include/linux/ceph/osd_client.h

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 #ifndef _FS_CEPH_OSD_CLIENT_H
   3 #define _FS_CEPH_OSD_CLIENT_H
   4
   5 #include <linux/bitrev.h>
   6 #include <linux/completion.h>
   7 #include <linux/kref.h>
   8 #include <linux/mempool.h>
   9 #include <linux/rbtree.h>
  10 #include <linux/refcount.h>
  11
  12 #include <linux/ceph/types.h>
  13 #include <linux/ceph/osdmap.h>
  14 #include <linux/ceph/messenger.h>
  15 #include <linux/ceph/msgpool.h>
  16 #include <linux/ceph/auth.h>
  17 #include <linux/ceph/pagelist.h>
  18
  19 struct ceph_msg;
  20 struct ceph_snap_context;
  21 struct ceph_osd_request;
  22 struct ceph_osd_client;
  23
  24 /*
  25  * completion callback for async writepages
  26  */
  27 typedef void (*ceph_osdc_callback_t)(struct ceph_osd_request *);
  28
  29 #define CEPH_HOMELESS_OSD       -1
  30
  31 /* a given osd we're communicating with */
  32 struct ceph_osd {
  33         refcount_t o_ref;
  34         struct ceph_osd_client *o_osdc;
  35         int o_osd;
  36         int o_incarnation;
  37         struct rb_node o_node;
  38         struct ceph_connection o_con;
  39         struct rb_root o_requests;
  40         struct rb_root o_linger_requests;
  41         struct rb_root o_backoff_mappings;
  42         struct rb_root o_backoffs_by_id;
  43         struct list_head o_osd_lru;
  44         struct ceph_auth_handshake o_auth;
  45         unsigned long lru_ttl;
  46         struct list_head o_keepalive_item;
  47         struct mutex lock;
  48 };
  49
  50 #define CEPH_OSD_SLAB_OPS       2
  51 #define CEPH_OSD_MAX_OPS        16
  52
  53 enum ceph_osd_data_type {
  54         CEPH_OSD_DATA_TYPE_NONE = 0,
  55         CEPH_OSD_DATA_TYPE_PAGES,
  56         CEPH_OSD_DATA_TYPE_PAGELIST,
  57 #ifdef CONFIG_BLOCK
  58         CEPH_OSD_DATA_TYPE_BIO,
  59 #endif /* CONFIG_BLOCK */
  60 };
  61
  62 struct ceph_osd_data {
  63         enum ceph_osd_data_type type;
  64         union {
  65                 struct {
  66                         struct page     **pages;
  67                         u64             length;
  68                         u32             alignment;
  69                         bool            pages_from_pool;
  70                         bool            own_pages;
  71                 };
  72                 struct ceph_pagelist    *pagelist;
  73 #ifdef CONFIG_BLOCK
  74                 struct {
  75                         struct bio      *bio;           /* list of bios */
  76                         size_t          bio_length;     /* total in list */
  77                 };
  78 #endif /* CONFIG_BLOCK */
  79         };
  80 };
  81
  82 struct ceph_osd_req_op {
  83         u16 op;           /* CEPH_OSD_OP_* */
  84         u32 flags;        /* CEPH_OSD_OP_FLAG_* */
  85         u32 indata_len;   /* request */
  86         u32 outdata_len;  /* reply */
  87         s32 rval;
  88
  89         union {
  90                 struct ceph_osd_data raw_data_in;
  91                 struct {
  92                         u64 offset, length;
  93                         u64 truncate_size;
  94                         u32 truncate_seq;
  95                         struct ceph_osd_data osd_data;
  96                 } extent;
  97                 struct {
  98                         u32 name_len;
  99                         u32 value_len;
 100                         __u8 cmp_op;       /* CEPH_OSD_CMPXATTR_OP_* */
 101                         __u8 cmp_mode;     /* CEPH_OSD_CMPXATTR_MODE_* */
 102                         struct ceph_osd_data osd_data;
 103                 } xattr;
 104                 struct {
 105                         const char *class_name;
 106                         const char *method_name;
 107                         struct ceph_osd_data request_info;
 108                         struct ceph_osd_data request_data;
 109                         struct ceph_osd_data response_data;
 110                         __u8 class_len;
 111                         __u8 method_len;
 112                         u32 indata_len;
 113                 } cls;
 114                 struct {
 115                         u64 cookie;
 116                         __u8 op;           /* CEPH_OSD_WATCH_OP_ */
 117                         u32 gen;
 118                 } watch;
 119                 struct {
 120                         struct ceph_osd_data request_data;
