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_calc_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_setup_c2h_data_pdu(struct nvmet_tcp_cmd *cmd) 417 { 418 struct nvme_tcp_data_pdu *pdu = cmd->data_pdu; 419 struct nvmet_tcp_queue *queue = cmd->queue; 420 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue); 421 u8 ddgst = nvmet_tcp_ddgst_len(cmd->queue); 422 423 cmd->offset = 0; 424 cmd->state = NVMET_TCP_SEND_DATA_PDU; 425 426 pdu->hdr.type = nvme_tcp_c2h_data; 427 pdu->hdr.flags = NVME_TCP_F_DATA_LAST | (queue->nvme_sq.sqhd_disabled ? 428 NVME_TCP_F_DATA_SUCCESS : 0); 429 pdu->hdr.hlen = sizeof(*pdu); 430 pdu->hdr.pdo = pdu->hdr.hlen + hdgst; 431 pdu->hdr.plen = 432 cpu_to_le32(pdu->hdr.hlen + hdgst + 433 cmd->req.transfer_len + ddgst); 434 pdu->command_id = cmd->req.cqe->command_id; 435 pdu->data_length = cpu_to_le32(cmd->req.transfer_len); 436 pdu->data_offset = cpu_to_le32(cmd->wbytes_done); 437 438 if (queue->data_digest) { 439 pdu->hdr.flags |= NVME_TCP_F_DDGST; 440 nvmet_tcp_calc_ddgst(queue->snd_hash, cmd); 441 } 442 443 if (cmd->queue->hdr_digest) { 444 pdu->hdr.flags |= NVME_TCP_F_HDGST; 445 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu)); 446 } 447 } 448 449 static void nvmet_setup_r2t_pdu(struct nvmet_tcp_cmd *cmd) 450 { 451 struct nvme_tcp_r2t_pdu *pdu = cmd->r2t_pdu; 452 struct nvmet_tcp_queue *queue = cmd->queue; 453 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue); 454 455 cmd->offset = 0; 456 cmd->state = NVMET_TCP_SEND_R2T; 457 458 pdu->hdr.type = nvme_tcp_r2t; 459 pdu->hdr.flags = 0; 460 pdu->hdr.hlen = sizeof(*pdu); 461 pdu->hdr.pdo = 0; 462 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst); 463 464 pdu->command_id = cmd->req.cmd->common.command_id; 465 pdu->ttag = nvmet_tcp_cmd_tag(cmd->queue, cmd); 466 pdu->r2t_length = cpu_to_le32(cmd->req.transfer_len - cmd->rbytes_done); 467 pdu->r2t_offset = cpu_to_le32(cmd->rbytes_done); 468 if (cmd->queue->hdr_digest) { 469 pdu->hdr.flags |= NVME_TCP_F_HDGST; 470 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu)); 471 } 472 } 473 474 static void nvmet_setup_response_pdu(struct nvmet_tcp_cmd *cmd) 475 { 476 struct nvme_tcp_rsp_pdu *pdu = cmd->rsp_pdu; 477 struct nvmet_tcp_queue *queue = cmd->queue; 478 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue); 479 480 cmd->offset = 0; 481 cmd->state = NVMET_TCP_SEND_RESPONSE; 482 483 pdu->hdr.type = nvme_tcp_rsp; 484 pdu->hdr.flags = 0; 485 pdu->hdr.hlen = sizeof(*pdu); 486 pdu->hdr.pdo = 0; 487 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst); 488 if (cmd->queue->hdr_digest) { 489 pdu->hdr.flags |= NVME_TCP_F_HDGST; 490 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu)); 491 } 492 } 493 494 static void nvmet_tcp_process_resp_list(struct nvmet_tcp_queue *queue) 495 { 496 struct llist_node *node; 497 struct nvmet_tcp_cmd *cmd; 498 499 for (node = llist_del_all(&queue->resp_list); node; node = node->next) { 500 cmd = llist_entry(node, struct nvmet_tcp_cmd, lentry); 501 list_add(&cmd->entry, &queue->resp_send_list); 502 queue->send_list_len++; 503 } 504 } 505 506 static struct nvmet_tcp_cmd *nvmet_tcp_fetch_cmd(struct nvmet_tcp_queue *queue) 507 { 508 queue->snd_cmd = list_first_entry_or_null(&queue->resp_send_list, 509 struct nvmet_tcp_cmd, entry); 510 if (!queue->snd_cmd) { 511 nvmet_tcp_process_resp_list(queue); 512 queue->snd_cmd = 513 list_first_entry_or_null(&queue->resp_send_list, 514 struct nvmet_tcp_cmd, entry); 515 if (unlikely(!queue->snd_cmd)) 516 return NULL; 517 } 518 519 list_del_init(&queue->snd_cmd->entry); 520 queue->send_list_len--; 521 522 if (nvmet_tcp_need_data_out(queue->snd_cmd)) 523 nvmet_setup_c2h_data_pdu(queue->snd_cmd); 524 else if (nvmet_tcp_need_data_in(queue->snd_cmd)) 525 nvmet_setup_r2t_pdu(queue->snd_cmd); 526 else 527 nvmet_setup_response_pdu(queue->snd_cmd); 528 529 return queue->snd_cmd; 530 } 531 532 static void nvmet_tcp_queue_response(struct nvmet_req *req) 533 { 534 struct nvmet_tcp_cmd *cmd = 535 container_of(req, struct nvmet_tcp_cmd, req); 536 struct nvmet_tcp_queue *queue = cmd->queue; 537 struct nvme_sgl_desc *sgl; 538 u32 len; 539 540 if (unlikely(cmd == queue->cmd)) { 541 sgl = &cmd->req.cmd->common.dptr.sgl; 542 len = le32_to_cpu(sgl->length); 543 544 /* 545 * Wait for inline data before processing the response. 546 * Avoid using helpers, this might happen before 547 * nvmet_req_init is completed. 