[SCM] Samba Shared Repository - branch v3-2-test updated -
initial-v3-2-unstable-1280-gc6609c0
Jeremy Allison
jra at samba.org
Fri Jan 11 23:11:39 GMT 2008
The branch, v3-2-test has been updated
via c6609c042b128e7d63eb64cfdfb0f6b65cb59d76 (commit)
from 75acdb54a454ffda9d422fcafb573c8f5581d2e8 (commit)
http://gitweb.samba.org/?p=samba.git;a=shortlog;h=v3-2-test
- Log -----------------------------------------------------------------
commit c6609c042b128e7d63eb64cfdfb0f6b65cb59d76
Author: Jeremy Allison <jra at samba.org>
Date: Fri Jan 11 15:08:37 2008 -0800
Sync tdb with the tdb changes in ctdb. Spoke to tridge about
this. Fixes insidious problem with order n^2 freelist merging.
Jeremy.
-----------------------------------------------------------------------
Summary of changes:
source/lib/tdb/common/freelist.c | 101 ++++++++---
source/lib/tdb/common/io.c | 20 +-
source/lib/tdb/common/lock.c | 7 +
source/lib/tdb/common/open.c | 16 ++-
source/lib/tdb/common/tdb.c | 72 +++++++-
source/lib/tdb/common/tdb_private.h | 8 +-
source/lib/tdb/common/transaction.c | 343 +++++++++++++++++------------------
source/lib/tdb/common/traverse.c | 3 +
source/lib/tdb/configure.ac | 2 +-
source/lib/tdb/include/tdb.h | 4 +
10 files changed, 358 insertions(+), 218 deletions(-)
Changeset truncated at 500 lines:
diff --git a/source/lib/tdb/common/freelist.c b/source/lib/tdb/common/freelist.c
index b109643..c086c15 100644
--- a/source/lib/tdb/common/freelist.c
+++ b/source/lib/tdb/common/freelist.c
@@ -27,6 +27,12 @@
#include "tdb_private.h"
+/* 'right' merges can involve O(n^2) cost when combined with a
+ traverse, so they are disabled until we find a way to do them in
+ O(1) time
+*/
+#define USE_RIGHT_MERGES 0
+
/* read a freelist record and check for simple errors */
int tdb_rec_free_read(struct tdb_context *tdb, tdb_off_t off, struct list_struct *rec)
{
@@ -56,7 +62,7 @@ int tdb_rec_free_read(struct tdb_context *tdb, tdb_off_t off, struct list_struct
}
-
+#if USE_RIGHT_MERGES
/* Remove an element from the freelist. Must have alloc lock. */
static int remove_from_freelist(struct tdb_context *tdb, tdb_off_t off, tdb_off_t next)
{
@@ -75,6 +81,7 @@ static int remove_from_freelist(struct tdb_context *tdb, tdb_off_t off, tdb_off_
TDB_LOG((tdb, TDB_DEBUG_FATAL,"remove_from_freelist: not on list at off=%d\n", off));
return TDB_ERRCODE(TDB_ERR_CORRUPT, -1);
}
+#endif
/* update a record tailer (must hold allocation lock) */
@@ -93,8 +100,6 @@ static int update_tailer(struct tdb_context *tdb, tdb_off_t offset,
neccessary. */
int tdb_free(struct tdb_context *tdb, tdb_off_t offset, struct list_struct *rec)
{
- tdb_off_t right, left;
-
/* Allocation and tailer lock */
if (tdb_lock(tdb, -1, F_WRLCK) != 0)
return -1;
@@ -105,9 +110,10 @@ int tdb_free(struct tdb_context *tdb, tdb_off_t offset, struct list_struct *rec)
goto fail;
}
+#if USE_RIGHT_MERGES
/* Look right first (I'm an Australian, dammit) */
- right = offset + sizeof(*rec) + rec->rec_len;
- if (right + sizeof(*rec) <= tdb->map_size) {
+ if (offset + sizeof(*rec) + rec->rec_len + sizeof(*rec) <= tdb->map_size) {
+ tdb_off_t right = offset + sizeof(*rec) + rec->rec_len;
struct list_struct r;
if (tdb->methods->tdb_read(tdb, right, &r, sizeof(r), DOCONV()) == -1) {
@@ -122,13 +128,18 @@ int tdb_free(struct tdb_context *tdb, tdb_off_t offset, struct list_struct *rec)
goto left;
}
rec->rec_len += sizeof(r) + r.rec_len;
+ if (update_tailer(tdb, offset, rec) == -1) {
+ TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: update_tailer failed at %u\n", offset));
+ goto fail;
+ }
}
}
-
left:
+#endif
+
