messenger.h (18501B)
1/* SPDX-License-Identifier: GPL-2.0 */ 2#ifndef __FS_CEPH_MESSENGER_H 3#define __FS_CEPH_MESSENGER_H 4 5#include <linux/bvec.h> 6#include <linux/crypto.h> 7#include <linux/kref.h> 8#include <linux/mutex.h> 9#include <linux/net.h> 10#include <linux/radix-tree.h> 11#include <linux/uio.h> 12#include <linux/workqueue.h> 13#include <net/net_namespace.h> 14 15#include <linux/ceph/types.h> 16#include <linux/ceph/buffer.h> 17 18struct ceph_msg; 19struct ceph_connection; 20 21/* 22 * Ceph defines these callbacks for handling connection events. 23 */ 24struct ceph_connection_operations { 25 struct ceph_connection *(*get)(struct ceph_connection *); 26 void (*put)(struct ceph_connection *); 27 28 /* handle an incoming message. */ 29 void (*dispatch) (struct ceph_connection *con, struct ceph_msg *m); 30 31 /* authorize an outgoing connection */ 32 struct ceph_auth_handshake *(*get_authorizer) ( 33 struct ceph_connection *con, 34 int *proto, int force_new); 35 int (*add_authorizer_challenge)(struct ceph_connection *con, 36 void *challenge_buf, 37 int challenge_buf_len); 38 int (*verify_authorizer_reply) (struct ceph_connection *con); 39 int (*invalidate_authorizer)(struct ceph_connection *con); 40 41 /* there was some error on the socket (disconnect, whatever) */ 42 void (*fault) (struct ceph_connection *con); 43 44 /* a remote host as terminated a message exchange session, and messages 45 * we sent (or they tried to send us) may be lost. */ 46 void (*peer_reset) (struct ceph_connection *con); 47 48 struct ceph_msg * (*alloc_msg) (struct ceph_connection *con, 49 struct ceph_msg_header *hdr, 50 int *skip); 51 52 void (*reencode_message) (struct ceph_msg *msg); 53 54 int (*sign_message) (struct ceph_msg *msg); 55 int (*check_message_signature) (struct ceph_msg *msg); 56 57 /* msgr2 authentication exchange */ 58 int (*get_auth_request)(struct ceph_connection *con, 59 void *buf, int *buf_len, 60 void **authorizer, int *authorizer_len); 61 int (*handle_auth_reply_more)(struct ceph_connection *con, 62 void *reply, int reply_len, 63 void *buf, int *buf_len, 64 void **authorizer, int *authorizer_len); 65 int (*handle_auth_done)(struct ceph_connection *con, 66 u64 global_id, void *reply, int reply_len, 67 u8 *session_key, int *session_key_len, 68 u8 *con_secret, int *con_secret_len); 69 int (*handle_auth_bad_method)(struct ceph_connection *con, 70 int used_proto, int result, 71 const int *allowed_protos, int proto_cnt, 72 const int *allowed_modes, int mode_cnt); 73}; 74 75/* use format string %s%lld */ 76#define ENTITY_NAME(n) ceph_entity_type_name((n).type), le64_to_cpu((n).num) 77 78struct ceph_messenger { 79 struct ceph_entity_inst inst; /* my name+address */ 80 struct ceph_entity_addr my_enc_addr; 81 82 atomic_t stopping; 83 possible_net_t net; 84 85 /* 86 * the global_seq counts connections i (attempt to) initiate 87 * in order to disambiguate certain connect race conditions. 