2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
13 /* these are now appended by the error table compiler */
15 /* ubik error codes */
16 #define UMINCODE 100000 /* lowest ubik error code */
17 #define UNOQUORUM 100000 /* no quorum elected */
18 #define UNOTSYNC 100001 /* not synchronization site (should work on sync site) */
19 #define UNHOSTS 100002 /* too many hosts */
20 #define UIOERROR 100003 /* I/O error writing dbase or log */
21 #define UINTERNAL 100004 /* mysterious internal error */
22 #define USYNC 100005 /* major synchronization error */
23 #define UNOENT 100006 /* file not found when processing dbase */
24 #define UBADLOCK 100007 /* bad lock range size (must be 1) */
25 #define UBADLOG 100008 /* read error reprocessing log */
26 #define UBADHOST 100009 /* problems with host name */
27 #define UBADTYPE 100010 /* bad operation for this transaction type */
28 #define UTWOENDS 100011 /* two commits or aborts done to transaction */
29 #define UDONE 100012 /* operation done after abort (or commmit) */
30 #define UNOSERVERS 100013 /* no servers appear to be up */
31 #define UEOF 100014 /* premature EOF */
32 #define ULOGIO 100015 /* error writing log file */
33 #define UMAXCODE 100100 /* largest ubik error code */
39 #else /* defined(UKERNEL) */
41 #endif /* defined(UKERNEL) */
43 /* ubik_trans types */
44 #define UBIK_READTRANS 0
45 #define UBIK_WRITETRANS 1
50 #if !defined(UBIK_PAUSE)
52 #endif /* UBIK_PAUSE */
54 /* ubik client flags */
55 #define UPUBIKONLY 1 /* only check servers presumed functional */
57 /* RX services types */
58 #define VOTE_SERVICE_ID 50
59 #define DISK_SERVICE_ID 51
60 #define USER_SERVICE_ID 52 /* Since most applications use same port! */
62 #define UBIK_MAGIC 0x354545
64 /* global ubik parameters */
65 #define MAXSERVERS 20 /* max number of servers */
67 /* version comparison macro */
68 #define vcmp(a,b) ((a).epoch == (b).epoch? ((a).counter - (b).counter) : ((a).epoch - (b).epoch))
70 /* ubik_client state bits */
71 #define CFLastFailed 1 /* last call failed to this guy (to detect down hosts) */
73 #ifdef AFS_PTHREAD_ENV
80 /* per-client structure for ubik */
82 short initializationState; /* ubik client init state */
83 short states[MAXSERVERS]; /* state bits */
84 struct rx_connection *conns[MAXSERVERS];
86 #ifdef AFS_PTHREAD_ENV
91 #ifdef AFS_PTHREAD_ENV
92 #define LOCK_UBIK_CLIENT(client) assert(pthread_mutex_lock(&client->cm)==0)
93 #define UNLOCK_UBIK_CLIENT(client) assert(pthread_mutex_unlock(&client->cm)==0)
95 #define LOCK_UBIK_CLIENT(client)
96 #define UNLOCK_UBIK_CLIENT(client)
99 #define ubik_GetRPCConn(astr,aindex) ((aindex) >= MAXSERVERS? 0 : (astr)->conns[aindex])
100 #define ubik_GetRPCHost(astr,aindex) ((aindex) >= MAXSERVERS? 0 : (astr)->hosts[aindex])
102 /* ubik header file structure */
104 afs_int32 magic; /* magic number */
105 short pad1; /* some 0-initd padding */
106 short size; /* header allocation size */
107 struct ubik_version version; /* the version for this file */
110 /* representation of a ubik transaction */
112 struct ubik_dbase *dbase; /* corresponding database */
113 struct ubik_trans *next; /* in the list */
114 afs_int32 locktype; /* transaction lock */
115 struct ubik_trunc *activeTruncs; /* queued truncates */
116 struct ubik_tid tid; /* transaction id of this trans (if write trans.) */
117 afs_int32 minCommitTime; /* time before which this trans can't commit */
118 afs_int32 seekFile; /* seek ptr: file number */
119 afs_int32 seekPos; /* seek ptr: offset therein */
120 short flags; /* trans flag bits */
121 char type; /* type of trans */
122 iovec_wrt iovec_info;
123 iovec_buf iovec_data;
126 /* representation of a truncation operation */
128 struct ubik_trunc *next;
129 afs_int32 file; /* file to truncate */
130 afs_int32 length; /* new size */
139 #include "afs/lock.h"
140 #else /* defined(UKERNEL) */
141 #include <lock.h> /* just to make sure we've go this */
142 #endif /* defined(UKERNEL) */
144 /* representation of a ubik database. Contains info on low-level disk access routines
145 for use by disk transaction module.