/* Look left */
- left = offset - sizeof(tdb_off_t);
- if (left > TDB_DATA_START(tdb->header.hash_size)) {
+ if (offset - sizeof(tdb_off_t) > TDB_DATA_START(tdb->header.hash_size)) {
+ tdb_off_t left = offset - sizeof(tdb_off_t);
struct list_struct l;
tdb_off_t leftsize;
@@ -145,7 +156,12 @@ left:
left = offset - leftsize;
- /* Now read in record */
+ if (leftsize > offset ||
+ left < TDB_DATA_START(tdb->header.hash_size)) {
+ goto update;
+ }
+
+ /* Now read in the left record */
if (tdb->methods->tdb_read(tdb, left, &l, sizeof(l), DOCONV()) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: left read failed at %u (%u)\n", left, leftsize));
goto update;
@@ -153,21 +169,24 @@ left:
/* If it's free, expand to include it. */
if (l.magic == TDB_FREE_MAGIC) {
- if (remove_from_freelist(tdb, left, l.next) == -1) {
- TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: left free failed at %u\n", left));
- goto update;
- } else {
- offset = left;
- rec->rec_len += leftsize;
+ /* we now merge the new record into the left record, rather than the other
+ way around. This makes the operation O(1) instead of O(n). This change
+ prevents traverse from being O(n^2) after a lot of deletes */
+ l.rec_len += sizeof(*rec) + rec->rec_len;
+ if (tdb_rec_write(tdb, left, &l) == -1) {
+ TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: update_left failed at %u\n", left));
+ goto fail;
}
+ if (update_tailer(tdb, left, &l) == -1) {
+ TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: update_tailer failed at %u\n", offset));
+ goto fail;
+ }
+ tdb_unlock(tdb, -1, F_WRLCK);
+ return 0;
}
}
update:
- if (update_tailer(tdb, offset, rec) == -1) {
- TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: update_tailer failed at %u\n", offset));
- goto fail;
- }
/* Now, prepend to free list */
rec->magic = TDB_FREE_MAGIC;
@@ -261,6 +280,7 @@ tdb_off_t tdb_allocate(struct tdb_context *tdb, tdb_len_t length, struct list_st
tdb_off_t rec_ptr, last_ptr;
tdb_len_t rec_len;
} bestfit;
+ float multiplier = 1.0;
if (tdb_lock(tdb, -1, F_WRLCK) == -1)
return 0;
@@ -295,18 +315,27 @@ tdb_off_t tdb_allocate(struct tdb_context *tdb, tdb_len_t length, struct list_st
bestfit.rec_len = rec->rec_len;
bestfit.rec_ptr = rec_ptr;
bestfit.last_ptr = last_ptr;
- /* consider a fit to be good enough if
- we aren't wasting more than half
- the space */
- if (bestfit.rec_len < 2*length) {
- break;
- }
}
}
/* move to the next record */
last_ptr = rec_ptr;
rec_ptr = rec->next;
+
+ /* if we've found a record that is big enough, then
+ stop searching if its also not too big. The
+ definition of 'too big' changes as we scan
+ through */
+ if (bestfit.rec_len > 0 &&
+ bestfit.rec_len < length * multiplier) {
+ break;
+ }
+
+ /* this multiplier means we only extremely rarely
+ search more than 50 or so records. At 50 records we
+ accept records up to 11 times larger than what we
+ want */
+ multiplier *= 1.05;
}
if (bestfit.rec_ptr != 0) {
@@ -328,3 +357,25 @@ tdb_off_t tdb_allocate(struct tdb_context *tdb, tdb_len_t length, struct list_st
return 0;
}
+
+
+/*
+ return the size of the freelist - used to decide if we should repack
+*/
+int tdb_freelist_size(struct tdb_context *tdb)
+{
+ tdb_off_t ptr;
+ int count=0;
+
+ if (tdb_lock(tdb, -1, F_RDLCK) == -1) {
+ return -1;
+ }
+
+ ptr = FREELIST_TOP;
+ while (tdb_ofs_read(tdb, ptr, &ptr) == 0 && ptr != 0) {
+ count++;
+ }
+
+ tdb_unlock(tdb, -1, F_RDLCK);
+ return count;
+}
diff --git a/source/lib/tdb/common/io.c b/source/lib/tdb/common/io.c
index 8ab0768..172ab69 100644
--- a/source/lib/tdb/common/io.c