88 */ 89 u32 global_seq; 90 spinlock_t global_seq_lock; 91}; 92 93enum ceph_msg_data_type { 94 CEPH_MSG_DATA_NONE, /* message contains no data payload */ 95 CEPH_MSG_DATA_PAGES, /* data source/destination is a page array */ 96 CEPH_MSG_DATA_PAGELIST, /* data source/destination is a pagelist */ 97#ifdef CONFIG_BLOCK 98 CEPH_MSG_DATA_BIO, /* data source/destination is a bio list */ 99#endif /* CONFIG_BLOCK */ 100 CEPH_MSG_DATA_BVECS, /* data source/destination is a bio_vec array */ 101}; 102 103#ifdef CONFIG_BLOCK 104 105struct ceph_bio_iter { 106 struct bio *bio; 107 struct bvec_iter iter; 108}; 109 110#define __ceph_bio_iter_advance_step(it, n, STEP) do { \ 111 unsigned int __n = (n), __cur_n; \ 112 \ 113 while (__n) { \ 114 BUG_ON(!(it)->iter.bi_size); \ 115 __cur_n = min((it)->iter.bi_size, __n); \ 116 (void)(STEP); \ 117 bio_advance_iter((it)->bio, &(it)->iter, __cur_n); \ 118 if (!(it)->iter.bi_size && (it)->bio->bi_next) { \ 119 dout("__ceph_bio_iter_advance_step next bio\n"); \ 120 (it)->bio = (it)->bio->bi_next; \ 121 (it)->iter = (it)->bio->bi_iter; \ 122 } \ 123 __n -= __cur_n; \ 124 } \ 125} while (0) 126 127/* 128 * Advance @it by @n bytes. 129 */ 130#define ceph_bio_iter_advance(it, n) \ 131 __ceph_bio_iter_advance_step(it, n, 0) 132 133/* 134 * Advance @it by @n bytes, executing BVEC_STEP for each bio_vec. 135 */ 136#define ceph_bio_iter_advance_step(it, n, BVEC_STEP) \ 137 __ceph_bio_iter_advance_step(it, n, ({ \ 138 struct bio_vec bv; \ 139 struct bvec_iter __cur_iter; \ 140 \ 141 __cur_iter = (it)->iter; \ 142 __cur_iter.bi_size = __cur_n; \ 143 __bio_for_each_segment(bv, (it)->bio, __cur_iter, __cur_iter) \ 144 (void)(BVEC_STEP); \ 145 })) 146 147#endif /* CONFIG_BLOCK */ 148 149struct ceph_bvec_iter { 150 struct bio_vec *bvecs; 151 struct bvec_iter iter; 152}; 153 154#define __ceph_bvec_iter_advance_step(it, n, STEP) do { \ 155 BUG_ON((n) > (it)->iter.bi_size); \ 156 (void)(STEP); \ 157 bvec_iter_advance((it)->bvecs, &(it)->iter, (n)); \ 158} while (0) 159 160/* 161 * Advance @it by @n bytes. 162 */ 163#define ceph_bvec_iter_advance(it, n) \ 164 __ceph_bvec_iter_advance_step(it, n, 0) 165 166/* 167 * Advance @it by @n bytes, executing BVEC_STEP for each bio_vec. 168 */ 169#define ceph_bvec_iter_advance_step(it, n, BVEC_STEP) \ 170 __ceph_bvec_iter_advance_step(it, n, ({ \ 171 struct bio_vec bv; \ 172 struct bvec_iter __cur_iter; \ 173 \ 174 __cur_iter = (it)->iter; \ 175 __cur_iter.bi_size = (n); \ 176 for_each_bvec(bv, (it)->bvecs, __cur_iter, __cur_iter) \ 177 (void)(BVEC_STEP); \ 178 })) 179 180#define ceph_bvec_iter_shorten(it, n) do { \ 181 BUG_ON((n) > (it)->iter.bi_size); \ 182 (it)->iter.bi_size = (n); \ 183} while (0) 184 185struct ceph_msg_data { 186 enum ceph_msg_data_type type; 187 union { 188#ifdef CONFIG_BLOCK 189 struct { 190 struct ceph_bio_iter bio_pos; 191 u32 bio_length; 192 }; 193#endif /* CONFIG_BLOCK */ 194 struct ceph_bvec_iter bvec_pos; 195 struct { 196 struct page **pages; 197 size_t length; /* total # bytes */ 198 unsigned int alignment; /* first page */ 199 bool own_pages; 200 }; 201 struct ceph_pagelist *pagelist; 202 }; 203}; 204 205struct ceph_msg_data_cursor { 206 size_t total_resid; /* across all data items */ 207 208 struct ceph_msg_data *data; /* current data item */ 209 size_t resid; /* bytes not yet consumed */ 210 bool last_piece; /* current is last piece */ 211 bool need_crc; /* crc update needed */ 212 union { 213#ifdef CONFIG_BLOCK 214 struct ceph_bio_iter bio_iter; 