548 */ 549 if (queue->rcv_state == NVMET_TCP_RECV_PDU && 550 len && len <= cmd->req.port->inline_data_size && 551 nvme_is_write(cmd->req.cmd)) 552 return; 553 } 554 555 llist_add(&cmd->lentry, &queue->resp_list); 556 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &cmd->queue->io_work); 557 } 558 559 static void nvmet_tcp_execute_request(struct nvmet_tcp_cmd *cmd) 560 { 561 if (unlikely(cmd->flags & NVMET_TCP_F_INIT_FAILED)) 562 nvmet_tcp_queue_response(&cmd->req); 563 else 564 cmd->req.execute(&cmd->req); 565 } 566 567 static int nvmet_try_send_data_pdu(struct nvmet_tcp_cmd *cmd) 568 { 569 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue); 570 int left = sizeof(*cmd->data_pdu) - cmd->offset + hdgst; 571 int ret; 572 573 ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->data_pdu), 574 offset_in_page(cmd->data_pdu) + cmd->offset, 575 left, MSG_DONTWAIT | MSG_MORE | MSG_SENDPAGE_NOTLAST); 576 if (ret <= 0) 577 return ret; 578 579 cmd->offset += ret; 580 left -= ret; 581 582 if (left) 583 return -EAGAIN; 584 585 cmd->state = NVMET_TCP_SEND_DATA; 586 cmd->offset = 0; 587 return 1; 588 } 589 590 static int nvmet_try_send_data(struct nvmet_tcp_cmd *cmd, bool last_in_batch) 591 { 592 struct nvmet_tcp_queue *queue = cmd->queue; 593 int ret; 594 595 while (cmd->cur_sg) { 596 struct page *page = sg_page(cmd->cur_sg); 597 u32 left = cmd->cur_sg->length - cmd->offset; 598 int flags = MSG_DONTWAIT; 599 600 if ((!last_in_batch && cmd->queue->send_list_len) || 601 cmd->wbytes_done + left < cmd->req.transfer_len || 602 queue->data_digest || !queue->nvme_sq.sqhd_disabled) 603 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST; 604 605 ret = kernel_sendpage(cmd->queue->sock, page, cmd->offset, 606 left, flags); 607 if (ret <= 0) 608 return ret; 609 610 cmd->offset += ret; 611 cmd->wbytes_done += ret; 612 613 /* Done with sg?*/ 614 if (cmd->offset == cmd->cur_sg->length) { 615 cmd->cur_sg = sg_next(cmd->cur_sg); 616 cmd->offset = 0; 617 } 618 } 619 620 if (queue->data_digest) { 621 cmd->state = NVMET_TCP_SEND_DDGST; 622 cmd->offset = 0; 623 } else { 624 if (queue->nvme_sq.sqhd_disabled) { 625 cmd->queue->snd_cmd = NULL; 626 nvmet_tcp_put_cmd(cmd); 627 } else { 628 nvmet_setup_response_pdu(cmd); 629 } 630 } 631 632 if (queue->nvme_sq.sqhd_disabled) 633 nvmet_tcp_free_cmd_buffers(cmd); 634 635 return 1; 636 637 } 638 639 static int nvmet_try_send_response(struct nvmet_tcp_cmd *cmd, 640 bool last_in_batch) 641 { 642 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue); 643 int left = sizeof(*cmd->rsp_pdu) - cmd->offset + hdgst; 644 int flags = MSG_DONTWAIT; 645 int ret; 646 647 if (!last_in_batch && cmd->queue->send_list_len) 648 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST; 649 else 650 flags |= MSG_EOR; 651 652 ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->rsp_pdu), 653 offset_in_page(cmd->rsp_pdu) + cmd->offset, left, flags); 654 if (ret <= 0) 655 return ret; 656 cmd->offset += ret; 657 left -= ret; 658 659 if (left) 660 return -EAGAIN; 661 662 nvmet_tcp_free_cmd_buffers(cmd); 663 cmd->queue->snd_cmd = NULL; 664 nvmet_tcp_put_cmd(cmd); 665 return 1; 666 } 667 668 static int nvmet_try_send_r2t(struct nvmet_tcp_cmd *cmd, bool last_in_batch) 669 { 670 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue); 671 int left = sizeof(*cmd->r2t_pdu) - cmd->offset + hdgst; 672 int flags = MSG_DONTWAIT; 673 int ret; 674 675 if (!last_in_batch && cmd->queue->send_list_len) 676 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST; 677 else 678 flags |= MSG_EOR; 679 680 ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->r2t_pdu), 681 offset_in_page(cmd->r2t_pdu) + cmd->offset, left, flags); 682 if (ret <= 0) 683 return ret; 684 cmd->offset += ret; 685 left -= ret; 686 687 if (left) 688 return -EAGAIN; 689 690 cmd->queue->snd_cmd = NULL; 691 return 1; 692 } 693 694 static int nvmet_try_send_ddgst(struct nvmet_tcp_cmd *cmd, bool last_in_batch) 695 { 696 struct nvmet_tcp_queue *queue = cmd->queue; 697 int left = NVME_TCP_DIGEST_LENGTH - cmd->offset; 698 struct msghdr msg = { .msg_flags = MSG_DONTWAIT }; 699 struct kvec iov = { 700 .iov_base = (u8 *)&cmd->exp_ddgst + cmd->offset, 701 .iov_len = left 702 }; 703 int ret; 704 705 if (!last_in_batch && cmd->queue->send_list_len) 706 msg.msg_flags |= MSG_MORE; 707 else 708 msg.msg_flags |= MSG_EOR; 709 710 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len); 711 if (unlikely(ret <= 0)) 712 return ret; 713 714 cmd->offset += ret; 715 left -= ret; 716 717 if (left) 718 return -EAGAIN; 719 720 if (queue->nvme_sq.sqhd_disabled) { 721 cmd->queue->snd_cmd = NULL; 722 nvmet_tcp_put_cmd(cmd); 723 } else { 724 nvmet_setup_response_pdu(cmd); 725 } 726 return 1; 727 } 728 729 static int nvmet_tcp_try_send_one(struct nvmet_tcp_queue *queue, 730 bool last_in_batch) 731 { 732 struct nvmet_tcp_cmd *cmd = queue->snd_cmd; 733 int ret = 0; 734 735 if (!cmd || queue->state == NVMET_TCP_Q_DISCONNECTING) { 736 cmd = nvmet_tcp_fetch_cmd(queue); 737 if (unlikely(!cmd)) 738 return 0; 739 } 740 741 if (cmd->state == NVMET_TCP_SEND_DATA_PDU) { 742 ret = nvmet_try_send_data_pdu(cmd); 743 if (ret <= 0) 744 goto done_send; 745 } 746 747 if (cmd->state == NVMET_TCP_SEND_DATA) { 748 ret = nvmet_try_send_data(cmd, last_in_batch); 749 if (ret <= 0) 750 goto done_send; 751 } 752 753 if (cmd->state == NVMET_TCP_SEND_DDGST) { 754 ret = nvmet_try_send_ddgst(cmd, last_in_batch); 755 if (ret <= 0) 756 goto done_send; 757 } 758 759 if (cmd->state == NVMET_TCP_SEND_R2T) { 760 ret = nvmet_try_send_r2t(cmd, last_in_batch); 761 if (ret <= 0) 762 goto done_send; 763 } 764 765 if (cmd->state == NVMET_TCP_SEND_RESPONSE) 766 ret = nvmet_try_send_response(cmd, last_in_batch); 767 768 done_send: 769 if (ret < 0) { 770 if (ret == -EAGAIN) 771 return 0; 772 return ret; 773 } 774 775 return 1; 776 } 777 778 static int nvmet_tcp_try_send(struct nvmet_tcp_queue *queue, 779 