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 a time period (in usecs) that io_work() shall sample an activated 33 * queue before determining it to be idle. This optional module behavior 34 * can enable NIC solutions that support socket optimized packet processing 35 * using advanced interrupt moderation techniques. 36 */ 37 static int idle_poll_period_usecs; 38 module_param(idle_poll_period_usecs, int, 0644); 39 MODULE_PARM_DESC(idle_poll_period_usecs, 40 "nvmet tcp io_work poll till idle time period in usecs"); 41 42 #define NVMET_TCP_RECV_BUDGET 8 43 #define NVMET_TCP_SEND_BUDGET 8 44 #define NVMET_TCP_IO_WORK_BUDGET 64 45 46 enum nvmet_tcp_send_state { 47 NVMET_TCP_SEND_DATA_PDU, 48 NVMET_TCP_SEND_DATA, 49 NVMET_TCP_SEND_R2T, 50 NVMET_TCP_SEND_DDGST, 51 NVMET_TCP_SEND_RESPONSE 52 }; 53 54 enum nvmet_tcp_recv_state { 55 NVMET_TCP_RECV_PDU, 56 NVMET_TCP_RECV_DATA, 57 NVMET_TCP_RECV_DDGST, 58 NVMET_TCP_RECV_ERR, 59 }; 60 61 enum { 62 NVMET_TCP_F_INIT_FAILED = (1 << 0), 63 }; 64 65 struct nvmet_tcp_cmd { 66 struct nvmet_tcp_queue *queue; 67 struct nvmet_req req; 68 69 struct nvme_tcp_cmd_pdu *cmd_pdu; 70 struct nvme_tcp_rsp_pdu *rsp_pdu; 71 struct nvme_tcp_data_pdu *data_pdu; 72 struct nvme_tcp_r2t_pdu *r2t_pdu; 73 74 u32 rbytes_done; 75 u32 wbytes_done; 76 77 u32 pdu_len; 78 u32 pdu_recv; 79 int sg_idx; 80 int nr_mapped; 81 struct msghdr recv_msg; 82 struct kvec *iov; 83 u32 flags; 84 85 struct list_head entry; 86 struct llist_node lentry; 87 88 /* send state */ 89 u32 offset; 90 struct scatterlist *cur_sg; 91 enum nvmet_tcp_send_state state; 92 93 __le32 exp_ddgst; 94 __le32 recv_ddgst; 95 }; 96 97 enum nvmet_tcp_queue_state { 98 NVMET_TCP_Q_CONNECTING, 99 NVMET_TCP_Q_LIVE, 100 NVMET_TCP_Q_DISCONNECTING, 101 }; 102 103 struct nvmet_tcp_queue { 104 struct socket *sock; 105 struct nvmet_tcp_port *port; 106 struct work_struct io_work; 107 struct nvmet_cq nvme_cq; 108 struct nvmet_sq nvme_sq; 109 110 /* send state */ 111 struct nvmet_tcp_cmd *cmds; 112 unsigned int nr_cmds; 113 struct list_head free_list; 114 struct llist_head resp_list; 115 struct list_head resp_send_list; 116 int send_list_len; 117 struct nvmet_tcp_cmd *snd_cmd; 118 119 /* recv state */ 120 int offset; 121 int left; 122 enum nvmet_tcp_recv_state rcv_state; 123 struct nvmet_tcp_cmd *cmd; 124 union nvme_tcp_pdu pdu; 125 126 /* digest state */ 127 bool hdr_digest; 128 bool data_digest; 129 struct ahash_request *snd_hash; 130 struct ahash_request *rcv_hash; 131 132 unsigned long poll_end; 133 134 spinlock_t state_lock; 135 enum nvmet_tcp_queue_state state; 136 137 struct sockaddr_storage sockaddr; 138 struct sockaddr_storage sockaddr_peer; 139 struct work_struct release_work; 140 141 int idx; 142 struct list_head queue_list; 143 144 struct nvmet_tcp_cmd connect; 145 146 struct page_frag_cache pf_cache; 147 148 void (*data_ready)(struct sock *); 149 void (*state_change)(struct sock *); 150 void (*write_space)(struct sock *); 151 }; 152 153 struct nvmet_tcp_port { 154 struct socket *sock; 155 struct work_struct accept_work; 156 struct nvmet_port *nport; 157 struct sockaddr_storage addr; 158 void (*data_ready)(struct sock *); 159 }; 160 161 static DEFINE_IDA(nvmet_tcp_queue_ida); 162 static LIST_HEAD(nvmet_tcp_queue_list); 163 static DEFINE_MUTEX(nvmet_tcp_queue_mutex); 164 165 static struct workqueue_struct *nvmet_tcp_wq; 166 static const struct nvmet_fabrics_ops nvmet_tcp_ops; 167 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c); 168 static void nvmet_tcp_finish_cmd(struct nvmet_tcp_cmd *cmd); 169 static void nvmet_tcp_free_cmd_buffers(struct nvmet_tcp_cmd *cmd); 170 static void nvmet_tcp_unmap_pdu_iovec(struct nvmet_tcp_cmd *cmd); 171 172 static inline u16 nvmet_tcp_cmd_tag(struct nvmet_tcp_queue *queue, 173 struct nvmet_tcp_cmd *cmd) 174 { 175 if (unlikely(!queue->nr_cmds)) { 176 /* We didn't allocate cmds yet, send 0xffff */ 177 return USHRT_MAX; 178 } 179 180 return cmd - queue->cmds; 181 } 182 183 static inline bool nvmet_tcp_has_data_in(struct nvmet_tcp_cmd *cmd) 184 { 185 return nvme_is_write(cmd->req.cmd) && 186 cmd->rbytes_done < cmd->req.transfer_len; 187 } 188 189 static inline bool nvmet_tcp_need_data_in(struct nvmet_tcp_cmd *cmd) 190 { 191 return nvmet_tcp_has_data_in(cmd) && !cmd->req.cqe->status; 192 } 193 194 static inline bool nvmet_tcp_need_data_out(struct nvmet_tcp_cmd *cmd) 195 { 196 return !nvme_is_write(cmd->req.cmd) && 197 cmd->req.transfer_len > 0 && 198 !cmd->req.cqe->status; 199 } 200 201 static inline bool nvmet_tcp_has_inline_data(struct nvmet_tcp_cmd *cmd) 202 { 203 return nvme_is_write(cmd->req.cmd) && cmd->pdu_len && 204 !cmd->rbytes_done; 205 } 206 207 static inline struct nvmet_tcp_cmd * 208 nvmet_tcp_get_cmd(struct nvmet_tcp_queue *queue) 209 { 210 struct nvmet_tcp_cmd *cmd; 211 212 cmd = list_first_entry_or_null(&queue->free_list, 213 struct nvmet_tcp_cmd, entry); 214 if (!cmd) 215 return NULL; 216 list_del_init(&cmd->entry); 217 218 cmd->rbytes_done = cmd->wbytes_done = 0; 219 cmd->pdu_len = 0; 220 cmd->pdu_recv = 0; 221 cmd->iov = NULL; 222 cmd->flags = 0; 223 return cmd; 224 } 225 226 static inline void nvmet_tcp_put_cmd(struct nvmet_tcp_cmd *cmd) 227 { 228 if (unlikely(cmd == &cmd->queue->connect)) 229 return; 230 231 list_add_tail(&cmd->entry, &cmd->queue->free_list); 232 } 233 234 static inline int queue_cpu(struct nvmet_tcp_queue *queue) 235 { 236 return queue->sock->sk->sk_incoming_cpu; 237 } 238 239 static inline u8 nvmet_tcp_hdgst_len(struct nvmet_tcp_queue *queue) 240 { 241 return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0; 242 } 243 244 static inline u8 nvmet_tcp_ddgst_len(struct nvmet_tcp_queue *queue) 245 { 246 return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0; 247 } 248 249 static inline void nvmet_tcp_hdgst(struct ahash_request *hash, 250 void *pdu, size_t len) 251 { 252 struct scatterlist sg; 253 254 sg_init_one(&sg, pdu, len); 255 ahash_request_set_crypt(hash, &sg, pdu + len, len); 256 crypto_ahash_digest(hash); 257 } 258 259 static int nvmet_tcp_verify_hdgst(struct nvmet_tcp_queue *queue, 260 void *pdu, size_t len) 261 { 262 struct nvme_tcp_hdr *hdr = pdu; 263 __le32 recv_digest; 264 __le32 exp_digest; 265 266 if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) { 267 pr_err("queue %d: header digest enabled but no header digest\n", 268 queue->idx); 269 return -EPROTO; 270 } 271 272 recv_digest = *(__le32 *)(pdu + hdr->hlen); 273 nvmet_tcp_hdgst(queue->rcv_hash, pdu, len); 274 exp_digest = *(__le32 *)(pdu + hdr->hlen); 275 if (recv_digest != exp_digest) { 276 pr_err("queue %d: header digest error: recv %#x expected %#x\n", 277 queue->idx, le32_to_cpu(recv_digest), 278 le32_to_cpu(exp_digest)); 279 return -EPROTO; 280 } 281 282 return 0; 283 } 284 285 static int nvmet_tcp_check_ddgst(struct nvmet_tcp_queue *queue, void *pdu) 286 { 287 struct nvme_tcp_hdr *hdr = pdu; 288 u8 digest_len = nvmet_tcp_hdgst_len(queue); 289 u32 len; 290 291 len = le32_to_cpu(hdr->plen) - hdr->hlen - 292 (hdr->flags & NVME_TCP_F_HDGST ? digest_len : 0); 293 294 if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) { 295 pr_err("queue %d: data digest flag is cleared\n", queue->idx); 296 return -EPROTO; 297 } 298 299 return 0; 300 } 301 302 static void nvmet_tcp_free_cmd_buffers(struct nvmet_tcp_cmd *cmd) 303 { 304 WARN_ON(unlikely(cmd->nr_mapped > 0)); 305 306 kfree(cmd->iov); 307 sgl_free(cmd->req.sg); 308 cmd->iov = NULL; 309 cmd->req.sg = NULL; 310 } 311 312 static void nvmet_tcp_unmap_pdu_iovec(struct nvmet_tcp_cmd *cmd) 313 { 314 struct scatterlist *sg; 315 int i; 316 317 sg = &cmd->req.sg[cmd->sg_idx]; 318 319 for (i = 0; i < cmd->nr_mapped; i++) 320 kunmap(sg_page(&sg[i])); 321 322 cmd->nr_mapped = 0; 323 } 324 325 static void nvmet_tcp_map_pdu_iovec(struct nvmet_tcp_cmd *cmd) 326 { 327 struct kvec *iov = cmd->iov; 328 struct scatterlist *sg; 329 u32 length, offset, sg_offset; 330 331 length = cmd->pdu_len; 332 cmd->nr_mapped = DIV_ROUND_UP(length, PAGE_SIZE); 333 offset = cmd->rbytes_done; 334 cmd->sg_idx = offset / PAGE_SIZE; 335 sg_offset = offset % PAGE_SIZE; 336 sg = &cmd->req.sg[cmd->sg_idx]; 337 338 while (length) { 339 u32 iov_len = min_t(u32, length, sg->length - sg_offset); 340 341 iov->iov_base = kmap(sg_page(sg)) + sg->offset + sg_offset; 342 iov->iov_len = iov_len; 343 344 length -= iov_len; 345 sg = sg_next(sg); 346 iov++; 347 sg_offset = 0; 348 } 349 350 iov_iter_kvec(&cmd->recv_msg.msg_iter, READ, cmd->iov, 351 cmd->nr_mapped, cmd->pdu_len); 352 } 353 354 static void nvmet_tcp_fatal_error(struct nvmet_tcp_queue *queue) 355 { 356 queue->rcv_state = NVMET_TCP_RECV_ERR; 357 if (queue->nvme_sq.ctrl) 358 nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl); 359 else 360 kernel_sock_shutdown(queue->sock, SHUT_RDWR); 361 } 362 363 static void nvmet_tcp_socket_error(struct nvmet_tcp_queue *queue, int status) 364 { 365 if (status == -EPIPE || status == -ECONNRESET) 366 kernel_sock_shutdown(queue->sock, SHUT_RDWR); 367 else 368 nvmet_tcp_fatal_error(queue); 369 } 370 371 static int nvmet_tcp_map_data(struct nvmet_tcp_cmd *cmd) 372 { 373 struct nvme_sgl_desc *sgl = &cmd->req.cmd->common.dptr.sgl; 374 u32 len = le32_to_cpu(sgl->length); 375 376 if (!len) 377 return 0; 378 379 if (sgl->type == ((NVME_SGL_FMT_DATA_DESC << 4) | 380 NVME_SGL_FMT_OFFSET)) { 381 if (!nvme_is_write(cmd->req.cmd)) 382 return NVME_SC_INVALID_FIELD | NVME_SC_DNR; 383 384 if (len > cmd->req.port->inline_data_size) 385 return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR; 386 cmd->pdu_len = len; 387 } 388 cmd->req.transfer_len += len; 389 390 cmd->req.sg = sgl_alloc(len, GFP_KERNEL, &cmd->req.sg_cnt); 391 if (!cmd->req.sg) 392 return NVME_SC_INTERNAL; 393 cmd->cur_sg = cmd->req.sg; 394 395 if (nvmet_tcp_has_data_in(cmd)) { 396 cmd->iov = kmalloc_array(cmd->req.sg_cnt, 397 sizeof(*cmd->iov), GFP_KERNEL); 398 if (!cmd->iov) 399 goto err; 400 } 401 402 return 0; 403 err: 404 nvmet_tcp_free_cmd_buffers(cmd); 405 return NVME_SC_INTERNAL; 406 } 407 408 static void nvmet_tcp_send_ddgst(struct ahash_request *hash, 409 struct nvmet_tcp_cmd *cmd) 410 { 411 ahash_request_set_crypt(hash, cmd->req.sg, 412 (void *)&cmd->exp_ddgst, cmd->req.transfer_len); 413 crypto_ahash_digest(hash); 414 } 415 416 static void nvmet_tcp_recv_ddgst(struct ahash_request *hash, 417 struct nvmet_tcp_cmd *cmd) 418 { 419 struct scatterlist sg; 420 struct kvec *iov; 421 int i; 422 423 crypto_ahash_init(hash); 424 for (i = 0, iov = cmd->iov; i < cmd->nr_mapped; i++, iov++) { 425 sg_init_one(&sg, iov->iov_base, iov->iov_len); 426 ahash_request_set_crypt(hash, &sg, NULL, iov->iov_len); 427 crypto_ahash_update(hash); 428 } 429 ahash_request_set_crypt(hash, NULL, (void *)&cmd->exp_ddgst, 0); 430 crypto_ahash_final(hash); 431 } 432 433 static void nvmet_setup_c2h_data_pdu(struct nvmet_tcp_cmd *cmd) 434 { 435 struct nvme_tcp_data_pdu *pdu = cmd->data_pdu; 436 struct nvmet_tcp_queue *queue = cmd->queue; 437 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue); 438 u8 ddgst = nvmet_tcp_ddgst_len(cmd->queue); 439 440 cmd->offset = 0; 441 cmd->state = NVMET_TCP_SEND_DATA_PDU; 442 443 pdu->hdr.type = nvme_tcp_c2h_data; 444 pdu->hdr.flags = NVME_TCP_F_DATA_LAST | (queue->nvme_sq.sqhd_disabled ? 445 NVME_TCP_F_DATA_SUCCESS : 0); 446 pdu->hdr.hlen = sizeof(*pdu); 447 pdu->hdr.pdo = pdu->hdr.hlen + hdgst; 448 pdu->hdr.plen = 449 cpu_to_le32(pdu->hdr.hlen + hdgst + 450 cmd->req.transfer_len + ddgst); 451 pdu->command_id = cmd->req.cqe->command_id; 452 pdu->data_length = cpu_to_le32(cmd->req.transfer_len); 453 pdu->data_offset = cpu_to_le32(cmd->wbytes_done); 454 455 if (queue->data_digest) { 456 pdu->hdr.flags |= NVME_TCP_F_DDGST; 457 nvmet_tcp_send_ddgst(queue->snd_hash, cmd); 458 } 459 460 if (cmd->queue->hdr_digest) { 461 pdu->hdr.flags |= NVME_TCP_F_HDGST; 462 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu)); 463 } 464 } 465 466 static void nvmet_setup_r2t_pdu(struct nvmet_tcp_cmd *cmd) 467 { 468 struct nvme_tcp_r2t_pdu *pdu = cmd->r2t_pdu; 469 struct nvmet_tcp_queue *queue = cmd->queue; 470 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue); 471 472 cmd->offset = 0; 473 cmd->state = NVMET_TCP_SEND_R2T; 474 475 pdu->hdr.type = nvme_tcp_r2t; 476 pdu->hdr.flags = 0; 477 pdu->hdr.hlen = sizeof(*pdu); 478 pdu->hdr.pdo = 0; 479 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst); 480 481 pdu->command_id = cmd->req.cmd->common.command_id; 482 pdu->ttag = nvmet_tcp_cmd_tag(cmd->queue, cmd); 483 pdu->r2t_length = cpu_to_le32(cmd->req.transfer_len - cmd->rbytes_done); 484 pdu->r2t_offset = cpu_to_le32(cmd->rbytes_done); 485 if (cmd->queue->hdr_digest) { 486 pdu->hdr.flags |= NVME_TCP_F_HDGST; 487 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu)); 488 } 489 } 490 491 static void nvmet_setup_response_pdu(struct nvmet_tcp_cmd *cmd) 492 { 493 struct nvme_tcp_rsp_pdu *pdu = cmd->rsp_pdu; 494 struct nvmet_tcp_queue *queue = cmd->queue; 495 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue); 496 497 cmd->offset = 