5 * Data types for the Ceph distributed object storage layer RADOS
6 * (Reliable Autonomic Distributed Object Store).
9 #include <linux/ceph/msgr.h>
15 unsigned char fsid[16];
18 static inline int ceph_fsid_compare(const struct ceph_fsid *a,
19 const struct ceph_fsid *b)
21 return memcmp(a, b, sizeof(*a));
27 typedef __le64 ceph_snapid_t;
28 #define CEPH_SNAPDIR ((__u64)(-1)) /* reserved for hidden .snap dir */
29 #define CEPH_NOSNAP ((__u64)(-2)) /* "head", "live" revision */
30 #define CEPH_MAXSNAP ((__u64)(-3)) /* largest valid snapid */
32 struct ceph_timespec {
35 } __attribute__ ((packed));
39 * object layout - how objects are mapped into PGs
41 #define CEPH_OBJECT_LAYOUT_HASH 1
42 #define CEPH_OBJECT_LAYOUT_LINEAR 2
43 #define CEPH_OBJECT_LAYOUT_HASHINO 3
46 * pg layout -- how PGs are mapped onto (sets of) OSDs
48 #define CEPH_PG_LAYOUT_CRUSH 0
49 #define CEPH_PG_LAYOUT_HASH 1
50 #define CEPH_PG_LAYOUT_LINEAR 2
51 #define CEPH_PG_LAYOUT_HYBRID 3
53 #define CEPH_PG_MAX_SIZE 16 /* max # osds in a single pg */
57 * we encode this into one __le64.
60 __le16 preferred; /* preferred primary osd */
61 __le16 ps; /* placement seed */
62 __le32 pool; /* object pool */
63 } __attribute__ ((packed));
66 * pg_pool is a set of pgs storing a pool of objects
68 * pg_num -- base number of pseudorandomly placed pgs
70 * pgp_num -- effective number when calculating pg placement. this
71 * is used for pg_num increases. new pgs result in data being "split"
72 * into new pgs. for this to proceed smoothly, new pgs are intiially
73 * colocated with their parents; that is, pgp_num doesn't increase
74 * until the new pgs have successfully split. only _then_ are the new
75 * pgs placed independently.
77 * lpg_num -- localized pg count (per device). replicas are randomly
80 * lpgp_num -- as above.
82 #define CEPH_NOPOOL ((__u64) (-1)) /* pool id not defined */
84 #define CEPH_POOL_TYPE_REP 1
85 #define CEPH_POOL_TYPE_RAID4 2 /* never implemented */
86 #define CEPH_POOL_TYPE_EC 3
89 * stable_mod func is used to control number of placement groups.
90 * similar to straight-up modulo, but produces a stable mapping as b
91 * increases over time. b is the number of bins, and bmask is the
92 * containing power of 2 minus 1.
94 * b <= bmask and bmask=(2**n)-1
95 * e.g., b=12 -> bmask=15, b=123 -> bmask=127
97 static inline int ceph_stable_mod(int x, int b, int bmask)
102 return x & (bmask >> 1);
106 * object layout - how a given object should be stored.
108 struct ceph_object_layout {
109 struct ceph_pg_v1 ol_pgid; /* raw pg, with _full_ ps precision. */
110 __le32 ol_stripe_unit; /* for per-object parity, if any */
111 } __attribute__ ((packed));
114 * compound epoch+version, used by storage layer to serialize mutations
116 struct ceph_eversion {
119 } __attribute__ ((packed));
126 #define CEPH_OSD_EXISTS (1<<0)
127 #define CEPH_OSD_UP (1<<1)
128 #define CEPH_OSD_AUTOOUT (1<<2) /* osd was automatically marked out */
129 #define CEPH_OSD_NEW (1<<3) /* osd is new, never marked in */
131 extern const char *ceph_osd_state_name(int s);
133 /* osd weights. fixed point value: 0x10000 == 1.0 ("in"), 0 == "out" */
134 #define CEPH_OSD_IN 0x10000
135 #define CEPH_OSD_OUT 0
137 /* osd primary-affinity. fixed point value: 0x10000 == baseline */
138 #define CEPH_OSD_MAX_PRIMARY_AFFINITY 0x10000
139 #define CEPH_OSD_DEFAULT_PRIMARY_AFFINITY 0x10000
145 #define CEPH_OSDMAP_NEARFULL (1<<0) /* sync writes (near ENOSPC) */
146 #define CEPH_OSDMAP_FULL (1<<1) /* no data writes (ENOSPC) */
147 #define CEPH_OSDMAP_PAUSERD (1<<2) /* pause all reads */
148 #define CEPH_OSDMAP_PAUSEWR (1<<3) /* pause all writes */
149 #define CEPH_OSDMAP_PAUSEREC (1<<4) /* pause recovery */
150 #define CEPH_OSDMAP_NOUP (1<<5) /* block osd boot */
151 #define CEPH_OSDMAP_NODOWN (1<<6) /* block osd mark-down/failure */
152 #define CEPH_OSDMAP_NOOUT (1<<7) /* block osd auto mark-out */
153 #define CEPH_OSDMAP_NOIN (1<<8) /* block osd auto mark-in */
154 #define CEPH_OSDMAP_NOBACKFILL (1<<9) /* block osd backfill */
155 #define CEPH_OSDMAP_NORECOVER (1<<10) /* block osd recovery and backfill */
158 * The error code to return when an OSD can't handle a write
159 * because it is too large.