int budget, int *sends) 780 { 781 int i, ret = 0; 782 783 for (i = 0; i < budget; i++) { 784 ret = nvmet_tcp_try_send_one(queue, i == budget - 1); 785 if (unlikely(ret < 0)) { 786 nvmet_tcp_socket_error(queue, ret); 787 goto done; 788 } else if (ret == 0) { 789 break; 790 } 791 (*sends)++; 792 } 793 done: 794 return ret; 795 } 796 797 static void nvmet_prepare_receive_pdu(struct nvmet_tcp_queue *queue) 798 { 799 queue->offset = 0; 800 queue->left = sizeof(struct nvme_tcp_hdr); 801 queue->cmd = NULL; 802 queue->rcv_state = NVMET_TCP_RECV_PDU; 803 } 804 805 static void nvmet_tcp_free_crypto(struct nvmet_tcp_queue *queue) 806 { 807 struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash); 808 809 ahash_request_free(queue->rcv_hash); 810 ahash_request_free(queue->snd_hash); 811 crypto_free_ahash(tfm); 812 } 813 814 static int nvmet_tcp_alloc_crypto(struct nvmet_tcp_queue *queue) 815 { 816 struct crypto_ahash *tfm; 817 818 tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC); 819 if (IS_ERR(tfm)) 820 return PTR_ERR(tfm); 821 822 queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL); 823 if (!queue->snd_hash) 824 goto free_tfm; 825 ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL); 826 827 queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL); 828 if (!queue->rcv_hash) 829 goto free_snd_hash; 830 ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL); 831 832 return 0; 833 free_snd_hash: 834 ahash_request_free(queue->snd_hash); 835 free_tfm: 836 crypto_free_ahash(tfm); 837 return -ENOMEM; 838 } 839 840 841 static int nvmet_tcp_handle_icreq(struct nvmet_tcp_queue *queue) 842 { 843 struct nvme_tcp_icreq_pdu *icreq = &queue->pdu.icreq; 844 struct nvme_tcp_icresp_pdu *icresp = &queue->pdu.icresp; 845 struct msghdr msg = {}; 846 struct kvec iov; 847 int ret; 848 849 if (le32_to_cpu(icreq->hdr.plen) != sizeof(struct nvme_tcp_icreq_pdu)) { 850 pr_err("bad nvme-tcp pdu length (%d)\n", 851 le32_to_cpu(icreq->hdr.plen)); 852 nvmet_tcp_fatal_error(queue); 853 } 854 855 if (icreq->pfv != NVME_TCP_PFV_1_0) { 856 pr_err("queue %d: bad pfv %d\n", queue->idx, icreq->pfv); 857 return -EPROTO; 858 } 859 860 if (icreq->hpda != 0) { 861 pr_err("queue %d: unsupported hpda %d\n", queue->idx, 862 icreq->hpda); 863 return -EPROTO; 864 } 865 866 queue->hdr_digest = !!(icreq->digest & NVME_TCP_HDR_DIGEST_ENABLE); 867 queue->data_digest = !!(icreq->digest & NVME_TCP_DATA_DIGEST_ENABLE); 868 if (queue->hdr_digest || queue->data_digest) { 869 ret = nvmet_tcp_alloc_crypto(queue); 870 if (ret) 871 return ret; 872 } 873 874 memset(icresp, 0, sizeof(*icresp)); 875 icresp->hdr.type = nvme_tcp_icresp; 876 icresp->hdr.hlen = sizeof(*icresp); 877 icresp->hdr.pdo = 0; 878 icresp->hdr.plen = cpu_to_le32(icresp->hdr.hlen); 879 icresp->pfv = cpu_to_le16(NVME_TCP_PFV_1_0); 880 icresp->maxdata = cpu_to_le32(0x400000); /* 16M arbitrary limit */ 881 icresp->cpda = 0; 882 if (queue->hdr_digest) 883 icresp->digest |= NVME_TCP_HDR_DIGEST_ENABLE; 884 if (queue->data_digest) 885 icresp->digest |= NVME_TCP_DATA_DIGEST_ENABLE; 886 887 iov.iov_base = icresp; 888 iov.iov_len = sizeof(*icresp); 889 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len); 890 if (ret < 0) 891 goto free_crypto; 892 893 queue->state = NVMET_TCP_Q_LIVE; 894 nvmet_prepare_receive_pdu(queue); 895 return 0; 896 free_crypto: 897 if (queue->hdr_digest || queue->data_digest) 898 nvmet_tcp_free_crypto(queue); 899 return ret; 900 } 901 902 static void nvmet_tcp_handle_req_failure(struct nvmet_tcp_queue *queue, 903 struct nvmet_tcp_cmd *cmd, struct nvmet_req *req) 904 { 905 size_t data_len = le32_to_cpu(req->cmd->common.dptr.sgl.length); 906 int ret; 907 908 /* 909 * This command has not been processed yet, hence we are trying to 910 * figure out if there is still pending data left to receive. If 911 * we don't, we can simply prepare for the next pdu and bail out, 912 * otherwise we will need to prepare a buffer and receive the 913 * stale data before continuing forward. 914 */ 915 if (!nvme_is_write(cmd->req.cmd) || !data_len || 916 data_len > cmd->req.port->inline_data_size) { 917 nvmet_prepare_receive_pdu(queue); 918 return; 919 } 920 921 ret = nvmet_tcp_map_data(cmd); 922 if (unlikely(ret)) { 923 pr_err("queue %d: failed to map data\n", queue->idx); 924 nvmet_tcp_fatal_error(queue); 925 return; 926 } 927 928 queue->rcv_state = NVMET_TCP_RECV_DATA; 929 nvmet_tcp_map_pdu_iovec(cmd); 930 cmd->flags |= NVMET_TCP_F_INIT_FAILED; 931 } 932 933 static int nvmet_tcp_handle_h2c_data_pdu(struct nvmet_tcp_queue *queue) 934 { 935 struct nvme_tcp_data_pdu *data = &queue->pdu.data; 936 struct nvmet_tcp_cmd *cmd; 937 938 if (likely(queue->nr_cmds)) 939 cmd = &queue->cmds[data->ttag]; 940 else 941 cmd = &queue->connect; 942 943 if (le32_to_cpu(data->data_offset) != cmd->rbytes_done) { 944 pr_err("ttag %u unexpected data offset %u (expected %u)\n", 945 data->ttag, le32_to_cpu(data->data_offset), 946 cmd->rbytes_done); 947 /* FIXME: use path and transport errors */ 948 nvmet_req_complete(&cmd->req, 949 NVME_SC_INVALID_FIELD | NVME_SC_DNR); 950 return -EPROTO; 951 } 952 953 cmd->pdu_len = le32_to_cpu(data->data_length); 954 