linux/ceph/osd_client.h

0001 /* SPDX-License-Identifier: GPL-2.0 */
0002 #ifndef _FS_CEPH_OSD_CLIENT_H
0003 #define _FS_CEPH_OSD_CLIENT_H
0004
0005 #include <linux/bitrev.h>
0006 #include <linux/completion.h>
0007 #include <linux/kref.h>
0008 #include <linux/mempool.h>
0009 #include <linux/rbtree.h>
0010 #include <linux/refcount.h>
0011 #include <linux/ktime.h>
0012
0013 #include <linux/ceph/types.h>
0014 #include <linux/ceph/osdmap.h>
0015 #include <linux/ceph/messenger.h>
0016 #include <linux/ceph/msgpool.h>
0017 #include <linux/ceph/auth.h>
0018 #include <linux/ceph/pagelist.h>
0019
0020 struct ceph_msg;
0021 struct ceph_snap_context;
0022 struct ceph_osd_request;
0023 struct ceph_osd_client;
0024
0025 /*
0026  * completion callback for async writepages
0027  */
0028 typedef void (*ceph_osdc_callback_t)(struct ceph_osd_request *);
0029
0030 #define CEPH_HOMELESS_OSD   -1
0031
0032 /* a given osd we're communicating with */
0033 struct ceph_osd {
0034     refcount_t o_ref;
0035     struct ceph_osd_client *o_osdc;
0036     int o_osd;
0037     int o_incarnation;
0038     struct rb_node o_node;
0039     struct ceph_connection o_con;
0040     struct rb_root o_requests;
0041     struct rb_root o_linger_requests;
0042     struct rb_root o_backoff_mappings;
0043     struct rb_root o_backoffs_by_id;
0044     struct list_head o_osd_lru;
0045     struct ceph_auth_handshake o_auth;
0046     unsigned long lru_ttl;
0047     struct list_head o_keepalive_item;
0048     struct mutex lock;
0049 };
0050
0051 #define CEPH_OSD_SLAB_OPS   2
0052 #define CEPH_OSD_MAX_OPS    16
0053
0054 enum ceph_osd_data_type {
0055     CEPH_OSD_DATA_TYPE_NONE = 0,
0056     CEPH_OSD_DATA_TYPE_PAGES,
0057     CEPH_OSD_DATA_TYPE_PAGELIST,
0058 #ifdef CONFIG_BLOCK
0059     CEPH_OSD_DATA_TYPE_BIO,
0060 #endif /* CONFIG_BLOCK */
0061     CEPH_OSD_DATA_TYPE_BVECS,
0062 };
0063
0064 struct ceph_osd_data {
0065     enum ceph_osd_data_type type;
0066     union {
0067         struct {
0068             struct page **pages;
0069             u64     length;
0070             u32     alignment;
0071             bool        pages_from_pool;
0072             bool        own_pages;
0073         };
0074         struct ceph_pagelist    *pagelist;
0075 #ifdef CONFIG_BLOCK
0076         struct {
0077             struct ceph_bio_iter    bio_pos;
0078             u32         bio_length;
0079         };
0080 #endif /* CONFIG_BLOCK */
0081         struct {
0082             struct ceph_bvec_iter   bvec_pos;
0083             u32         num_bvecs;
0084         };
0085     };
0086 };
0087
0088 struct ceph_osd_req_op {
0089     u16 op;           /* CEPH_OSD_OP_* */
0090     u32 flags;        /* CEPH_OSD_OP_FLAG_* */
0091     u32 indata_len;   /* request */
0092     u32 outdata_len;  /* reply */
0093     s32 rval;
0094
0095     union {
0096         struct ceph_osd_data raw_data_in;
0097         struct {
0098             u64 offset, length;
0099             u64 truncate_size;
0100             u32 truncate_seq;
0101             struct ceph_osd_data osd_data;
0102         } extent;
0103         struct {
0104             u32 name_len;
0105             u32 value_len;
0106             __u8 cmp_op;       /* CEPH_OSD_CMPXATTR_OP_* */
0107             __u8 cmp_mode;     /* CEPH_OSD_CMPXATTR_MODE_* */
0108             struct ceph_osd_data osd_data;
0109         } xattr;
0110         struct {
0111             const char *class_name;
0112             const char *method_name;
0113             struct ceph_osd_data request_info;
0114             struct ceph_osd_data request_data;
0115             struct ceph_osd_data response_data;
0116             __u8 class_len;
0117             __u8 method_len;
0118             u32 indata_len;
0119         } cls;
0120         struct {
0121             u64 cookie;
0122             __u8 op;           /* CEPH_OSD_WATCH_OP_ */
0123             u32 gen;
0124         } watch;
0125         struct {
0126             struct ceph_osd_data request_data;
0127         } notify_ack;
0128         struct {
0129             u64 cookie;
0130             struct ceph_osd_data request_data;
0131             struct ceph_osd_data response_data;
0132         } notify;
