]> git.ozlabs.org Git - ccan/blob - ccan/tdb/traverse.c
tdb2: limit coalescing based on how successful we are.
[ccan] / ccan / tdb / traverse.c
1  /* 
2    Unix SMB/CIFS implementation.
3
4    trivial database library
5
6    Copyright (C) Andrew Tridgell              1999-2005
7    Copyright (C) Paul `Rusty' Russell              2000
8    Copyright (C) Jeremy Allison                    2000-2003
9    
10      ** NOTE! The following LGPL license applies to the tdb
11      ** library. This does NOT imply that all of Samba is released
12      ** under the LGPL
13    
14    This library is free software; you can redistribute it and/or
15    modify it under the terms of the GNU Lesser General Public
16    License as published by the Free Software Foundation; either
17    version 3 of the License, or (at your option) any later version.
18
19    This library is distributed in the hope that it will be useful,
20    but WITHOUT ANY WARRANTY; without even the implied warranty of
21    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
22    Lesser General Public License for more details.
23
24    You should have received a copy of the GNU Lesser General Public
25    License along with this library; if not, see <http://www.gnu.org/licenses/>.
26 */
27
28 #include "tdb_private.h"
29
30 #define TDB_NEXT_LOCK_ERR ((tdb_off_t)-1)
31
32 /* Uses traverse lock: 0 = finish, TDB_NEXT_LOCK_ERR = error,
33    other = record offset */
34 static tdb_off_t tdb_next_lock(struct tdb_context *tdb, struct tdb_traverse_lock *tlock,
35                          struct tdb_record *rec)
36 {
37         int want_next = (tlock->off != 0);
38
39         /* Lock each chain from the start one. */
40         for (; tlock->hash < tdb->header.hash_size; tlock->hash++) {
41                 if (!tlock->off && tlock->hash != 0) {
42                         /* this is an optimization for the common case where
43                            the hash chain is empty, which is particularly
44                            common for the use of tdb with ldb, where large
45                            hashes are used. In that case we spend most of our
46                            time in tdb_brlock(), locking empty hash chains.
47                            
48                            To avoid this, we do an unlocked pre-check to see
49                            if the hash chain is empty before starting to look
50                            inside it. If it is empty then we can avoid that
51                            hash chain. If it isn't empty then we can't believe
52                            the value we get back, as we read it without a
53                            lock, so instead we get the lock and re-fetch the
54                            value below.
55                            
56                            Notice that not doing this optimization on the
57                            first hash chain is critical. We must guarantee
58                            that we have done at least one fcntl lock at the
59                            start of a search to guarantee that memory is
60                            coherent on SMP systems. If records are added by
61                            others during the search then thats OK, and we
62                            could possibly miss those with this trick, but we
63                            could miss them anyway without this trick, so the
64                            semantics don't change.
65                            
66                            With a non-indexed ldb search this trick gains us a
67                            factor of around 80 in speed on a linux 2.6.x
68                            system (testing using ldbtest).
69                         */
70                         tdb->methods->next_hash_chain(tdb, &tlock->hash);
71                         if (tlock->hash == tdb->header.hash_size) {
72                                 continue;
73                         }
74                 }
75
76                 if (tdb_lock(tdb, tlock->hash, tlock->lock_rw) == -1)
77                         return TDB_NEXT_LOCK_ERR;
78
79                 /* No previous record?  Start at top of chain. */
80                 if (!tlock->off) {
81                         if (tdb_ofs_read(tdb, TDB_HASH_TOP(tlock->hash),
82                                      &tlock->off) == -1)
83                                 goto fail;
84                 } else {
85                         /* Otherwise unlock the previous record. */
86                         if (tdb_unlock_record(tdb, tlock->off) != 0)
87                                 goto fail;
88                 }
89
90                 if (want_next) {
91                         /* We have offset of old record: grab next */
92                         if (tdb_rec_read(tdb, tlock->off, rec) == -1)
93                                 goto fail;
94                         tlock->off = rec->next;
95                 }
96
97                 /* Iterate through chain */
98                 while( tlock->off) {
99                         tdb_off_t current;
100                         if (tdb_rec_read(tdb, tlock->off, rec) == -1)
101                                 goto fail;
102
103                         /* Detect infinite loops. From "Shlomi Yaakobovich" <Shlomi@exanet.com>. */
104                         if (tlock->off == rec->next) {
105                                 tdb->ecode = TDB_ERR_CORRUPT;
106                                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_next_lock: loop detected.\n"));
107                                 goto fail;
108                         }
109
110                         if (!TDB_DEAD(rec)) {
111                                 /* Woohoo: we found one! */
112                                 if (tdb_lock_record(tdb, tlock->off) != 0)
113                                         goto fail;
114                                 return tlock->off;
115                         }
116
117                         /* Try to clean dead ones from old traverses */
118                         current = tlock->off;
119                         tlock->off = rec->next;
120                         if (!(tdb->read_only || tdb->traverse_read) && 
121                             tdb_do_delete(tdb, current, rec) != 0)
122                                 goto fail;
123                 }
124                 tdb_unlock(tdb, tlock->hash, tlock->lock_rw);
125                 want_next = 0;
126         }
127         /* We finished iteration without finding anything */
128         tdb->ecode = TDB_SUCCESS;
129         return 0;
130
131  fail:
132         tlock->off = 0;
133         if (tdb_unlock(tdb, tlock->hash, tlock->lock_rw) != 0)
134                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_next_lock: On error unlock failed!\n"));
135         return TDB_NEXT_LOCK_ERR;
136 }
137
138 /* traverse the entire database - calling fn(tdb, key, data) on each element.
