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 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 #define CEPH_LINGER_ID_START    0xffff000000000000ULL
 262
 263 struct ceph_osd_client {
 264         struct ceph_client     *client;
 265
 266         struct ceph_osdmap     *osdmap;       /* current map */
 267         struct rw_semaphore    lock;
 268
 269         struct rb_root         osds;          /* osds */
 270         struct list_head       osd_lru;       /* idle osds */
 271         spinlock_t             osd_lru_lock;
 272         struct ceph_osd        homeless_osd;
 273         atomic64_t             last_tid;      /* tid of last request */
 274         u64                    last_linger_id;
 275         struct rb_root         linger_requests; /* lingering requests */
 276         struct rb_root         map_checks;
 277         struct rb_root         linger_map_checks;
 278         atomic_t               num_requests;
 279         atomic_t               num_homeless;
 280         struct delayed_work    timeout_work;
 281         struct delayed_work    osds_timeout_work;
 282 #ifdef CONFIG_DEBUG_FS
 283         struct dentry          *debugfs_file;
 284 #endif
 285
 286         mempool_t              *req_mempool;
 287
 288         struct ceph_msgpool     msgpool_op;
 289         struct ceph_msgpool     msgpool_op_reply;
 290
 291         struct workqueue_struct *notify_wq;
 292 };
 293
 294 static inline bool ceph_osdmap_flag(struct ceph_osd_client *osdc, int flag)
 295 {
 296         return osdc->osdmap->flags & flag;
 297 }
 298
 299 extern int ceph_osdc_setup(void);
 300 extern void ceph_osdc_cleanup(void);
 301
 302 extern int ceph_osdc_init(struct ceph_osd_client *osdc,
 303                           struct ceph_client *client);
 304 extern void ceph_osdc_stop(struct ceph_osd_client *osdc);
 305
 306 extern void ceph_osdc_handle_reply(struct ceph_osd_client *osdc,
 307                                    struct ceph_msg *msg);
 308 extern void ceph_osdc_handle_map(struct ceph_osd_client *osdc,
 309                                  struct ceph_msg *msg);
 310
 311 extern void osd_req_op_init(struct ceph_osd_request *osd_req,
 312                             unsigned int which, u16 opcode, u32 flags);
 313
 314 extern void osd_req_op_raw_data_in_pages(struct ceph_osd_request *,
 315                                         unsigned int which,
 316                                         struct page **pages, u64 length,
 317                                         u32 alignment, bool pages_from_pool,
 318                                         bool own_pages);
 319
 320 extern void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
 321                                         unsigned int which, u16 opcode,
 322                                         u64 offset, u64 length,
 323                                         u64 truncate_size, u32 truncate_seq);
 324 extern void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
 325                                         unsigned int which, u64 length);
 326 extern void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req,
 327                                        unsigned int which, u64 offset_inc);
 328
 329 extern struct ceph_osd_data *osd_req_op_extent_osd_data(
 330                                         struct ceph_osd_request *osd_req,
 331                                         unsigned int which);
 332
 333 extern void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *,
 334                                         unsigned int which,
 335                                         struct page **pages, u64 length,
 336                                         u32 alignment, bool pages_from_pool,
 337                                         bool own_pages);
 338 extern void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *,
 339                                         unsigned int which,
 340                                         struct ceph_pagelist *pagelist);
 341 #ifdef CONFIG_BLOCK
 342 extern void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *,
 343                                         unsigned int which,
 344                                         struct bio *bio, size_t bio_length);
 345 #endif /* CONFIG_BLOCK */
 346
 347 extern void osd_req_op_cls_request_data_pagelist(struct ceph_osd_request *,
 348                                         unsigned int which,
 349                                         struct ceph_pagelist *pagelist);
 350 extern void osd_req_op_cls_request_data_pages(struct ceph_osd_request *,
 351                                         unsigned int which,
 352                                         struct page **pages, u64 length,
 353                                         u32 alignment, bool pages_from_pool,
 354                                         bool own_pages);
 355 extern void osd_req_op_cls_response_data_pages(struct ceph_osd_request *,
 356                                         unsigned int which,
 357                                         struct page **pages, u64 length,
 358                                         u32 alignment, bool pages_from_pool,
 359                                         bool own_pages);
 360 extern void osd_req_op_cls_init(struct ceph_osd_request *osd_req,
 361                                         unsigned int which, u16 opcode,
 362                                         const char *class, const char *method);
