include/linux/ceph/osd_client.h

   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         bool recovery_deletes;
 152
 153         unsigned int flags;                /* CEPH_OSD_FLAG_* */
 154         bool paused;
 155
 156         u32 epoch;
 157         u32 last_force_resend;
 158
 159         int osd;
 160 };
 161
 162 /* an in-flight request */
 163 struct ceph_osd_request {
 164         u64             r_tid;              /* unique for this client */
 165         struct rb_node  r_node;
 166         struct rb_node  r_mc_node;          /* map check */
 167         struct ceph_osd *r_osd;
 168
 169         struct ceph_osd_request_target r_t;
 170 #define r_base_oid      r_t.base_oid
 171 #define r_base_oloc     r_t.base_oloc
 172 #define r_flags         r_t.flags
 173
 174         struct ceph_msg  *r_request, *r_reply;
 175         u32               r_sent;      /* >0 if r_request is sending/sent */
 176
 177         /* request osd ops array  */
 178         unsigned int            r_num_ops;
 179
 180         int               r_result;
 181
 182         struct ceph_osd_client *r_osdc;
 183         struct kref       r_kref;
 184         bool              r_mempool;
 185         struct completion r_completion;       /* private to osd_client.c */
 186         ceph_osdc_callback_t r_callback;
 187         struct list_head  r_unsafe_item;
 188
 189         struct inode *r_inode;                /* for use by callbacks */
 190         void *r_priv;                         /* ditto */
 191
 192         /* set by submitter */
 193         u64 r_snapid;                         /* for reads, CEPH_NOSNAP o/w */
 194         struct ceph_snap_context *r_snapc;    /* for writes */
 195         struct timespec r_mtime;              /* ditto */
 196         u64 r_data_offset;                    /* ditto */
 197         bool r_linger;                        /* don't resend on failure */
 198         bool r_abort_on_full;                 /* return ENOSPC when full */
 199
 200         /* internal */
 201         unsigned long r_stamp;                /* jiffies, send or check time */
 202         unsigned long r_start_stamp;          /* jiffies */
 203         int r_attempts;
 204         u32 r_map_dne_bound;
 205
 206         struct ceph_osd_req_op r_ops[];
 207 };
 208
 209 struct ceph_request_redirect {
 210         struct ceph_object_locator oloc;
 211 };
 212
 213 /*
 214  * osd request identifier
 215  *
 216  * caller name + incarnation# + tid to unique identify this request
 217  */
 218 struct ceph_osd_reqid {
 219         struct ceph_entity_name name;
 220         __le64 tid;
 221         __le32 inc;
 222 } __packed;
 223
 224 struct ceph_blkin_trace_info {
 225         __le64 trace_id;
 226         __le64 span_id;
 227         __le64 parent_span_id;
 228 } __packed;
 229
 230 typedef void (*rados_watchcb2_t)(void *arg, u64 notify_id, u64 cookie,
 231                                  u64 notifier_id, void *data, size_t data_len);
 232 typedef void (*rados_watcherrcb_t)(void *arg, u64 cookie, int err);
 233
 234 struct ceph_osd_linger_request {
 235         struct ceph_osd_client *osdc;
 236         u64 linger_id;
 237         bool committed;
 238         bool is_watch;                  /* watch or notify */
 239
 240         struct ceph_osd *osd;
 241         struct ceph_osd_request *reg_req;
 242         struct ceph_osd_request *ping_req;
 243         unsigned long ping_sent;
 244         unsigned long watch_valid_thru;
 245         struct list_head pending_lworks;
 246
 247         struct ceph_osd_request_target t;
 248         u32 map_dne_bound;
 249
 250         struct timespec mtime;
 251
 252         struct kref kref;
 253         struct mutex lock;
 254         struct rb_node node;            /* osd */
 255         struct rb_node osdc_node;       /* osdc */
 256         struct rb_node mc_node;         /* map check */
 257         struct list_head scan_item;
 258
 259         struct completion reg_commit_wait;
 260         struct completion notify_finish_wait;
 261         int reg_commit_error;
 262         int notify_finish_error;
 263         int last_error;
 264
 265         u32 register_gen;
 266         u64 notify_id;
 267
 268         rados_watchcb2_t wcb;
 269         rados_watcherrcb_t errcb;
 270         void *data;
 271
 272         struct page ***preply_pages;
 273         size_t *preply_len;
 274 };
 275
 276 struct ceph_watch_item {
 277         struct ceph_entity_name name;
 278         u64 cookie;
 279         struct ceph_entity_addr addr;
 280 };
 281
 282 struct ceph_spg_mapping {
 283         struct rb_node node;
 284         struct ceph_spg spgid;
 285
 286         struct rb_root backoffs;
 287 };
 288
 289 struct ceph_hobject_id {
 290         void *key;
 291         size_t key_len;
 292         void *oid;
 293         size_t oid_len;
 294         u64 snapid;
 295         u32 hash;
 296         u8 is_max;
 297         void *nspace;
 298         size_t nspace_len;
 299         s64 pool;
 300
 301         /* cache */
 302         u32 hash_reverse_bits;
 303 };
 304
 305 static inline void ceph_hoid_build_hash_cache(struct ceph_hobject_id *hoid)
 306 {
 307         hoid->hash_reverse_bits = bitrev32(hoid->hash);
 308 }
 309
 310 /*
 311  * PG-wide backoff: [begin, end)
 312  * per-object backoff: begin == end
 313  */
 314 struct ceph_osd_backoff {
 315         struct rb_node spg_node;
 316         struct rb_node id_node;
 317
 318         struct ceph_spg spgid;
 319         u64 id;