+++ b/source/lib/tdb/common/io.c
@@ -101,8 +101,8 @@ static int tdb_write(struct tdb_context *tdb, tdb_off_t off,
off+written);
}
if (written == -1) {
- /* Ensure ecode is set for log fn. */
- tdb->ecode = TDB_ERR_IO;
+ /* Ensure ecode is set for log fn. */
+ tdb->ecode = TDB_ERR_IO;
TDB_LOG((tdb, TDB_DEBUG_FATAL,"tdb_write failed at %d "
"len=%d (%s)\n", off, len, strerror(errno)));
return TDB_ERRCODE(TDB_ERR_IO, -1);
@@ -111,8 +111,8 @@ static int tdb_write(struct tdb_context *tdb, tdb_off_t off,
"write %d bytes at %d in two attempts\n",
len, off));
errno = ENOSPC;
- return TDB_ERRCODE(TDB_ERR_IO, -1);
- }
+ return TDB_ERRCODE(TDB_ERR_IO, -1);
+ }
}
return 0;
}
@@ -230,7 +230,7 @@ void tdb_mmap(struct tdb_context *tdb)
says to use for mmap expansion */
static int tdb_expand_file(struct tdb_context *tdb, tdb_off_t size, tdb_off_t addition)
{
- char buf[1024];
+ char buf[8192];
if (tdb->read_only || tdb->traverse_read) {
tdb->ecode = TDB_ERR_RDONLY;
@@ -294,7 +294,7 @@ static int tdb_expand_file(struct tdb_context *tdb, tdb_off_t size, tdb_off_t ad
int tdb_expand(struct tdb_context *tdb, tdb_off_t size)
{
struct list_struct rec;
- tdb_off_t offset;
+ tdb_off_t offset, new_size;
if (tdb_lock(tdb, -1, F_WRLCK) == -1) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "lock failed in tdb_expand\n"));
@@ -304,9 +304,11 @@ int tdb_expand(struct tdb_context *tdb, tdb_off_t size)
/* must know about any previous expansions by another process */
tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
- /* always make room for at least 10 more records, and round
- the database up to a multiple of the page size */
- size = TDB_ALIGN(tdb->map_size + size*10, tdb->page_size) - tdb->map_size;
+ /* always make room for at least 100 more records, and at
+ least 25% more space. Round the database up to a multiple
+ of the page size */
+ new_size = MAX(tdb->map_size + size*100, tdb->map_size * 1.25);
+ size = TDB_ALIGN(new_size, tdb->page_size) - tdb->map_size;
if (!(tdb->flags & TDB_INTERNAL))
tdb_munmap(tdb);
diff --git a/source/lib/tdb/common/lock.c b/source/lib/tdb/common/lock.c
index e3fe888..f156c0f 100644
--- a/source/lib/tdb/common/lock.c
+++ b/source/lib/tdb/common/lock.c
@@ -505,6 +505,9 @@ int tdb_chainunlock_read(struct tdb_context *tdb, TDB_DATA key)
/* record lock stops delete underneath */
int tdb_lock_record(struct tdb_context *tdb, tdb_off_t off)
{
+ if (tdb->global_lock.count) {
+ return 0;
+ }
return off ? tdb->methods->tdb_brlock(tdb, off, F_RDLCK, F_SETLKW, 0, 1) : 0;
}
@@ -537,6 +540,10 @@ int tdb_unlock_record(struct tdb_context *tdb, tdb_off_t off)
struct tdb_traverse_lock *i;
uint32_t count = 0;
+ if (tdb->global_lock.count) {
+ return 0;
+ }
+
if (off == 0)
return 0;
for (i = &tdb->travlocks; i; i = i->next)
diff --git a/source/lib/tdb/common/open.c b/source/lib/tdb/common/open.c
index 6efa482..6bd8fda 100644
--- a/source/lib/tdb/common/open.c
+++ b/source/lib/tdb/common/open.c
@@ -35,7 +35,7 @@ static struct tdb_context *tdbs = NULL;
static unsigned int default_tdb_hash(TDB_DATA *key)
{
uint32_t value; /* Used to compute the hash value. */
- uint32_t i; /* Used to cycle through random values. */
+ uint32_t i; /* Used to cycle through random values. */
/* Set the initial value from the key size. */
for (value = 0x238F13AF * key->dsize, i=0; i < key->dsize; i++)
@@ -90,7 +90,7 @@ static int tdb_new_database(struct tdb_context *tdb, int hash_size)
size -= written;
written = write(tdb->fd, newdb+written, size);
if (written == size) {
- ret = 0;
+ ret = 0;
} else if (written >= 0) {
/* a second incomplete write - we give up.