139    return -1 on error or the record count traversed
140    if fn is NULL then it is not called
141    a non-zero return value from fn() indicates that the traversal should stop
142   */
143 static int tdb_traverse_internal(struct tdb_context *tdb, 
144                                  tdb_traverse_func fn, void *private_data,
145                                  struct tdb_traverse_lock *tl)
146 {
147         TDB_DATA key, dbuf;
148         struct tdb_record rec;
149         int ret = 0, count = 0;
150         tdb_off_t off;
151
152         /* This was in the initializaton, above, but the IRIX compiler
153          * did not like it.  crh
154          */
155         tl->next = tdb->travlocks.next;
156
157         /* fcntl locks don't stack: beware traverse inside traverse */
158         tdb->travlocks.next = tl;
159
160         /* tdb_next_lock places locks on the record returned, and its chain */
161         while ((off = tdb_next_lock(tdb, tl, &rec)) != 0) {
162                 if (off == TDB_NEXT_LOCK_ERR) {
163                         ret = -1;
164                         goto out;
165                 }
166                 count++;
167                 /* now read the full record */
168                 key.dptr = tdb_alloc_read(tdb, tl->off + sizeof(rec), 
169                                           rec.key_len + rec.data_len);
170                 if (!key.dptr) {
171                         ret = -1;
172                         if (tdb_unlock(tdb, tl->hash, tl->lock_rw) != 0)
173                                 goto out;
174                         if (tdb_unlock_record(tdb, tl->off) != 0)
175                                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_traverse: key.dptr == NULL and unlock_record failed!\n"));
176                         goto out;
177                 }
178                 key.dsize = rec.key_len;
179                 dbuf.dptr = key.dptr + rec.key_len;
180                 dbuf.dsize = rec.data_len;
181
182                 tdb_trace_1rec_retrec(tdb, "traverse", key, dbuf);
183
184                 /* Drop chain lock, call out */
185                 if (tdb_unlock(tdb, tl->hash, tl->lock_rw) != 0) {
186                         ret = -1;
187                         SAFE_FREE(key.dptr);
188                         goto out;
189                 }
190                 if (fn && fn(tdb, key, dbuf, private_data)) {
191                         /* They want us to terminate traversal */
192                         tdb_trace_ret(tdb, "tdb_traverse_end", count);
193                         if (tdb_unlock_record(tdb, tl->off) != 0) {
194                                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_traverse: unlock_record failed!\n"));;
195                                 ret = -1;
196                         }
197                         SAFE_FREE(key.dptr);
198                         goto out;
199                 }
200                 SAFE_FREE(key.dptr);
201         }
202         tdb_trace(tdb, "tdb_traverse_end");
203 out:
204         tdb->travlocks.next = tl->next;
205         if (ret < 0)
206                 return -1;
207         else
208                 return count;
209 }
210
211
212 /*
213   a write style traverse - temporarily marks the db read only
214 */
215 int tdb_traverse_read(struct tdb_context *tdb, 
216                       tdb_traverse_func fn, void *private_data)
217 {
218         struct tdb_traverse_lock tl = { NULL, 0, 0, F_RDLCK };
219         int ret;
220
221         /* we need to get a read lock on the transaction lock here to
222            cope with the lock ordering semantics of solaris10 */
223         if (tdb_transaction_lock(tdb, F_RDLCK)) {
224                 return -1;
225         }
226
227         tdb->traverse_read++;
228         tdb_trace(tdb, "tdb_traverse_read_start");
229         ret = tdb_traverse_internal(tdb, fn, private_data, &tl);
230         tdb->traverse_read--;
231
232         tdb_transaction_unlock(tdb, F_RDLCK);
233
234         return ret;
235 }
236
237 /*
238   a write style traverse - needs to get the transaction lock to
239   prevent deadlocks
240
241   WARNING: The data buffer given to the callback fn does NOT meet the
242   alignment restrictions malloc gives you.