0; 498 cmd->state = NVMET_TCP_SEND_RESPONSE; 499 500 pdu->hdr.type = nvme_tcp_rsp; 501 pdu->hdr.flags = 0; 502 pdu->hdr.hlen = sizeof(*pdu); 503 pdu->hdr.pdo = 0; 504 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst); 505 if (cmd->queue->hdr_digest) { 506 pdu->hdr.flags |= NVME_TCP_F_HDGST; 507 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu)); 508 } 509 } 510 511 static void nvmet_tcp_process_resp_list(struct nvmet_tcp_queue *queue) 512 { 513 struct llist_node *node; 514 struct nvmet_tcp_cmd *cmd; 515 516 for (node = llist_del_all(&queue->resp_list); node; node = node->next) { 517 cmd = llist_entry(node, struct nvmet_tcp_cmd, lentry); 518 list_add(&cmd->entry, &queue->resp_send_list); 519 queue->send_list_len++; 520 } 521 } 522 523 static struct nvmet_tcp_cmd *nvmet_tcp_fetch_cmd(struct nvmet_tcp_queue *queue) 524 { 525 queue->snd_cmd = list_first_entry_or_null(&queue->resp_send_list, 526 struct nvmet_tcp_cmd, entry); 527 if (!queue->snd_cmd) { 528 nvmet_tcp_process_resp_list(queue); 529 queue->snd_cmd = 530 list_first_entry_or_null(&queue->resp_send_list, 531 struct nvmet_tcp_cmd, entry); 532 if (unlikely(!queue->snd_cmd)) 533 return NULL; 534 } 535 536 list_del_init(&queue->snd_cmd->entry); 537 queue->send_list_len--; 538 539 if (nvmet_tcp_need_data_out(queue->snd_cmd)) 540 nvmet_setup_c2h_data_pdu(queue->snd_cmd); 541 else if (nvmet_tcp_need_data_in(queue->snd_cmd)) 542 nvmet_setup_r2t_pdu(queue->snd_cmd); 543 else 544 nvmet_setup_response_pdu(queue->snd_cmd); 545 546 return queue->snd_cmd; 547 } 548 549 static void nvmet_tcp_queue_response(struct nvmet_req *req) 550 { 551 struct nvmet_tcp_cmd *cmd = 552 container_of(req, struct nvmet_tcp_cmd, req); 553 struct nvmet_tcp_queue *queue = cmd->queue; 554 struct nvme_sgl_desc *sgl; 555 u32 len; 556 557 if (unlikely(cmd == queue->cmd)) { 558 sgl = &cmd->req.cmd->common.dptr.sgl; 559 len = le32_to_cpu(sgl->length); 560 561 /* 562 * Wait for inline data before processing the response. 563 * Avoid using helpers, this might happen before 564 * nvmet_req_init is completed. 565 */ 566 if (queue->rcv_state == NVMET_TCP_RECV_PDU && 567 len && len <= cmd->req.port->inline_data_size && 568 nvme_is_write(cmd->req.cmd)) 569 return; 570 } 571 572 llist_add(&cmd->lentry, &queue->resp_list); 573 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &cmd->queue->io_work); 574 } 575 576 static void nvmet_tcp_execute_request(struct nvmet_tcp_cmd *cmd) 577 { 578 if (unlikely(cmd->flags & NVMET_TCP_F_INIT_FAILED)) 579 nvmet_tcp_queue_response(&cmd->req); 580 else 581 cmd->req.execute(&cmd->req); 582 } 583 584 static int nvmet_try_send_data_pdu(struct nvmet_tcp_cmd *cmd) 585 { 586 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue); 587 int left = sizeof(*cmd->data_pdu) - cmd->offset + hdgst; 588 int ret; 589 590 ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->data_pdu), 591 offset_in_page(cmd->data_pdu) + cmd->offset, 592 left, MSG_DONTWAIT | MSG_MORE | MSG_SENDPAGE_NOTLAST); 593 if (ret <= 0) 594 return ret; 595 596 cmd->offset += ret; 597 left -= ret; 598 599 if (left) 600 return -EAGAIN; 601 602 cmd->state = NVMET_TCP_SEND_DATA; 603 cmd->offset = 0; 604 return 1; 605 } 606 607 static int nvmet_try_send_data(struct nvmet_tcp_cmd *cmd, bool last_in_batch) 608 { 609 struct nvmet_tcp_queue *queue = cmd->queue; 610 int ret; 611 612 while (cmd->cur_sg) { 613 struct page *page = sg_page(cmd->cur_sg); 614 u32 left = cmd->cur_sg->length - cmd->offset; 615 int flags = MSG_DONTWAIT; 616 617 if ((!last_in_batch && cmd->queue->send_list_len) || 618 cmd->wbytes_done + left < cmd->req.transfer_len || 619 queue->data_digest || !queue->nvme_sq.sqhd_disabled) 620 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST; 621 622 ret = kernel_sendpage(cmd->queue->sock, page, cmd->offset, 623 left, flags); 624 if (ret <= 0) 625 return ret; 626 627 cmd->offset += ret; 628 cmd->wbytes_done += ret; 629 630 /* Done with sg?*/ 631 if (cmd->offset == cmd->cur_sg->length) { 632 cmd->cur_sg = sg_next(cmd->cur_sg); 633 cmd->offset = 0; 634 } 635 } 636 637 if (queue->data_digest) { 638 cmd->state = NVMET_TCP_SEND_DDGST; 639 cmd->offset = 0; 640 } else { 641 if (queue->nvme_sq.sqhd_disabled) { 642 cmd->queue->snd_cmd = NULL; 643 nvmet_tcp_put_cmd(cmd); 644 } else { 645 nvmet_setup_response_pdu(cmd); 646 } 647 } 648 649 if (queue->nvme_sq.sqhd_disabled) 650 nvmet_tcp_free_cmd_buffers(cmd); 651 652 return 1; 653 654 } 655 656 static int nvmet_try_send_response(struct nvmet_tcp_cmd *cmd, 657 bool last_in_batch) 658 { 659 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue); 660 int left = sizeof(*cmd->rsp_pdu) - cmd->offset + hdgst; 661 int flags = MSG_DONTWAIT; 662 int ret; 663 664 if (!last_in_batch && cmd->queue->send_list_len) 665 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST; 666 else 667 flags |= MSG_EOR; 668 669 ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->rsp_pdu), 670 offset_in_page(cmd->rsp_pdu) + cmd->offset, left, flags); 671 if (ret <= 0) 672 return ret; 673 cmd->offset += ret; 674 left -= ret; 675 676 if (left) 677 return -EAGAIN; 678 679 nvmet_tcp_free_cmd_buffers(cmd); 680 cmd->queue->snd_cmd = NULL; 681 nvmet_tcp_put_cmd(cmd); 682 return 1; 683 } 684 685 static int nvmet_try_send_r2t(struct nvmet_tcp_cmd *cmd, bool last_in_batch) 686 { 687 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue); 688 int left = sizeof(*cmd->r2t_pdu) - cmd->offset + hdgst; 689 int flags = MSG_DONTWAIT; 690 int ret; 691 692 if (!last_in_batch && cmd->queue->send_list_len) 693 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST; 694 else 695 flags |= MSG_EOR; 696 697 ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->r2t_pdu), 698 offset_in_page(cmd->r2t_pdu) + cmd->offset, left, flags); 699 if (ret <= 0) 700 return ret; 701 cmd->offset += ret; 702 left -= ret; 703 704 if (left) 705 return -EAGAIN; 706 707 cmd->queue->snd_cmd = NULL; 708 return 1; 709 } 710 711 static int nvmet_try_send_ddgst(struct nvmet_tcp_cmd *cmd, bool last_in_batch) 712 { 713 struct nvmet_tcp_queue *queue = cmd->queue; 714 int left = NVME_TCP_DIGEST_LENGTH - cmd->offset; 715 struct msghdr msg = { .msg_flags = MSG_DONTWAIT }; 716 struct kvec iov = { 717 .iov_base = (u8 *)&cmd->exp_ddgst + cmd->offset, 718 .iov_len = left 719 }; 720 int ret; 721 722 if (!last_in_batch && cmd->queue->send_list_len) 723 msg.msg_flags |= MSG_MORE; 724 else 725 msg.msg_flags |= MSG_EOR; 726 727 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len); 728 