 121                 } notify_ack;
 122                 struct {
 123                         u64 cookie;
 124                         struct ceph_osd_data request_data;
 125                         struct ceph_osd_data response_data;
 126                 } notify;
 127                 struct {
 128                         struct ceph_osd_data response_data;
 129                 } list_watchers;
 130                 struct {
 131                         u64 expected_object_size;
 132                         u64 expected_write_size;
 133                 } alloc_hint;
 134         };
 135 };
 136
 137 struct ceph_osd_request_target {
 138         struct ceph_object_id base_oid;
 139         struct ceph_object_locator base_oloc;
 140         struct ceph_object_id target_oid;
 141         struct ceph_object_locator target_oloc;
 142
 143         struct ceph_pg pgid;               /* last raw pg we mapped to */
 144         struct ceph_spg spgid;             /* last actual spg we mapped to */
 145         u32 pg_num;
 146         u32 pg_num_mask;
 147         struct ceph_osds acting;
 148         struct ceph_osds up;
 149         int size;
 150         int min_size;
 151         bool sort_bitwise;
 152         bool recovery_deletes;
 153
 154         unsigned int flags;                /* CEPH_OSD_FLAG_* */
 155         bool paused;
 156
 157         u32 epoch;
 158         u32 last_force_resend;
 159
 160         int osd;
 161 };
 162
 163 /* an in-flight request */
 164 struct ceph_osd_request {
 165         u64             r_tid;              /* unique for this client */
 166         struct rb_node  r_node;
 167         struct rb_node  r_mc_node;          /* map check */
 168         struct ceph_osd *r_osd;
 169
 170         struct ceph_osd_request_target r_t;
 171 #define r_base_oid      r_t.base_oid
 172 #define r_base_oloc     r_t.base_oloc
 173 #define r_flags         r_t.flags
 174
 175         struct ceph_msg  *r_request, *r_reply;
 176         u32               r_sent;      /* >0 if r_request is sending/sent */
 177
 178         /* request osd ops array  */
 179         unsigned int            r_num_ops;
 180
 181         int               r_result;
 182
 183         struct ceph_osd_client *r_osdc;
 184         struct kref       r_kref;
 185         bool              r_mempool;
 186         struct completion r_completion;       /* private to osd_client.c */
 187         ceph_osdc_callback_t r_callback;
 188         struct list_head  r_unsafe_item;
 189
 190         struct inode *r_inode;                /* for use by callbacks */
 191         void *r_priv;                         /* ditto */
 192
 193         /* set by submitter */
 194         u64 r_snapid;                         /* for reads, CEPH_NOSNAP o/w */
 195         struct ceph_snap_context *r_snapc;    /* for writes */
 196         struct timespec r_mtime;              /* ditto */
 197         u64 r_data_offset;                    /* ditto */
 198         bool r_linger;                        /* don't resend on failure */
 199         bool r_abort_on_full;                 /* return ENOSPC when full */
 200
 201         /* internal */
 202         unsigned long r_stamp;                /* jiffies, send or check time */
 203         unsigned long r_start_stamp;          /* jiffies */
 204         int r_attempts;
 205         u32 r_map_dne_bound;
 206
 207         struct ceph_osd_req_op r_ops[];
 208 };
 209
 210 struct ceph_request_redirect {
 211         struct ceph_object_locator oloc;
 212 };
 213
 214 /*
 215  * osd request identifier
 216  *
 217  * caller name + incarnation# + tid to unique identify this request
 218  */
 219 struct ceph_osd_reqid {
 220         struct ceph_entity_name name;
 221         __le64 tid;
 222         __le32 inc;
 223 } __packed;
 224
 225 struct ceph_blkin_trace_info {
 226         __le64 trace_id;
 227         __le64 span_id;
 228         __le64 parent_span_id;
 229 } __packed;
 230
 231 typedef void (*rados_watchcb2_t)(void *arg, u64 notify_id, u64 cookie,
 232                                  u64 notifier_id, void *data, size_t data_len);
 233 typedef void (*rados_watcherrcb_t)(void *arg, u64 cookie, int err);
 234
 235 struct ceph_osd_linger_request {
 236         struct ceph_osd_client *osdc;
 237         u64 linger_id;
 238         bool committed;
 239         bool is_watch;                  /* watch or notify */
 240