cmd->pdu_recv = 0; 955 nvmet_tcp_map_pdu_iovec(cmd); 956 queue->cmd = cmd; 957 queue->rcv_state = NVMET_TCP_RECV_DATA; 958 959 return 0; 960 } 961 962 static int nvmet_tcp_done_recv_pdu(struct nvmet_tcp_queue *queue) 963 { 964 struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr; 965 struct nvme_command *nvme_cmd = &queue->pdu.cmd.cmd; 966 struct nvmet_req *req; 967 int ret; 968 969 if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) { 970 if (hdr->type != nvme_tcp_icreq) { 971 pr_err("unexpected pdu type (%d) before icreq\n", 972 hdr->type); 973 nvmet_tcp_fatal_error(queue); 974 return -EPROTO; 975 } 976 return nvmet_tcp_handle_icreq(queue); 977 } 978 979 if (hdr->type == nvme_tcp_h2c_data) { 980 ret = nvmet_tcp_handle_h2c_data_pdu(queue); 981 if (unlikely(ret)) 982 return ret; 983 return 0; 984 } 985 986 queue->cmd = nvmet_tcp_get_cmd(queue); 987 if (unlikely(!queue->cmd)) { 988 /* This should never happen */ 989 pr_err("queue %d: out of commands (%d) send_list_len: %d, opcode: %d", 990 queue->idx, queue->nr_cmds, queue->send_list_len, 991 nvme_cmd->common.opcode); 992 nvmet_tcp_fatal_error(queue); 993 return -ENOMEM; 994 } 995 996 req = &queue->cmd->req; 997 memcpy(req->cmd, nvme_cmd, sizeof(*nvme_cmd)); 998 999 if (unlikely(!nvmet_req_init(req, &queue->nvme_cq, 1000 &queue->nvme_sq, &nvmet_tcp_ops))) { 1001 pr_err("failed cmd %p id %d opcode %d, data_len: %d\n", 1002 req->cmd, req->cmd->common.command_id, 1003 req->cmd->common.opcode, 1004 le32_to_cpu(req->cmd->common.dptr.sgl.length)); 1005 1006 nvmet_tcp_handle_req_failure(queue, queue->cmd, req); 1007 return 0; 1008 } 1009 1010 ret = nvmet_tcp_map_data(queue->cmd); 1011 if (unlikely(ret)) { 1012 pr_err("queue %d: failed to map data\n", queue->idx); 1013 if (nvmet_tcp_has_inline_data(queue->cmd)) 1014 nvmet_tcp_fatal_error(queue); 1015 else 1016 nvmet_req_complete(req, ret); 1017 ret = -EAGAIN; 1018 goto out; 1019 } 1020 1021 if (nvmet_tcp_need_data_in(queue->cmd)) { 1022 if (nvmet_tcp_has_inline_data(queue->cmd)) { 1023 queue->rcv_state = NVMET_TCP_RECV_DATA; 1024 nvmet_tcp_map_pdu_iovec(queue->cmd); 1025 return 0; 1026 } 1027 /* send back R2T */ 1028 nvmet_tcp_queue_response(&queue->cmd->req); 1029 goto out; 1030 } 1031 1032 queue->cmd->req.execute(&queue->cmd->req); 1033 out: 1034 nvmet_prepare_receive_pdu(queue); 1035 return ret; 1036 } 1037 1038 static const u8 nvme_tcp_pdu_sizes[] = { 1039 [nvme_tcp_icreq] = sizeof(struct nvme_tcp_icreq_pdu), 1040 [nvme_tcp_cmd] = sizeof(struct nvme_tcp_cmd_pdu), 1041 [nvme_tcp_h2c_data] = sizeof(struct nvme_tcp_data_pdu), 1042 }; 1043 1044 static inline u8 nvmet_tcp_pdu_size(u8 type) 1045 { 1046 size_t idx = type; 1047 1048 return (idx < ARRAY_SIZE(nvme_tcp_pdu_sizes) && 1049 nvme_tcp_pdu_sizes[idx]) ? 1050 nvme_tcp_pdu_sizes[idx] : 0; 1051 } 1052 1053 static inline bool nvmet_tcp_pdu_valid(u8 type) 1054 { 1055 switch (type) { 1056 case nvme_tcp_icreq: 1057 case nvme_tcp_cmd: 1058 case nvme_tcp_h2c_data: 1059 /* fallthru */ 1060 return true; 1061 } 1062 1063 return false; 1064 } 1065 1066 static int nvmet_tcp_try_recv_pdu(struct nvmet_tcp_queue *queue) 1067 { 1068 struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr; 1069 int len; 1070 struct kvec iov; 1071 struct msghdr msg = { .msg_flags = MSG_DONTWAIT }; 1072 1073 recv: 1074 iov.iov_base = (void *)&queue->pdu + queue->offset; 1075 iov.iov_len = queue->left; 1076 len = kernel_recvmsg(queue->sock, &msg, &iov, 1, 1077 iov.iov_len, msg.msg_flags); 1078 if (unlikely(len < 0)) 1079 return len; 1080 1081 queue->offset += len; 1082 queue->left -= len; 1083 if (queue->left) 1084 return -EAGAIN; 1085 1086 if (queue->offset == sizeof(struct nvme_tcp_hdr)) { 1087 u8 hdgst = nvmet_tcp_hdgst_len(queue); 1088 1089 if (unlikely(!nvmet_tcp_pdu_valid(hdr->type))) { 1090 pr_err("unexpected pdu type %d\n", hdr->type); 1091 nvmet_tcp_fatal_error(queue); 1092 return -EIO; 1093 } 1094 1095 if (unlikely(hdr->hlen != nvmet_tcp_pdu_size(hdr->type))) { 1096 pr_err("pdu %d bad hlen %d\n", hdr->type, hdr->hlen); 1097 return -EIO; 1098 } 1099 1100 queue->left = hdr->hlen - queue->offset + hdgst; 1101 goto recv; 1102 } 1103 1104 if (queue->hdr_digest && 1105 nvmet_tcp_verify_hdgst(queue, &queue->pdu, hdr->hlen)) { 1106 nvmet_tcp_fatal_error(queue); /* fatal */ 1107 return -EPROTO; 1108 } 1109 1110 if (queue->data_digest && 1111 nvmet_tcp_check_ddgst(queue, &queue->pdu)) { 1112 nvmet_tcp_fatal_error(queue); /* fatal */ 1113 return -EPROTO; 1114 } 1115 1116 return nvmet_tcp_done_recv_pdu(queue); 1117 } 1118 1119 static void nvmet_tcp_prep_recv_ddgst(struct nvmet_tcp_cmd *cmd) 1120 { 1121 struct nvmet_tcp_queue *queue = cmd->queue; 1122 1123 nvmet_tcp_calc_ddgst(queue->rcv_hash, cmd); 1124 queue->offset = 0; 1125 queue->left = NVME_TCP_DIGEST_LENGTH; 1126 queue->rcv_state = NVMET_TCP_RECV_DDGST; 1127 } 1128 1129 static int nvmet_tcp_try_recv_data(struct nvmet_tcp_queue *queue) 1130 { 1131 struct nvmet_tcp_cmd *cmd = queue->cmd; 1132 int ret; 1133 1134 while (msg_data_left(&cmd->recv_msg)) { 1135 ret = sock_recvmsg(cmd->queue->sock, &cmd->recv_msg, 1136 cmd->recv_msg.msg_flags); 1137 if (ret <= 0) 1138 return ret; 1139 1140 