if (unlikely(ret <= 0)) 729 return ret; 730 731 cmd->offset += ret; 732 left -= ret; 733 734 if (left) 735 return -EAGAIN; 736 737 if (queue->nvme_sq.sqhd_disabled) { 738 cmd->queue->snd_cmd = NULL; 739 nvmet_tcp_put_cmd(cmd); 740 } else { 741 nvmet_setup_response_pdu(cmd); 742 } 743 return 1; 744 } 745 746 static int nvmet_tcp_try_send_one(struct nvmet_tcp_queue *queue, 747 bool last_in_batch) 748 { 749 struct nvmet_tcp_cmd *cmd = queue->snd_cmd; 750 int ret = 0; 751 752 if (!cmd || queue->state == NVMET_TCP_Q_DISCONNECTING) { 753 cmd = nvmet_tcp_fetch_cmd(queue); 754 if (unlikely(!cmd)) 755 return 0; 756 } 757 758 if (cmd->state == NVMET_TCP_SEND_DATA_PDU) { 759 ret = nvmet_try_send_data_pdu(cmd); 760 if (ret <= 0) 761 goto done_send; 762 } 763 764 if (cmd->state == NVMET_TCP_SEND_DATA) { 765 ret = nvmet_try_send_data(cmd, last_in_batch); 766 if (ret <= 0) 767 goto done_send; 768 } 769 770 if (cmd->state == NVMET_TCP_SEND_DDGST) { 771 ret = nvmet_try_send_ddgst(cmd, last_in_batch); 772 if (ret <= 0) 773 goto done_send; 774 } 775 776 if (cmd->state == NVMET_TCP_SEND_R2T) { 777 ret = nvmet_try_send_r2t(cmd, last_in_batch); 778 if (ret <= 0) 779 goto done_send; 780 } 781 782 if (cmd->state == NVMET_TCP_SEND_RESPONSE) 783 ret = nvmet_try_send_response(cmd, last_in_batch); 784 785 done_send: 786 if (ret < 0) { 787 if (ret == -EAGAIN) 788 return 0; 789 return ret; 790 } 791 792 return 1; 793 } 794 795 static int nvmet_tcp_try_send(struct nvmet_tcp_queue *queue, 796 int budget, int *sends) 797 { 798 int i, ret = 0; 799 800 for (i = 0; i < budget; i++) { 801 ret = nvmet_tcp_try_send_one(queue, i == budget - 1); 802 if (unlikely(ret < 0)) { 803 nvmet_tcp_socket_error(queue, ret); 804 goto done; 805 } else if (ret == 0) { 806 break; 807 } 808 (*sends)++; 809 } 810 done: 811 return ret; 812 } 813 814 static void nvmet_prepare_receive_pdu(struct nvmet_tcp_queue *queue) 815 { 816 queue->offset = 0; 817 queue->left = sizeof(struct nvme_tcp_hdr); 818 queue->cmd = NULL; 819 queue->rcv_state = NVMET_TCP_RECV_PDU; 820 } 821 822 static void nvmet_tcp_free_crypto(struct nvmet_tcp_queue *queue) 823 { 824 struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash); 825 826 ahash_request_free(queue->rcv_hash); 827 ahash_request_free(queue->snd_hash); 828 crypto_free_ahash(tfm); 829 } 830 831 static int nvmet_tcp_alloc_crypto(struct nvmet_tcp_queue *queue) 832 { 833 struct crypto_ahash *tfm; 834 835 tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC); 836 if (IS_ERR(tfm)) 837 return PTR_ERR(tfm); 838 839 queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL); 840 if (!queue->snd_hash) 841 goto free_tfm; 842 ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL); 843 844 queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL); 845 if (!queue->rcv_hash) 846 goto free_snd_hash; 847 ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL); 848 849 return 0; 850 free_snd_hash: 851 ahash_request_free(queue->snd_hash); 852 free_tfm: 853 crypto_free_ahash(tfm); 854 return -ENOMEM; 855 } 856 857 858 static int nvmet_tcp_handle_icreq(struct nvmet_tcp_queue *queue) 859 { 860 struct nvme_tcp_icreq_pdu *icreq = &queue->pdu.icreq; 861 struct nvme_tcp_icresp_pdu *icresp = &queue->pdu.icresp; 862 struct msghdr msg = {}; 863 struct kvec iov; 864 int ret; 865 866 if (le32_to_cpu(icreq->hdr.plen) != sizeof(struct nvme_tcp_icreq_pdu)) { 867 pr_err("bad nvme-tcp pdu length (%d)\n", 868 le32_to_cpu(icreq->hdr.plen)); 869 nvmet_tcp_fatal_error(queue); 870 } 871 872 if (icreq->pfv != NVME_TCP_PFV_1_0) { 873 pr_err("queue %d: bad pfv %d\n", queue->idx, icreq->pfv); 874 return -EPROTO; 875 } 876 877 if (icreq->hpda != 0) { 878 pr_err("queue %d: unsupported hpda %d\n", queue->idx, 879 icreq->hpda); 880 return -EPROTO; 881 } 882 883 queue->hdr_digest = !!(icreq->digest & NVME_TCP_HDR_DIGEST_ENABLE); 884 queue->data_digest = !!(icreq->digest & NVME_TCP_DATA_DIGEST_ENABLE); 885 if (queue->hdr_digest || queue->data_digest) { 886 ret = nvmet_tcp_alloc_crypto(queue); 887 if (ret) 888 return ret; 889 } 890 891 memset(icresp, 0, sizeof(*icresp)); 892 icresp->hdr.type = nvme_tcp_icresp; 893 icresp->hdr.hlen = sizeof(*icresp); 894 icresp->hdr.pdo = 0; 895 icresp->hdr.plen = cpu_to_le32(icresp->hdr.hlen); 896 icresp->pfv = cpu_to_le16(NVME_TCP_PFV_1_0); 897 icresp->maxdata = cpu_to_le32(0x400000); /* 16M arbitrary limit */ 898 icresp->cpda = 0; 899 if (queue->hdr_digest) 900 icresp->digest |= NVME_TCP_HDR_DIGEST_ENABLE; 901 if (queue->data_digest) 902 icresp->digest |= NVME_TCP_DATA_DIGEST_ENABLE; 903 904 iov.iov_base = icresp; 905 iov.iov_len = sizeof(*icresp); 906 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len); 907 if (ret < 0) 908 goto free_crypto; 909 910 queue->state = NVMET_TCP_Q_LIVE; 911 nvmet_prepare_receive_pdu(queue); 912 return 0; 913 free_crypto: 914 if (queue->hdr_digest || queue->data_digest) 915 nvmet_tcp_free_crypto(queue); 916 return ret; 917 } 918 919 static void nvmet_tcp_handle_req_failure(struct nvmet_tcp_queue *queue, 920 struct nvmet_tcp_cmd *cmd, struct nvmet_req *req) 921 { 922 size_t data_len = le32_to_cpu(req->cmd->common.dptr.sgl.length); 923 int ret; 924 925 /* 926 * This command has not been processed yet, hence we are trying to 927 * figure out if there is still pending data left to receive. If 928 * we don't, we can simply prepare for the next pdu and bail out, 929 * otherwise we will need to prepare a buffer and receive the 930 * stale data before continuing forward. 931 */ 932 if (!nvme_is_write(cmd->req.cmd) || !data_len || 933 data_len > cmd->req.port->inline_data_size) { 934 nvmet_prepare_receive_pdu(queue); 935 return; 936 } 937 938 ret = nvmet_tcp_map_data(cmd); 939 if (unlikely(ret)) { 940 pr_err("queue %d: failed to map data\n", queue->idx); 941 nvmet_tcp_fatal_error(queue); 942 return; 943 } 944 945 queue->rcv_state = NVMET_TCP_RECV_DATA; 946 nvmet_tcp_map_pdu_iovec(cmd); 947 cmd->flags |= NVMET_TCP_F_INIT_FAILED; 948 } 949 950 static int nvmet_tcp_handle_h2c_data_pdu(struct nvmet_tcp_queue *queue) 951 { 952 struct nvme_tcp_data_pdu *data = &queue->pdu.data; 953 struct nvmet_tcp_cmd *cmd; 954 955 if (likely(queue->nr_cmds)) 956 cmd = &queue->cmds[data->ttag]; 957 else 958 cmd = &queue->connect; 959 960 if (le32_to_cpu(data->data_offset) != cmd->rbytes_done) { 961 pr_err("ttag %u unexpected data offset %u (expected %u)\n", 962 data->ttag, le32_to_cpu(data->data_offset), 963 cmd->rbytes_done); 964 /* FIXME: use path and transport errors */ 965 