1 #ifndef _FS_CEPH_OSD_CLIENT_H 2 #define _FS_CEPH_OSD_CLIENT_H 3 4 #include <linux/bitrev.h> 5 #include <linux/completion.h> 6 #include <linux/kref.h> 7 #include <linux/mempool.h> 8 #include <linux/rbtree.h> 9 #include <linux/refcount.h> 10 11 #include <linux/ceph/types.h> 12 #include <linux/ceph/osdmap.h> 13 #include <linux/ceph/messenger.h> 14 #include <linux/ceph/msgpool.h> 15 #include <linux/ceph/auth.h> 16 #include <linux/ceph/pagelist.h> 17 18 struct ceph_msg; 19 struct ceph_snap_context; 20 struct ceph_osd_request; 21 struct ceph_osd_client; 22 23 /* 24 * completion callback for async writepages 25 */ 26 typedef void (*ceph_osdc_callback_t)(struct ceph_osd_request *); 27 28 #define CEPH_HOMELESS_OSD -1 29 30 /* a given osd we're communicating with */ 31 struct ceph_osd { 32 refcount_t o_ref; 33 struct ceph_osd_client *o_osdc; 34 int o_osd; 35 int o_incarnation; 36 struct rb_node o_node; 37 struct ceph_connection o_con; 38 struct rb_root o_requests; 39 struct rb_root o_linger_requests; 40 struct rb_root o_backoff_mappings; 41 struct rb_root o_backoffs_by_id; 42 struct list_head o_osd_lru; 43 struct ceph_auth_handshake o_auth; 44 unsigned long lru_ttl; 45 struct list_head o_keepalive_item; 46 struct mutex lock; 47 }; 48 49 #define CEPH_OSD_SLAB_OPS 2 50 #define CEPH_OSD_MAX_OPS 16 51 52 enum ceph_osd_data_type { 53 CEPH_OSD_DATA_TYPE_NONE = 0, 54 CEPH_OSD_DATA_TYPE_PAGES, 55 CEPH_OSD_DATA_TYPE_PAGELIST, 56 #ifdef CONFIG_BLOCK 57 CEPH_OSD_DATA_TYPE_BIO, 58 #endif /* CONFIG_BLOCK */ 59 }; 60 61 struct ceph_osd_data { 62 enum ceph_osd_data_type type; 63 union { 64 struct { 65 struct page **pages; 66 u64 length; 67 u32 alignment; 68 bool pages_from_pool; 69 bool own_pages; 70 }; 71 struct ceph_pagelist *pagelist; 72 #ifdef CONFIG_BLOCK 73 struct { 74 struct bio *bio; /* list of bios */ 75 size_t bio_length; /* total in list */ 76 }; 77 #endif /* CONFIG_BLOCK */ 78 }; 79 }; 80 81 struct ceph_osd_req_op { 82 u16 op; /* CEPH_OSD_OP_* */ 83 u32 flags; /* CEPH_OSD_OP_FLAG_* */ 84 u32 indata_len; /* request */ 85 u32 outdata_len; /* reply */ 86 s32 rval; 87 88 union { 89 struct ceph_osd_data raw_data_in; 90 struct { 91 u64 offset, length; 92 u64 truncate_size; 93 u32 truncate_seq; 94 struct ceph_osd_data osd_data; 95 } extent; 96 struct { 97 u32 name_len; 98 u32 value_len; 99 __u8 cmp_op; /* CEPH_OSD_CMPXATTR_OP_* */ 100 __u8 cmp_mode; /* CEPH_OSD_CMPXATTR_MODE_* */ 101 struct ceph_osd_data osd_data; 102 } xattr; 103 struct { 104 const char *class_name; 105 const char *method_name; 106 struct ceph_osd_data request_info; 107 struct ceph_osd_data request_data; 108 struct ceph_osd_data response_data; 109 __u8 class_len; 110 __u8 method_len; 111 u32 indata_len; 112 } cls; 113 struct { 114 u64 cookie; 115 __u8 op; /* CEPH_OSD_WATCH_OP_ */ 116 u32 gen; 117 } watch; 118 struct { 119 struct ceph_osd_data request_data; 120 } notify_ack; 121 struct { 122 u64 cookie; 123 struct ceph_osd_data request_data; 124 struct ceph_osd_data response_data; 125 } notify; 126 struct { 127 struct ceph_osd_data response_data; 128 } list_watchers; 129 struct { 130 u64 expected_object_size; 131 u64 expected_write_size; 132 } alloc_hint; 133 }; 134 }; 135 136 struct ceph_osd_request_target { 137 struct ceph_object_id base_oid; 138 struct ceph_object_locator base_oloc; 139 struct ceph_object_id target_oid; 140 struct ceph_object_locator target_oloc; 141 142 struct ceph_pg pgid; /* last raw pg we mapped to */ 143 struct ceph_spg spgid; /* last actual spg we mapped to */ 144 u32 pg_num; 145 u32 pg_num_mask; 146 struct ceph_osds acting; 147 struct ceph_osds up; 148 int size; 149 int min_size; 150 bool sort_bitwise; 151 152 unsigned int flags; /* CEPH_OSD_FLAG_* */ 153 bool paused; 154 155 u32 epoch; 156 u32 