 241         struct ceph_osd *osd;
 242         struct ceph_osd_request *reg_req;
 243         struct ceph_osd_request *ping_req;
 244         unsigned long ping_sent;
 245         unsigned long watch_valid_thru;
 246         struct list_head pending_lworks;
 247
 248         struct ceph_osd_request_target t;
 249         u32 map_dne_bound;
 250
 251         struct timespec mtime;
 252
 253         struct kref kref;
 254         struct mutex lock;
 255         struct rb_node node;            /* osd */
 256         struct rb_node osdc_node;       /* osdc */
 257         struct rb_node mc_node;         /* map check */
 258         struct list_head scan_item;
 259
 260         struct completion reg_commit_wait;
 261         struct completion notify_finish_wait;
 262         int reg_commit_error;
 263         int notify_finish_error;
 264         int last_error;
 265
 266         u32 register_gen;
 267         u64 notify_id;
 268
 269         rados_watchcb2_t wcb;
 270         rados_watcherrcb_t errcb;
 271         void *data;
 272
 273         struct page ***preply_pages;
 274         size_t *preply_len;
 275 };
 276
 277 struct ceph_watch_item {
 278         struct ceph_entity_name name;
 279         u64 cookie;
 280         struct ceph_entity_addr addr;
 281 };
 282
 283 struct ceph_spg_mapping {
 284         struct rb_node node;
 285         struct ceph_spg spgid;
 286
 287         struct rb_root backoffs;
 288 };
 289
 290 struct ceph_hobject_id {
 291         void *key;
 292         size_t key_len;
 293         void *oid;
 294         size_t oid_len;
 295         u64 snapid;
 296         u32 hash;
 297         u8 is_max;
 298         void *nspace;
 299         size_t nspace_len;
 300         s64 pool;
 301
 302         /* cache */
 303         u32 hash_reverse_bits;
 304 };
 305
 306 static inline void ceph_hoid_build_hash_cache(struct ceph_hobject_id *hoid)
 307 {
 308         hoid->hash_reverse_bits = bitrev32(hoid->hash);
 309 }
 310
 311 /*
 312  * PG-wide backoff: [begin, end)
 313  * per-object backoff: begin == end
 314  */
 315 struct ceph_osd_backoff {
 316         struct rb_node spg_node;
 317         struct rb_node id_node;
 318
 319         struct ceph_spg spgid;
 320         u64 id;
 321         struct ceph_hobject_id *begin;
 322         struct ceph_hobject_id *end;
 323 };
 324
 325 #define CEPH_LINGER_ID_START    0xffff000000000000ULL
 326
 327 struct ceph_osd_client {
 328         struct ceph_client     *client;
 329
 330         struct ceph_osdmap     *osdmap;       /* current map */
 331         struct rw_semaphore    lock;
 332
 333         struct rb_root         osds;          /* osds */
 334         struct list_head       osd_lru;       /* idle osds */
 335         spinlock_t             osd_lru_lock;
 336         u32                    epoch_barrier;
 337         struct ceph_osd        homeless_osd;
 338         atomic64_t             last_tid;      /* tid of last request */
 339         u64                    last_linger_id;
 340         struct rb_root         linger_requests; /* lingering requests */
 341         struct rb_root         map_checks;
 342         struct rb_root         linger_map_checks;
 343         atomic_t               num_requests;
 344         atomic_t               num_homeless;
 345         struct delayed_work    timeout_work;
 346         struct delayed_work    osds_timeout_work;
 347 #ifdef CONFIG_DEBUG_FS
 348         struct dentry          *debugfs_file;
 349 #endif
 350
 351         mempool_t              *req_mempool;
 352
 353         struct ceph_msgpool     msgpool_op;
 354         struct ceph_msgpool     msgpool_op_reply;
 355
 356         struct workqueue_struct *notify_wq;
 357 };
 358
 359 static inline bool ceph_osdmap_flag(struct ceph_osd_client *osdc, int flag)
 360 {
 361         return osdc->osdmap->flags & flag;
 362 }
 363
 364 extern int ceph_osdc_setup(void);
 365 extern void ceph_osdc_cleanup(void);
 366
 367 extern int ceph_osdc_init(struct ceph_osd_client *osdc,
 368                           struct ceph_client *client);
 369 extern void ceph_osdc_stop(struct ceph_osd_client *osdc);
 370
 371 extern void ceph_osdc_handle_reply(struct ceph_osd_client *osdc,
 372                                    struct ceph_msg *msg);
 373 extern void ceph_osdc_handle_map(struct ceph_osd_client *osdc,
 374                                  struct ceph_msg *msg);
 375 void ceph_osdc_update_epoch_barrier(struct ceph_osd_client *osdc, u32 eb);
