queue_sock(struct nvmet_tcp_queue *queue) 1442 { 1443 struct socket *sock = queue->sock; 1444 struct inet_sock *inet = inet_sk(sock->sk); 1445 int ret; 1446 1447 ret = kernel_getsockname(sock, 1448 (struct sockaddr *)&queue->sockaddr); 1449 if (ret < 0) 1450 return ret; 1451 1452 ret = kernel_getpeername(sock, 1453 (struct sockaddr *)&queue->sockaddr_peer); 1454 if (ret < 0) 1455 return ret; 1456 1457 /* 1458 * Cleanup whatever is sitting in the TCP transmit queue on socket 1459 * close. This is done to prevent stale data from being sent should 1460 * the network connection be restored before TCP times out. 1461 */ 1462 sock_no_linger(sock->sk); 1463 1464 if (so_priority > 0) 1465 sock_set_priority(sock->sk, so_priority); 1466 1467 /* Set socket type of service */ 1468 if (inet->rcv_tos > 0) 1469 ip_sock_set_tos(sock->sk, inet->rcv_tos); 1470 1471 write_lock_bh(&sock->sk->sk_callback_lock); 1472 sock->sk->sk_user_data = queue; 1473 queue->data_ready = sock->sk->sk_data_ready; 1474 sock->sk->sk_data_ready = nvmet_tcp_data_ready; 1475 queue->state_change = sock->sk->sk_state_change; 1476 sock->sk->sk_state_change = nvmet_tcp_state_change; 1477 queue->write_space = sock->sk->sk_write_space; 1478 sock->sk->sk_write_space = nvmet_tcp_write_space; 1479 write_unlock_bh(&sock->sk->sk_callback_lock); 1480 1481 return 0; 1482 } 1483 1484 static int nvmet_tcp_alloc_queue(struct nvmet_tcp_port *port, 1485 struct socket *newsock) 1486 { 1487 struct nvmet_tcp_queue *queue; 1488 int ret; 1489 1490 queue = kzalloc(sizeof(*queue), GFP_KERNEL); 1491 if (!queue) 1492 return -ENOMEM; 1493 1494 INIT_WORK(&queue->release_work, nvmet_tcp_release_queue_work); 1495 INIT_WORK(&queue->io_work, nvmet_tcp_io_work); 1496 queue->sock = newsock; 1497 queue->port = port; 1498 queue->nr_cmds = 0; 1499 spin_lock_init(&queue->state_lock); 1500 queue->state = NVMET_TCP_Q_CONNECTING; 1501 INIT_LIST_HEAD(&queue->free_list); 1502 init_llist_head(&queue->resp_list); 1503 INIT_LIST_HEAD(&queue->resp_send_list); 1504 1505 queue->idx = ida_simple_get(&nvmet_tcp_queue_ida, 0, 0, GFP_KERNEL); 1506 if (queue->idx < 0) { 1507 ret = queue->idx; 1508 goto out_free_queue; 1509 } 1510 1511 ret = nvmet_tcp_alloc_cmd(queue, &queue->connect); 1512 if (ret) 1513 goto out_ida_remove; 1514 1515 ret = nvmet_sq_init(&queue->nvme_sq); 1516 if (ret) 1517 goto out_free_connect; 1518 1519 nvmet_prepare_receive_pdu(queue); 1520 1521 mutex_lock(&nvmet_tcp_queue_mutex); 1522 list_add_tail(&queue->queue_list, &nvmet_tcp_queue_list); 1523 mutex_unlock(&nvmet_tcp_queue_mutex); 1524 1525 ret = nvmet_tcp_set_queue_sock(queue); 1526 if (ret) 1527 goto out_destroy_sq; 1528 1529 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work); 1530 1531 return 0; 1532 out_destroy_sq: 1533 mutex_lock(&nvmet_tcp_queue_mutex); 1534 list_del_init(&queue->queue_list); 1535 mutex_unlock(&nvmet_tcp_queue_mutex); 1536 nvmet_sq_destroy(&queue->nvme_sq); 1537 out_free_connect: 1538 nvmet_tcp_free_cmd(&queue->connect); 1539 out_ida_remove: 1540 ida_simple_remove(&nvmet_tcp_queue_ida, queue->idx); 1541 out_free_queue: 1542 kfree(queue); 1543 return ret; 