 320         struct ceph_hobject_id *begin;
 321         struct ceph_hobject_id *end;
 322 };
 323
 324 #define CEPH_LINGER_ID_START    0xffff000000000000ULL
 325
 326 struct ceph_osd_client {
 327         struct ceph_client     *client;
 328
 329         struct ceph_osdmap     *osdmap;       /* current map */
 330         struct rw_semaphore    lock;
 331
 332         struct rb_root         osds;          /* osds */
 333         struct list_head       osd_lru;       /* idle osds */
 334         spinlock_t             osd_lru_lock;
 335         u32                    epoch_barrier;
 336         struct ceph_osd        homeless_osd;
 337         atomic64_t             last_tid;      /* tid of last request */
 338         u64                    last_linger_id;
 339         struct rb_root         linger_requests; /* lingering requests */
 340         struct rb_root         map_checks;
 341         struct rb_root         linger_map_checks;
 342         atomic_t               num_requests;
 343         atomic_t               num_homeless;
 344         struct delayed_work    timeout_work;
 345         struct delayed_work    osds_timeout_work;
 346 #ifdef CONFIG_DEBUG_FS
 347         struct dentry          *debugfs_file;
 348 #endif
 349
 350         mempool_t              *req_mempool;
 351
 352         struct ceph_msgpool     msgpool_op;
 353         struct ceph_msgpool     msgpool_op_reply;
 354
 355         struct workqueue_struct *notify_wq;
 356 };
 357
 358 static inline bool ceph_osdmap_flag(struct ceph_osd_client *osdc, int flag)
 359 {
 360         return osdc->osdmap->flags & flag;
 361 }
 362
 363 extern int ceph_osdc_setup(void);
 364 extern void ceph_osdc_cleanup(void);
 365
 366 extern int ceph_osdc_init(struct ceph_osd_client *osdc,
 367                           struct ceph_client *client);
 368 extern void ceph_osdc_stop(struct ceph_osd_client *osdc);
 369
 370 extern void ceph_osdc_handle_reply(struct ceph_osd_client *osdc,
 371                                    struct ceph_msg *msg);
 372 extern void ceph_osdc_handle_map(struct ceph_osd_client *osdc,
 373                                  struct ceph_msg *msg);
 374 void ceph_osdc_update_epoch_barrier(struct ceph_osd_client *osdc, u32 eb);
 375
 376 extern void osd_req_op_init(struct ceph_osd_request *osd_req,
 377                             unsigned int which, u16 opcode, u32 flags);
 378
 379 extern void osd_req_op_raw_data_in_pages(struct ceph_osd_request *,
 380                                         unsigned int which,
 381                                         struct page **pages, u64 length,
 382                                         u32 alignment, bool pages_from_pool,
 383                                         bool own_pages);
 384
 385 extern void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
 386                                         unsigned int which, u16 opcode,
 387                                         u64 offset, u64 length,
 388                                         u64 truncate_size, u32 truncate_seq);
 389 extern void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
 390                                         unsigned int which, u64 length);
 391 extern void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req,
 392                                        unsigned int which, u64 offset_inc);
 393
 394 extern struct ceph_osd_data *osd_req_op_extent_osd_data(
 395                                         struct ceph_osd_request *osd_req,
 396                                         unsigned int which);
 397
 398 extern void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *,
 399                                         unsigned int which,
 400                                         struct page **pages, u64 length,
 401                                         u32 alignment, bool pages_from_pool,
 402                                         bool own_pages);
 403 extern void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *,
 404                                         unsigned int which,
 405                                         struct ceph_pagelist *pagelist);
 406 #ifdef CONFIG_BLOCK
 407 extern void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *,
 408                                         unsigned int which,
 409                                         struct bio *bio, size_t bio_length);
 410 #endif /* CONFIG_BLOCK */
 411
 412 extern void osd_req_op_cls_request_data_pagelist(struct ceph_osd_request *,
 413                                         unsigned int which,
 414                                         struct ceph_pagelist *pagelist);
 415 extern void osd_req_op_cls_request_data_pages(struct ceph_osd_request *,
 416                                         unsigned int which,
 417                                         struct page **pages, u64 length,
 418                                         u32 alignment, bool pages_from_pool,
 419                                         bool own_pages);
 420 extern void osd_req_op_cls_response_data_pages(struct ceph_osd_request *,
 421                                         unsigned int which,
 422                                         struct page **pages, u64 length,
 423                                         u32 alignment, bool pages_from_pool,
 424                                         bool own_pages);
 425 extern void osd_req_op_cls_init(struct ceph_osd_request *osd_req,
 426                                         unsigned int which, u16 opcode,
