X-Git-Url: https://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ftdb%2Fcheck.c;h=53fb286613a3e829e8f2aa159109124c3e182503;hp=6eca5537372992a8e5a4c2a02aee5a9a6308ef43;hb=a1ace0cd114e0c588d6ce1b9e7386af11716e0bd;hpb=6f05c2b8ad782f7f602bddad8b72b42c55c32bf9 diff --git a/ccan/tdb/check.c b/ccan/tdb/check.c index 6eca5537..53fb2866 100644 --- a/ccan/tdb/check.c +++ b/ccan/tdb/check.c @@ -1,14 +1,14 @@ - /* + /* Unix SMB/CIFS implementation. trivial database library Copyright (C) Rusty Russell 2009 - + ** NOTE! The following LGPL license applies to the tdb ** library. This does NOT imply that all of Samba is released ** under the LGPL - + This library is free software; you can redistribute it and/or modify it under the terms of the GNU Lesser General Public License as published by the Free Software Foundation; either @@ -25,76 +25,12 @@ #include "tdb_private.h" #include -/* - For each value, we flip F bits in a bitmap of size 2^B. So we can think - of this as a F*B bit hash (this isn't quite true due to hash collisions, - but it seems good enough for F << B). - - Assume that we only have a single error; this is *not* the birthday - problem, since the question is: "does that error hash to the same as - the correct value", ie. a simple 1 in 2^F*B. The chances of detecting - multiple errors is even higher (since we only need to detect one of - them). - - Given that ldb uses a hash size of 10000, using 512 bytes per hash chain - (5M) seems reasonable. With 128 hashes, that's about 1 in a million chance - of missing a single linked list error. -*/ -#define NUM_HASHES 128 -#define BITMAP_BITS (512 * CHAR_BIT) - -/* We use the excellent Jenkins lookup3 hash; this is based on hash_word2. - * See: http://burtleburtle.net/bob/c/lookup3.c - */ -#define rot(x,k) (((x)<<(k)) | ((x)>>(32-(k)))) - -#define final(a,b,c) \ -{ \ - c ^= b; c -= rot(b,14); \ - a ^= c; a -= rot(c,11); \ - b ^= a; b -= rot(a,25); \ - c ^= b; c -= rot(b,16); \ - a ^= c; a -= rot(c,4); \ - b ^= a; b -= rot(a,14); \ - c ^= b; c -= rot(b,24); \ -} - -static void hash(uint32_t key, uint32_t *pc, uint32_t *pb) -{ - uint32_t a,b,c; - - /* Set up the internal state */ - a = b = c = 0xdeadbeef + *pc; - c += *pb; - - a += key; - final(a,b,c); - *pc=c; *pb=b; -} - -static void bit_flip(unsigned char bits[], unsigned int idx) -{ - bits[idx / CHAR_BIT] ^= (1 << (idx % CHAR_BIT)); -} - -static void add_to_hash(unsigned char bits[], tdb_off_t off) -{ - uint32_t h1 = off, h2 = 0; - unsigned int i; - for (i = 0; i < NUM_HASHES / 2; i++) { - hash(off, &h1, &h2); - bit_flip(bits, h1 % BITMAP_BITS); - bit_flip(bits, h2 % BITMAP_BITS); - h2++; - } -} - /* Since we opened it, these shouldn't fail unless it's recent corruption. */ static bool tdb_check_header(struct tdb_context *tdb, tdb_off_t *recovery) { struct tdb_header hdr; - if (tdb->methods->tdb_read(tdb, 0, &hdr, sizeof(hdr), DOCONV()) == -1) + if (tdb->methods->tdb_read(tdb, 0, &hdr, sizeof(hdr), 0) == -1) return false; if (strcmp(hdr.magic_food, TDB_MAGIC_FOOD) != 0) goto corrupt; @@ -103,7 +39,7 @@ static bool tdb_check_header(struct tdb_context *tdb, tdb_off_t *recovery) if (hdr.version != TDB_VERSION) goto corrupt; - if (hdr.rwlocks != 0) + if (hdr.hashcheck != hashcheck(tdb)) goto corrupt; if (hdr.hash_size == 0) @@ -121,41 +57,66 @@ static bool tdb_check_header(struct tdb_context *tdb, tdb_off_t *recovery) corrupt: tdb->ecode = TDB_ERR_CORRUPT; + TDB_LOG((tdb, TDB_DEBUG_ERROR, "Header is corrupt\n")); return false; } +/* Generic record header check. */ static bool tdb_check_record(struct tdb_context *tdb, tdb_off_t off, - const struct list_struct *rec) + const struct tdb_record *rec) { tdb_off_t tailer; /* Check rec->next: 0 or points to record offset, aligned. */ - if (rec->next > 0 && rec->next < TDB_DATA_START(tdb->header.hash_size)) + if (rec->next > 0 && rec->next < TDB_DATA_START(tdb->header.hash_size)){ + TDB_LOG((tdb, TDB_DEBUG_ERROR, + "Record offset %d too small next %d\n", + off, rec->next)); goto corrupt; - if (rec->next + sizeof(*rec) < rec->next) + } + if (rec->next + sizeof(*rec) < rec->next) { + TDB_LOG((tdb, TDB_DEBUG_ERROR, + "Record offset %d too large next %d\n", + off, rec->next)); goto corrupt; - if ((rec->next % TDB_ALIGNMENT) != 0) + } + if ((rec->next % TDB_ALIGNMENT) != 0) { + TDB_LOG((tdb, TDB_DEBUG_ERROR, + "Record offset %d misaligned next %d\n", + off, rec->next)); goto corrupt; - if (tdb->methods->tdb_oob(tdb, rec->next+sizeof(*rec), 1)) + } + if (tdb->methods->tdb_oob(tdb, rec->next+sizeof(*rec), 0)) goto corrupt; /* Check rec_len: similar to rec->next, implies next record. */ - if ((rec->rec_len % TDB_ALIGNMENT) != 0) + if ((rec->rec_len % TDB_ALIGNMENT) != 0) { + TDB_LOG((tdb, TDB_DEBUG_ERROR, + "Record offset %d misaligned length %d\n", + off, rec->rec_len)); goto corrupt; + } /* Must fit tailer. */ - if (rec->rec_len < sizeof(tailer)) + if (rec->rec_len < sizeof(tailer)) { + TDB_LOG((tdb, TDB_DEBUG_ERROR, + "Record offset %d too short length %d\n", + off, rec->rec_len)); goto corrupt; + } /* OOB allows "right at the end" access, so this works for last rec. */ - if (tdb->methods->tdb_oob(tdb, off+sizeof(*rec)+rec->rec_len, 1)) + if (tdb->methods->tdb_oob(tdb, off+sizeof(*rec)+rec->rec_len, 0)) goto corrupt; /* Check tailer. */ if (tdb_ofs_read(tdb, off+sizeof(*rec)+rec->rec_len-sizeof(tailer), &tailer) == -1) goto corrupt; - if (tailer != sizeof(*rec) + rec->rec_len) + if (tailer != sizeof(*rec) + rec->rec_len) { + TDB_LOG((tdb, TDB_DEBUG_ERROR, + "Record offset %d invalid tailer\n", off)); goto corrupt; + } return true; @@ -164,13 +125,15 @@ corrupt: return false; } -static TDB_DATA get_data(struct tdb_context *tdb, tdb_off_t off, tdb_len_t len) +/* Grab some bytes: may copy if can't use mmap. + Caller has already done bounds check. */ +static TDB_DATA get_bytes(struct tdb_context *tdb, + tdb_off_t off, tdb_len_t len) { TDB_DATA d; d.dsize = len; - /* We've already done bounds check here. */ if (tdb->transaction == NULL && tdb->map_ptr != NULL) d.dptr = (unsigned char *)tdb->map_ptr + off; else @@ -178,19 +141,98 @@ static TDB_DATA get_data(struct tdb_context *tdb, tdb_off_t off, tdb_len_t len) return d; } -static void put_data(struct tdb_context *tdb, TDB_DATA d) +/* Frees data if we're not able to simply use mmap. */ +static void put_bytes(struct tdb_context *tdb, TDB_DATA d) { if (tdb->transaction == NULL && tdb->map_ptr != NULL) return; free(d.dptr); } +/* We use the excellent Jenkins lookup3 hash; this is based on hash_word2. + * See: http://burtleburtle.net/bob/c/lookup3.c + */ +#define rot(x,k) (((x)<<(k)) | ((x)>>(32-(k)))) +static void hash(uint32_t key, uint32_t *pc, uint32_t *pb) +{ + uint32_t a,b,c; + + /* Set up the internal state */ + a = b = c = 0xdeadbeef + *pc; + c += *pb; + a += key; + c ^= b; c -= rot(b,14); + a ^= c; a -= rot(c,11); + b ^= a; b -= rot(a,25); + c ^= b; c -= rot(b,16); + a ^= c; a -= rot(c,4); + b ^= a; b -= rot(a,14); + c ^= b; c -= rot(b,24); + *pc=c; *pb=b; +} + +/* + We want to check that all free records are in the free list + (only once), and all free list entries are free records. Similarly + for each hash chain of used records. + + Doing that naively (without walking hash chains, since we want to be + linear) means keeping a list of records which have been seen in each + hash chain, and another of records pointed to (ie. next pointers + from records and the initial hash chain heads). These two lists + should be equal. This will take 8 bytes per record, and require + sorting at the end. + + So instead, we record each offset in a bitmap such a way that + recording it twice will cancel out. Since each offset should appear + exactly twice, the bitmap should be zero at the end. + + The approach was inspired by Bloom Filters (see Wikipedia). For + each value, we flip K bits in a bitmap of size N. The number of + distinct arrangements is: + + N! / (K! * (N-K)!) + + Of course, not all arrangements are actually distinct, but testing + shows this formula to be close enough. + + So, if K == 8 and N == 256, the probability of two things flipping the same + bits is 1 in 409,663,695,276,000. + + Given that ldb uses a hash size of 10000, using 32 bytes per hash chain + (320k) seems reasonable. +*/ +#define NUM_HASHES 8 +#define BITMAP_BITS 256 + +static void bit_flip(unsigned char bits[], unsigned int idx) +{ + bits[idx / CHAR_BIT] ^= (1 << (idx % CHAR_BIT)); +} + +/* We record offsets in a bitmap for the particular chain it should be in. */ +static void record_offset(unsigned char bits[], tdb_off_t off) +{ + uint32_t h1 = off, h2 = 0; + unsigned int i; + + /* We get two good hash values out of jhash2, so we use both. Then + * we keep going to produce further hash values. */ + for (i = 0; i < NUM_HASHES / 2; i++) { + hash(off, &h1, &h2); + bit_flip(bits, h1 % BITMAP_BITS); + bit_flip(bits, h2 % BITMAP_BITS); + h2++; + } +} + +/* Check that an in-use record is valid. */ static bool tdb_check_used_record(struct tdb_context *tdb, tdb_off_t off, - const struct list_struct *rec, + const struct tdb_record *rec, unsigned char **hashes, int (*check)(TDB_DATA, TDB_DATA, void *), - void *private) + void *private_data) { TDB_DATA key, data; @@ -198,83 +240,122 @@ static bool tdb_check_used_record(struct tdb_context *tdb, return false; /* key + data + tailer must fit in record */ - if (rec->key_len + rec->data_len + sizeof(tdb_off_t) > rec->rec_len) + if (rec->key_len + rec->data_len + sizeof(tdb_off_t) > rec->rec_len) { + TDB_LOG((tdb, TDB_DEBUG_ERROR, + "Record offset %d too short for contents\n", off)); return