148 char *pathName; /* root name for dbase */
149 struct ubik_trans *activeTrans; /* active transaction list */
150 struct ubik_version version; /* version number */
152 struct afs_lock versionLock; /* lock on version number */
153 #else /* defined(UKERNEL) */
154 struct Lock versionLock; /* lock on version number */
155 #endif /* defined(UKERNEL) */
156 afs_int32 tidCounter; /* last RW or RO trans tid counter */
157 afs_int32 writeTidCounter; /* last write trans tid counter */
158 afs_int32 flags; /* flags */
159 /* physio procedures */
160 int (*read) (struct ubik_dbase * adbase, afs_int32 afile, char *abuffer,
161 afs_int32 apos, afs_int32 alength);
162 int (*write) (struct ubik_dbase * adbase, afs_int32 afile, char *abuffer,
163 afs_int32 apos, afs_int32 alength);
164 int (*truncate) (struct ubik_dbase * adbase, afs_int32 afile,
166 int (*sync) (struct ubik_dbase * adbase, afs_int32 afile);
167 int (*stat) (struct ubik_dbase * adbase, afs_int32 afid,
168 struct ubik_stat * astat);
169 int (*open) (struct ubik_dbase * adbase, afs_int32 afid);
170 int (*setlabel) (struct ubik_dbase * adbase, afs_int32 afile, struct ubik_version * aversion); /* set the version label */
171 int (*getlabel) (struct ubik_dbase * adbase, afs_int32 afile, struct ubik_version * aversion); /* retrieve the version label */
172 int (*getnfiles) (struct ubik_dbase * adbase); /* find out number of files */
173 short readers; /* number of current read transactions */
174 struct ubik_version cachedVersion; /* version of caller's cached data */
175 #ifdef AFS_PTHREAD_ENV
176 pthread_cond_t version_cond; /* condition variable to manage changes to version */
177 pthread_cond_t flags_cond; /* condition variable to manage changes to flags */
178 pthread_mutex_t version_mutex;
179 pthread_mutex_t flags_mutex;
183 /* procedures for automatically authenticating ubik connections */
184 extern int (*ubik_CRXSecurityProc) ();
185 extern char *ubik_CRXSecurityRock;
186 extern int (*ubik_SRXSecurityProc) ();
187 extern char *ubik_SRXSecurityRock;
188 extern int (*ubik_CheckRXSecurityProc) ();
189 extern char *ubik_CheckRXSecurityRock;
191 /****************INTERNALS BELOW ****************/
193 #ifdef UBIK_INTERNALS
194 /* some ubik parameters */
195 #define UBIK_PAGESIZE 1024 /* fits in current r packet */
196 #define UBIK_LOGPAGESIZE 10 /* base 2 log thereof */
197 #define NBUFFERS 20 /* number of 1K buffers */
198 #define HDRSIZE 64 /* bytes of header per dbfile */
200 /* ubik_dbase flags */
201 #define DBWRITING 1 /* are any write trans. in progress */
202 #if defined(UBIK_PAUSE)
203 #define DBVOTING 2 /* the beacon task is polling */
204 #endif /* UBIK_PAUSE */
206 /* ubik trans flags */
207 #define TRDONE 1 /* commit or abort done */
208 #define TRABORT 2 /* if TRDONE, tells if aborted */
209 #define TRREADANY 4 /* read any data available in trans */
210 #if defined(UBIK_PAUSE)
211 #define TRSETLOCK 8 /* SetLock is using trans */
212 #define TRSTALE 16 /* udisk_end during getLock */
213 #endif /* UBIK_PAUSE */
215 /* ubik_lock flags */
218 /* ubik system database numbers */
221 /* define log opcodes */
222 #define LOGNEW 100 /* start transaction */
223 #define LOGEND 101 /* commit (good) end transaction */
224 #define LOGABORT 102 /* abort (fail) transaction */
225 #define LOGDATA 103 /* data */
226 #define LOGTRUNCATE 104 /* truncate operation */
228 /* time constant for replication algorithms: the R time period is 20 seconds. Both SMALLTIME
229 and BIGTIME must be larger than RPCTIMEOUT+max(RPCTIMEOUT,POLLTIME),
230 so that timeouts do not prevent us from getting through to our servers in time.