0133         struct {
0134             struct ceph_osd_data response_data;
0135         } list_watchers;
0136         struct {
0137             u64 expected_object_size;
0138             u64 expected_write_size;
0139             u32 flags;  /* CEPH_OSD_OP_ALLOC_HINT_FLAG_* */
0140         } alloc_hint;
0141         struct {
0142             u64 snapid;
0143             u64 src_version;
0144             u8 flags;
0145             u32 src_fadvise_flags;
0146             struct ceph_osd_data osd_data;
0147         } copy_from;
0148     };
0149 };
0150
0151 struct ceph_osd_request_target {
0152     struct ceph_object_id base_oid;
0153     struct ceph_object_locator base_oloc;
0154     struct ceph_object_id target_oid;
0155     struct ceph_object_locator target_oloc;
0156
0157     struct ceph_pg pgid;               /* last raw pg we mapped to */
0158     struct ceph_spg spgid;             /* last actual spg we mapped to */
0159     u32 pg_num;
0160     u32 pg_num_mask;
0161     struct ceph_osds acting;
0162     struct ceph_osds up;
0163     int size;
0164     int min_size;
0165     bool sort_bitwise;
0166     bool recovery_deletes;
0167
0168     unsigned int flags;                /* CEPH_OSD_FLAG_* */
0169     bool used_replica;
0170     bool paused;
0171
0172     u32 epoch;
0173     u32 last_force_resend;
0174
0175     int osd;
0176 };
0177
0178 /* an in-flight request */
0179 struct ceph_osd_request {
0180     u64             r_tid;              /* unique for this client */
0181     struct rb_node  r_node;
0182     struct rb_node  r_mc_node;          /* map check */
0183     struct work_struct r_complete_work;
0184     struct ceph_osd *r_osd;
0185
0186     struct ceph_osd_request_target r_t;
0187 #define r_base_oid  r_t.base_oid
0188 #define r_base_oloc r_t.base_oloc
0189 #define r_flags     r_t.flags
0190
0191     struct ceph_msg  *r_request, *r_reply;
0192     u32               r_sent;      /* >0 if r_request is sending/sent */
0193
0194     /* request osd ops array  */
0195     unsigned int        r_num_ops;
0196
0197     int               r_result;
0198
0199     struct ceph_osd_client *r_osdc;
0200     struct kref       r_kref;
0201     bool              r_mempool;
0202     struct completion r_completion;       /* private to osd_client.c */
0203     ceph_osdc_callback_t r_callback;
0204
0205     struct inode *r_inode;                /* for use by callbacks */
0206     struct list_head r_private_item;      /* ditto */
0207     void *r_priv;                 /* ditto */
0208
0209     /* set by submitter */
0210     u64 r_snapid;                         /* for reads, CEPH_NOSNAP o/w */
0211     struct ceph_snap_context *r_snapc;    /* for writes */
0212     struct timespec64 r_mtime;            /* ditto */
0213     u64 r_data_offset;                    /* ditto */
0214     bool r_linger;                        /* don't resend on failure */
0215
0216     /* internal */
0217     unsigned long r_stamp;                /* jiffies, send or check time */
0218     unsigned long r_start_stamp;          /* jiffies */
0219     ktime_t r_start_latency;              /* ktime_t */
0220     ktime_t r_end_latency;                /* ktime_t */
0221     int r_attempts;
0222     u32 r_map_dne_bound;
0223
0224     struct ceph_osd_req_op r_ops[];
0225 };
0226
0227 struct ceph_request_redirect {
0228     struct ceph_object_locator oloc;
0229 };
0230
0231 /*
0232  * osd request identifier
0233  *
0234  * caller name + incarnation# + tid to unique identify this request
0235  */
0236 struct ceph_osd_reqid {
0237     struct ceph_entity_name name;
0238     __le64 tid;
0239     __le32 inc;
0240 } __packed;
0241
0242 struct ceph_blkin_trace_info {
0243     __le64 trace_id;
0244     __le64 span_id;
0245     __le64 parent_span_id;
0246 } __packed;
0247
0248 typedef void (*rados_watchcb2_t)(void *arg, u64 notify_id, u64 cookie,
0249                  u64 notifier_id, void *data, size_t data_len);
0250 typedef void (*rados_watcherrcb_t)(void *arg, u64 cookie, int err);
0251
0252 struct ceph_osd_linger_request {
0253     struct ceph_osd_client *osdc;
0254     u64 linger_id;
0255     bool committed;
0256     bool is_watch;                  /* watch or notify */