nvmet_req_complete(&cmd->req, 966 NVME_SC_INVALID_FIELD | NVME_SC_DNR); 967 return -EPROTO; 968 } 969 970 cmd->pdu_len = le32_to_cpu(data->data_length); 971 cmd->pdu_recv = 0; 972 nvmet_tcp_map_pdu_iovec(cmd); 973 queue->cmd = cmd; 974 queue->rcv_state = NVMET_TCP_RECV_DATA; 975 976 return 0; 977 } 978 979 static int nvmet_tcp_done_recv_pdu(struct nvmet_tcp_queue *queue) 980 { 981 struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr; 982 struct nvme_command *nvme_cmd = &queue->pdu.cmd.cmd; 983 struct nvmet_req *req; 984 int ret; 985 986 if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) { 987 if (hdr->type != nvme_tcp_icreq) { 988 pr_err("unexpected pdu type (%d) before icreq\n", 989 hdr->type); 990 nvmet_tcp_fatal_error(queue); 991 return -EPROTO; 992 } 993 return nvmet_tcp_handle_icreq(queue); 994 } 995 996 if (hdr->type == nvme_tcp_h2c_data) { 997 ret = nvmet_tcp_handle_h2c_data_pdu(queue); 998 if (unlikely(ret)) 999 return ret; 1000 return 0; 1001 } 1002 1003 queue->cmd = nvmet_tcp_get_cmd(queue); 1004 if (unlikely(!queue->cmd)) { 1005 /* This should never happen */ 1006 pr_err("queue %d: out of commands (%d) send_list_len: %d, opcode: %d", 1007 queue->idx, queue->nr_cmds, queue->send_list_len, 1008 nvme_cmd->common.opcode); 1009 nvmet_tcp_fatal_error(queue); 1010 return -ENOMEM; 1011 } 1012 1013 req = &queue->cmd->req; 1014 memcpy(req->cmd, nvme_cmd, sizeof(*nvme_cmd)); 1015 1016 if (unlikely(!nvmet_req_init(req, &queue->nvme_cq, 1017 &queue->nvme_sq, &nvmet_tcp_ops))) { 1018 pr_err("failed cmd %p id %d opcode %d, data_len: %d\n", 1019 req->cmd, req->cmd->common.command_id, 1020 req->cmd->common.opcode, 1021 le32_to_cpu(req->cmd->common.dptr.sgl.length)); 1022 1023 nvmet_tcp_handle_req_failure(queue, queue->cmd, req); 1024 return 0; 1025 } 1026 1027 ret = nvmet_tcp_map_data(queue->cmd); 1028 if (unlikely(ret)) { 1029 pr_err("queue %d: failed to map data\n", queue->idx); 1030 if (nvmet_tcp_has_inline_data(queue->cmd)) 1031 nvmet_tcp_fatal_error(queue); 1032 else 1033 nvmet_req_complete(req, ret); 1034 ret = -EAGAIN; 1035 goto out; 1036 } 1037 1038 if (nvmet_tcp_need_data_in(queue->cmd)) { 1039 if (nvmet_tcp_has_inline_data(queue->cmd)) { 1040 queue->rcv_state = NVMET_TCP_RECV_DATA; 1041 nvmet_tcp_map_pdu_iovec(queue->cmd); 1042 return 0; 1043 } 1044 /* send back R2T */ 1045 nvmet_tcp_queue_response(&queue->cmd->req); 1046 goto out; 1047 } 1048 1049 queue->cmd->req.execute(&queue->cmd->req); 1050 out: 1051 nvmet_prepare_receive_pdu(queue); 1052 return ret; 1053 } 1054 1055 static const u8 nvme_tcp_pdu_sizes[] = { 1056 [nvme_tcp_icreq] = sizeof(struct nvme_tcp_icreq_pdu), 1057 [nvme_tcp_cmd] = sizeof(struct nvme_tcp_cmd_pdu), 1058 [nvme_tcp_h2c_data] = sizeof(struct nvme_tcp_data_pdu), 1059 }; 1060 1061 static inline u8 nvmet_tcp_pdu_size(u8 type) 1062 { 1063 size_t idx = type; 1064 1065 return (idx < ARRAY_SIZE(nvme_tcp_pdu_sizes) && 1066 nvme_tcp_pdu_sizes[idx]) ? 1067 nvme_tcp_pdu_sizes[idx] : 0; 1068 } 1069 1070 static inline bool nvmet_tcp_pdu_valid(u8 type) 1071 { 1072 switch (type) { 1073 case nvme_tcp_icreq: 1074 case nvme_tcp_cmd: 1075 case nvme_tcp_h2c_data: 1076 /* fallthru */ 1077 return true; 1078 } 1079 1080 return false; 1081 } 1082 1083 static int nvmet_tcp_try_recv_pdu(struct nvmet_tcp_queue *queue) 1084 { 1085 struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr; 1086 int len; 1087 struct kvec iov; 1088 struct msghdr msg = { .msg_flags = MSG_DONTWAIT }; 1089 1090 recv: 1091 iov.iov_base = (void *)&queue->pdu + queue->offset; 1092 iov.iov_len = queue->left; 1093 len = kernel_recvmsg(queue->sock, &msg, &iov, 1, 1094 iov.iov_len, msg.msg_flags); 1095 if (unlikely(len < 0)) 1096 return len; 1097 1098 queue->offset += len; 1099 queue->left -= len; 1100 if (queue->left) 1101 return -EAGAIN; 1102 1103 if (queue->offset == sizeof(struct nvme_tcp_hdr)) { 1104 u8 hdgst = nvmet_tcp_hdgst_len(queue); 1105 1106 if (unlikely(!nvmet_tcp_pdu_valid(hdr->type))) { 1107 pr_err("unexpected pdu type %d\n", hdr->type); 1108 nvmet_tcp_fatal_error(queue); 1109 return -EIO; 1110 } 1111 1112 if (unlikely(hdr->hlen != nvmet_tcp_pdu_size(hdr->type))) { 1113 pr_err("pdu %d bad hlen %d\n", hdr->type, hdr->hlen); 1114 return -EIO; 1115 } 1116 1117 queue->left = hdr->hlen - queue->offset + hdgst; 1118 goto recv; 1119 } 1120 1121 if (queue->hdr_digest && 1122 nvmet_tcp_verify_hdgst(queue, &queue->pdu, hdr->hlen)) { 1123 nvmet_tcp_fatal_error(queue); /* fatal */ 1124 return -EPROTO; 1125 } 1126 1127 if (queue->data_digest && 1128 nvmet_tcp_check_ddgst(queue, &queue->pdu)) { 1129 nvmet_tcp_fatal_error(queue); /* fatal */ 1130 return -EPROTO; 1131 } 1132 1133 return nvmet_tcp_done_recv_pdu(queue); 1134 } 1135 1136 static void nvmet_tcp_prep_recv_ddgst(struct nvmet_tcp_cmd *cmd) 1137 { 1138 struct nvmet_tcp_queue *queue = cmd->queue; 1139 1140 nvmet_tcp_recv_ddgst(queue->rcv_hash, cmd); 1141 queue->offset = 0; 1142 queue->left = NVME_TCP_DIGEST_LENGTH; 1143 queue->rcv_state = NVMET_TCP_RECV_DDGST; 1144 } 1145 1146 static int nvmet_tcp_try_recv_data(struct nvmet_tcp_queue *queue) 1147 { 1148 struct nvmet_tcp_cmd *cmd = queue->cmd; 1149 int ret; 1150 1151 while (msg_data_left(&cmd->recv_msg)) { 1152 ret = sock_recvmsg(cmd->queue->sock, &cmd->recv_msg, 1153 cmd->recv_msg.msg_flags); 1154 if (ret <= 0) 1155 return ret; 1156 1157 cmd->pdu_recv += ret; 1158 cmd->rbytes_done += ret; 1159 } 1160 1161 nvmet_tcp_unmap_pdu_iovec(cmd); 1162 if (queue->data_digest) { 1163 nvmet_tcp_prep_recv_ddgst(cmd); 1164 return 0; 1165 } 1166 1167 if (cmd->rbytes_done == cmd->req.transfer_len) 1168 nvmet_tcp_execute_request(cmd); 1169 1170 nvmet_prepare_receive_pdu(queue); 1171 return 0; 1172 } 1173 1174 static int nvmet_tcp_try_recv_ddgst(struct nvmet_tcp_queue *queue) 1175 { 1176 struct nvmet_tcp_cmd *cmd = queue->cmd; 1177 int ret; 1178 struct msghdr msg = { .msg_flags = MSG_DONTWAIT }; 1179 struct kvec iov = { 1180 .iov_base = (void *)&cmd->recv_ddgst + queue->offset, 1181 .iov_len = queue->left 1182 }; 1183 1184 ret = kernel_recvmsg(queue->sock, &msg, &iov, 1, 1185 iov.iov_len, msg.msg_flags); 1186 if (unlikely(ret < 0)) 1187 return ret; 1188 1189 queue->offset += ret; 1190 queue->left -= ret; 1191 if (queue->left) 1192 return -EAGAIN; 1193 1194 if (queue->data_digest && cmd->exp_ddgst != cmd->recv_ddgst) { 1195 pr_err("queue %d: cmd %d pdu (%d) data digest error: recv %#x expected %#x\n", 1196 queue->idx, cmd->req.cmd->common.command_id, 1197 queue->pdu.cmd.hdr.type, le32_to_cpu(cmd->recv_ddgst), 