 363 extern int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
 364                                  u16 opcode, const char *name, const void *value,
 365                                  size_t size, u8 cmp_op, u8 cmp_mode);
 366 extern void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
 367                                        unsigned int which,
 368                                        u64 expected_object_size,
 369                                        u64 expected_write_size);
 370
 371 extern struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
 372                                                struct ceph_snap_context *snapc,
 373                                                unsigned int num_ops,
 374                                                bool use_mempool,
 375                                                gfp_t gfp_flags);
 376 int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp);
 377
 378 extern struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *,
 379                                       struct ceph_file_layout *layout,
 380                                       struct ceph_vino vino,
 381                                       u64 offset, u64 *len,
 382                                       unsigned int which, int num_ops,
 383                                       int opcode, int flags,
 384                                       struct ceph_snap_context *snapc,
 385                                       u32 truncate_seq, u64 truncate_size,
 386                                       bool use_mempool);
 387
 388 extern void ceph_osdc_get_request(struct ceph_osd_request *req);
 389 extern void ceph_osdc_put_request(struct ceph_osd_request *req);
 390
 391 extern int ceph_osdc_start_request(struct ceph_osd_client *osdc,
 392                                    struct ceph_osd_request *req,
 393                                    bool nofail);
 394 extern void ceph_osdc_cancel_request(struct ceph_osd_request *req);
 395 extern int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
 396                                   struct ceph_osd_request *req);
 397 extern void ceph_osdc_sync(struct ceph_osd_client *osdc);
 398
 399 extern void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc);
 400 void ceph_osdc_maybe_request_map(struct ceph_osd_client *osdc);
 401
 402 int ceph_osdc_call(struct ceph_osd_client *osdc,
 403                    struct ceph_object_id *oid,
 404                    struct ceph_object_locator *oloc,
 405                    const char *class, const char *method,
 406                    unsigned int flags,
 407                    struct page *req_page, size_t req_len,
 408                    struct page *resp_page, size_t *resp_len);
 409
 410 extern int ceph_osdc_readpages(struct ceph_osd_client *osdc,
 411                                struct ceph_vino vino,
 412                                struct ceph_file_layout *layout,
 413                                u64 off, u64 *plen,
 414                                u32 truncate_seq, u64 truncate_size,
 415                                struct page **pages, int nr_pages,
 416                                int page_align);
 417
 418 extern int ceph_osdc_writepages(struct ceph_osd_client *osdc,
 419                                 struct ceph_vino vino,
 420                                 struct ceph_file_layout *layout,
 421                                 struct ceph_snap_context *sc,
 422                                 u64 off, u64 len,
 423                                 u32 truncate_seq, u64 truncate_size,
 424                                 struct timespec *mtime,
 425                                 struct page **pages, int nr_pages);
 426
 427 /* watch/notify */
 428 struct ceph_osd_linger_request *
 429 ceph_osdc_watch(struct ceph_osd_client *osdc,
 430                 struct ceph_object_id *oid,
 431                 struct ceph_object_locator *oloc,
 432                 rados_watchcb2_t wcb,
 433                 rados_watcherrcb_t errcb,
 434                 void *data);
 435 int ceph_osdc_unwatch(struct ceph_osd_client *osdc,
 436                       struct ceph_osd_linger_request *lreq);
 437
 438 int ceph_osdc_notify_ack(struct ceph_osd_client *osdc,
 439                          struct ceph_object_id *oid,
 440                          struct ceph_object_locator *oloc,
 441                          u64 notify_id,
 442                          u64 cookie,
 443                          void *payload,
 444                          size_t payload_len);
 445 int ceph_osdc_notify(struct ceph_osd_client *osdc,
 446                      struct ceph_object_id *oid,
 447                      struct ceph_object_locator *oloc,
 448                      void *payload,
 449                      size_t payload_len,
 450                      u32 timeout,
 451                      struct page ***preply_pages,
 452                      size_t *preply_len);
 453 int ceph_osdc_watch_check(struct ceph_osd_client *osdc,
 454                           struct ceph_osd_linger_request *lreq);
 455 int ceph_osdc_list_watchers(struct ceph_osd_client *osdc,
 456                             struct ceph_object_id *oid,
 457                             struct ceph_object_locator *oloc,
 458                             struct ceph_watch_item **watchers,
 459                             u32 *num_watchers);
 460 #endif
 461