 376
 377 extern void osd_req_op_init(struct ceph_osd_request *osd_req,
 378                             unsigned int which, u16 opcode, u32 flags);
 379
 380 extern void osd_req_op_raw_data_in_pages(struct ceph_osd_request *,
 381                                         unsigned int which,
 382                                         struct page **pages, u64 length,
 383                                         u32 alignment, bool pages_from_pool,
 384                                         bool own_pages);
 385
 386 extern void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
 387                                         unsigned int which, u16 opcode,
 388                                         u64 offset, u64 length,
 389                                         u64 truncate_size, u32 truncate_seq);
 390 extern void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
 391                                         unsigned int which, u64 length);
 392 extern void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req,
 393                                        unsigned int which, u64 offset_inc);
 394
 395 extern struct ceph_osd_data *osd_req_op_extent_osd_data(
 396                                         struct ceph_osd_request *osd_req,
 397                                         unsigned int which);
 398
 399 extern void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *,
 400                                         unsigned int which,
 401                                         struct page **pages, u64 length,
 402                                         u32 alignment, bool pages_from_pool,
 403                                         bool own_pages);
 404 extern void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *,
 405                                         unsigned int which,
 406                                         struct ceph_pagelist *pagelist);
 407 #ifdef CONFIG_BLOCK
 408 extern void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *,
 409                                         unsigned int which,
 410                                         struct bio *bio, size_t bio_length);
 411 #endif /* CONFIG_BLOCK */
 412
 413 extern void osd_req_op_cls_request_data_pagelist(struct ceph_osd_request *,
 414                                         unsigned int which,
 415                                         struct ceph_pagelist *pagelist);
 416 extern void osd_req_op_cls_request_data_pages(struct ceph_osd_request *,
 417                                         unsigned int which,
 418                                         struct page **pages, u64 length,
 419                                         u32 alignment, bool pages_from_pool,
 420                                         bool own_pages);
 421 extern void osd_req_op_cls_response_data_pages(struct ceph_osd_request *,
 422                                         unsigned int which,
 423                                         struct page **pages, u64 length,
 424                                         u32 alignment, bool pages_from_pool,
 425                                         bool own_pages);
 426 extern void osd_req_op_cls_init(struct ceph_osd_request *osd_req,
 427                                         unsigned int which, u16 opcode,
 428                                         const char *class, const char *method);
 429 extern int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
 430                                  u16 opcode, const char *name, const void *value,
 431                                  size_t size, u8 cmp_op, u8 cmp_mode);
 432 extern void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
 433                                        unsigned int which,
 434                                        u64 expected_object_size,
 435                                        u64 expected_write_size);
 436
 437 extern struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
 438                                                struct ceph_snap_context *snapc,
 439                                                unsigned int num_ops,
 440                                                bool use_mempool,
 441                                                gfp_t gfp_flags);
 442 int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp);
 443
 444 extern struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *,