 427                                         const char *class, const char *method);
 428 extern int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
 429                                  u16 opcode, const char *name, const void *value,
 430                                  size_t size, u8 cmp_op, u8 cmp_mode);
 431 extern void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
 432                                        unsigned int which,
 433                                        u64 expected_object_size,
 434                                        u64 expected_write_size);
 435
 436 extern struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
 437                                                struct ceph_snap_context *snapc,
 438                                                unsigned int num_ops,
 439                                                bool use_mempool,
 440                                                gfp_t gfp_flags);
 441 int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp);
 442
 443 extern struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *,
 444                                       struct ceph_file_layout *layout,
 445                                       struct ceph_vino vino,
 446                                       u64 offset, u64 *len,
 447                                       unsigned int which, int num_ops,
 448                                       int opcode, int flags,
 449                                       struct ceph_snap_context *snapc,
 450                                       u32 truncate_seq, u64 truncate_size,
 451                                       bool use_mempool);
 452
 453 extern void ceph_osdc_get_request(struct ceph_osd_request *req);
 454 extern void ceph_osdc_put_request(struct ceph_osd_request *req);
 455
 456 extern int ceph_osdc_start_request(struct ceph_osd_client *osdc,
 457                                    struct ceph_osd_request *req,
 458                                    bool nofail);
 459 extern void ceph_osdc_cancel_request(struct ceph_osd_request *req);
 460 extern int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
 461                                   struct ceph_osd_request *req);
 462 extern void ceph_osdc_sync(struct ceph_osd_client *osdc);
 463
 464 extern void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc);
 465 void ceph_osdc_maybe_request_map(struct ceph_osd_client *osdc);
 466
 467 int ceph_osdc_call(struct ceph_osd_client *osdc,
 468                    struct ceph_object_id *oid,
 469                    struct ceph_object_locator *oloc,
 470                    const char *class, const char *method,
 471                    unsigned int flags,
 472                    struct page *req_page, size_t req_len,
 473                    struct page *resp_page, size_t *resp_len);
 474
 475 extern int ceph_osdc_readpages(struct ceph_osd_client *osdc,
 476                                struct ceph_vino vino,
 477                                struct ceph_file_layout *layout,
 478                                u64 off, u64 *plen,
 479                                u32 truncate_seq, u64 truncate_size,
 480                                struct page **pages, int nr_pages,
 481                                int page_align);
 482
 483 extern int ceph_osdc_writepages(struct ceph_osd_client *osdc,
 484                                 struct ceph_vino vino,
 485                                 struct ceph_file_layout *layout,
 486                                 struct ceph_snap_context *sc,
 487                                 u64 off, u64 len,
 488                                 u32 truncate_seq, u64 truncate_size,
 489                                 struct timespec *mtime,
 490                                 struct page **pages, int nr_pages);
 491
 492 /* watch/notify */
 493 struct ceph_osd_linger_request *
 494 ceph_osdc_watch(struct ceph_osd_client *osdc,
 495                 struct ceph_object_id *oid,
 496                 struct ceph_object_locator *oloc,
 497                 rados_watchcb2_t wcb,
 498                 rados_watcherrcb_t errcb,
 499                 void *data);
 500 int ceph_osdc_unwatch(struct ceph_osd_client *osdc,
 501                       struct ceph_osd_linger_request *lreq);
 502
 503 int ceph_osdc_notify_ack(struct ceph_osd_client *osdc,
 504                          struct ceph_object_id *oid,
 505                          struct ceph_object_locator *oloc,
 506                          u64 notify_id,
 507                          u64 cookie,
 508                          void *payload,
 509                          size_t payload_len);
 510 int ceph_osdc_notify(struct ceph_osd_client *osdc,
 511                      struct ceph_object_id *oid,
 512                      struct ceph_object_locator *oloc,
 513                      void *payload,
 514                      size_t payload_len,
 515                      u32 timeout,
 516                      struct page ***preply_pages,
 517                      size_t *preply_len);
 518 int ceph_osdc_watch_check(struct ceph_osd_client *osdc,
 519                           struct ceph_osd_linger_request *lreq);
 520 int ceph_osdc_list_watchers(struct ceph_osd_client *osdc,
 521                             struct ceph_object_id *oid,
 522                             struct ceph_object_locator *oloc,
 523                             struct ceph_watch_item **watchers,
 524                             u32 *num_watchers);
 525 #endif
 526