215#endif /* CONFIG_BLOCK */ 216 struct bvec_iter bvec_iter; 217 struct { /* pages */ 218 unsigned int page_offset; /* offset in page */ 219 unsigned short page_index; /* index in array */ 220 unsigned short page_count; /* pages in array */ 221 }; 222 struct { /* pagelist */ 223 struct page *page; /* page from list */ 224 size_t offset; /* bytes from list */ 225 }; 226 }; 227}; 228 229/* 230 * a single message. it contains a header (src, dest, message type, etc.), 231 * footer (crc values, mainly), a "front" message body, and possibly a 232 * data payload (stored in some number of pages). 233 */ 234struct ceph_msg { 235 struct ceph_msg_header hdr; /* header */ 236 union { 237 struct ceph_msg_footer footer; /* footer */ 238 struct ceph_msg_footer_old old_footer; /* old format footer */ 239 }; 240 struct kvec front; /* unaligned blobs of message */ 241 struct ceph_buffer *middle; 242 243 size_t data_length; 244 struct ceph_msg_data *data; 245 int num_data_items; 246 int max_data_items; 247 struct ceph_msg_data_cursor cursor; 248 249 struct ceph_connection *con; 250 struct list_head list_head; /* links for connection lists */ 251 252 struct kref kref; 253 bool more_to_follow; 254 bool needs_out_seq; 255 int front_alloc_len; 256 257 struct ceph_msgpool *pool; 258}; 259 260/* 261 * connection states 262 */ 263#define CEPH_CON_S_CLOSED 1 264#define CEPH_CON_S_PREOPEN 2 265#define CEPH_CON_S_V1_BANNER 3 266#define CEPH_CON_S_V1_CONNECT_MSG 4 267#define CEPH_CON_S_V2_BANNER_PREFIX 5 268#define CEPH_CON_S_V2_BANNER_PAYLOAD 6 269#define CEPH_CON_S_V2_HELLO 7 270#define CEPH_CON_S_V2_AUTH 8 271#define CEPH_CON_S_V2_AUTH_SIGNATURE 9 272#define CEPH_CON_S_V2_SESSION_CONNECT 10 273#define CEPH_CON_S_V2_SESSION_RECONNECT 11 274#define CEPH_CON_S_OPEN 12 275#define CEPH_CON_S_STANDBY 13 276 277/* 278 * ceph_connection flag bits 279 */ 280#define CEPH_CON_F_LOSSYTX 0 /* we can close channel or drop 281 messages on errors */ 282#define CEPH_CON_F_KEEPALIVE_PENDING 1 /* we need to send a keepalive */ 283#define CEPH_CON_F_WRITE_PENDING 2 /* we have data ready to send */ 284#define CEPH_CON_F_SOCK_CLOSED 3 /* socket state changed to closed */ 285#define CEPH_CON_F_BACKOFF 4 /* need to retry queuing delayed 286 work */ 287 288/* ceph connection fault delay defaults, for exponential backoff */ 289#define BASE_DELAY_INTERVAL (HZ / 4) 290#define MAX_DELAY_INTERVAL (15 * HZ) 291 292struct ceph_connection_v1_info { 293 struct kvec out_kvec[8], /* sending header/footer data */ 294 *out_kvec_cur; 295 int out_kvec_left; /* kvec's left in out_kvec */ 296 int out_skip; /* skip this many bytes */ 297 int out_kvec_bytes; /* total bytes left */ 298 bool out_more; /* there is more data after the kvecs */ 299 bool out_msg_done; 300 301 struct ceph_auth_handshake *auth; 302 int auth_retry; /* true if we need a newer authorizer */ 303 304 /* connection negotiation temps */ 305 u8 in_banner[CEPH_BANNER_MAX_LEN]; 306 struct ceph_entity_addr actual_peer_addr; 307 struct ceph_entity_addr peer_addr_for_me; 308 struct ceph_msg_connect out_connect; 309 struct ceph_msg_connect_reply in_reply; 310 311 int in_base_pos; /* bytes read */ 312 313 /* message in temps */ 314 u8 in_tag; /* protocol control