 445                                       struct ceph_file_layout *layout,
 446                                       struct ceph_vino vino,
 447                                       u64 offset, u64 *len,
 448                                       unsigned int which, int num_ops,
 449                                       int opcode, int flags,
 450                                       struct ceph_snap_context *snapc,
 451                                       u32 truncate_seq, u64 truncate_size,
 452                                       bool use_mempool);
 453
 454 extern void ceph_osdc_get_request(struct ceph_osd_request *req);
 455 extern void ceph_osdc_put_request(struct ceph_osd_request *req);
 456
 457 extern int ceph_osdc_start_request(struct ceph_osd_client *osdc,
 458                                    struct ceph_osd_request *req,
 459                                    bool nofail);
 460 extern void ceph_osdc_cancel_request(struct ceph_osd_request *req);
 461 extern int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
 462                                   struct ceph_osd_request *req);
 463 extern void ceph_osdc_sync(struct ceph_osd_client *osdc);
 464
 465 extern void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc);
 466 void ceph_osdc_maybe_request_map(struct ceph_osd_client *osdc);
 467
 468 int ceph_osdc_call(struct ceph_osd_client *osdc,
 469                    struct ceph_object_id *oid,
 470                    struct ceph_object_locator *oloc,
 471                    const char *class, const char *method,
 472                    unsigned int flags,
 473                    struct page *req_page, size_t req_len,
 474                    struct page *resp_page, size_t *resp_len);
 475
 476 extern int ceph_osdc_readpages(struct ceph_osd_client *osdc,
 477                                struct ceph_vino vino,
 478                                struct ceph_file_layout *layout,
 479                                u64 off, u64 *plen,
 480                                u32 truncate_seq, u64 truncate_size,
 481                                struct page **pages, int nr_pages,
 482                                int page_align);
 483
 484 extern int ceph_osdc_writepages(struct ceph_osd_client *osdc,
 485                                 struct ceph_vino vino,
 486                                 struct ceph_file_layout *layout,
 487                                 struct ceph_snap_context *sc,
 488                                 u64 off, u64 len,
 489                                 u32 truncate_seq, u64 truncate_size,
 490                                 struct timespec *mtime,
 491                                 struct page **pages, int nr_pages);
 492
 493 /* watch/notify */
 494 struct ceph_osd_linger_request *
 495 ceph_osdc_watch(struct ceph_osd_client *osdc,
 496                 struct ceph_object_id *oid,
 497                 struct ceph_object_locator *oloc,
 498                 rados_watchcb2_t wcb,
 499                 rados_watcherrcb_t errcb,
 500                 void *data);
 501 int ceph_osdc_unwatch(struct ceph_osd_client *osdc,
 502                       struct ceph_osd_linger_request *lreq);
 503
 504 int ceph_osdc_notify_ack(struct ceph_osd_client *osdc,
 505                          struct ceph_object_id *oid,
 506                          struct ceph_object_locator *oloc,
 507                          u64 notify_id,
 508                          u64 cookie,
 509                          void *payload,
 510                          size_t payload_len);
 511 int ceph_osdc_notify(struct ceph_osd_client *osdc,
 512                      struct ceph_object_id *oid,
 513                      struct ceph_object_locator *oloc,
 514                      void *payload,
 515                      size_t payload_len,
 516                      u32 timeout,
 517                      struct page ***preply_pages,
 518                      size_t *preply_len);
 519 int ceph_osdc_watch_check(struct ceph_osd_client *osdc,
 520                           struct ceph_osd_linger_request *lreq);
 521 int ceph_osdc_list_watchers(struct ceph_osd_client *osdc,
 522                             struct ceph_object_id *oid,
 523                             struct ceph_object_locator *oloc,
 524                             struct ceph_watch_item **watchers,
 525                             u32 *num_watchers);
 526 #endif
 527