0257
0258     struct ceph_osd *osd;
0259     struct ceph_osd_request *reg_req;
0260     struct ceph_osd_request *ping_req;
0261     unsigned long ping_sent;
0262     unsigned long watch_valid_thru;
0263     struct list_head pending_lworks;
0264
0265     struct ceph_osd_request_target t;
0266     u32 map_dne_bound;
0267
0268     struct timespec64 mtime;
0269
0270     struct kref kref;
0271     struct mutex lock;
0272     struct rb_node node;            /* osd */
0273     struct rb_node osdc_node;       /* osdc */
0274     struct rb_node mc_node;         /* map check */
0275     struct list_head scan_item;
0276
0277     struct completion reg_commit_wait;
0278     struct completion notify_finish_wait;
0279     int reg_commit_error;
0280     int notify_finish_error;
0281     int last_error;
0282
0283     u32 register_gen;
0284     u64 notify_id;
0285
0286     rados_watchcb2_t wcb;
0287     rados_watcherrcb_t errcb;
0288     void *data;
0289
0290     struct ceph_pagelist *request_pl;
0291     struct page **notify_id_pages;
0292
0293     struct page ***preply_pages;
0294     size_t *preply_len;
0295 };
0296
0297 struct ceph_watch_item {
0298     struct ceph_entity_name name;
0299     u64 cookie;
0300     struct ceph_entity_addr addr;
0301 };
0302
0303 struct ceph_spg_mapping {
0304     struct rb_node node;
0305     struct ceph_spg spgid;
0306
0307     struct rb_root backoffs;
0308 };
0309
0310 struct ceph_hobject_id {
0311     void *key;
0312     size_t key_len;
0313     void *oid;
0314     size_t oid_len;
0315     u64 snapid;
0316     u32 hash;
0317     u8 is_max;
0318     void *nspace;
0319     size_t nspace_len;
0320     s64 pool;
0321
0322     /* cache */
0323     u32 hash_reverse_bits;
0324 };
0325
0326 static inline void ceph_hoid_build_hash_cache(struct ceph_hobject_id *hoid)
0327 {
0328     hoid->hash_reverse_bits = bitrev32(hoid->hash);
0329 }
0330
0331 /*
0332  * PG-wide backoff: [begin, end)
0333  * per-object backoff: begin == end
0334  */
0335 struct ceph_osd_backoff {
0336     struct rb_node spg_node;
0337     struct rb_node id_node;
0338
0339     struct ceph_spg spgid;
0340     u64 id;
0341     struct ceph_hobject_id *begin;
0342     struct ceph_hobject_id *end;
0343 };
0344
0345 #define CEPH_LINGER_ID_START    0xffff000000000000ULL
0346
0347 struct ceph_osd_client {
0348     struct ceph_client     *client;
0349
0350     struct ceph_osdmap     *osdmap;       /* current map */
0351     struct rw_semaphore    lock;
0352
0353     struct rb_root         osds;          /* osds */
0354     struct list_head       osd_lru;       /* idle osds */
0355     spinlock_t             osd_lru_lock;
0356     u32            epoch_barrier;
0357     struct ceph_osd        homeless_osd;
0358     atomic64_t             last_tid;      /* tid of last request */
0359     u64                    last_linger_id;
0360     struct rb_root         linger_requests; /* lingering requests */
0361     struct rb_root         map_checks;
0362     struct rb_root         linger_map_checks;
0363     atomic_t               num_requests;
0364     atomic_t               num_homeless;
0365     int                    abort_err;
0366     struct delayed_work    timeout_work;
0367     struct delayed_work    osds_timeout_work;
0368 #ifdef CONFIG_DEBUG_FS
0369     struct dentry          *debugfs_file;
0370 #endif
0371
0372     mempool_t              *req_mempool;
0373
0374     struct ceph_msgpool msgpool_op;
0375     struct ceph_msgpool msgpool_op_reply;
0376
0377     struct workqueue_struct *notify_wq;
0378     struct workqueue_struct *completion_wq;
0379 };
0380
0381 static inline bool ceph_osdmap_flag(struct ceph_osd_client *osdc, int flag)
0382 {
0383     return osdc->osdmap->flags & flag;
0384 }
0385
0386 extern int ceph_osdc_setup(void);
0387 extern void ceph_osdc_cleanup(void);
0388
0389 extern int ceph_osdc_init(struct ceph_osd_client *osdc,
0390               struct ceph_client *client);
0391 extern void ceph_osdc_stop(struct ceph_osd_client *osdc);
0392 extern void ceph_osdc_reopen_osds(struct ceph_osd_client *osdc);
0393
0394 extern void ceph_osdc_handle_reply(struct ceph_osd_client *osdc,
0395                    struct ceph_msg *msg);