byte */ 315 struct ceph_msg_header in_hdr; 316 __le64 in_temp_ack; /* for reading an ack */ 317 318 /* message out temps */ 319 struct ceph_msg_header out_hdr; 320 __le64 out_temp_ack; /* for writing an ack */ 321 struct ceph_timespec out_temp_keepalive2; /* for writing keepalive2 322 stamp */ 323 324 u32 connect_seq; /* identify the most recent connection 325 attempt for this session */ 326 u32 peer_global_seq; /* peer's global seq for this connection */ 327}; 328 329#define CEPH_CRC_LEN 4 330#define CEPH_GCM_KEY_LEN 16 331#define CEPH_GCM_IV_LEN sizeof(struct ceph_gcm_nonce) 332#define CEPH_GCM_BLOCK_LEN 16 333#define CEPH_GCM_TAG_LEN 16 334 335#define CEPH_PREAMBLE_LEN 32 336#define CEPH_PREAMBLE_INLINE_LEN 48 337#define CEPH_PREAMBLE_PLAIN_LEN CEPH_PREAMBLE_LEN 338#define CEPH_PREAMBLE_SECURE_LEN (CEPH_PREAMBLE_LEN + \ 339 CEPH_PREAMBLE_INLINE_LEN + \ 340 CEPH_GCM_TAG_LEN) 341#define CEPH_EPILOGUE_PLAIN_LEN (1 + 3 * CEPH_CRC_LEN) 342#define CEPH_EPILOGUE_SECURE_LEN (CEPH_GCM_BLOCK_LEN + CEPH_GCM_TAG_LEN) 343 344#define CEPH_FRAME_MAX_SEGMENT_COUNT 4 345 346struct ceph_frame_desc { 347 int fd_tag; /* FRAME_TAG_* */ 348 int fd_seg_cnt; 349 int fd_lens[CEPH_FRAME_MAX_SEGMENT_COUNT]; /* logical */ 350 int fd_aligns[CEPH_FRAME_MAX_SEGMENT_COUNT]; 351}; 352 353struct ceph_gcm_nonce { 354 __le32 fixed; 355 __le64 counter __packed; 356}; 357 358struct ceph_connection_v2_info { 359 struct iov_iter in_iter; 360 struct kvec in_kvecs[5]; /* recvmsg */ 361 struct bio_vec in_bvec; /* recvmsg (in_cursor) */ 362 int in_kvec_cnt; 363 int in_state; /* IN_S_* */ 364 365 struct iov_iter out_iter; 366 struct kvec out_kvecs[8]; /* sendmsg */ 367 struct bio_vec out_bvec; /* sendpage (out_cursor, out_zero), 368 sendmsg (out_enc_pages) */ 369 int out_kvec_cnt; 370 int out_state; /* OUT_S_* */ 371 372 int out_zero; /* # of zero bytes to send */ 373 bool out_iter_sendpage; /* use sendpage if possible */ 374 375 struct ceph_frame_desc in_desc; 376 struct ceph_msg_data_cursor in_cursor; 377 struct ceph_msg_data_cursor out_cursor; 378 379 struct crypto_shash *hmac_tfm; /* post-auth signature */ 380 struct crypto_aead *gcm_tfm; /* on-wire encryption */ 381 struct aead_request *gcm_req; 382 struct crypto_wait gcm_wait; 383 struct ceph_gcm_nonce in_gcm_nonce; 384 struct ceph_gcm_nonce out_gcm_nonce; 385 386 struct page **in_enc_pages; 387 int in_enc_page_cnt; 388 int in_enc_resid; 389 int in_enc_i; 390 struct page **out_enc_pages; 391 int out_enc_page_cnt; 392 int out_enc_resid; 393 int out_enc_i; 394 395 int con_mode; /* CEPH_CON_MODE_* */ 396 397 void *conn_bufs[16]; 398 int conn_buf_cnt; 399 400 struct kvec in_sign_kvecs[8]; 401 struct kvec out_sign_kvecs[8]; 402 int in_sign_kvec_cnt; 403 int out_sign_kvec_cnt; 404 405 u64 client_cookie; 406 u64 server_cookie; 407 u64 global_seq; 408 u64 connect_seq; 409 u64 peer_global_seq; 410 411 u8 in_buf[CEPH_PREAMBLE_SECURE_LEN]; 412 u8 out_buf[CEPH_PREAMBLE_SECURE_LEN]; 413 struct { 414 u8 late_status; /* FRAME_LATE_STATUS_* */ 415 union { 416 struct { 417 u32 front_crc; 418 u32 middle_crc; 419 u32 data_crc; 420 } __packed; 421 u8 pad[CEPH_GCM_BLOCK_LEN - 1]; 422 }; 423 } out_epil; 424}; 425 426/* 427 * A single connection with another host. 428 * 429 * We maintain a queue of outgoing messages, and some session state to 430 * ensure that we can preserve the lossless, ordered delivery of 431 * messages in the case of a TCP disconnect. 