* guessing the errno... */
@@ -152,6 +152,7 @@ struct tdb_context *tdb_open_ex(const char *name, int hash_size, int tdb_flags,
int rev = 0, locked = 0;
unsigned char *vp;
uint32_t vertest;
+ unsigned v;
if (!(tdb = (struct tdb_context *)calloc(1, sizeof *tdb))) {
/* Can't log this */
@@ -178,7 +179,9 @@ struct tdb_context *tdb_open_ex(const char *name, int hash_size, int tdb_flags,
tdb->page_size = 0x2000;
}
- tdb->max_dead_records = (open_flags & TDB_VOLATILE) ? 5 : 0;
+ if (open_flags & TDB_VOLATILE) {
+ tdb->max_dead_records = 5;
+ }
if ((open_flags & O_ACCMODE) == O_WRONLY) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_open_ex: can't open tdb %s write-only\n",
@@ -213,6 +216,10 @@ struct tdb_context *tdb_open_ex(const char *name, int hash_size, int tdb_flags,
goto fail; /* errno set by open(2) */
}
+ /* on exec, don't inherit the fd */
+ v = fcntl(tdb->fd, F_GETFD, 0);
+ fcntl(tdb->fd, F_SETFD, v | FD_CLOEXEC);
+
/* ensure there is only one process initialising at once */
if (tdb->methods->tdb_brlock(tdb, GLOBAL_LOCK, F_WRLCK, F_SETLKW, 0, 1) == -1) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_open_ex: failed to get global lock on %s: %s\n",
@@ -240,7 +247,7 @@ struct tdb_context *tdb_open_ex(const char *name, int hash_size, int tdb_flags,
/* its not a valid database - possibly initialise it */
if (!(open_flags & O_CREAT) || tdb_new_database(tdb, hash_size) == -1) {
if (errno == 0) {
- errno = EIO; /* ie bad format or something */
+ errno = EIO; /* ie bad format or something */
}
goto fail;
}
@@ -281,6 +288,7 @@ struct tdb_context *tdb_open_ex(const char *name, int hash_size, int tdb_flags,
tdb->map_size = st.st_size;
tdb->device = st.st_dev;
tdb->inode = st.st_ino;
+ tdb->max_dead_records = 0;
tdb_mmap(tdb);
if (locked) {
if (tdb->methods->tdb_brlock(tdb, ACTIVE_LOCK, F_UNLCK, F_SETLK, 0, 1) == -1) {
diff --git a/source/lib/tdb/common/tdb.c b/source/lib/tdb/common/tdb.c
index 0e9d1db..bf3abb7 100644
--- a/source/lib/tdb/common/tdb.c
+++ b/source/lib/tdb/common/tdb.c
@@ -102,8 +102,7 @@ static tdb_off_t tdb_find(struct tdb_context *tdb, TDB_DATA key, uint32_t hash,
}
/* As tdb_find, but if you succeed, keep the lock */
-tdb_off_t tdb_find_lock_hash(struct tdb_context *tdb, TDB_DATA key,
- uint32_t hash, int locktype,
+tdb_off_t tdb_find_lock_hash(struct tdb_context *tdb, TDB_DATA key, uint32_t hash, int locktype,
struct list_struct *rec)
{
uint32_t rec_ptr;
@@ -237,14 +236,15 @@ int tdb_exists(struct tdb_context *tdb, TDB_DATA key)
}
/* actually delete an entry in the database given the offset */
-int tdb_do_delete(struct tdb_context *tdb, tdb_off_t rec_ptr, struct list_struct*rec)
+int tdb_do_delete(struct tdb_context *tdb, tdb_off_t rec_ptr, struct list_struct *rec)
{
tdb_off_t last_ptr, i;
struct list_struct lastrec;
if (tdb->read_only || tdb->traverse_read) return -1;
- if (tdb_write_lock_record(tdb, rec_ptr) == -1) {
+ if (tdb->traverse_write != 0 ||
+ tdb_write_lock_record(tdb, rec_ptr) == -1) {
/* Someone traversing here: mark it as dead */
rec->magic = TDB_DEAD_MAGIC;
return tdb_rec_write(tdb, rec_ptr, rec);
@@ -666,6 +666,16 @@ int tdb_get_flags(struct tdb_context *tdb)
return tdb->flags;
}
+void tdb_add_flags(struct tdb_context *tdb, unsigned flags)
+{
+ tdb->flags |= flags;
+}
+
+void tdb_remove_flags(struct tdb_context *tdb, unsigned flags)
+{
+ tdb->flags &= ~flags;
+}
+
/*
enable sequence number handling on an open tdb
@@ -674,3 +684,57 @@ void tdb_enable_seqnum(struct tdb_context *tdb)
{
tdb->flags |= TDB_SEQNUM;
}
+
+
+/*
+ wipe the entire database, deleting all records. This can be done
+ very fast by using a global lock. The entire data portion of the
+ file becomes a single entry in the freelist.