232 We use multi-R to time out multiple down hosts concurrently.
233 The only other restrictions: BIGTIME > SMALLTIME and
234 BIGTIME-SMALLTIME > MAXSKEW (the clock skew).
238 #define RPCTIMEOUT 20
242 /* the per-server state, used by the sync site to keep track of its charges */
244 struct ubik_server *next; /* next ptr */
245 afs_uint32 addr[UBIK_MAX_INTERFACE_ADDR]; /* network order, addr[0] is primary */
246 afs_int32 lastVoteTime; /* last time yes vote received */
247 afs_int32 lastBeaconSent; /* last time beacon attempted */
248 struct ubik_version version; /* version, only used during recovery */
249 struct rx_connection *vote_rxcid; /* cid to use to contact dude for votes */
250 struct rx_connection *disk_rxcid; /* cid to use to contact dude for disk reqs */
251 char lastVote; /* true if last vote was yes */
252 char up; /* is it up? */
253 char beaconSinceDown; /* did beacon get through since last crash? */
254 char currentDB; /* is dbase up-to-date */
255 char magic; /* the one whose vote counts twice */
256 char isClone; /* is only a clone, doesn't vote */
259 /* hold and release functions on a database */
260 #define DBHOLD(a) ObtainWriteLock(&((a)->versionLock))
261 #define DBRELE(a) ReleaseWriteLock(&((a)->versionLock))
265 /* list of all servers in the system */
266 extern struct ubik_server *ubik_servers;
267 extern char amIClone;
269 /* network port info */
270 extern short ubik_callPortal;
272 /* urecovery state bits for sync site */
273 #define UBIK_RECSYNCSITE 1 /* am sync site */
274 #define UBIK_RECFOUNDDB 2 /* found acceptable dbase from quorum */
275 #define UBIK_RECHAVEDB 4 /* fetched best dbase */
276 #define UBIK_RECLABELDB 8 /* relabelled dbase */
277 #define UBIK_RECSENTDB 0x10 /* sent best db to *everyone* */
278 #define UBIK_RECSBETTER UBIK_RECLABELDB /* last state */
280 extern afs_int32 ubik_quorum; /* min hosts in quorum */
281 extern struct ubik_dbase *ubik_dbase; /* the database handled by this server */
282 extern afs_uint32 ubik_host[UBIK_MAX_INTERFACE_ADDR]; /* this host addr, in net order */
283 extern int ubik_amSyncSite; /* sleep on this waiting to be sync site */
284 extern struct ubik_stats { /* random stats */
287 extern afs_int32 ubik_epochTime; /* time when this site started */
288 extern afs_int32 urecovery_state; /* sync site recovery process state */
289 extern struct ubik_trans *ubik_currentTrans; /* current trans */
290 extern struct ubik_version ubik_dbVersion; /* sync site's dbase version */
291 extern afs_int32 ubik_debugFlag; /* ubik debug flag */
292 extern int ubikPrimaryAddrOnly; /* use only primary address */
294 /* this extern gives the sync site's db version, with epoch of 0 if none yet */
297 extern int uphys_close(register int afd);
298 extern int uphys_stat(struct ubik_dbase *adbase, afs_int32 afid,
299 struct ubik_stat *astat);
300 extern int uphys_read(register struct ubik_dbase *adbase, afs_int32 afile,