432 */ 433struct ceph_connection { 434 void *private; 435 436 const struct ceph_connection_operations *ops; 437 438 struct ceph_messenger *msgr; 439 440 int state; /* CEPH_CON_S_* */ 441 atomic_t sock_state; 442 struct socket *sock; 443 444 unsigned long flags; /* CEPH_CON_F_* */ 445 const char *error_msg; /* error message, if any */ 446 447 struct ceph_entity_name peer_name; /* peer name */ 448 struct ceph_entity_addr peer_addr; /* peer address */ 449 u64 peer_features; 450 451 struct mutex mutex; 452 453 /* out queue */ 454 struct list_head out_queue; 455 struct list_head out_sent; /* sending or sent but unacked */ 456 u64 out_seq; /* last message queued for send */ 457 458 u64 in_seq, in_seq_acked; /* last message received, acked */ 459 460 struct ceph_msg *in_msg; 461 struct ceph_msg *out_msg; /* sending message (== tail of 462 out_sent) */ 463 464 struct page *bounce_page; 465 u32 in_front_crc, in_middle_crc, in_data_crc; /* calculated crc */ 466 467 struct timespec64 last_keepalive_ack; /* keepalive2 ack stamp */ 468 469 struct delayed_work work; /* send|recv work */ 470 unsigned long delay; /* current delay interval */ 471 472 union { 473 struct ceph_connection_v1_info v1; 474 struct ceph_connection_v2_info v2; 475 }; 476}; 477 478extern struct page *ceph_zero_page; 479 480void ceph_con_flag_clear(struct ceph_connection *con, unsigned long con_flag); 481void ceph_con_flag_set(struct ceph_connection *con, unsigned long con_flag); 482bool ceph_con_flag_test(struct ceph_connection *con, unsigned long con_flag); 483bool ceph_con_flag_test_and_clear(struct ceph_connection *con, 484 unsigned long con_flag); 485bool ceph_con_flag_test_and_set(struct ceph_connection *con, 486 unsigned long con_flag); 487 488void ceph_encode_my_addr(struct ceph_messenger *msgr); 489 490int ceph_tcp_connect(struct ceph_connection *con); 491int ceph_con_close_socket(struct ceph_connection *con); 492void ceph_con_reset_session(struct ceph_connection *con); 493 494u32 ceph_get_global_seq(struct ceph_messenger *msgr, u32 gt); 495void ceph_con_discard_sent(struct ceph_connection *con, u64 ack_seq); 496void ceph_con_discard_requeued(struct ceph_connection *con, u64 reconnect_seq); 497 498void ceph_msg_data_cursor_init(struct ceph_msg_data_cursor *cursor, 499 struct ceph_msg *msg, size_t length); 500struct page *ceph_msg_data_next(struct ceph_msg_data_cursor *cursor, 501 size_t *page_offset, size_t *length, 502 bool *last_piece); 503void ceph_msg_data_advance(struct ceph_msg_data_cursor *cursor, size_t bytes); 504 505u32 ceph_crc32c_page(u32 crc, struct page *page, unsigned int page_offset, 506 unsigned int length); 507 508bool ceph_addr_is_blank(const struct ceph_entity_addr *addr); 509int ceph_addr_port(const struct ceph_entity_addr *addr); 510void ceph_addr_set_port(struct ceph_entity_addr *addr, int p); 511 512void ceph_con_process_message(struct ceph_connection *con); 513int ceph_con_in_msg_alloc(struct ceph_connection *con, 514 struct ceph_msg_header *hdr, int *skip); 515void