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