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/auth.h>
  13 #include <linux/ceph/pagelist.h>
  14
  15 struct ceph_msg;
  16 struct ceph_snap_context;
  17 struct ceph_osd_request;
  18 struct ceph_osd_client;
  19
  20 /*
  21  * completion callback for async writepages
  22  */
  23 typedef void (*ceph_osdc_callback_t)(struct ceph_osd_request *);
  24 typedef void (*ceph_osdc_unsafe_callback_t)(struct ceph_osd_request *, bool);
  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                         u64 expected_object_size;
 124                         u64 expected_write_size;
 125                 } alloc_hint;
 126         };
 127 };
 128
 129 struct ceph_osd_request_target {
 130         struct ceph_object_id base_oid;
 131         struct ceph_object_locator base_oloc;
 132         struct ceph_object_id target_oid;
 133         struct ceph_object_locator target_oloc;
 134
 135         struct ceph_pg pgid;
 136         u32 pg_num;
 137         u32 pg_num_mask;
 138         struct ceph_osds acting;
 139         struct ceph_osds up;
 140         int size;
 141         int min_size;
 142         bool sort_bitwise;
 143
 144         unsigned int flags;                /* CEPH_OSD_FLAG_* */
 145         bool paused;
 146
 147         int osd;
 148 };
 149
 150 /* an in-flight request */
 151 struct ceph_osd_request {
 152         u64             r_tid;              /* unique for this client */
 153         struct rb_node  r_node;
 154         struct ceph_osd *r_osd;
 155
 156         struct ceph_osd_request_target r_t;
 157 #define r_base_oid      r_t.base_oid
 158 #define r_base_oloc     r_t.base_oloc
 159 #define r_flags         r_t.flags
 160
 161         struct ceph_msg  *r_request, *r_reply;
 162         u32               r_sent;      /* >0 if r_request is sending/sent */
 163
 164         /* request osd ops array  */
 165         unsigned int            r_num_ops;
 166
 167         int               r_result;
 168         bool              r_got_reply;
 169
 170         struct ceph_osd_client *r_osdc;
 171         struct kref       r_kref;
 172         bool              r_mempool;
 173         struct completion r_completion;
 174         struct completion r_safe_completion;  /* fsync waiter */
 175         ceph_osdc_callback_t r_callback;
 176         ceph_osdc_unsafe_callback_t r_unsafe_callback;
 177         struct list_head  r_unsafe_item;
 178
 179         struct inode *r_inode;                /* for use by callbacks */
 180         void *r_priv;                         /* ditto */
 181
 182         /* set by submitter */
 183         u64 r_snapid;                         /* for reads, CEPH_NOSNAP o/w */
 184         struct ceph_snap_context *r_snapc;    /* for writes */
 185         struct timespec r_mtime;              /* ditto */
 186         u64 r_data_offset;                    /* ditto */
 187         bool r_linger;                        /* don't resend on failure */
 188
 189         /* internal */
 190         unsigned long r_stamp;                /* jiffies, send or check time */
 191         int r_attempts;
 192         struct ceph_eversion r_replay_version; /* aka reassert_version */
 193         u32 r_last_force_resend;
 194
 195         struct ceph_osd_req_op r_ops[];
 196 };
 197
 198 struct ceph_request_redirect {
 199         struct ceph_object_locator oloc;
 200 };
 201
 202 typedef void (*rados_watchcb2_t)(void *arg, u64 notify_id, u64 cookie,
 203                                  u64 notifier_id, void *data, size_t data_len);
 204 typedef void (*rados_watcherrcb_t)(void *arg, u64 cookie, int err);
 205
 206 struct ceph_osd_linger_request {
 207         struct ceph_osd_client *osdc;
 208         u64 linger_id;
 209         bool committed;
 210         bool is_watch;                  /* watch or notify */
 211
 212         struct ceph_osd *osd;
 213         struct ceph_osd_request *reg_req;
 214         struct ceph_osd_request *ping_req;
 215         unsigned long ping_sent;
 216
 217         struct ceph_osd_request_target t;
 218         u32 last_force_resend;
 219
 220         struct timespec mtime;
 221
 222         struct kref kref;
 223         struct mutex lock;
 224         struct rb_node node;            /* osd */
 225         struct rb_node osdc_node;       /* osdc */
 226         struct list_head scan_item;
 227
 228         struct completion reg_commit_wait;
 229         struct completion notify_finish_wait;
 230         int reg_commit_error;
 231         int notify_finish_error;
 232         int last_error;
 233
 234         u32 register_gen;
 235         u64 notify_id;
 236
 237         rados_watchcb2_t wcb;
 238         rados_watcherrcb_t errcb;
 239         void *data;
 240
 241         struct page ***preply_pages;
 242         size_t *preply_len;
 243 };
 244