cmd->pdu_recv += ret; 1141 cmd->rbytes_done += ret; 1142 } 1143 1144 nvmet_tcp_unmap_pdu_iovec(cmd); 1145 if (queue->data_digest) { 1146 nvmet_tcp_prep_recv_ddgst(cmd); 1147 return 0; 1148 } 1149 1150 if (cmd->rbytes_done == cmd->req.transfer_len) 1151 nvmet_tcp_execute_request(cmd); 1152 1153 nvmet_prepare_receive_pdu(queue); 1154 return 0; 1155 } 1156 1157 static int nvmet_tcp_try_recv_ddgst(struct nvmet_tcp_queue *queue) 1158 { 1159 struct nvmet_tcp_cmd *cmd = queue->cmd; 1160 int ret; 1161 struct msghdr msg = { .msg_flags = MSG_DONTWAIT }; 1162 struct kvec iov = { 1163 .iov_base = (void *)&cmd->recv_ddgst + queue->offset, 1164 .iov_len = queue->left 1165 }; 1166 1167 ret = kernel_recvmsg(queue->sock, &msg, &iov, 1, 1168 iov.iov_len, msg.msg_flags); 1169 if (unlikely(ret < 0)) 1170 return ret; 1171 1172 queue->offset += ret; 1173 queue->left -= ret; 1174 if (queue->left) 1175 return -EAGAIN; 1176 1177 if (queue->data_digest && cmd->exp_ddgst != cmd->recv_ddgst) { 1178 pr_err("queue %d: cmd %d pdu (%d) data digest error: recv %#x expected %#x\n", 1179 queue->idx, cmd->req.cmd->common.command_id, 1180 queue->pdu.cmd.hdr.type, le32_to_cpu(cmd->recv_ddgst), 1181 le32_to_cpu(cmd->exp_ddgst)); 1182 nvmet_tcp_finish_cmd(cmd); 1183 nvmet_tcp_fatal_error(queue); 1184 ret = -EPROTO; 1185 goto out; 1186 } 1187 1188 if (cmd->rbytes_done == cmd->req.transfer_len) 1189 nvmet_tcp_execute_request(cmd); 1190 1191 ret = 0; 1192 out: 1193 nvmet_prepare_receive_pdu(queue); 1194 return ret; 1195 } 1196 1197 static int nvmet_tcp_try_recv_one(struct nvmet_tcp_queue *queue) 1198 { 1199 int result = 0; 1200 1201 if (unlikely(queue->rcv_state == NVMET_TCP_RECV_ERR)) 1202 return 0; 1203 1204 if (queue->rcv_state == NVMET_TCP_RECV_PDU) { 1205 result = nvmet_tcp_try_recv_pdu(queue); 1206 if (result != 0) 1207 goto done_recv; 1208 } 1209 1210 if (queue->rcv_state == NVMET_TCP_RECV_DATA) { 1211 result = nvmet_tcp_try_recv_data(queue); 1212 if (result != 0) 1213 goto done_recv; 1214 } 1215 1216 if (queue->rcv_state == NVMET_TCP_RECV_DDGST) { 1217 result = nvmet_tcp_try_recv_ddgst(queue); 1218 if (result != 0) 1219 goto done_recv; 1220 } 1221 1222 done_recv: 1223 if (result < 0) { 1224 if (result == -EAGAIN) 1225 return 0; 1226 return result; 1227 } 1228 return 1; 1229 } 1230 1231 static int nvmet_tcp_try_recv(struct nvmet_tcp_queue *queue, 1232 int budget, int *recvs) 1233 { 1234 int i, ret = 0; 1235 1236 for (i = 0; i < budget; i++) { 1237 ret = nvmet_tcp_try_recv_one(queue); 1238 if (unlikely(ret < 0)) { 1239 nvmet_tcp_socket_error(queue, ret); 1240 goto done; 1241 } else if (ret == 0) { 1242 break; 1243 } 1244 (*recvs)++; 1245 } 1246 done: 1247 return ret; 1248 } 1249 1250 static void nvmet_tcp_schedule_release_queue(struct nvmet_tcp_queue *queue) 1251 { 1252 spin_lock(&queue->state_lock); 1253 if (queue->state != NVMET_TCP_Q_DISCONNECTING) { 1254 queue->state = NVMET_TCP_Q_DISCONNECTING; 1255 queue_work(nvmet_wq, &queue->release_work); 1256 } 1257 spin_unlock(&queue->state_lock); 1258 } 1259 1260 static inline void nvmet_tcp_arm_queue_deadline(struct nvmet_tcp_queue *queue) 1261 { 1262 queue->poll_end = jiffies + usecs_to_jiffies(idle_poll_period_usecs); 1263 } 1264 1265 static bool nvmet_tcp_check_queue_deadline(struct nvmet_tcp_queue *queue, 1266 int ops) 1267 { 1268 if (!idle_poll_period_usecs) 1269 return false; 1270 1271 if (ops) 1272 nvmet_tcp_arm_queue_deadline(queue); 1273 1274 return !time_after(jiffies, queue->poll_end); 1275 } 1276 1277 static void nvmet_tcp_io_work(struct work_struct *w) 1278 { 1279 struct nvmet_tcp_queue *queue = 1280 container_of(w, struct nvmet_tcp_queue, io_work); 1281 bool pending; 1282 int ret, ops = 0; 1283 1284 do { 1285 pending = false; 1286 1287 ret = nvmet_tcp_try_recv(queue, NVMET_TCP_RECV_BUDGET, &ops); 1288 if (ret > 0) 1289 pending = true; 1290 else if (ret < 0) 1291 return; 1292 1293 ret = nvmet_tcp_try_send(queue, NVMET_TCP_SEND_BUDGET, &ops); 1294 if (ret > 0) 1295 pending = true; 1296 else if (ret < 0) 1297 return; 1298 1299 } while (pending && ops < NVMET_TCP_IO_WORK_BUDGET); 1300 1301 /* 1302 * Requeue the worker if idle deadline period is in progress or any 1303 * ops activity was recorded during the do-while loop above. 1304 */ 1305 if (nvmet_tcp_check_queue_deadline(queue, ops) || pending) 1306 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work); 1307 } 1308 1309 static int nvmet_tcp_alloc_cmd(struct nvmet_tcp_queue *queue, 1310 struct nvmet_tcp_cmd *c) 1311 { 1312 u8 hdgst = nvmet_tcp_hdgst_len(queue); 1313 1314 c->queue = queue; 1315 c->req.port = queue->port->nport; 1316 1317 c->cmd_pdu = page_frag_alloc(&queue->pf_cache, 1318 sizeof(*c->cmd_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO); 1319 if (!c->cmd_pdu) 1320 return -ENOMEM; 1321 c->req.cmd = &c->cmd_pdu->cmd; 1322 1323 c->rsp_pdu = page_frag_alloc(&queue->pf_cache, 1324 sizeof(*c->rsp_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO); 1325 if (!c->rsp_pdu) 1326 goto out_free_cmd; 1327 c->req.cqe = &c->rsp_pdu->cqe; 1328 1329 c->data_pdu = page_frag_alloc(&queue->pf_cache, 1330 sizeof(*c->data_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO); 1331 if (!c->data_pdu) 1332 goto out_free_rsp; 1333 1334 c->r2t_pdu = page_frag_alloc(&queue->pf_cache, 1335 sizeof(*c->r2t_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO); 1336 if (!c->r2t_pdu) 1337 goto out_free_data; 1338 1339 c->recv_msg.