243 */
244 int tdb_traverse(struct tdb_context *tdb, 
245                  tdb_traverse_func fn, void *private_data)
246 {
247         struct tdb_traverse_lock tl = { NULL, 0, 0, F_WRLCK };
248         int ret;
249
250         if (tdb->read_only || tdb->traverse_read) {
251                 return tdb_traverse_read(tdb, fn, private_data);
252         }
253
254         if (tdb_transaction_lock(tdb, F_WRLCK)) {
255                 return -1;
256         }
257
258         tdb->traverse_write++;
259         tdb_trace(tdb, "tdb_traverse_start");
260         ret = tdb_traverse_internal(tdb, fn, private_data, &tl);
261         tdb->traverse_write--;
262
263         tdb_transaction_unlock(tdb, F_WRLCK);
264
265         return ret;
266 }
267
268
269 /* find the first entry in the database and return its key */
270 TDB_DATA tdb_firstkey(struct tdb_context *tdb)
271 {
272         TDB_DATA key;
273         struct tdb_record rec;
274         tdb_off_t off;
275
276         /* release any old lock */
277         if (tdb_unlock_record(tdb, tdb->travlocks.off) != 0)
278                 return tdb_null;
279         tdb->travlocks.off = tdb->travlocks.hash = 0;
280         tdb->travlocks.lock_rw = F_RDLCK;
281
282         /* Grab first record: locks chain and returned record. */
283         off = tdb_next_lock(tdb, &tdb->travlocks, &rec);
284         if (off == 0 || off == TDB_NEXT_LOCK_ERR) {
285                 tdb_trace_retrec(tdb, "tdb_firstkey", tdb_null);
286                 return tdb_null;
287         }
288         /* now read the key */
289         key.dsize = rec.key_len;
290         key.dptr =tdb_alloc_read(tdb,tdb->travlocks.off+sizeof(rec),key.dsize);
291
292         tdb_trace_retrec(tdb, "tdb_firstkey", key);
293
294         /* Unlock the hash chain of the record we just read. */
295         if (tdb_unlock(tdb, tdb->travlocks.hash, tdb->travlocks.lock_rw) != 0)
296                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_firstkey: error occurred while tdb_unlocking!\n"));
297         return key;
298 }
299
300 /* find the next entry in the database, returning its key */
301 TDB_DATA tdb_nextkey(struct tdb_context *tdb, TDB_DATA oldkey)
302 {
303         uint32_t oldhash;
304         TDB_DATA key = tdb_null;
305         struct tdb_record rec;
306         unsigned char *k = NULL;
307         tdb_off_t off;
308
309         /* Is locked key the old key?  If so, traverse will be reliable. */
310         if (tdb->travlocks.off) {
311                 if (tdb_lock(tdb,tdb->travlocks.hash,tdb->travlocks.lock_rw))
312                         return tdb_null;
313                 if (tdb_rec_read(tdb, tdb->travlocks.off, &rec) == -1
314                     || !(k = tdb_alloc_read(tdb,tdb->travlocks.off+sizeof(rec),
315                                             rec.key_len))
316                     || memcmp(k, oldkey.dptr, oldkey.dsize) != 0) {
317                         /* No, it wasn't: unlock it and start from scratch */
318                         if (tdb_unlock_record(tdb, tdb->travlocks.off) != 0) {
319                                 tdb_trace_1rec_retrec(tdb, "tdb_nextkey",
320                                                       oldkey, tdb_null);
321                                 SAFE_FREE(k);
322                                 return tdb_null;
323                         }
324                         if (tdb_unlock(tdb, tdb->travlocks.hash, tdb->travlocks.lock_rw) != 0) {
325                                 SAFE_FREE(k);
326                                 return tdb_null;
327                         }
328                         tdb->travlocks.off = 0;
329                 }
330
331                 SAFE_FREE(k);
332         }
333
334         if (!tdb->travlocks.off) {
335                 /* No previous element: do normal find, and lock record */
336                 tdb->travlocks.off = tdb_find_lock_hash(tdb, oldkey, tdb->hash_fn(&oldkey), tdb->travlocks.lock_rw, &rec);
337                 if (!tdb->travlocks.off) {
338                         tdb_trace_1rec_retrec(tdb, "tdb_nextkey", oldkey, tdb_null);
339                         return tdb_null;
340                 }
341                 tdb->travlocks.hash = BUCKET(rec.full_hash);
342                 if (tdb_lock_record(tdb, tdb->travlocks.off) != 0) {
343                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_nextkey: lock_record failed (%s)!\n", strerror(errno)));
344                         return tdb_null;
345                 }
346         }
347         oldhash = tdb->travlocks.hash;
348
349         /* Grab next record: locks chain and returned record,
350            unlocks old record */
351         off = tdb_next_lock(tdb, &tdb->travlocks, &rec);
352         if (off != TDB_NEXT_LOCK_ERR && off != 0) {
353                 key.dsize = rec.key_len;
354                 key.dptr = tdb_alloc_read(tdb, tdb->travlocks.off+sizeof(rec),
355                                           key.dsize);
356                 /* Unlock the chain of this new record */
357                 if (tdb_unlock(tdb, tdb->travlocks.hash, tdb->travlocks.lock_rw) != 0)
358                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_nextkey: WARNING tdb_unlock failed!\n"));
359         }
360         /* Unlock the chain of old record */
361         if (tdb_unlock(tdb, BUCKET(oldhash), tdb->travlocks.lock_rw) != 0)
362                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_nextkey: WARNING tdb_unlock failed!\n"));
363         tdb_trace_1rec_retrec(tdb, "tdb_nextkey", oldkey, key);
364         return key;
365 }
366