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                         u64 expected_object_size;
 125                         u64 expected_write_size;
 126                 } alloc_hint;
 127         };
 128 };
 129
 130 struct ceph_osd_request_target {
 131         struct ceph_object_id base_oid;
 132         struct ceph_object_locator base_oloc;
 133         struct ceph_object_id target_oid;
 134         struct ceph_object_locator target_oloc;
 135
 136         struct ceph_pg pgid;
 137         u32 pg_num;
 138         u32 pg_num_mask;
 139         struct ceph_osds acting;
 140         struct ceph_osds up;
 141         int size;
 142         int min_size;
 143         bool sort_bitwise;
 144
 145         unsigned int flags;                /* CEPH_OSD_FLAG_* */
 146         bool paused;
 147
 148         int osd;
 149 };
 150
 151 /* an in-flight request */
 152 struct ceph_osd_request {
 153         u64             r_tid;              /* unique for this client */
 154         struct rb_node  r_node;
 155         struct rb_node  r_mc_node;          /* map check */
 156         struct ceph_osd *r_osd;
 157
 158         struct ceph_osd_request_target r_t;
 159 #define r_base_oid      r_t.base_oid
 160 #define r_base_oloc     r_t.base_oloc
 161 #define r_flags         r_t.flags
 162
 163         struct ceph_msg  *r_request, *r_reply;
 164         u32               r_sent;      /* >0 if r_request is sending/sent */
 165
 166         /* request osd ops array  */
 167         unsigned int            r_num_ops;
 168
 169         int               r_result;
 170         bool              r_got_reply;
 171
 172         struct ceph_osd_client *r_osdc;
 173         struct kref       r_kref;
 174         bool              r_mempool;
 175         struct completion r_completion;
 176         struct completion r_safe_completion;  /* fsync waiter */
 177         ceph_osdc_callback_t r_callback;
 178         ceph_osdc_unsafe_callback_t r_unsafe_callback;
 179         struct list_head  r_unsafe_item;
 180
 181         struct inode *r_inode;                /* for use by callbacks */
 182         void *r_priv;                         /* ditto */
 183
 184         /* set by submitter */
 185         u64 r_snapid;                         /* for reads, CEPH_NOSNAP o/w */
 186         struct ceph_snap_context *r_snapc;    /* for writes */
 187         struct timespec r_mtime;              /* ditto */
 188         u64 r_data_offset;                    /* ditto */
 189         bool r_linger;                        /* don't resend on failure */
 190
 191         /* internal */
 192         unsigned long r_stamp;                /* jiffies, send or check time */
 193         int r_attempts;
 194         struct ceph_eversion r_replay_version; /* aka reassert_version */
 195         u32 r_last_force_resend;
 196         u32 r_map_dne_bound;
 197
 198         struct ceph_osd_req_op r_ops[];
 199 };
 200
 201 struct ceph_request_redirect {
 202         struct ceph_object_locator oloc;
 203 };
 204
 205 typedef void (*rados_watchcb2_t)(void *arg, u64 notify_id, u64 cookie,
 206                                  u64 notifier_id, void *data, size_t data_len);
 207 typedef void (*rados_watcherrcb_t)(void *arg, u64 cookie, int err);
 208
 209 struct ceph_osd_linger_request {
 210         struct ceph_osd_client *osdc;
 211         u64 linger_id;
 212         bool committed;
 213         bool is_watch;                  /* watch or notify */
 214
 215         struct ceph_osd *osd;
 216         struct ceph_osd_request *reg_req;
 217         struct ceph_osd_request *ping_req;
 218         unsigned long ping_sent;
 219         unsigned long watch_valid_thru;
 220         struct list_head pending_lworks;
 221
 222         struct ceph_osd_request_target t;
 223         u32 last_force_resend;
 224         u32 map_dne_bound;
 225
 226         struct timespec mtime;
 227
 228         struct kref kref;
 229         struct mutex lock;
 230         struct rb_node node;            /* osd */
 231         struct rb_node osdc_node;       /* osdc */
 232         struct rb_node mc_node;         /* map check */
 233         struct list_head scan_item;
 234
 235         struct completion reg_commit_wait;
 236         struct completion notify_finish_wait;
 237         int reg_commit_error;
 238         int notify_finish_error;
 239         int last_error;
 240
 241         u32 register_gen;
 242         u64 notify_id;
 243
 244         rados_watchcb2_t wcb;
 245         rados_watcherrcb_t errcb;
 246         void *data;
 247
 248         struct page ***preply_pages;
 249         size_t *preply_len;
 250 };
 251
 252 struct ceph_osd_client {
 253         struct ceph_client     *client;
 254
 255         struct ceph_osdmap     *osdmap;       /* current map */
 256         struct rw_semaphore    lock;
 257
 258         struct rb_root         osds;          /* osds */