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL; 1340 1341 list_add_tail(&c->entry, &queue->free_list); 1342 1343 return 0; 1344 out_free_data: 1345 page_frag_free(c->data_pdu); 1346 out_free_rsp: 1347 page_frag_free(c->rsp_pdu); 1348 out_free_cmd: 1349 page_frag_free(c->cmd_pdu); 1350 return -ENOMEM; 1351 } 1352 1353 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c) 1354 { 1355 page_frag_free(c->r2t_pdu); 1356 page_frag_free(c->data_pdu); 1357 page_frag_free(c->rsp_pdu); 1358 page_frag_free(c->cmd_pdu); 1359 } 1360 1361 static int nvmet_tcp_alloc_cmds(struct nvmet_tcp_queue *queue) 1362 { 1363 struct nvmet_tcp_cmd *cmds; 1364 int i, ret = -EINVAL, nr_cmds = queue->nr_cmds; 1365 1366 cmds = kcalloc(nr_cmds, sizeof(struct nvmet_tcp_cmd), GFP_KERNEL); 1367 if (!cmds) 1368 goto out; 1369 1370 for (i = 0; i < nr_cmds; i++) { 1371 ret = nvmet_tcp_alloc_cmd(queue, cmds + i); 1372 if (ret) 1373 goto out_free; 1374 } 1375 1376 queue->cmds = cmds; 1377 1378 return 0; 1379 out_free: 1380 while (--i >= 0) 1381 nvmet_tcp_free_cmd(cmds + i); 1382 kfree(cmds); 1383 out: 1384 return ret; 1385 } 1386 1387 static void nvmet_tcp_free_cmds(struct nvmet_tcp_queue *queue) 1388 { 1389 struct nvmet_tcp_cmd *cmds = queue->cmds; 1390 int i; 1391 1392 for (i = 0; i < queue->nr_cmds; i++) 1393 nvmet_tcp_free_cmd(cmds + i); 1394 1395 nvmet_tcp_free_cmd(&queue->connect); 1396 kfree(cmds); 1397 } 1398 1399 static void nvmet_tcp_restore_socket_callbacks(struct nvmet_tcp_queue *queue) 1400 { 1401 struct socket *sock = queue->sock; 1402 1403 write_lock_bh(&sock->sk->sk_callback_lock); 1404 sock->sk->sk_data_ready = queue->data_ready; 1405 sock->sk->sk_state_change = queue->state_change; 1406 sock->sk->sk_write_space = queue->write_space; 1407 sock->sk->sk_user_data = NULL; 1408 write_unlock_bh(&sock->sk->sk_callback_lock); 1409 } 1410 1411 static void nvmet_tcp_finish_cmd(struct nvmet_tcp_cmd *cmd) 1412 { 1413 nvmet_req_uninit(&cmd->req); 1414 nvmet_tcp_unmap_pdu_iovec(cmd); 1415 nvmet_tcp_free_cmd_buffers(cmd); 1416 } 1417 1418 static void nvmet_tcp_uninit_data_in_cmds(struct nvmet_tcp_queue *queue) 1419 { 1420 struct nvmet_tcp_cmd *cmd = queue->cmds; 1421 int i; 1422 1423 for (i = 0; i < queue->nr_cmds; i++, cmd++) { 1424 if (nvmet_tcp_need_data_in(cmd)) 1425 nvmet_req_uninit(&cmd->req); 1426 1427 nvmet_tcp_unmap_pdu_iovec(cmd); 1428 nvmet_tcp_free_cmd_buffers(cmd); 1429 } 1430 1431 if (!queue->nr_cmds && nvmet_tcp_need_data_in(&queue->connect)) { 1432 /* failed in connect */ 1433 nvmet_tcp_finish_cmd(&queue->connect); 1434 } 1435 } 1436 1437 static void nvmet_tcp_release_queue_work(struct work_struct *w) 1438 { 1439 struct page *page; 1440 struct nvmet_tcp_queue *queue = 1441 container_of(w, struct nvmet_tcp_queue, release_work); 1442 1443 mutex_lock(&nvmet_tcp_queue_mutex); 1444 list_del_init(&queue->queue_list); 1445 mutex_unlock(&nvmet_tcp_queue_mutex); 1446 1447 nvmet_tcp_restore_socket_callbacks(queue); 1448 cancel_work_sync(&queue->io_work); 1449 /* stop accepting incoming data */ 1450 queue->rcv_state = NVMET_TCP_RECV_ERR; 1451 1452 nvmet_tcp_uninit_data_in_cmds(queue); 1453 nvmet_sq_destroy(&queue->nvme_sq); 1454 cancel_work_sync(&queue->io_work); 1455 sock_release(queue->sock); 1456 nvmet_tcp_free_cmds(queue); 1457 if (queue->hdr_digest || queue->data_digest) 1458 nvmet_tcp_free_crypto(queue); 1459 ida_free(&nvmet_tcp_queue_ida, queue->idx); 1460 1461 page = virt_to_head_page(queue->pf_cache.va); 1462 __page_frag_cache_drain(page, queue->pf_cache.pagecnt_bias); 1463 kfree(queue); 1464 } 1465 1466 static void nvmet_tcp_data_ready(struct sock *sk) 1467 { 1468 struct nvmet_tcp_queue *queue; 1469 1470 read_lock_bh(&sk->sk_callback_lock); 1471 queue = sk->sk_user_data; 1472 if (likely(queue)) 1473 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work); 1474 read_unlock_bh(&sk->sk_callback_lock); 1475 } 1476 1477 static void nvmet_tcp_write_space(struct sock *sk) 1478 { 1479 struct nvmet_tcp_queue *queue; 1480 1481 read_lock_bh(&sk->sk_callback_lock); 1482 queue = sk->sk_user_data; 1483 if (unlikely(!queue)) 1484 goto out; 1485 1486 if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) { 1487 queue->write_space(sk); 1488 goto out; 1489 } 1490 1491 if (sk_stream_is_writeable(sk)) { 1492 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1493 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work); 1494 } 1495 out: 1496 read_unlock_bh(&sk->sk_callback_lock); 1497 } 1498 1499 static void nvmet_tcp_state_change(struct sock *sk) 1500 { 1501 struct nvmet_tcp_queue *queue; 1502 1503 read_lock_bh(&sk->sk_callback_lock); 1504 queue = sk->sk_user_data; 1505 if (!queue) 1506 goto done; 1507 1508 switch (sk->sk_state) { 1509 case TCP_FIN_WAIT1: 1510 case TCP_CLOSE_WAIT: 1511 case TCP_CLOSE: 1512 /* FALLTHRU */ 1513 nvmet_tcp_schedule_release_queue(queue); 1514 break; 1515 default: 1516 pr_warn("queue %d unhandled