false; + } - key = get_data(tdb, off + sizeof(*rec), rec->key_len); + key = get_bytes(tdb, off + sizeof(*rec), rec->key_len); if (!key.dptr) return false; - if (tdb->hash_fn(&key) != rec->full_hash) + if (tdb->hash_fn(&key) != rec->full_hash) { + TDB_LOG((tdb, TDB_DEBUG_ERROR, + "Record offset %d has incorrect hash\n", off)); goto fail_put_key; + } - add_to_hash(hashes[BUCKET(rec->full_hash)+1], off); + /* Mark this offset as a known value for this hash bucket. */ + record_offset(hashes[BUCKET(rec->full_hash)+1], off); + /* And similarly if the next pointer is valid. */ if (rec->next) - add_to_hash(hashes[BUCKET(rec->full_hash)+1], rec->next); + record_offset(hashes[BUCKET(rec->full_hash)+1], rec->next); - /* If they supply a check function, get data. */ - if (check) { - data = get_data(tdb, off + sizeof(*rec) + rec->key_len, - rec->data_len); + /* If they supply a check function and this record isn't dead, + get data and feed it. */ + if (check && rec->magic != TDB_DEAD_MAGIC) { + data = get_bytes(tdb, off + sizeof(*rec) + rec->key_len, + rec->data_len); if (!data.dptr) goto fail_put_key; - if (check(key, data, private) == -1) + if (check(key, data, private_data) == -1) goto fail_put_data; - put_data(tdb, data); + put_bytes(tdb, data); } - put_data(tdb, key); + put_bytes(tdb, key); return true; fail_put_data: - put_data(tdb, data); + put_bytes(tdb, data); fail_put_key: - put_data(tdb, key); + put_bytes(tdb, key); return false; } +/* Check that an unused record is valid. */ static bool tdb_check_free_record(struct tdb_context *tdb, tdb_off_t off, - const struct list_struct *rec, + const struct tdb_record *rec, unsigned char **hashes) { if (!tdb_check_record(tdb, off, rec)) return false; - add_to_hash(hashes[0], off); + /* Mark this offset as a known value for the free list. */ + record_offset(hashes[0], off); + /* And similarly if the next pointer is valid. */ if (rec->next) - add_to_hash(hashes[0], rec->next); + record_offset(hashes[0], rec->next); return true; } -/* We do this via linear scan, even though it's not 100% accurate. */ +/* Slow, but should be very rare. */ +static size_t dead_space(struct tdb_context *tdb, tdb_off_t off) +{ + size_t len; + + for (len = 0; off + len < tdb->map_size; len++) { + char c; + if (tdb->methods->tdb_read(tdb, off, &c, 1, 0)) + return 0; + if (c != 0 && c != 0x42) + break; + } + return len; +} + int tdb_check(struct tdb_context *tdb, - int (*check)(TDB_DATA key, TDB_DATA data, void *private), - void *private) + int (*check)(TDB_DATA key, TDB_DATA data, void *private_data), + void *private_data) { unsigned int h; unsigned char **hashes; tdb_off_t off, recovery_start; - struct list_struct rec; + struct tdb_record rec; bool found_recovery = false; - - if (tdb_lockall(tdb) == -1) - return -1; + tdb_len_t dead; + bool locked; + + /* Read-only databases use no locking at all: it's best-effort. + * We may have a write lock already, so skip that case too. */ + if (tdb->read_only || tdb->allrecord_lock.count != 0) { + locked = false; + } else { + if (tdb_lockall_read(tdb) == -1) + return -1; + locked = true; + } /* Make sure we know true size of the underlying file. */ tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1); + /* Header must be OK: also gets us the recovery ptr, if any. */ if (!tdb_check_header(tdb, &recovery_start)) goto unlock; + /* We should have the whole header, too. */ if (tdb->map_size < TDB_DATA_START(tdb->header.hash_size)) { tdb->ecode = TDB_ERR_CORRUPT; + TDB_LOG((tdb, TDB_DEBUG_ERROR, "File too short for hashes\n")); goto unlock; } /* One big malloc: pointers then bit arrays. */ - hashes = calloc(1, sizeof(hashes[0]) * (1+tdb->header.hash_size) + hashes = (unsigned char **)calloc( + 1, sizeof(hashes[0]) * (1+tdb->header.hash_size) + BITMAP_BITS / CHAR_BIT * (1+tdb->header.hash_size)); if (!hashes) { tdb->ecode = TDB_ERR_OOM; @@ -286,15 +367,16 @@ int tdb_check(struct tdb_context *tdb, for (h = 1; h < 1+tdb->header.hash_size; h++) hashes[h] = hashes[h-1] + BITMAP_BITS / CHAR_BIT; - /* Freelist and hash headers are all in a row. */ + /* Freelist and hash headers are all in a row: read them. */ for (h = 0; h < 1+tdb->header.hash_size; h++) { if (tdb_ofs_read(tdb, FREELIST_TOP + h*sizeof(tdb_off_t), &off) == -1) goto free; if (off) - add_to_hash(hashes[h], off); + record_offset(hashes[h], off); } + /* For each record, read it in and check it's ok. */ for (off = TDB_DATA_START(tdb->header.hash_size); off < tdb->map_size; off += sizeof(rec) + rec.rec_len) { @@ -305,20 +387,44 @@ int tdb_check(struct tdb_context *tdb, case TDB_MAGIC: case TDB_DEAD_MAGIC: if (!tdb_check_used_record(tdb, off, &rec, hashes, - check, private)) + check, private_data)) goto free; break; case TDB_FREE_MAGIC: if (!tdb_check_free_record(tdb, off, &rec, hashes)) goto free; break; + /* If we crash after ftruncate, we can get zeroes or fill. */ + case TDB_RECOVERY_INVALID_MAGIC: + case 0x42424242: + if (recovery_start == off) { + found_recovery = true; + break; + } + dead = dead_space(tdb, off); + if (dead < sizeof(rec)) + goto corrupt; + + TDB_LOG((tdb, TDB_DEBUG_ERROR, + "Dead space at %d-%d (of %u)\n", + off, off + dead, tdb->map_size)); + rec.rec_len = dead - sizeof(rec); + break; case TDB_RECOVERY_MAGIC: - if (recovery_start != off) + if (recovery_start != off) { + TDB_LOG((tdb, TDB_DEBUG_ERROR, + "Unexpected recovery record at offset %d\n", + off)); goto free; + } found_recovery = true; break; - default: + default: ; + corrupt: tdb->ecode = TDB_ERR_CORRUPT; + TDB_LOG((tdb, TDB_DEBUG_ERROR, + "Bad magic 0x%x at offset %d\n", + rec.magic, off)); goto free; } } @@ -330,28 +436,32 @@ int tdb_check(struct tdb_context *tdb, for (i = 0; i < BITMAP_BITS / CHAR_BIT; i++) { if (hashes[h][i] != 0) { tdb->ecode = TDB_ERR_CORRUPT; + TDB_LOG((tdb, TDB_DEBUG_ERROR, + "Hashes do not match records\n")); goto free; } } } - /* We must have found recovery area. */ - if (recovery_start != 0 && !found_recovery) + /* We must have found recovery area if there was one. */ + if (recovery_start != 0 && !found_recovery) { + TDB_LOG((tdb, TDB_DEBUG_ERROR, + "Expected a recovery area at %u\n", + recovery_start)); goto free; + } free(hashes); - tdb_unlockall(tdb); + if (locked) { + tdb_unlockall_read(tdb); + } return 0; free: free(hashes); unlock: - tdb_unlockall(tdb); + if (locked) { + tdb_unlockall_read(tdb); + } return -1; } - - - - - -