1198 le32_to_cpu(cmd->exp_ddgst)); 1199 nvmet_tcp_finish_cmd(cmd); 1200 nvmet_tcp_fatal_error(queue); 1201 ret = -EPROTO; 1202 goto out; 1203 } 1204 1205 if (cmd->rbytes_done == cmd->req.transfer_len) 1206 nvmet_tcp_execute_request(cmd); 1207 1208 ret = 0; 1209 out: 1210 nvmet_prepare_receive_pdu(queue); 1211 return ret; 1212 } 1213 1214 static int nvmet_tcp_try_recv_one(struct nvmet_tcp_queue *queue) 1215 { 1216 int result = 0; 1217 1218 if (unlikely(queue->rcv_state == NVMET_TCP_RECV_ERR)) 1219 return 0; 1220 1221 if (queue->rcv_state == NVMET_TCP_RECV_PDU) { 1222 result = nvmet_tcp_try_recv_pdu(queue); 1223 if (result != 0) 1224 goto done_recv; 1225 } 1226 1227 if (queue->rcv_state == NVMET_TCP_RECV_DATA) { 1228 result = nvmet_tcp_try_recv_data(queue); 1229 if (result != 0) 1230 goto done_recv; 1231 } 1232 1233 if (queue->rcv_state == NVMET_TCP_RECV_DDGST) { 1234 result = nvmet_tcp_try_recv_ddgst(queue); 1235 if (result != 0) 1236 goto done_recv; 1237 } 1238 1239 done_recv: 1240 if (result < 0) { 1241 if (result == -EAGAIN) 1242 return 0; 1243 return result; 1244 } 1245 return 1; 1246 } 1247 1248 static int nvmet_tcp_try_recv(struct nvmet_tcp_queue *queue, 1249 int budget, int *recvs) 1250 { 1251 int i, ret = 0; 1252 1253 for (i = 0; i < budget; i++) { 1254 ret = nvmet_tcp_try_recv_one(queue); 1255 if (unlikely(ret < 0)) { 1256 nvmet_tcp_socket_error(queue, ret); 1257 goto done; 1258 } else if (ret == 0) { 1259 break; 1260 } 1261 (*recvs)++; 1262 } 1263 done: 1264 return ret; 1265 } 1266 1267 static void nvmet_tcp_schedule_release_queue(struct nvmet_tcp_queue *queue) 1268 { 1269 spin_lock(&queue->state_lock); 1270 if (queue->state != NVMET_TCP_Q_DISCONNECTING) { 1271 queue->state = NVMET_TCP_Q_DISCONNECTING; 1272 schedule_work(&queue->release_work); 1273 } 1274 spin_unlock(&queue->state_lock); 1275 } 1276 1277 static inline void nvmet_tcp_arm_queue_deadline(struct nvmet_tcp_queue *queue) 1278 { 1279 queue->poll_end = jiffies + usecs_to_jiffies(idle_poll_period_usecs); 1280 } 1281 1282 static bool nvmet_tcp_check_queue_deadline(struct nvmet_tcp_queue *queue, 1283 int ops) 1284 { 1285 if (!idle_poll_period_usecs) 1286 return false; 1287 1288 if (ops) 1289 nvmet_tcp_arm_queue_deadline(queue); 1290 1291 return !time_after(jiffies, queue->poll_end); 1292 } 1293 1294 static void nvmet_tcp_io_work(struct work_struct *w) 1295 { 1296 struct nvmet_tcp_queue *queue = 1297 container_of(w, struct nvmet_tcp_queue, io_work); 1298 bool pending; 1299 int ret, ops = 0; 1300 1301 do { 1302 pending = false; 1303 1304 ret = nvmet_tcp_try_recv(queue, NVMET_TCP_RECV_BUDGET, &ops); 1305 if (ret > 0) 1306 pending = true; 1307 else if (ret < 0) 1308 return; 1309 1310 ret = nvmet_tcp_try_send(queue, NVMET_TCP_SEND_BUDGET, &ops); 1311 if (ret > 0) 1312 pending = true; 1313 else if (ret < 0) 1314 return; 1315 1316 } while (pending && ops < NVMET_TCP_IO_WORK_BUDGET); 1317 1318 /* 1319 * Requeue the worker if idle deadline period is in progress or any 1320 * ops activity was recorded during the do-while loop above. 1321 */ 1322 if (nvmet_tcp_check_queue_deadline(queue, ops) || pending) 1323 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work); 1324 } 1325 1326 static int nvmet_tcp_alloc_cmd(struct nvmet_tcp_queue *queue, 1327 struct nvmet_tcp_cmd *c) 1328 { 1329 u8 hdgst = nvmet_tcp_hdgst_len(queue); 1330 1331 c->queue = queue; 1332 c->req.port = queue->port->nport; 1333 1334 c->cmd_pdu = page_frag_alloc(&queue->pf_cache, 1335 sizeof(*c->cmd_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO); 1336 if (!c->cmd_pdu) 1337 return -ENOMEM; 1338 c->req.cmd = &c->cmd_pdu->cmd; 1339 1340 c->rsp_pdu = page_frag_alloc(&queue->pf_cache, 1341 sizeof(*c->rsp_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO); 1342 if (!c->rsp_pdu) 1343 goto out_free_cmd; 1344 c->req.cqe = &c->rsp_pdu->cqe; 1345 1346 c->data_pdu = page_frag_alloc(&queue->pf_cache, 1347 sizeof(*c->data_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO); 1348 if (!c->data_pdu) 1349 goto out_free_rsp; 1350 1351 c->r2t_pdu = page_frag_alloc(&queue->pf_cache, 1352 sizeof(*c->r2t_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO); 1353 if (!c->r2t_pdu) 1354 goto out_free_data; 1355 1356 c->recv_msg.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL; 1357 1358 list_add_tail(&c->entry, &queue->free_list); 1359 1360 return 0; 1361 out_free_data: 1362 page_frag_free(c->data_pdu); 1363 out_free_rsp: 1364 page_frag_free(c->rsp_pdu); 1365 out_free_cmd: 1366 page_frag_free(c->cmd_pdu); 1367 return -ENOMEM; 1368 } 1369 1370 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c) 1371 { 1372 page_frag_free(c->r2t_pdu); 1373 page_frag_free(c->data_pdu); 1374 page_frag_free(c->rsp_pdu); 1375 page_frag_free(c->cmd_pdu); 1376 } 1377 1378 static int nvmet_tcp_alloc_cmds(struct nvmet_tcp_queue *queue) 1379 { 1380 struct nvmet_tcp_cmd *cmds; 1381 int i, ret = -EINVAL, nr_cmds = queue->nr_cmds; 1382 1383 cmds = kcalloc(nr_cmds, sizeof(struct nvmet_tcp_cmd), GFP_KERNEL); 1384 if (!cmds) 1385 goto out; 1386 1387 for (i = 0; i < nr_cmds; i++) { 1388 ret = nvmet_tcp_alloc_cmd(queue, cmds + i); 1389 if (ret) 1390 goto out_free; 1391 } 1392 1393 queue->cmds = cmds; 1394 1395 return 0; 1396 out_free: 1397 while (--i >= 0) 1398 nvmet_tcp_free_cmd(cmds + i); 1399 kfree(cmds); 1400 out: 1401 return ret; 1402 } 1403 1404 static void nvmet_tcp_free_cmds(struct nvmet_tcp_queue *queue) 1405 { 1406 struct nvmet_tcp_cmd *cmds = queue->cmds; 1407 int i; 1408 1409 for (i = 0; i < queue->nr_cmds; i++) 1410 nvmet_tcp_free_cmd(cmds + i); 1411 1412 nvmet_tcp_free_cmd(&queue->connect); 1413 kfree(cmds); 1414 } 1415 1416 static void nvmet_tcp_restore_socket_callbacks(struct nvmet_tcp_queue *queue) 1417 { 1418 struct socket *sock = queue->sock; 1419 1420 write_lock_bh(&sock->sk->sk_callback_lock); 1421 sock->sk->sk_data_ready = queue->data_ready; 1422 sock->sk->sk_state_change = queue->state_change; 1423 sock->sk->sk_write_space = queue->write_space; 1424 sock->sk->sk_user_data = NULL; 1425 write_unlock_bh(&sock->sk->sk_callback_lock); 1426 } 1427 1428 static void nvmet_tcp_finish_cmd(struct nvmet_tcp_cmd *cmd) 1429 { 1430 nvmet_req_uninit(&cmd->req); 1431 nvmet_tcp_unmap_pdu_iovec(cmd); 1432 nvmet_tcp_free_cmd_buffers(cmd); 1433 } 1434 1435 static void nvmet_tcp_uninit_data_in_cmds(struct nvmet_tcp_queue *queue) 1436 { 1437 struct nvmet_tcp_cmd *cmd = queue->cmds; 1438 int i; 1439 1440 for (i = 0; i < queue->nr_cmds; i++, cmd++) { 1441 