 245 struct ceph_osd_client {
 246         struct ceph_client     *client;
 247
 248         struct ceph_osdmap     *osdmap;       /* current map */
 249         struct rw_semaphore    lock;
 250
 251         struct rb_root         osds;          /* osds */
 252         struct list_head       osd_lru;       /* idle osds */
 253         spinlock_t             osd_lru_lock;
 254         struct ceph_osd        homeless_osd;
 255         atomic64_t             last_tid;      /* tid of last request */
 256         u64                    last_linger_id;
 257         struct rb_root         linger_requests; /* lingering requests */
 258         atomic_t               num_requests;
 259         atomic_t               num_homeless;
 260         struct delayed_work    timeout_work;
 261         struct delayed_work    osds_timeout_work;
 262 #ifdef CONFIG_DEBUG_FS
 263         struct dentry          *debugfs_file;
 264 #endif
 265
 266         mempool_t              *req_mempool;
 267
 268         struct ceph_msgpool     msgpool_op;
 269         struct ceph_msgpool     msgpool_op_reply;
 270
 271         struct workqueue_struct *notify_wq;
 272 };
 273
 274 extern int ceph_osdc_setup(void);
 275 extern void ceph_osdc_cleanup(void);
 276
 277 extern int ceph_osdc_init(struct ceph_osd_client *osdc,
 278                           struct ceph_client *client);
 279 extern void ceph_osdc_stop(struct ceph_osd_client *osdc);
 280
 281 extern void ceph_osdc_handle_reply(struct ceph_osd_client *osdc,
 282                                    struct ceph_msg *msg);
 283 extern void ceph_osdc_handle_map(struct ceph_osd_client *osdc,
 284                                  struct ceph_msg *msg);
 285
 286 extern void osd_req_op_init(struct ceph_osd_request *osd_req,
 287                             unsigned int which, u16 opcode, u32 flags);
 288
 289 extern void osd_req_op_raw_data_in_pages(struct ceph_osd_request *,
 290                                         unsigned int which,
 291                                         struct page **pages, u64 length,
 292                                         u32 alignment, bool pages_from_pool,
 293                                         bool own_pages);
 294
 295 extern void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
 296                                         unsigned int which, u16 opcode,
 297                                         u64 offset, u64 length,
 298                                         u64 truncate_size, u32 truncate_seq);
 299 extern void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
 300                                         unsigned int which, u64 length);
 301 extern void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req,
 302                                        unsigned int which, u64 offset_inc);
 303
 304 extern struct ceph_osd_data *osd_req_op_extent_osd_data(
 305                                         struct ceph_osd_request *osd_req,
 306                                         unsigned int which);
 307
 308 extern void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *,
 309                                         unsigned int which,
 310                                         struct page **pages, u64 length,
 311                                         u32 alignment, bool pages_from_pool,
 312                                         bool own_pages);
 313 extern void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *,
 314                                         unsigned int which,
 315                                         struct ceph_pagelist *pagelist);
 316 #ifdef CONFIG_BLOCK
 317 extern void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *,
 318                                         unsigned int which,
 319                                         struct bio *bio, size_t bio_length);
 320 #endif /* CONFIG_BLOCK */
 321
 322 extern void osd_req_op_cls_request_data_pagelist(struct ceph_osd_request *,
 323                                         unsigned int which,
 324                                         struct ceph_pagelist *pagelist);
 325 extern void osd_req_op_cls_request_data_pages(struct ceph_osd_request *,
 326                                         unsigned int which,
 327                                         struct page **pages, u64 length,
 328                                         u32 alignment, bool pages_from_pool,
 329                                         bool own_pages);
 330 extern void osd_req_op_cls_response_data_pages(struct ceph_osd_request *,
 331                                         unsigned int which,
 332                                         struct page **pages, u64 length,
 333                                         u32 alignment, bool pages_from_pool,