161 #define OSD_WRITETOOBIG EMSGSIZE
166 * WARNING: do not use these op codes directly. Use the helpers
167 * defined below instead. In certain cases, op code behavior was
168 * redefined, resulting in special-cases in the helpers.
170 #define CEPH_OSD_OP_MODE 0xf000
171 #define CEPH_OSD_OP_MODE_RD 0x1000
172 #define CEPH_OSD_OP_MODE_WR 0x2000
173 #define CEPH_OSD_OP_MODE_RMW 0x3000
174 #define CEPH_OSD_OP_MODE_SUB 0x4000
176 #define CEPH_OSD_OP_TYPE 0x0f00
177 #define CEPH_OSD_OP_TYPE_LOCK 0x0100
178 #define CEPH_OSD_OP_TYPE_DATA 0x0200
179 #define CEPH_OSD_OP_TYPE_ATTR 0x0300
180 #define CEPH_OSD_OP_TYPE_EXEC 0x0400
181 #define CEPH_OSD_OP_TYPE_PG 0x0500
182 #define CEPH_OSD_OP_TYPE_MULTI 0x0600 /* multiobject */
187 CEPH_OSD_OP_READ = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_DATA | 1,
188 CEPH_OSD_OP_STAT = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_DATA | 2,
189 CEPH_OSD_OP_MAPEXT = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_DATA | 3,
192 CEPH_OSD_OP_MASKTRUNC = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_DATA | 4,
193 CEPH_OSD_OP_SPARSE_READ = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_DATA | 5,
195 CEPH_OSD_OP_NOTIFY = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_DATA | 6,
196 CEPH_OSD_OP_NOTIFY_ACK = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_DATA | 7,
199 CEPH_OSD_OP_ASSERT_VER = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_DATA | 8,
202 CEPH_OSD_OP_WRITE = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_DATA | 1,
203 CEPH_OSD_OP_WRITEFULL = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_DATA | 2,
204 CEPH_OSD_OP_TRUNCATE = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_DATA | 3,
205 CEPH_OSD_OP_ZERO = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_DATA | 4,
206 CEPH_OSD_OP_DELETE = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_DATA | 5,
209 CEPH_OSD_OP_APPEND = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_DATA | 6,
210 CEPH_OSD_OP_STARTSYNC = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_DATA | 7,
211 CEPH_OSD_OP_SETTRUNC = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_DATA | 8,
212 CEPH_OSD_OP_TRIMTRUNC = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_DATA | 9,
214 CEPH_OSD_OP_TMAPUP = CEPH_OSD_OP_MODE_RMW | CEPH_OSD_OP_TYPE_DATA | 10,
215 CEPH_OSD_OP_TMAPPUT = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_DATA | 11,
216 CEPH_OSD_OP_TMAPGET = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_DATA | 12,
218 CEPH_OSD_OP_CREATE = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_DATA | 13,
219 CEPH_OSD_OP_ROLLBACK= CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_DATA | 14,
221 CEPH_OSD_OP_WATCH = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_DATA | 15,
224 CEPH_OSD_OP_OMAPGETKEYS = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_DATA | 17,
225 CEPH_OSD_OP_OMAPGETVALS = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_DATA | 18,
226 CEPH_OSD_OP_OMAPGETHEADER = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_DATA | 19,
227 CEPH_OSD_OP_OMAPGETVALSBYKEYS =
228 CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_DATA | 20,
229 CEPH_OSD_OP_OMAPSETVALS = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_DATA | 21,
230 CEPH_OSD_OP_OMAPSETHEADER = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_DATA | 22,
231 CEPH_OSD_OP_OMAPCLEAR = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_DATA | 23,
232 CEPH_OSD_OP_OMAPRMKEYS = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_DATA | 24,
233 CEPH_OSD_OP_OMAP_CMP = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_DATA | 25,
236 CEPH_OSD_OP_SETALLOCHINT = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_DATA | 35,