 259         struct list_head       osd_lru;       /* idle osds */
 260         spinlock_t             osd_lru_lock;
 261         struct ceph_osd        homeless_osd;
 262         atomic64_t             last_tid;      /* tid of last request */
 263         u64                    last_linger_id;
 264         struct rb_root         linger_requests; /* lingering requests */
 265         struct rb_root         map_checks;
 266         struct rb_root         linger_map_checks;
 267         atomic_t               num_requests;
 268         atomic_t               num_homeless;
 269         struct delayed_work    timeout_work;
 270         struct delayed_work    osds_timeout_work;
 271 #ifdef CONFIG_DEBUG_FS
 272         struct dentry          *debugfs_file;
 273 #endif
 274
 275         mempool_t              *req_mempool;
 276
 277         struct ceph_msgpool     msgpool_op;
 278         struct ceph_msgpool     msgpool_op_reply;
 279
 280         struct workqueue_struct *notify_wq;
 281 };
 282
 283 static inline bool ceph_osdmap_flag(struct ceph_osd_client *osdc, int flag)
 284 {
 285         return osdc->osdmap->flags & flag;
 286 }
 287
 288 extern int ceph_osdc_setup(void);
 289 extern void ceph_osdc_cleanup(void);
 290
 291 extern int ceph_osdc_init(struct ceph_osd_client *osdc,
 292                           struct ceph_client *client);
 293 extern void ceph_osdc_stop(struct ceph_osd_client *osdc);
 294
 295 extern void ceph_osdc_handle_reply(struct ceph_osd_client *osdc,
 296                                    struct ceph_msg *msg);
 297 extern void ceph_osdc_handle_map(struct ceph_osd_client *osdc,
 298                                  struct ceph_msg *msg);
 299
 300 extern void osd_req_op_init(struct ceph_osd_request *osd_req,
 301                             unsigned int which, u16 opcode, u32 flags);
 302
 303 extern void osd_req_op_raw_data_in_pages(struct ceph_osd_request *,
 304                                         unsigned int which,
 305                                         struct page **pages, u64 length,
 306                                         u32 alignment, bool pages_from_pool,
 307                                         bool own_pages);
 308
 309 extern void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
 310                                         unsigned int which, u16 opcode,
 311                                         u64 offset, u64 length,
 312                                         u64 truncate_size, u32 truncate_seq);
 313 extern void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
 314                                         unsigned int which, u64 length);
 315 extern void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req,
 316                                        unsigned int which, u64 offset_inc);
 317
 318 extern struct ceph_osd_data *osd_req_op_extent_osd_data(
 319                                         struct ceph_osd_request *osd_req,
 320                                         unsigned int which);
 321
 322 extern void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *,
 323                                         unsigned int which,
 324                                         struct page **pages, u64 length,
 325                                         u32 alignment, bool pages_from_pool,
 326                                         bool own_pages);
 327 extern void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *,
 328                                         unsigned int which,
 329                                         struct ceph_pagelist *pagelist);
 330 #ifdef CONFIG_BLOCK
 331 extern void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *,
 332                                         unsigned int which,
 333                                         struct bio *bio, size_t bio_length);
 334 #endif /* CONFIG_BLOCK */
 335
 336 extern void osd_req_op_cls_request_data_pagelist(struct ceph_osd_request *,
 337                                         unsigned int which,
 338                                         struct ceph_pagelist *pagelist);
 339 extern void osd_req_op_cls_request_data_pages(struct ceph_osd_request *,
 340                                         unsigned int which,
 341                                         struct page **pages, u64 length,
 342                                         u32 alignment, bool pages_from_pool,
 343                                         bool own_pages);
 344 extern void osd_req_op_cls_response_data_pages(struct ceph_osd_request *,
 345                                         unsigned int which,
 346                                         struct page **pages, u64 length,
 347                                         u32 alignment, bool pages_from_pool,
 348                                         bool own_pages);
 349
 350 extern void osd_req_op_cls_init(struct ceph_osd_request *osd_req,