0396 extern void ceph_osdc_handle_map(struct ceph_osd_client *osdc,
0397                  struct ceph_msg *msg);
0398 void ceph_osdc_update_epoch_barrier(struct ceph_osd_client *osdc, u32 eb);
0399 void ceph_osdc_abort_requests(struct ceph_osd_client *osdc, int err);
0400 void ceph_osdc_clear_abort_err(struct ceph_osd_client *osdc);
0401
0402 #define osd_req_op_data(oreq, whch, typ, fld)               \
0403 ({                                  \
0404     struct ceph_osd_request *__oreq = (oreq);           \
0405     unsigned int __whch = (whch);                   \
0406     BUG_ON(__whch >= __oreq->r_num_ops);                \
0407     &__oreq->r_ops[__whch].typ.fld;                 \
0408 })
0409
0410 struct ceph_osd_req_op *osd_req_op_init(struct ceph_osd_request *osd_req,
0411                 unsigned int which, u16 opcode, u32 flags);
0412
0413 extern void osd_req_op_raw_data_in_pages(struct ceph_osd_request *,
0414                     unsigned int which,
0415                     struct page **pages, u64 length,
0416                     u32 alignment, bool pages_from_pool,
0417                     bool own_pages);
0418
0419 extern void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
0420                     unsigned int which, u16 opcode,
0421                     u64 offset, u64 length,
0422                     u64 truncate_size, u32 truncate_seq);
0423 extern void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
0424                     unsigned int which, u64 length);
0425 extern void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req,
0426                        unsigned int which, u64 offset_inc);
0427
0428 extern struct ceph_osd_data *osd_req_op_extent_osd_data(
0429                     struct ceph_osd_request *osd_req,
0430                     unsigned int which);
0431
0432 extern void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *,
0433                     unsigned int which,
0434                     struct page **pages, u64 length,
0435                     u32 alignment, bool pages_from_pool,
0436                     bool own_pages);
0437 extern void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *,
0438                     unsigned int which,
0439                     struct ceph_pagelist *pagelist);
0440 #ifdef CONFIG_BLOCK
0441 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req,
0442                     unsigned int which,
0443                     struct ceph_bio_iter *bio_pos,
0444                     u32 bio_length);
0445 #endif /* CONFIG_BLOCK */
0446 void osd_req_op_extent_osd_data_bvecs(struct ceph_osd_request *osd_req,
0447                       unsigned int which,
0448                       struct bio_vec *bvecs, u32 num_bvecs,
0449                       u32 bytes);
0450 void osd_req_op_extent_osd_data_bvec_pos(struct ceph_osd_request *osd_req,
0451                      unsigned int which,
0452                      struct ceph_bvec_iter *bvec_pos);
0453
0454 extern void osd_req_op_cls_request_data_pagelist(struct ceph_osd_request *,
0455                     unsigned int which,
0456                     struct ceph_pagelist *pagelist);
0457 extern void osd_req_op_cls_request_data_pages(struct ceph_osd_request *,
0458                     unsigned int which,
0459                     struct page **pages, u64 length,
0460                     u32 alignment, bool pages_from_pool,
0461                     bool own_pages);
0462 void osd_req_op_cls_request_data_bvecs(struct ceph_osd_request *osd_req,
0463                        unsigned int which,
0464                        struct bio_vec *bvecs, u32 num_bvecs,
0465                        u32 bytes);
0466 extern void osd_req_op_cls_response_data_pages(struct ceph_osd_request *,
0467                     unsigned int which,
0468                     struct page **pages, u64 length,
0469                     u32 alignment, bool pages_from_pool,
0470                     bool own_pages);
0471 int osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
0472             const char *class, const char *method);
0473 extern int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
0474                  u16 opcode, const char *name, const void *value,
