include/linux/ceph/osd_client.h

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