if (nvmet_tcp_need_data_in(cmd)) 1442 nvmet_req_uninit(&cmd->req); 1443 1444 nvmet_tcp_unmap_pdu_iovec(cmd); 1445 nvmet_tcp_free_cmd_buffers(cmd); 1446 } 1447 1448 if (!queue->nr_cmds && nvmet_tcp_need_data_in(&queue->connect)) { 1449 /* failed in connect */ 1450 nvmet_tcp_finish_cmd(&queue->connect); 1451 } 1452 } 1453 1454 static void nvmet_tcp_release_queue_work(struct work_struct *w) 1455 { 1456 struct page *page; 1457 struct nvmet_tcp_queue *queue = 1458 container_of(w, struct nvmet_tcp_queue, release_work); 1459 1460 mutex_lock(&nvmet_tcp_queue_mutex); 1461 list_del_init(&queue->queue_list); 1462 mutex_unlock(&nvmet_tcp_queue_mutex); 1463 1464 nvmet_tcp_restore_socket_callbacks(queue); 1465 cancel_work_sync(&queue->io_work); 1466 /* stop accepting incoming data */ 1467 queue->rcv_state = NVMET_TCP_RECV_ERR; 1468 1469 nvmet_tcp_uninit_data_in_cmds(queue); 1470 nvmet_sq_destroy(&queue->nvme_sq); 1471 cancel_work_sync(&queue->io_work); 1472 sock_release(queue->sock); 1473 nvmet_tcp_free_cmds(queue); 1474 if (queue->hdr_digest || queue->data_digest) 1475 nvmet_tcp_free_crypto(queue); 1476 ida_simple_remove(&nvmet_tcp_queue_ida, queue->idx); 1477 1478 page = virt_to_head_page(queue->pf_cache.va); 1479 __page_frag_cache_drain(page, queue->pf_cache.pagecnt_bias); 1480 kfree(queue); 1481 } 1482 1483 static void nvmet_tcp_data_ready(struct sock *sk) 1484 { 1485 struct nvmet_tcp_queue *queue; 1486 1487 read_lock_bh(&sk->sk_callback_lock); 1488 queue = sk->sk_user_data; 1489 if (likely(queue)) 1490 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work); 1491 read_unlock_bh(&sk->sk_callback_lock); 1492 } 1493 1494 static void nvmet_tcp_write_space(struct sock *sk) 1495 { 1496 struct nvmet_tcp_queue *queue; 1497 1498 read_lock_bh(&sk->sk_callback_lock); 1499 queue = sk->sk_user_data; 1500 if (unlikely(!queue)) 1501 goto out; 1502 1503 if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) { 1504 queue->write_space(sk); 1505 goto out; 1506 } 1507 1508 if (sk_stream_is_writeable(sk)) { 1509 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1510 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work); 1511 } 1512 out: 1513 read_unlock_bh(&sk->sk_callback_lock); 1514 } 1515 1516 static void nvmet_tcp_state_change(struct sock *sk) 1517 { 1518 struct nvmet_tcp_queue *queue; 1519 1520 read_lock_bh(&sk->sk_callback_lock); 1521 queue = sk->sk_user_data; 1522 if (!queue) 1523 goto done; 1524 1525 switch (sk->sk_state) { 1526 case TCP_FIN_WAIT1: 1527 case TCP_CLOSE_WAIT: 1528 case TCP_CLOSE: 1529 /* FALLTHRU */ 1530 nvmet_tcp_schedule_release_queue(queue); 1531 break; 1532 default: 1533 pr_warn("queue %d unhandled state %d\n", 1534 queue->idx, sk->sk_state); 1535 } 1536 done: 1537 read_unlock_bh(&sk->sk_callback_lock); 1538 } 1539 1540 static int nvmet_tcp_set_queue_sock(struct nvmet_tcp_queue *queue) 1541 { 1542 struct socket *sock = queue->sock; 1543 struct inet_sock *inet = inet_sk(sock->sk); 1544 int ret; 1545 1546 ret = kernel_getsockname(sock, 1547 (struct sockaddr *)&queue->sockaddr); 1548 if (ret < 0) 1549 return ret; 1550 1551 ret = kernel_getpeername(sock, 1552 (struct sockaddr *)&queue->sockaddr_peer); 1553 if (ret < 0) 1554 return ret; 1555 1556 /* 1557 * Cleanup whatever is sitting in the TCP transmit queue on socket 1558 * close. This is done to prevent stale data from being sent should 1559 * the network connection be restored before TCP times out. 1560 */ 1561 sock_no_linger(sock->sk); 1562 1563 if (so_priority > 0) 1564 sock_set_priority(sock->sk, so_priority); 1565 1566 /* Set socket type of service */ 1567 if (inet->rcv_tos > 0) 1568 ip_sock_set_tos(sock->sk, inet->rcv_tos); 1569 1570 ret = 0; 1571 write_lock_bh(&sock->sk->sk_callback_lock); 1572 if (sock->sk->sk_state != TCP_ESTABLISHED) { 1573 /* 1574 * If the socket is already closing, don't even start 1575 * consuming it 1576 */ 1577 ret = -ENOTCONN; 1578 } else { 1579 sock->sk->sk_user_data = queue; 1580 queue->data_ready = sock->sk->sk_data_ready; 1581 sock->sk->sk_data_ready = nvmet_tcp_data_ready; 1582 queue->state_change = sock->sk->sk_state_change; 1583 sock->sk->sk_state_change = nvmet_tcp_state_change; 1584 queue->write_space = sock->sk->sk_write_space; 1585 sock->sk->sk_write_space = nvmet_tcp_write_space; 1586 if (idle_poll_period_usecs) 1587 nvmet_tcp_arm_queue_deadline(queue); 1588 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work); 1589 } 1590 write_unlock_bh(&sock->sk->sk_callback_lock); 1591 1592 return ret; 1593 } 1594 1595 static int nvmet_tcp_alloc_queue(struct nvmet_tcp_port *port, 1596 struct socket *newsock) 1597 { 1598 struct nvmet_tcp_queue *queue; 1599 int ret; 1600 1601 queue = kzalloc(sizeof(*queue), GFP_KERNEL); 1602 if (!queue) 1603 return -ENOMEM; 1604 1605 INIT_WORK(&queue->release_work, nvmet_tcp_release_queue_work); 1606 INIT_WORK(&queue->io_work, nvmet_tcp_io_work); 1607 queue->sock = newsock; 1608 queue->port = port; 1609 queue->nr_cmds = 0; 1610 spin_lock_init(&queue->state_lock); 1611 queue->state = NVMET_TCP_Q_CONNECTING; 1612 INIT_LIST_HEAD(&queue->free_list); 1613 init_llist_head(&queue->resp_list); 1614 INIT_LIST_HEAD(&queue->resp_send_list); 1615 1616 queue->idx = ida_simple_get(&nvmet_tcp_queue_ida, 0, 0, GFP_KERNEL); 1617 if (queue->idx < 0) { 1618 ret = queue->idx; 1619 goto out_free_queue; 1620 } 1621 1622 ret = nvmet_tcp_alloc_cmd(queue, &queue->connect); 1623 if (ret) 1624 goto out_ida_remove; 1625 1626 ret = nvmet_sq_init(&queue->nvme_sq); 1627 if (ret) 1628 goto out_free_connect; 1629 1630 nvmet_prepare_receive_pdu(queue); 1631 1632 mutex_lock(&nvmet_tcp_queue_mutex); 1633 list_add_tail(&queue->queue_list, &nvmet_tcp_queue_list); 1634 mutex_unlock(&nvmet_tcp_queue_mutex); 1635 1636 ret = nvmet_tcp_set_queue_sock(queue); 1637 if (ret) 1638 goto out_destroy_sq; 1639 1640 return 0; 1641 out_destroy_sq: 1642 mutex_lock(&nvmet_tcp_queue_mutex); 1643 list_del_init(&queue->queue_list); 1644 mutex_unlock(&nvmet_tcp_queue_mutex); 1645 nvmet_sq_destroy(&queue->nvme_sq); 1646 out_free_connect: 1647 nvmet_tcp_free_cmd(&queue->connect); 1648 out_ida_remove: 1649 ida_simple_remove(&nvmet_tcp_queue_ida, queue->idx); 1650 out_free_queue: 1651 kfree(queue); 1652 return ret; 1653 } 1654 1655 static void nvmet_tcp_accept_work(struct work_struct *w) 1656 { 1657 struct nvmet_tcp_port *port = 1658 container_of(w, struct nvmet_tcp_port, accept_work); 1659 struct socket *newsock; 1660 int ret; 1661 1662 while (true) { 1663 ret = kernel_accept(port->sock, &newsock, O_NONBLOCK); 