state %d\n", 1517 queue->idx, sk->sk_state); 1518 } 1519 done: 1520 read_unlock_bh(&sk->sk_callback_lock); 1521 } 1522 1523 static int nvmet_tcp_set_queue_sock(struct nvmet_tcp_queue *queue) 1524 { 1525 struct socket *sock = queue->sock; 1526 struct inet_sock *inet = inet_sk(sock->sk); 1527 int ret; 1528 1529 ret = kernel_getsockname(sock, 1530 (struct sockaddr *)&queue->sockaddr); 1531 if (ret < 0) 1532 return ret; 1533 1534 ret = kernel_getpeername(sock, 1535 (struct sockaddr *)&queue->sockaddr_peer); 1536 if (ret < 0) 1537 return ret; 1538 1539 /* 1540 * Cleanup whatever is sitting in the TCP transmit queue on socket 1541 * close. This is done to prevent stale data from being sent should 1542 * the network connection be restored before TCP times out. 1543 */ 1544 sock_no_linger(sock->sk); 1545 1546 if (so_priority > 0) 1547 sock_set_priority(sock->sk, so_priority); 1548 1549 /* Set socket type of service */ 1550 if (inet->rcv_tos > 0) 1551 ip_sock_set_tos(sock->sk, inet->rcv_tos); 1552 1553 ret = 0; 1554 write_lock_bh(&sock->sk->sk_callback_lock); 1555 if (sock->sk->sk_state != TCP_ESTABLISHED) { 1556 /* 1557 * If the socket is already closing, don't even start 1558 * consuming it 1559 */ 1560 ret = -ENOTCONN; 1561 } else { 1562 sock->sk->sk_user_data = queue; 1563 queue->data_ready = sock->sk->sk_data_ready; 1564 sock->sk->sk_data_ready = nvmet_tcp_data_ready; 1565 queue->state_change = sock->sk->sk_state_change; 1566 sock->sk->sk_state_change = nvmet_tcp_state_change; 1567 queue->write_space = sock->sk->sk_write_space; 1568 sock->sk->sk_write_space = nvmet_tcp_write_space; 1569 if (idle_poll_period_usecs) 1570 nvmet_tcp_arm_queue_deadline(queue); 1571 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work); 1572 } 1573 write_unlock_bh(&sock->sk->sk_callback_lock); 1574 1575 return ret; 1576 } 1577 1578 static int nvmet_tcp_alloc_queue(struct nvmet_tcp_port *port, 1579 struct socket *newsock) 1580 { 1581 struct nvmet_tcp_queue *queue; 1582 int ret; 1583 1584 queue = kzalloc(sizeof(*queue), GFP_KERNEL); 1585 if (!queue) 1586 return -ENOMEM; 1587 1588 INIT_WORK(&queue->release_work, nvmet_tcp_release_queue_work); 1589 INIT_WORK(&queue->io_work, nvmet_tcp_io_work); 1590 queue->sock = newsock; 1591 queue->port = port; 1592 queue->nr_cmds = 0; 1593 spin_lock_init(&queue->state_lock); 1594 queue->state = NVMET_TCP_Q_CONNECTING; 1595 INIT_LIST_HEAD(&queue->free_list); 1596 init_llist_head(&queue->resp_list); 1597 INIT_LIST_HEAD(&queue->resp_send_list); 1598 1599 queue->idx = ida_alloc(&nvmet_tcp_queue_ida, GFP_KERNEL); 1600 if (queue->idx < 0) { 1601 ret = queue->idx; 1602 goto out_free_queue; 1603 } 1604 1605 ret = nvmet_tcp_alloc_cmd(queue, &queue->connect); 1606 if (ret) 1607 goto out_ida_remove; 1608 1609 ret = nvmet_sq_init(&queue->nvme_sq); 1610 if (ret) 1611 goto out_free_connect; 1612 1613 nvmet_prepare_receive_pdu(queue); 1614 1615 mutex_lock(&nvmet_tcp_queue_mutex); 1616 list_add_tail(&queue->queue_list, &nvmet_tcp_queue_list); 1617 mutex_unlock(&nvmet_tcp_queue_mutex); 1618 1619 ret = nvmet_tcp_set_queue_sock(queue); 1620 if (ret) 1621 goto out_destroy_sq; 1622 1623 return 0; 1624 out_destroy_sq: 1625 mutex_lock(&nvmet_tcp_queue_mutex); 1626 list_del_init(&queue->queue_list); 1627 mutex_unlock(&nvmet_tcp_queue_mutex); 1628 nvmet_sq_destroy(&queue->nvme_sq); 1629 out_free_connect: 1630 nvmet_tcp_free_cmd(&queue->connect); 1631 out_ida_remove: 1632 ida_free(&nvmet_tcp_queue_ida, queue->idx); 1633 out_free_queue: 1634 kfree(queue); 1635 return ret; 1636 } 1637 1638 static void nvmet_tcp_accept_work(struct work_struct *w) 1639 { 1640 struct nvmet_tcp_port *port = 1641 container_of(w, struct nvmet_tcp_port, accept_work); 1642 struct socket *newsock; 1643 int ret; 1644 1645 while (true) { 1646 ret = kernel_accept(port->sock, &newsock, O_NONBLOCK); 1647 if (ret < 0) { 1648 if (ret != -EAGAIN) 1649 pr_warn("failed to accept err=%d\n", ret); 1650 return; 1651 } 1652 ret = nvmet_tcp_alloc_queue(port, newsock); 1653 if (ret) { 1654 pr_err("failed to allocate queue\n"); 1655 sock_release(newsock); 1656 } 1657 } 1658 } 1659 1660 static void nvmet_tcp_listen_data_ready(struct sock *sk) 1661 { 1662 struct nvmet_tcp_port *port; 1663 1664 read_lock_bh(&sk->sk_callback_lock); 1665 port = sk->sk_user_data; 1666 if (!port) 1667 goto out; 1668 1669 if (sk->sk_state == TCP_LISTEN) 1670 queue_work(nvmet_wq, &port->accept_work); 1671 out: 1672 read_unlock_bh(&sk->sk_callback_lock); 1673 } 1674 1675 static int nvmet_tcp_add_port(struct nvmet_port *nport) 1676 { 1677 struct nvmet_tcp_port *port; 1678 __kernel_sa_family_t af; 1679 int ret; 1680 1681 port = kzalloc(sizeof(*port), GFP_KERNEL); 1682 if (!port) 1683 return -ENOMEM; 1684 1685 switch (nport->disc_addr.adrfam) { 1686 case NVMF_ADDR_FAMILY_IP4: 1687 af = AF_INET; 1688 break; 1689 case NVMF_ADDR_FAMILY_IP6: 1690 af = AF_INET6; 1691 break; 1692 default: 1693 pr_err("address family %d not supported\n", 1694 nport->disc_addr.adrfam); 1695 ret = -EINVAL; 1696 goto err_port; 1697 } 1698 1699 ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr, 1700 nport->disc_addr.trsvcid, &port->addr); 1701 if (ret) { 1702 