239 CEPH_OSD_OP_CLONERANGE = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_MULTI | 1,
240 CEPH_OSD_OP_ASSERT_SRC_VERSION = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_MULTI | 2,
241 CEPH_OSD_OP_SRC_CMPXATTR = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_MULTI | 3,
245 CEPH_OSD_OP_GETXATTR = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_ATTR | 1,
246 CEPH_OSD_OP_GETXATTRS = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_ATTR | 2,
247 CEPH_OSD_OP_CMPXATTR = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_ATTR | 3,
250 CEPH_OSD_OP_SETXATTR = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_ATTR | 1,
251 CEPH_OSD_OP_SETXATTRS = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_ATTR | 2,
252 CEPH_OSD_OP_RESETXATTRS = CEPH_OSD_OP_MODE_WR|CEPH_OSD_OP_TYPE_ATTR | 3,
253 CEPH_OSD_OP_RMXATTR = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_ATTR | 4,
256 CEPH_OSD_OP_PULL = CEPH_OSD_OP_MODE_SUB | 1,
257 CEPH_OSD_OP_PUSH = CEPH_OSD_OP_MODE_SUB | 2,
258 CEPH_OSD_OP_BALANCEREADS = CEPH_OSD_OP_MODE_SUB | 3,
259 CEPH_OSD_OP_UNBALANCEREADS = CEPH_OSD_OP_MODE_SUB | 4,
260 CEPH_OSD_OP_SCRUB = CEPH_OSD_OP_MODE_SUB | 5,
261 CEPH_OSD_OP_SCRUB_RESERVE = CEPH_OSD_OP_MODE_SUB | 6,
262 CEPH_OSD_OP_SCRUB_UNRESERVE = CEPH_OSD_OP_MODE_SUB | 7,
263 CEPH_OSD_OP_SCRUB_STOP = CEPH_OSD_OP_MODE_SUB | 8,
264 CEPH_OSD_OP_SCRUB_MAP = CEPH_OSD_OP_MODE_SUB | 9,
267 CEPH_OSD_OP_WRLOCK = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_LOCK | 1,
268 CEPH_OSD_OP_WRUNLOCK = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_LOCK | 2,
269 CEPH_OSD_OP_RDLOCK = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_LOCK | 3,
270 CEPH_OSD_OP_RDUNLOCK = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_LOCK | 4,
271 CEPH_OSD_OP_UPLOCK = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_LOCK | 5,
272 CEPH_OSD_OP_DNLOCK = CEPH_OSD_OP_MODE_WR | CEPH_OSD_OP_TYPE_LOCK | 6,
275 /* note: the RD bit here is wrong; see special-case below in helper */
276 CEPH_OSD_OP_CALL = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_EXEC | 1,
279 CEPH_OSD_OP_PGLS = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_PG | 1,
280 CEPH_OSD_OP_PGLS_FILTER = CEPH_OSD_OP_MODE_RD | CEPH_OSD_OP_TYPE_PG | 2,
283 static inline int ceph_osd_op_type_lock(int op)
285 return (op & CEPH_OSD_OP_TYPE) == CEPH_OSD_OP_TYPE_LOCK;
287 static inline int ceph_osd_op_type_data(int op)
289 return (op & CEPH_OSD_OP_TYPE) == CEPH_OSD_OP_TYPE_DATA;
291 static inline int ceph_osd_op_type_attr(int op)
293 return (op & CEPH_OSD_OP_TYPE) == CEPH_OSD_OP_TYPE_ATTR;
295 static inline int ceph_osd_op_type_exec(int op)
297 return (op & CEPH_OSD_OP_TYPE) == CEPH_OSD_OP_TYPE_EXEC;
299 static inline int ceph_osd_op_type_pg(int op)
301 return (op & CEPH_OSD_OP_TYPE) == CEPH_OSD_OP_TYPE_PG;
303 static inline int ceph_osd_op_type_multi(int op)
305 return (op & CEPH_OSD_OP_TYPE) == CEPH_OSD_OP_TYPE_MULTI;
308 static inline int ceph_osd_op_mode_subop(int op)
310 return (op & CEPH_OSD_OP_MODE) == CEPH_OSD_OP_MODE_SUB;
312 static inline int ceph_osd_op_mode_read(int op)
314 return (op & CEPH_OSD_OP_MODE_RD) &&
315 op != CEPH_OSD_OP_CALL;
317 static inline int ceph_osd_op_mode_modify(int op)
319 return op & CEPH_OSD_OP_MODE_WR;
323 * note that the following tmap stuff is also defined in the ceph librados.h
324 * any modification here needs to be updated there
326 #define CEPH_OSD_TMAP_HDR 'h'
327 #define CEPH_OSD_TMAP_SET 's'
328 #define CEPH_OSD_TMAP_CREATE 'c' /* create key */
329 #define CEPH_OSD_TMAP_RM 'r'
330 #define CEPH_OSD_TMAP_RMSLOPPY 'R'
332 extern const char *ceph_osd_op_name(int op);
337 * An op may be READ, WRITE, or READ|WRITE.