last_force_resend; 157 158 int osd; 159 }; 160 161 /* an in-flight request */ 162 struct ceph_osd_request { 163 u64 r_tid; /* unique for this client */ 164 struct rb_node r_node; 165 struct rb_node r_mc_node; /* map check */ 166 struct ceph_osd *r_osd; 167 168 struct ceph_osd_request_target r_t; 169 #define r_base_oid r_t.base_oid 170 #define r_base_oloc r_t.base_oloc 171 #define r_flags r_t.flags 172 173 struct ceph_msg *r_request, *r_reply; 174 u32 r_sent; /* >0 if r_request is sending/sent */ 175 176 /* request osd ops array */ 177 unsigned int r_num_ops; 178 179 int r_result; 180 181 struct ceph_osd_client *r_osdc; 182 struct kref r_kref; 183 bool r_mempool; 184 struct completion r_completion; /* private to osd_client.c */ 185 ceph_osdc_callback_t r_callback; 186 struct list_head r_unsafe_item; 187 188 struct inode *r_inode; /* for use by callbacks */ 189 void *r_priv; /* ditto */ 190 191 /* set by submitter */ 192 u64 r_snapid; /* for reads, CEPH_NOSNAP o/w */ 193 struct ceph_snap_context *r_snapc; /* for writes */ 194 struct timespec r_mtime; /* ditto */ 195 u64 r_data_offset; /* ditto */ 196 bool r_linger; /* don't resend on failure */ 197 bool r_abort_on_full; /* return ENOSPC when full */ 198 199 /* internal */ 200 unsigned long r_stamp; /* jiffies, send or check time */ 201 unsigned long r_start_stamp; /* jiffies */ 202 int r_attempts; 203 u32 r_map_dne_bound; 204 205 struct ceph_osd_req_op r_ops[]; 206 }; 207 208 struct ceph_request_redirect { 209 struct ceph_object_locator oloc; 210 }; 211 212 /* 213 * osd request identifier 214 * 215 * caller name + incarnation# + tid to unique identify this request 216 */ 217 struct ceph_osd_reqid { 218 struct ceph_entity_name name; 219 __le64 tid; 220 __le32 inc; 221 } __packed; 222 223 struct ceph_blkin_trace_info { 224 __le64 trace_id; 225 __le64 span_id; 226 __le64 parent_span_id; 227 } __packed; 228 229 typedef void (*rados_watchcb2_t)(void *arg, u64 notify_id, u64 cookie, 230 u64 notifier_id, void *data, size_t data_len); 231 typedef void (*rados_watcherrcb_t)(void *arg, u64 cookie, int err); 232 233 struct ceph_osd_linger_request { 234 struct ceph_osd_client *osdc; 235 u64 linger_id; 236 bool committed; 237 bool is_watch; /* watch or notify */ 238 239 struct ceph_osd *osd; 240 struct ceph_osd_request *reg_req; 241 struct ceph_osd_request *ping_req; 242 unsigned long ping_sent; 243 unsigned long watch_valid_thru; 244 struct list_head pending_lworks; 245 246 struct ceph_osd_request_target t; 247 u32 map_dne_bound; 248 249 struct timespec mtime; 250 251 struct kref kref; 252 struct mutex lock; 253 struct rb_node node; /* osd */ 254 struct rb_node osdc_node; /* osdc */ 255 struct rb_node mc_node; /* map check */ 256 struct list_head scan_item; 257 258 struct completion reg_commit_wait; 259 struct completion notify_finish_wait; 260 int reg_commit_error; 261 int notify_finish_error; 262 int last_error; 263 264 u32 register_gen; 265 u64 notify_id; 266 267 rados_watchcb2_t wcb; 268 rados_watcherrcb_t errcb; 269 void *data; 270 271 struct page ***preply_pages; 272 size_t *preply_len; 273 }; 274 275 struct ceph_watch_item { 276 struct ceph_entity_name name; 277 u64 cookie; 278 struct ceph_entity_addr addr; 279 }; 280 281 struct ceph_spg_mapping { 282 struct rb_node node; 283 struct ceph_spg spgid; 284 285 struct rb_root backoffs; 286 }; 287 288 struct ceph_hobject_id { 289 void *key; 290 size_t key_len; 291 void *oid; 292 size_t oid_len; 293 u64 snapid; 294 u32 hash; 295 u8 is_max; 296 void *nspace; 297 size_t nspace_len; 298 s64 pool; 299 300 /* cache */ 301 u32 hash_reverse_bits; 302 }; 303 304 static inline void ceph_hoid_build_hash_cache(struct ceph_hobject_id *hoid) 305 { 306 hoid->hash_reverse_bits = bitrev32(hoid->hash); 307 } 