1664 if (ret < 0) { 1665 if (ret != -EAGAIN) 1666 pr_warn("failed to accept err=%d\n", ret); 1667 return; 1668 } 1669 ret = nvmet_tcp_alloc_queue(port, newsock); 1670 if (ret) { 1671 pr_err("failed to allocate queue\n"); 1672 sock_release(newsock); 1673 } 1674 } 1675 } 1676 1677 static void nvmet_tcp_listen_data_ready(struct sock *sk) 1678 { 1679 struct nvmet_tcp_port *port; 1680 1681 read_lock_bh(&sk->sk_callback_lock); 1682 port = sk->sk_user_data; 1683 if (!port) 1684 goto out; 1685 1686 if (sk->sk_state == TCP_LISTEN) 1687 schedule_work(&port->accept_work); 1688 out: 1689 read_unlock_bh(&sk->sk_callback_lock); 1690 } 1691 1692 static int nvmet_tcp_add_port(struct nvmet_port *nport) 1693 { 1694 struct nvmet_tcp_port *port; 1695 __kernel_sa_family_t af; 1696 int ret; 1697 1698 port = kzalloc(sizeof(*port), GFP_KERNEL); 1699 if (!port) 1700 return -ENOMEM; 1701 1702 switch (nport->disc_addr.adrfam) { 1703 case NVMF_ADDR_FAMILY_IP4: 1704 af = AF_INET; 1705 break; 1706 case NVMF_ADDR_FAMILY_IP6: 1707 af = AF_INET6; 1708 break; 1709 default: 1710 pr_err("address family %d not supported\n", 1711 nport->disc_addr.adrfam); 1712 ret = -EINVAL; 1713 goto err_port; 1714 } 1715 1716 ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr, 1717 nport->disc_addr.trsvcid, &port->addr); 1718 if (ret) { 1719 pr_err("malformed ip/port passed: %s:%s\n", 1720 nport->disc_addr.traddr, nport->disc_addr.trsvcid); 1721 goto err_port; 1722 } 1723 1724 port->nport = nport; 1725 INIT_WORK(&port->accept_work, nvmet_tcp_accept_work); 1726 if (port->nport->inline_data_size < 0) 1727 port->nport->inline_data_size = NVMET_TCP_DEF_INLINE_DATA_SIZE; 1728 1729 ret = sock_create(port->addr.ss_family, SOCK_STREAM, 1730 IPPROTO_TCP, &port->sock); 1731 if (ret) { 1732 pr_err("failed to create a socket\n"); 1733 goto err_port; 1734 } 1735 1736 port->sock->sk->sk_user_data = port; 1737 port->data_ready = port->sock->sk->sk_data_ready; 1738 port->sock->sk->sk_data_ready = nvmet_tcp_listen_data_ready; 1739 sock_set_reuseaddr(port->sock->sk); 1740 tcp_sock_set_nodelay(port->sock->sk); 1741 if (so_priority > 0) 1742 sock_set_priority(port->sock->sk, so_priority); 1743 1744 ret = kernel_bind(port->sock, (struct sockaddr *)&port->addr, 1745 sizeof(port->addr)); 1746 if (ret) { 1747 pr_err("failed to bind port socket %d\n", ret); 1748 goto err_sock; 1749 } 1750 1751 ret = kernel_listen(port->sock, 128); 1752 if (ret) { 1753 pr_err("failed to listen %d on port sock\n", ret); 1754 goto err_sock; 1755 } 1756 1757 nport->priv = port; 1758 pr_info("enabling port %d (%pISpc)\n", 1759 le16_to_cpu(nport->disc_addr.portid), &port->addr); 1760 1761 return 0; 1762 1763 err_sock: 1764 sock_release(port->sock); 1765 err_port: 1766 kfree(port); 1767 return ret; 1768 } 1769 1770 static void nvmet_tcp_destroy_port_queues(struct nvmet_tcp_port *port) 1771 { 1772 struct nvmet_tcp_queue *queue; 1773 1774 mutex_lock(&nvmet_tcp_queue_mutex); 1775 list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list) 1776 if (queue->port == port) 1777 kernel_sock_shutdown(queue->sock, SHUT_RDWR); 1778 mutex_unlock(&nvmet_tcp_queue_mutex); 1779 } 1780 1781 static void nvmet_tcp_remove_port(struct nvmet_port *nport) 1782 { 1783 struct nvmet_tcp_port *port = nport->priv; 1784 1785 write_lock_bh(&port->sock->sk->sk_callback_lock); 1786 port->sock->sk->sk_data_ready = port->data_ready; 1787 port->sock->sk->sk_user_data = NULL; 1788 write_unlock_bh(&port->sock->sk->sk_callback_lock); 1789 cancel_work_sync(&port->accept_work); 1790 /* 1791 * Destroy the remaining queues, which are not belong to any 1792 * controller yet. 1793 */ 1794 nvmet_tcp_destroy_port_queues(port); 1795 1796 sock_release(port->sock); 1797 kfree(port); 1798 } 1799 1800 static void nvmet_tcp_delete_ctrl(struct nvmet_ctrl *ctrl) 1801 { 1802 struct nvmet_tcp_queue *queue; 1803 1804 mutex_lock(&nvmet_tcp_queue_mutex); 1805 list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list) 1806 if (queue->nvme_sq.ctrl == ctrl) 1807 kernel_sock_shutdown(queue->sock, SHUT_RDWR); 1808 mutex_unlock(&nvmet_tcp_queue_mutex); 1809 } 1810 1811 static u16 nvmet_tcp_install_queue(struct nvmet_sq *sq) 1812 { 1813 struct nvmet_tcp_queue *queue = 1814 container_of(sq, struct nvmet_tcp_queue, nvme_sq); 1815 1816 if (sq->qid == 0) { 1817 /* Let inflight controller teardown complete */ 1818 flush_scheduled_work(); 1819 } 1820 1821 queue->nr_cmds = sq->size * 2; 1822 if (nvmet_tcp_alloc_cmds(queue)) 1823 return NVME_SC_INTERNAL; 1824 return 0; 1825 } 1826 1827 static void nvmet_tcp_disc_port_addr(struct nvmet_req *req, 1828 struct nvmet_port *nport, char *traddr) 1829 { 1830 struct nvmet_tcp_port *port = nport->priv; 1831 1832 if (inet_addr_is_any((struct sockaddr *)&port->addr)) { 1833 struct nvmet_tcp_cmd *cmd = 1834 container_of(req, struct nvmet_tcp_cmd, req); 1835 struct nvmet_tcp_queue *queue = cmd->queue; 1836 1837 sprintf(traddr, "%pISc", (struct sockaddr *)&queue->sockaddr); 1838 } else { 1839 memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE); 1840 } 1841 } 1842 1843 static const struct nvmet_fabrics_ops nvmet_tcp_ops = { 1844 .owner = THIS_MODULE, 1845 .type = NVMF_TRTYPE_TCP, 1846 .msdbd = 1, 1847 .add_port = nvmet_tcp_add_port, 1848 .remove_port = nvmet_tcp_remove_port, 1849 .queue_response = nvmet_tcp_queue_response, 1850 .delete_ctrl = nvmet_tcp_delete_ctrl, 1851 .install_queue = nvmet_tcp_install_queue, 1852 .disc_traddr = nvmet_tcp_disc_port_addr, 1853 }; 1854 1855 static int __init nvmet_tcp_init(void) 1856 { 1857 int ret; 1858 1859 nvmet_tcp_wq = alloc_workqueue("nvmet_tcp_wq", WQ_HIGHPRI, 0); 1860 if (!nvmet_tcp_wq) 1861 return -ENOMEM; 1862 1863 ret = nvmet_register_transport(&nvmet_tcp_ops); 1864 if (ret) 1865 goto err; 1866 1867 return 0; 1868 err: 1869 destroy_workqueue(nvmet_tcp_wq); 1870 return ret; 1871 } 1872 1873 static void __exit nvmet_tcp_exit(void) 1874 { 1875 struct nvmet_tcp_queue *queue; 1876 1877 nvmet_unregister_transport(&nvmet_tcp_ops); 1878 1879 flush_scheduled_work(); 1880 mutex_lock(&nvmet_tcp_queue_mutex); 1881 list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list) 1882 kernel_sock_shutdown(queue->sock, SHUT_RDWR); 1883 mutex_unlock(&nvmet_tcp_queue_mutex); 1884 flush_scheduled_work(); 1885 1886 destroy_workqueue(nvmet_tcp_wq); 1887 } 1888 1889 module_init(nvmet_tcp_init); 1890 module_exit(nvmet_tcp_exit); 1891 1892 MODULE_LICENSE("GPL v2"); 1893 MODULE_ALIAS("nvmet-transport-3"); /* 3 == NVMF_TRTYPE_TCP */ 1894