 334                                         bool own_pages);
 335
 336 extern void osd_req_op_cls_init(struct ceph_osd_request *osd_req,
 337                                         unsigned int which, u16 opcode,
 338                                         const char *class, const char *method);
 339 extern int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
 340                                  u16 opcode, const char *name, const void *value,
 341                                  size_t size, u8 cmp_op, u8 cmp_mode);
 342 extern void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
 343                                        unsigned int which,
 344                                        u64 expected_object_size,
 345                                        u64 expected_write_size);
 346
 347 extern struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
 348                                                struct ceph_snap_context *snapc,
 349                                                unsigned int num_ops,
 350                                                bool use_mempool,
 351                                                gfp_t gfp_flags);
 352 int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp);
 353
 354 extern struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *,
 355                                       struct ceph_file_layout *layout,
 356                                       struct ceph_vino vino,
 357                                       u64 offset, u64 *len,
 358                                       unsigned int which, int num_ops,
 359                                       int opcode, int flags,
 360                                       struct ceph_snap_context *snapc,
 361                                       u32 truncate_seq, u64 truncate_size,
 362                                       bool use_mempool);
 363
 364 extern void ceph_osdc_get_request(struct ceph_osd_request *req);
 365 extern void ceph_osdc_put_request(struct ceph_osd_request *req);
 366
 367 extern int ceph_osdc_start_request(struct ceph_osd_client *osdc,
 368                                    struct ceph_osd_request *req,
 369                                    bool nofail);
 370 extern void ceph_osdc_cancel_request(struct ceph_osd_request *req);
 371 extern int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
 372                                   struct ceph_osd_request *req);
 373 extern void ceph_osdc_sync(struct ceph_osd_client *osdc);
 374
 375 extern void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc);
 376
 377 extern int ceph_osdc_readpages(struct ceph_osd_client *osdc,
 378                                struct ceph_vino vino,
 379                                struct ceph_file_layout *layout,
 380                                u64 off, u64 *plen,
 381                                u32 truncate_seq, u64 truncate_size,
 382                                struct page **pages, int nr_pages,
 383                                int page_align);
 384
 385 extern int ceph_osdc_writepages(struct ceph_osd_client *osdc,
 386                                 struct ceph_vino vino,
 387                                 struct ceph_file_layout *layout,
 388                                 struct ceph_snap_context *sc,
 389                                 u64 off, u64 len,
 390                                 u32 truncate_seq, u64 truncate_size,
 391                                 struct timespec *mtime,
 392                                 struct page **pages, int nr_pages);
 393
 394 /* watch/notify */
 395 struct ceph_osd_linger_request *
 396 ceph_osdc_watch(struct ceph_osd_client *osdc,
 397                 struct ceph_object_id *oid,
 398                 struct ceph_object_locator *oloc,
 399                 rados_watchcb2_t wcb,
 400                 rados_watcherrcb_t errcb,
 401                 void *data);
 402 int ceph_osdc_unwatch(struct ceph_osd_client *osdc,
 403                       struct ceph_osd_linger_request *lreq);
 404
 405 int ceph_osdc_notify_ack(struct ceph_osd_client *osdc,
 406                          struct ceph_object_id *oid,
 407                          struct ceph_object_locator *oloc,
 408                          u64 notify_id,
 409                          u64 cookie,
 410                          void *payload,
 411                          size_t payload_len);
 412 int ceph_osdc_notify(struct ceph_osd_client *osdc,
 413                      struct ceph_object_id *oid,
 414                      struct ceph_object_locator *oloc,
 415                      void *payload,
 416                      size_t payload_len,
 417                      u32 timeout,
 418                      struct page ***preply_pages,
 419                      size_t *preply_len);
 420 #endif
 421