308 309 /* 310 * PG-wide backoff: [begin, end) 311 * per-object backoff: begin == end 312 */ 313 struct ceph_osd_backoff { 314 struct rb_node spg_node; 315 struct rb_node id_node; 316 317 struct ceph_spg spgid; 318 u64 id; 319 struct ceph_hobject_id *begin; 320 struct ceph_hobject_id *end; 321 }; 322 323 #define CEPH_LINGER_ID_START 0xffff000000000000ULL 324 325 struct ceph_osd_client { 326 struct ceph_client *client; 327 328 struct ceph_osdmap *osdmap; /* current map */ 329 struct rw_semaphore lock; 330 331 struct rb_root osds; /* osds */ 332 struct list_head osd_lru; /* idle osds */ 333 spinlock_t osd_lru_lock; 334 u32 epoch_barrier; 335 struct ceph_osd homeless_osd; 336 atomic64_t last_tid; /* tid of last request */ 337 u64 last_linger_id; 338 struct rb_root linger_requests; /* lingering requests */ 339 struct rb_root map_checks; 340 struct rb_root linger_map_checks; 341 atomic_t num_requests; 342 atomic_t num_homeless; 343 struct delayed_work timeout_work; 344 struct delayed_work osds_timeout_work; 345 #ifdef CONFIG_DEBUG_FS 346 struct dentry *debugfs_file; 347 #endif 348 349 mempool_t *req_mempool; 350 351 struct ceph_msgpool msgpool_op; 352 struct ceph_msgpool msgpool_op_reply; 353 354 struct workqueue_struct *notify_wq; 355 }; 356 357 static inline bool ceph_osdmap_flag(struct ceph_osd_client *osdc, int flag) 358 { 359 return osdc->osdmap->flags & flag; 360 } 361 362 extern int ceph_osdc_setup(void); 363 extern void ceph_osdc_cleanup(void); 364 365 extern int ceph_osdc_init(struct ceph_osd_client *osdc, 366 struct ceph_client *client); 367 extern void ceph_osdc_stop(struct ceph_osd_client *osdc); 368 369 extern void ceph_osdc_handle_reply(struct ceph_osd_client *osdc, 370 struct ceph_msg *msg); 371 extern void ceph_osdc_handle_map(struct ceph_osd_client *osdc, 372 struct ceph_msg *msg); 373 void ceph_osdc_update_epoch_barrier(struct ceph_osd_client *osdc, u32 eb); 374 375 extern void osd_req_op_init(struct ceph_osd_request *osd_req, 376 unsigned int which, u16 opcode, u32 flags); 377 378 extern void osd_req_op_raw_data_in_pages(struct ceph_osd_request *, 379 unsigned int which, 380 struct page **pages, u64 length, 381 u32 alignment, bool pages_from_pool, 382 bool own_pages); 383 384 extern void osd_req_op_extent_init(struct ceph_osd_request *osd_req, 385 unsigned int which, u16 opcode, 386 u64 offset, u64 length, 387 u64 truncate_size, u32 truncate_seq); 388 extern void osd_req_op_extent_update(struct ceph_osd_request *osd_req, 389 unsigned int which, u64 length); 390 extern void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req, 391 unsigned int which, u64 offset_inc); 392 393 extern struct ceph_osd_data *osd_req_op_extent_osd_data( 394 struct ceph_osd_request *osd_req, 395 unsigned int which); 396 397 extern void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *, 398 unsigned int which, 399 struct page **pages, u64 length, 400 u32 alignment, bool pages_from_pool, 401 bool own_pages); 402 extern void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *, 403 unsigned int which, 404 struct ceph_pagelist *pagelist); 405 #ifdef CONFIG_BLOCK 406 extern void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *, 407 unsigned int which, 408 struct bio *bio, size_t bio_length); 409 #endif /* CONFIG_BLOCK */ 410 411 extern void osd_req_op_cls_request_data_pagelist(struct ceph_osd_request *, 412 unsigned int which, 413 struct ceph_pagelist *pagelist); 414 extern void osd_req_op_cls_request_data_pages(struct ceph_osd_request *, 415 unsigned int which, 416 struct page **pages, u64 length, 417 u32 alignment, bool pages_from_pool, 418 bool own_pages); 419 extern void osd_req_op_cls_response_data_pages(struct ceph_osd_request *, 420 unsigned int which, 421 struct page **pages, u64 length, 422 u32 alignment, bool pages_from_pool, 423 bool own_pages); 424 extern void osd_req_op_cls_init(struct ceph_osd_request *osd_req, 425 unsigned int which, u16 opcode, 426 const char *class, const char *method); 427 extern int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which, 428 u16 opcode, const char *name, const void *value, 429 size_t size, u8 cmp_op, u8 cmp_mode); 430 extern void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req, 431 unsigned int which, 432 u64 expected_object_size, 433 u64 expected_write_size); 434 435 extern struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 436 struct ceph_snap_context *snapc, 437 unsigned int num_ops, 438 bool use_mempool, 439 gfp_t gfp_flags); 440 int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp); 441 442 extern struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *, 443 struct ceph_file_layout *layout, 444 struct ceph_vino vino, 445 u64 offset, u64 *len, 446 unsigned int which, int num_ops, 447 int opcode, int flags, 448 struct ceph_snap_context *snapc, 449 u32 truncate_seq, u64 truncate_size, 450 bool use_mempool); 451 452 extern void ceph_osdc_get_request(struct ceph_osd_request *req); 453 extern void ceph_osdc_put_request(struct ceph_osd_request *req); 454 455 extern int ceph_osdc_start_request(struct ceph_osd_client *osdc, 456 struct ceph_osd_request *req, 457 bool nofail); 458 extern void ceph_osdc_cancel_request(struct ceph_osd_request *req); 459 extern int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 460 struct ceph_osd_request *req); 461 extern void ceph_osdc_sync(struct ceph_osd_client *osdc); 462 463 extern void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc); 464 void ceph_osdc_maybe_request_map(struct ceph_osd_client *osdc); 465 466 int ceph_osdc_call(struct ceph_osd_client *osdc, 467 struct ceph_object_id *oid, 468 struct ceph_object_locator *oloc, 469 const char *class, const char *method, 470 unsigned int flags, 471 struct page *req_page, size_t req_len, 472 struct page *resp_page, size_t *resp_len); 473 474 extern int ceph_osdc_readpages(struct ceph_osd_client *osdc, 475 struct ceph_vino vino, 476 struct ceph_file_layout *layout, 477 u64 off, u64 *plen, 478 u32 truncate_seq, u64 truncate_size, 479 struct page **pages, int nr_pages, 480 int page_align); 481 482 extern int ceph_osdc_writepages(struct ceph_osd_client *osdc, 483 struct ceph_vino vino, 484 struct ceph_file_layout *layout, 485 struct ceph_snap_context *sc, 486 u64 off, u64 len, 487 u32 truncate_seq, u64 truncate_size, 488 struct timespec *mtime, 489 struct page **pages, int nr_pages); 490 491 /* watch/notify */ 492 struct ceph_osd_linger_request * 493 ceph_osdc_watch(struct ceph_osd_client *osdc, 494 struct ceph_object_id *oid, 495 struct ceph_object_locator *oloc, 496 rados_watchcb2_t wcb, 497 rados_watcherrcb_t errcb, 498 void *data); 499 int ceph_osdc_unwatch(struct ceph_osd_client *osdc, 500 struct ceph_osd_linger_request *lreq); 501 502 int ceph_osdc_notify_ack(struct ceph_osd_client *osdc, 503 struct ceph_object_id *oid, 504 struct ceph_object_locator *oloc, 505 u64 notify_id, 506 u64 cookie, 507 void *payload, 508 size_t payload_len); 509 int ceph_osdc_notify(struct ceph_osd_client *osdc, 510 struct ceph_object_id *oid, 511 struct ceph_object_locator *oloc, 512 void *payload, 513 size_t payload_len, 514 u32 timeout, 515 struct page ***preply_pages, 516 size_t *preply_len); 517 int ceph_osdc_watch_check(struct ceph_osd_client *osdc, 518 struct ceph_osd_linger_request *lreq); 519 int ceph_osdc_list_watchers(struct ceph_osd_client *osdc, 520 struct ceph_object_id *oid, 521 struct ceph_object_locator *oloc, 522 struct ceph_watch_item **watchers, 523 u32 *num_watchers); 524 #endif 525 526