pr_err("malformed ip/port passed: %s:%s\n", 1703 nport->disc_addr.traddr, nport->disc_addr.trsvcid); 1704 goto err_port; 1705 } 1706 1707 port->nport = nport; 1708 INIT_WORK(&port->accept_work, nvmet_tcp_accept_work); 1709 if (port->nport->inline_data_size < 0) 1710 port->nport->inline_data_size = NVMET_TCP_DEF_INLINE_DATA_SIZE; 1711 1712 ret = sock_create(port->addr.ss_family, SOCK_STREAM, 1713 IPPROTO_TCP, &port->sock); 1714 if (ret) { 1715 pr_err("failed to create a socket\n"); 1716 goto err_port; 1717 } 1718 1719 port->sock->sk->sk_user_data = port; 1720 port->data_ready = port->sock->sk->sk_data_ready; 1721 port->sock->sk->sk_data_ready = nvmet_tcp_listen_data_ready; 1722 sock_set_reuseaddr(port->sock->sk); 1723 tcp_sock_set_nodelay(port->sock->sk); 1724 if (so_priority > 0) 1725 sock_set_priority(port->sock->sk, so_priority); 1726 1727 ret = kernel_bind(port->sock, (struct sockaddr *)&port->addr, 1728 sizeof(port->addr)); 1729 if (ret) { 1730 pr_err("failed to bind port socket %d\n", ret); 1731 goto err_sock; 1732 } 1733 1734 ret = kernel_listen(port->sock, 128); 1735 if (ret) { 1736 pr_err("failed to listen %d on port sock\n", ret); 1737 goto err_sock; 1738 } 1739 1740 nport->priv = port; 1741 pr_info("enabling port %d (%pISpc)\n", 1742 le16_to_cpu(nport->disc_addr.portid), &port->addr); 1743 1744 return 0; 1745 1746 err_sock: 1747 sock_release(port->sock); 1748 err_port: 1749 kfree(port); 1750 return ret; 1751 } 1752 1753 static void nvmet_tcp_destroy_port_queues(struct nvmet_tcp_port *port) 1754 { 1755 struct nvmet_tcp_queue *queue; 1756 1757 mutex_lock(&nvmet_tcp_queue_mutex); 1758 list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list) 1759 if (queue->port == port) 1760 kernel_sock_shutdown(queue->sock, SHUT_RDWR); 1761 mutex_unlock(&nvmet_tcp_queue_mutex); 1762 } 1763 1764 static void nvmet_tcp_remove_port(struct nvmet_port *nport) 1765 { 1766 struct nvmet_tcp_port *port = nport->priv; 1767 1768 write_lock_bh(&port->sock->sk->sk_callback_lock); 1769 port->sock->sk->sk_data_ready = port->data_ready; 1770 port->sock->sk->sk_user_data = NULL; 1771 write_unlock_bh(&port->sock->sk->sk_callback_lock); 1772 cancel_work_sync(&port->accept_work); 1773 /* 1774 * Destroy the remaining queues, which are not belong to any 1775 * controller yet. 1776 */ 1777 nvmet_tcp_destroy_port_queues(port); 1778 1779 sock_release(port->sock); 1780 kfree(port); 1781 } 1782 1783 static void nvmet_tcp_delete_ctrl(struct nvmet_ctrl *ctrl) 1784 { 1785 struct nvmet_tcp_queue *queue; 1786 1787 mutex_lock(&nvmet_tcp_queue_mutex); 1788 list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list) 1789 if (queue->nvme_sq.ctrl == ctrl) 1790 kernel_sock_shutdown(queue->sock, SHUT_RDWR); 1791 mutex_unlock(&nvmet_tcp_queue_mutex); 1792 } 1793 1794 static u16 nvmet_tcp_install_queue(struct nvmet_sq *sq) 1795 { 1796 struct nvmet_tcp_queue *queue = 1797 container_of(sq, struct nvmet_tcp_queue, nvme_sq); 1798 1799 if (sq->qid == 0) { 1800 /* Let inflight controller teardown complete */ 1801 flush_workqueue(nvmet_wq); 1802 } 1803 1804 queue->nr_cmds = sq->size * 2; 1805 if (nvmet_tcp_alloc_cmds(queue)) 1806 return NVME_SC_INTERNAL; 1807 return 0; 1808 } 1809 1810 static void nvmet_tcp_disc_port_addr(struct nvmet_req *req, 1811 struct nvmet_port *nport, char *traddr) 1812 { 1813 struct nvmet_tcp_port *port = nport->priv; 1814 1815 if (inet_addr_is_any((struct sockaddr *)&port->addr)) { 1816 struct nvmet_tcp_cmd *cmd = 1817 container_of(req, struct nvmet_tcp_cmd, req); 1818 struct nvmet_tcp_queue *queue = cmd->queue; 1819 1820 sprintf(traddr, "%pISc", (struct sockaddr *)&queue->sockaddr); 1821 } else { 1822 memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE); 1823 } 1824 } 1825 1826 static const struct nvmet_fabrics_ops nvmet_tcp_ops = { 1827 .owner = THIS_MODULE, 1828 .type = NVMF_TRTYPE_TCP, 1829 .msdbd = 1, 1830 .add_port = nvmet_tcp_add_port, 1831 .remove_port = nvmet_tcp_remove_port, 1832 .queue_response = nvmet_tcp_queue_response, 1833 .delete_ctrl = nvmet_tcp_delete_ctrl, 1834 .install_queue = nvmet_tcp_install_queue, 1835 .disc_traddr = nvmet_tcp_disc_port_addr, 1836 }; 1837 1838 static int __init nvmet_tcp_init(void) 1839 { 1840 int ret; 1841 1842 nvmet_tcp_wq = alloc_workqueue("nvmet_tcp_wq", WQ_HIGHPRI, 0); 1843 if (!nvmet_tcp_wq) 1844 return -ENOMEM; 1845 1846 ret = nvmet_register_transport(&nvmet_tcp_ops); 1847 if (ret) 1848 goto err; 1849 1850 return 0; 1851 err: 1852 destroy_workqueue(nvmet_tcp_wq); 1853 return ret; 1854 } 1855 1856 static void __exit nvmet_tcp_exit(void) 1857 { 1858 struct nvmet_tcp_queue *queue; 1859 1860 nvmet_unregister_transport(&nvmet_tcp_ops); 1861 1862 flush_workqueue(nvmet_wq); 1863 mutex_lock(&nvmet_tcp_queue_mutex); 1864 list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list) 1865 kernel_sock_shutdown(queue->sock, SHUT_RDWR); 1866 mutex_unlock(&nvmet_tcp_queue_mutex); 1867 flush_workqueue(nvmet_wq); 1868 1869 destroy_workqueue(nvmet_tcp_wq); 1870 } 1871 1872 module_init(nvmet_tcp_init); 1873 module_exit(nvmet_tcp_exit); 1874 1875 MODULE_LICENSE("GPL v2"); 1876 MODULE_ALIAS("nvmet-transport-3"); /* 3 == NVMF_TRTYPE_TCP */ 1877