 351                                         unsigned int which, u16 opcode,
 352                                         const char *class, const char *method);
 353 extern int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
 354                                  u16 opcode, const char *name, const void *value,
 355                                  size_t size, u8 cmp_op, u8 cmp_mode);
 356 extern void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
 357                                        unsigned int which,
 358                                        u64 expected_object_size,
 359                                        u64 expected_write_size);
 360
 361 extern struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
 362                                                struct ceph_snap_context *snapc,
 363                                                unsigned int num_ops,
 364                                                bool use_mempool,
 365                                                gfp_t gfp_flags);
 366 int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp);
 367
 368 extern struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *,
 369                                       struct ceph_file_layout *layout,
 370                                       struct ceph_vino vino,
 371                                       u64 offset, u64 *len,
 372                                       unsigned int which, int num_ops,
 373                                       int opcode, int flags,
 374                                       struct ceph_snap_context *snapc,
 375                                       u32 truncate_seq, u64 truncate_size,
 376                                       bool use_mempool);
 377
 378 extern void ceph_osdc_get_request(struct ceph_osd_request *req);
 379 extern void ceph_osdc_put_request(struct ceph_osd_request *req);
 380
 381 extern int ceph_osdc_start_request(struct ceph_osd_client *osdc,
 382                                    struct ceph_osd_request *req,
 383                                    bool nofail);
 384 extern void ceph_osdc_cancel_request(struct ceph_osd_request *req);
 385 extern int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
 386                                   struct ceph_osd_request *req);
 387 extern void ceph_osdc_sync(struct ceph_osd_client *osdc);
 388
 389 extern void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc);
 390 void ceph_osdc_maybe_request_map(struct ceph_osd_client *osdc);
 391
 392 extern int ceph_osdc_readpages(struct ceph_osd_client *osdc,
 393                                struct ceph_vino vino,
 394                                struct ceph_file_layout *layout,
 395                                u64 off, u64 *plen,
 396                                u32 truncate_seq, u64 truncate_size,
 397                                struct page **pages, int nr_pages,
 398                                int page_align);
 399
 400 extern int ceph_osdc_writepages(struct ceph_osd_client *osdc,
 401                                 struct ceph_vino vino,
 402                                 struct ceph_file_layout *layout,
 403                                 struct ceph_snap_context *sc,
 404                                 u64 off, u64 len,
 405                                 u32 truncate_seq, u64 truncate_size,
 406                                 struct timespec *mtime,
 407                                 struct page **pages, int nr_pages);
 408
 409 /* watch/notify */
 410 struct ceph_osd_linger_request *
 411 ceph_osdc_watch(struct ceph_osd_client *osdc,
 412                 struct ceph_object_id *oid,
 413                 struct ceph_object_locator *oloc,
 414                 rados_watchcb2_t wcb,
 415                 rados_watcherrcb_t errcb,
 416                 void *data);
 417 int ceph_osdc_unwatch(struct ceph_osd_client *osdc,
 418                       struct ceph_osd_linger_request *lreq);
 419
 420 int ceph_osdc_notify_ack(struct ceph_osd_client *osdc,
 421                          struct ceph_object_id *oid,
 422                          struct ceph_object_locator *oloc,
 423                          u64 notify_id,
 424                          u64 cookie,
 425                          void *payload,
 426                          size_t payload_len);
 427 int ceph_osdc_notify(struct ceph_osd_client *osdc,
 428                      struct ceph_object_id *oid,
 429                      struct ceph_object_locator *oloc,
 430                      void *payload,
 431                      size_t payload_len,
 432                      u32 timeout,
 433                      struct page ***preply_pages,
 434                      size_t *preply_len);
 435 int ceph_osdc_watch_check(struct ceph_osd_client *osdc,
 436                           struct ceph_osd_linger_request *lreq);
 437 #endif
 438