1544 } 1545 1546 static void nvmet_tcp_accept_work(struct work_struct *w) 1547 { 1548 struct nvmet_tcp_port *port = 1549 container_of(w, struct nvmet_tcp_port, accept_work); 1550 struct socket *newsock; 1551 int ret; 1552 1553 while (true) { 1554 ret = kernel_accept(port->sock, &newsock, O_NONBLOCK); 1555 if (ret < 0) { 1556 if (ret != -EAGAIN) 1557 pr_warn("failed to accept err=%d\n", ret); 1558 return; 1559 } 1560 ret = nvmet_tcp_alloc_queue(port, newsock); 1561 if (ret) { 1562 pr_err("failed to allocate queue\n"); 1563 sock_release(newsock); 1564 } 1565 } 1566 } 1567 1568 static void nvmet_tcp_listen_data_ready(struct sock *sk) 1569 { 1570 struct nvmet_tcp_port *port; 1571 1572 read_lock_bh(&sk->sk_callback_lock); 1573 port = sk->sk_user_data; 1574 if (!port) 1575 goto out; 1576 1577 if (sk->sk_state == TCP_LISTEN) 1578 schedule_work(&port->accept_work); 1579 out: 1580 read_unlock_bh(&sk->sk_callback_lock); 1581 } 1582 1583 static int nvmet_tcp_add_port(struct nvmet_port *nport) 1584 { 1585 struct nvmet_tcp_port *port; 1586 __kernel_sa_family_t af; 1587 int ret; 1588 1589 port = kzalloc(sizeof(*port), GFP_KERNEL); 1590 if (!port) 1591 return -ENOMEM; 1592 1593 switch (nport->disc_addr.adrfam) { 1594 case NVMF_ADDR_FAMILY_IP4: 1595 af = AF_INET; 1596 break; 1597 case NVMF_ADDR_FAMILY_IP6: 1598 af = AF_INET6; 1599 break; 1600 default: 1601 pr_err("address family %d not supported\n", 1602 nport->disc_addr.adrfam); 1603 ret = -EINVAL; 1604 goto err_port; 1605 } 1606 1607 ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr, 1608 nport->disc_addr.trsvcid, &port->addr); 1609 if (ret) { 1610 pr_err("malformed ip/port passed: %s:%s\n", 1611 nport->disc_addr.traddr, nport->disc_addr.trsvcid); 1612 goto err_port; 1613 } 1614 1615 port->nport = nport; 1616 INIT_WORK(&port->accept_work, nvmet_tcp_accept_work); 1617 if (port->nport->inline_data_size < 0) 1618 port->nport->inline_data_size = NVMET_TCP_DEF_INLINE_DATA_SIZE; 1619 1620 ret = sock_create(port->addr.ss_family, SOCK_STREAM, 1621 IPPROTO_TCP, &port->sock); 1622 if (ret) { 1623 pr_err("failed to create a socket\n"); 1624 goto err_port; 1625 } 1626 1627 port->sock->sk->sk_user_data = port; 1628 port->data_ready = port->sock->sk->sk_data_ready; 1629 port->sock->sk->sk_data_ready = nvmet_tcp_listen_data_ready; 1630 sock_set_reuseaddr(port->sock->sk); 1631 tcp_sock_set_nodelay(port->sock->sk); 1632 if (so_priority > 0) 1633 sock_set_priority(port->sock->sk, so_priority); 1634 1635 ret = kernel_bind(port->sock, (struct sockaddr *)&port->addr, 1636 sizeof(port->addr)); 1637 if (ret) { 1638 pr_err("failed to bind port socket %d\n", ret); 1639 goto err_sock; 1640 } 1641 1642 ret = kernel_listen(port->sock, 128); 1643 if (ret) { 1644 pr_err("failed to listen %d on port sock\n", ret); 1645 goto err_sock; 1646 } 1647 1648 nport->priv = port; 1649 pr_info("enabling port %d (%pISpc)\n", 1650 le16_to_cpu(nport->disc_addr.portid), &port->addr); 1651 1652 return 0; 1653 1654 err_sock: 1655 sock_release(port->sock); 1656 err_port: 1657 kfree(port); 1658 return ret; 1659 } 1660 1661 static void nvmet_tcp_remove_port(struct nvmet_port *nport) 1662 { 1663 struct nvmet_tcp_port *port = nport->priv; 1664 1665 write_lock_bh(&port->sock->sk->sk_callback_lock); 1666 port->sock->sk->sk_data_ready = port->data_ready; 1667 port->sock->sk->sk_user_data = NULL; 1668 write_unlock_bh(&port->sock->sk->sk_callback_lock); 1669 cancel_work_sync(&port->accept_work); 1670 1671 sock_release(port->sock); 1672 kfree(port); 1673 } 1674 1675 static void nvmet_tcp_delete_ctrl(struct nvmet_ctrl *ctrl) 1676 { 1677 struct nvmet_tcp_queue *queue; 1678 1679 mutex_lock(&nvmet_tcp_queue_mutex); 1680 list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list) 1681 if (queue->nvme_sq.ctrl == ctrl) 1682 kernel_sock_shutdown(queue->sock, SHUT_RDWR); 1683 mutex_unlock(&nvmet_tcp_queue_mutex); 1684 } 1685 1686 static u16 nvmet_tcp_install_queue(struct nvmet_sq *sq) 1687 { 1688 struct nvmet_tcp_queue *queue = 1689 container_of(sq, struct nvmet_tcp_queue, nvme_sq); 1690 1691 if (sq->qid == 0) { 1692 /* Let inflight controller teardown complete */ 1693 flush_scheduled_work(); 1694 } 1695 1696 queue->nr_cmds = sq->size * 2; 1697 if (nvmet_tcp_alloc_cmds(queue)) 1698 return NVME_SC_INTERNAL; 1699 return 0; 1700 } 1701 1702 static void nvmet_tcp_disc_port_addr(struct nvmet_req *req, 1703 struct nvmet_port *nport, char *traddr) 1704 { 1705 struct nvmet_tcp_port *port = nport->priv; 1706 1707 if (inet_addr_is_any((struct sockaddr *)&port->addr)) { 1708 struct nvmet_tcp_cmd *cmd = 1709 container_of(req, struct nvmet_tcp_cmd, req); 1710 struct nvmet_tcp_queue *queue = cmd->queue; 1711 1712 sprintf(traddr, "%pISc", (struct sockaddr *)&queue->sockaddr); 1713 } else { 1714 memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE); 1715 } 1716 } 1717 1718 static const struct nvmet_fabrics_ops nvmet_tcp_ops = { 1719 .owner = THIS_MODULE, 1720 .type = NVMF_TRTYPE_TCP, 1721 .msdbd = 1, 1722 .add_port = nvmet_tcp_add_port, 1723 .remove_port = nvmet_tcp_remove_port, 1724 .queue_response = nvmet_tcp_queue_response, 1725 .delete_ctrl = nvmet_tcp_delete_ctrl, 1726 .install_queue = nvmet_tcp_install_queue, 1727 .disc_traddr = nvmet_tcp_disc_port_addr, 1728 }; 1729 1730 static int __init nvmet_tcp_init(void) 1731 { 1732 int ret; 1733 1734 nvmet_tcp_wq = alloc_workqueue("nvmet_tcp_wq", WQ_HIGHPRI, 0); 1735 if (!nvmet_tcp_wq) 1736 return -ENOMEM; 1737 1738 ret = nvmet_register_transport(&nvmet_tcp_ops); 1739 if (ret) 1740 goto err; 1741 1742 return 0; 1743 err: 1744 destroy_workqueue(nvmet_tcp_wq); 1745 return ret; 1746 } 1747 1748 static void __exit nvmet_tcp_exit(void) 1749 { 1750 struct nvmet_tcp_queue *queue; 1751 1752 nvmet_unregister_transport(&nvmet_tcp_ops); 1753 1754 flush_scheduled_work(); 1755 mutex_lock(&nvmet_tcp_queue_mutex); 1756 list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list) 1757 kernel_sock_shutdown(queue->sock, SHUT_RDWR); 1758 mutex_unlock(&nvmet_tcp_queue_mutex); 1759 flush_scheduled_work(); 1760 1761 destroy_workqueue(nvmet_tcp_wq); 1762 } 1763 1764 module_init(nvmet_tcp_init); 1765 module_exit(nvmet_tcp_exit); 1766 1767 MODULE_LICENSE("GPL v2"); 1768 MODULE_ALIAS("nvmet-transport-3"); /* 3 == NVMF_TRTYPE_TCP */ 1769