340 CEPH_OSD_FLAG_ACK = 0x0001, /* want (or is) "ack" ack */
341 CEPH_OSD_FLAG_ONNVRAM = 0x0002, /* want (or is) "onnvram" ack */
342 CEPH_OSD_FLAG_ONDISK = 0x0004, /* want (or is) "ondisk" ack */
343 CEPH_OSD_FLAG_RETRY = 0x0008, /* resend attempt */
344 CEPH_OSD_FLAG_READ = 0x0010, /* op may read */
345 CEPH_OSD_FLAG_WRITE = 0x0020, /* op may write */
346 CEPH_OSD_FLAG_ORDERSNAP = 0x0040, /* EOLDSNAP if snapc is out of order */
347 CEPH_OSD_FLAG_PEERSTAT_OLD = 0x0080, /* DEPRECATED msg includes osd_peer_stat */
348 CEPH_OSD_FLAG_BALANCE_READS = 0x0100,
349 CEPH_OSD_FLAG_PARALLELEXEC = 0x0200, /* execute op in parallel */
350 CEPH_OSD_FLAG_PGOP = 0x0400, /* pg op, no object */
351 CEPH_OSD_FLAG_EXEC = 0x0800, /* op may exec */
352 CEPH_OSD_FLAG_EXEC_PUBLIC = 0x1000, /* DEPRECATED op may exec (public) */
353 CEPH_OSD_FLAG_LOCALIZE_READS = 0x2000, /* read from nearby replica, if any */
354 CEPH_OSD_FLAG_RWORDERED = 0x4000, /* order wrt concurrent reads */
355 CEPH_OSD_FLAG_IGNORE_CACHE = 0x8000, /* ignore cache logic */
356 CEPH_OSD_FLAG_SKIPRWLOCKS = 0x10000, /* skip rw locks */
357 CEPH_OSD_FLAG_IGNORE_OVERLAY = 0x20000, /* ignore pool overlay */
358 CEPH_OSD_FLAG_FLUSH = 0x40000, /* this is part of flush */
362 CEPH_OSD_OP_FLAG_EXCL = 1, /* EXCL object create */
363 CEPH_OSD_OP_FLAG_FAILOK = 2, /* continue despite failure */
366 #define EOLDSNAPC ERESTART /* ORDERSNAP flag set; writer has old snapc*/
367 #define EBLACKLISTED ESHUTDOWN /* blacklisted */
369 /* xattr comparison */
371 CEPH_OSD_CMPXATTR_OP_NOP = 0,
372 CEPH_OSD_CMPXATTR_OP_EQ = 1,
373 CEPH_OSD_CMPXATTR_OP_NE = 2,
374 CEPH_OSD_CMPXATTR_OP_GT = 3,
375 CEPH_OSD_CMPXATTR_OP_GTE = 4,
376 CEPH_OSD_CMPXATTR_OP_LT = 5,
377 CEPH_OSD_CMPXATTR_OP_LTE = 6
381 CEPH_OSD_CMPXATTR_MODE_STRING = 1,
382 CEPH_OSD_CMPXATTR_MODE_U64 = 2
385 #define RADOS_NOTIFY_VER 1
388 * an individual object operation. each may be accompanied by some data
392 __le16 op; /* CEPH_OSD_OP_* */
393 __le32 flags; /* CEPH_OSD_OP_FLAG_* */
396 __le64 offset, length;
397 __le64 truncate_size;
399 } __attribute__ ((packed)) extent;
403 __u8 cmp_op; /* CEPH_OSD_CMPXATTR_OP_* */
404 __u8 cmp_mode; /* CEPH_OSD_CMPXATTR_MODE_* */
405 } __attribute__ ((packed)) xattr;
411 } __attribute__ ((packed)) cls;
413 __le64 cookie, count;
414 } __attribute__ ((packed)) pgls;
417 } __attribute__ ((packed)) snap;
421 __u8 flag; /* 0 = unwatch, 1 = watch */
422 } __attribute__ ((packed)) watch;
424 __le64 offset, length;
426 } __attribute__ ((packed)) clonerange;
428 __le64 expected_object_size;
429 __le64 expected_write_size;
430 } __attribute__ ((packed)) alloc_hint;
433 } __attribute__ ((packed));