0475                  size_t size, u8 cmp_op, u8 cmp_mode);
0476 extern void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
0477                        unsigned int which,
0478                        u64 expected_object_size,
0479                        u64 expected_write_size,
0480                        u32 flags);
0481 extern int osd_req_op_copy_from_init(struct ceph_osd_request *req,
0482                      u64 src_snapid, u64 src_version,
0483                      struct ceph_object_id *src_oid,
0484                      struct ceph_object_locator *src_oloc,
0485                      u32 src_fadvise_flags,
0486                      u32 dst_fadvise_flags,
0487                      u32 truncate_seq, u64 truncate_size,
0488                      u8 copy_from_flags);
0489
0490 extern struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
0491                            struct ceph_snap_context *snapc,
0492                            unsigned int num_ops,
0493                            bool use_mempool,
0494                            gfp_t gfp_flags);
0495 int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp);
0496
0497 extern struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *,
0498                       struct ceph_file_layout *layout,
0499                       struct ceph_vino vino,
0500                       u64 offset, u64 *len,
0501                       unsigned int which, int num_ops,
0502                       int opcode, int flags,
0503                       struct ceph_snap_context *snapc,
0504                       u32 truncate_seq, u64 truncate_size,
0505                       bool use_mempool);
0506
0507 extern void ceph_osdc_get_request(struct ceph_osd_request *req);
0508 extern void ceph_osdc_put_request(struct ceph_osd_request *req);
0509
0510 void ceph_osdc_start_request(struct ceph_osd_client *osdc,
0511                  struct ceph_osd_request *req);
0512 extern void ceph_osdc_cancel_request(struct ceph_osd_request *req);
0513 extern int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
0514                   struct ceph_osd_request *req);
0515 extern void ceph_osdc_sync(struct ceph_osd_client *osdc);
0516
0517 extern void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc);
0518 void ceph_osdc_maybe_request_map(struct ceph_osd_client *osdc);
0519
0520 int ceph_osdc_call(struct ceph_osd_client *osdc,
0521            struct ceph_object_id *oid,
0522            struct ceph_object_locator *oloc,
0523            const char *class, const char *method,
0524            unsigned int flags,
0525            struct page *req_page, size_t req_len,
0526            struct page **resp_pages, size_t *resp_len);
0527
0528 /* watch/notify */
0529 struct ceph_osd_linger_request *
0530 ceph_osdc_watch(struct ceph_osd_client *osdc,
0531         struct ceph_object_id *oid,
0532         struct ceph_object_locator *oloc,
0533         rados_watchcb2_t wcb,
0534         rados_watcherrcb_t errcb,
0535         void *data);
0536 int ceph_osdc_unwatch(struct ceph_osd_client *osdc,
0537               struct ceph_osd_linger_request *lreq);
0538
0539 int ceph_osdc_notify_ack(struct ceph_osd_client *osdc,
0540              struct ceph_object_id *oid,
0541              struct ceph_object_locator *oloc,
0542              u64 notify_id,
0543              u64 cookie,
0544              void *payload,
0545              u32 payload_len);
0546 int ceph_osdc_notify(struct ceph_osd_client *osdc,
0547              struct ceph_object_id *oid,
0548              struct ceph_object_locator *oloc,
0549              void *payload,
0550              u32 payload_len,
0551              u32 timeout,
0552              struct page ***preply_pages,
0553              size_t *preply_len);
0554 int ceph_osdc_watch_check(struct ceph_osd_client *osdc,
0555               struct ceph_osd_linger_request *lreq);
0556 int ceph_osdc_list_watchers(struct ceph_osd_client *osdc,
0557                 struct ceph_object_id *oid,
0558                 struct ceph_object_locator *oloc,
0559                 struct ceph_watch_item **watchers,
0560                 u32 *num_watchers);
0561 #endif
0562