ceph_con_get_out_msg(struct ceph_connection *con); 516 517/* messenger_v1.c */ 518int ceph_con_v1_try_read(struct ceph_connection *con); 519int ceph_con_v1_try_write(struct ceph_connection *con); 520void ceph_con_v1_revoke(struct ceph_connection *con); 521void ceph_con_v1_revoke_incoming(struct ceph_connection *con); 522bool ceph_con_v1_opened(struct ceph_connection *con); 523void ceph_con_v1_reset_session(struct ceph_connection *con); 524void ceph_con_v1_reset_protocol(struct ceph_connection *con); 525 526/* messenger_v2.c */ 527int ceph_con_v2_try_read(struct ceph_connection *con); 528int ceph_con_v2_try_write(struct ceph_connection *con); 529void ceph_con_v2_revoke(struct ceph_connection *con); 530void ceph_con_v2_revoke_incoming(struct ceph_connection *con); 531bool ceph_con_v2_opened(struct ceph_connection *con); 532void ceph_con_v2_reset_session(struct ceph_connection *con); 533void ceph_con_v2_reset_protocol(struct ceph_connection *con); 534 535 536extern const char *ceph_pr_addr(const struct ceph_entity_addr *addr); 537 538extern int ceph_parse_ips(const char *c, const char *end, 539 struct ceph_entity_addr *addr, 540 int max_count, int *count, char delim); 541 542extern int ceph_msgr_init(void); 543extern void ceph_msgr_exit(void); 544extern void ceph_msgr_flush(void); 545 546extern void ceph_messenger_init(struct ceph_messenger *msgr, 547 struct ceph_entity_addr *myaddr); 548extern void ceph_messenger_fini(struct ceph_messenger *msgr); 549extern void ceph_messenger_reset_nonce(struct ceph_messenger *msgr); 550 551extern void ceph_con_init(struct ceph_connection *con, void *private, 552 const struct ceph_connection_operations *ops, 553 struct ceph_messenger *msgr); 554extern void ceph_con_open(struct ceph_connection *con, 555 __u8 entity_type, __u64 entity_num, 556 struct ceph_entity_addr *addr); 557extern bool ceph_con_opened(struct ceph_connection *con); 558extern void ceph_con_close(struct ceph_connection *con); 559extern void ceph_con_send(struct ceph_connection *con, struct ceph_msg *msg); 560 561extern void ceph_msg_revoke(struct ceph_msg *msg); 562extern void ceph_msg_revoke_incoming(struct ceph_msg *msg); 563 564extern void ceph_con_keepalive(struct ceph_connection *con); 565extern bool ceph_con_keepalive_expired(struct ceph_connection *con, 566 unsigned long interval); 567 568void ceph_msg_data_add_pages(struct ceph_msg *msg, struct page **pages, 569 size_t length, size_t alignment, bool own_pages); 570extern void ceph_msg_data_add_pagelist(struct ceph_msg *msg, 571 struct ceph_pagelist *pagelist); 572#ifdef CONFIG_BLOCK 573void ceph_msg_data_add_bio(struct ceph_msg *msg, struct ceph_bio_iter *bio_pos, 574 u32 length); 575#endif /* CONFIG_BLOCK */ 576void ceph_msg_data_add_bvecs(struct ceph_msg *msg, 577 struct ceph_bvec_iter *bvec_pos); 578 579struct ceph_msg *ceph_msg_new2(int type, int front_len, int max_data_items, 580 gfp_t flags, bool can_fail); 581extern struct ceph_msg *ceph_msg_new(int type, int front_len, gfp_t flags, 582 bool can_fail); 583 584extern struct ceph_msg *ceph_msg_get(struct ceph_msg *msg); 585extern void ceph_msg_put(struct ceph_msg *msg); 586 587extern void ceph_msg_dump(struct ceph_msg *msg); 588 589#endif