301 register char *abuffer, afs_int32 apos,
303 extern int uphys_write(register struct ubik_dbase *adbase, afs_int32 afile,
304 register char *abuffer, afs_int32 apos,
306 extern int uphys_truncate(register struct ubik_dbase *adbase, afs_int32 afile,
308 extern int uphys_getnfiles(register struct ubik_dbase *adbase);
309 extern int uphys_getlabel(register struct ubik_dbase *adbase, afs_int32 afile,
310 struct ubik_version *aversion);
311 extern int uphys_setlabel(register struct ubik_dbase *adbase, afs_int32 afile,
312 struct ubik_version *aversion);
313 extern int uphys_sync(register struct ubik_dbase *adbase, afs_int32 afile);
314 extern void uphys_invalidate(register struct ubik_dbase *adbase,
318 extern int urecovery_ResetState(void);
319 extern int urecovery_LostServer(void);
320 extern int urecovery_AllBetter(register struct ubik_dbase *adbase,
322 extern int urecovery_AbortAll(struct ubik_dbase *adbase);
323 extern int urecovery_CheckTid(register struct ubik_tid *atid);
324 extern int urecovery_Initialize(register struct ubik_dbase *adbase);
325 extern void *urecovery_Interact(void *);
326 extern int DoProbe(struct ubik_server *server);
329 extern void *ubeacon_Interact(void *);
330 extern int sdisk_Interact();
331 extern int uvote_Interact();
332 extern int DISK_Abort();
333 extern int DISK_Begin();
334 extern int DISK_ReleaseLocks();
335 extern int DISK_Commit();
336 extern int DISK_Lock();
337 extern int DISK_Write();
338 extern int DISK_WriteV();
339 extern int DISK_Truncate();
340 extern int DISK_SetVersion();
343 extern int udisk_abort(struct ubik_trans *atrans);
346 extern void ulock_relLock(struct ubik_trans *atrans);
348 #endif /* UBIK_INTERNALS */
350 extern afs_int32 ubik_nBuffers;
353 * Public function prototypes
356 extern int ubik_ParseClientList(int argc, char **argv, afs_int32 * aothers);
358 extern unsigned int afs_random(void
361 extern int ubik_ClientInit(register struct rx_connection **serverconns,
362 struct ubik_client **aclient);
364 extern afs_int32 ubik_ClientDestroy(struct ubik_client *aclient);
366 extern afs_int32 ubik_CallIter(int (*aproc) (), struct ubik_client *aclient,
367 afs_int32 aflags, int *apos, long p1, long p2,
368 long p3, long p4, long p5, long p6, long p7,
369 long p8, long p9, long p10, long p11, long p12,
370 long p13, long p14, long p15, long p16);
372 extern struct rx_connection *ubik_RefreshConn(struct rx_connection *tc);
375 extern int ubik_BeginTrans(register struct ubik_dbase *dbase,
376 afs_int32 transMode, struct ubik_trans **transPtr);
377 extern int ubik_EndTrans(register struct ubik_trans *transPtr);
381 extern afs_int32 ugen_ClientInit(int noAuthFlag, char *confDir, char *cellName,
383 struct ubik_client **uclientp,
384 int (*secproc) (), char *funcName,
385 afs_int32 gen_rxkad_level,
386 afs_int32 maxservers, char *serviceid,
387 afs_int32 deadtime, afs_uint32 server,
388 afs_uint32 port, afs_int32 usrvid);