+ */
+int tdb_wipe_all(struct tdb_context *tdb)
+{
+ int i;
+ tdb_off_t offset = 0;
+ ssize_t data_len;
+
+ if (tdb_lockall(tdb) != 0) {
+ return -1;
+ }
+
+ /* wipe the hashes */
+ for (i=0;i<tdb->header.hash_size;i++) {
+ if (tdb_ofs_write(tdb, TDB_HASH_TOP(i), &offset) == -1) {
+ TDB_LOG((tdb, TDB_DEBUG_FATAL,"tdb_wipe_all: failed to write hash %d\n", i));
+ goto failed;
+ }
+ }
+
+ /* wipe the freelist */
+ if (tdb_ofs_write(tdb, FREELIST_TOP, &offset) == -1) {
+ TDB_LOG((tdb, TDB_DEBUG_FATAL,"tdb_wipe_all: failed to write freelist\n"));
+ goto failed;
+ }
+
+ /* add all the rest of the file to the freelist */
+ data_len = (tdb->map_size - TDB_DATA_START(tdb->header.hash_size)) - sizeof(struct list_struct);
+ if (data_len > 0) {
+ struct list_struct rec;
+ memset(&rec,'\0',sizeof(rec));
+ rec.rec_len = data_len;
+ if (tdb_free(tdb, TDB_DATA_START(tdb->header.hash_size), &rec) == -1) {
+ TDB_LOG((tdb, TDB_DEBUG_FATAL,"tdb_wipe_all: failed to add free record\n"));
+ goto failed;
+ }
+ }
+
+ if (tdb_unlockall(tdb) != 0) {
+ TDB_LOG((tdb, TDB_DEBUG_FATAL,"tdb_wipe_all: failed to unlock\n"));
+ goto failed;
+ }
+
+ return 0;
+
+failed:
+ tdb_unlockall(tdb);
+ return -1;
+}
diff --git a/source/lib/tdb/common/tdb_private.h b/source/lib/tdb/common/tdb_private.h
index 58c30c1..63a6d04 100644
--- a/source/lib/tdb/common/tdb_private.h
+++ b/source/lib/tdb/common/tdb_private.h
@@ -38,6 +38,10 @@
typedef uint32_t tdb_len_t;
typedef uint32_t tdb_off_t;
+#ifndef offsetof
+#define offsetof(t,f) ((unsigned int)&((t *)0)->f)
+#endif
+
#define TDB_MAGIC_FOOD "TDB file\n"
#define TDB_VERSION (0x26011967 + 6)
#define TDB_MAGIC (0x26011999U)
@@ -54,7 +58,7 @@ typedef uint32_t tdb_off_t;
#define TDB_BAD_MAGIC(r) ((r)->magic != TDB_MAGIC && !TDB_DEAD(r))
#define TDB_HASH_TOP(hash) (FREELIST_TOP + (BUCKET(hash)+1)*sizeof(tdb_off_t))
#define TDB_HASHTABLE_SIZE(tdb) ((tdb->header.hash_size+1)*sizeof(tdb_off_t))
-#define TDB_DATA_START(hash_size) TDB_HASH_TOP(hash_size-1)
+#define TDB_DATA_START(hash_size) (TDB_HASH_TOP(hash_size-1) + sizeof(tdb_off_t))
#define TDB_RECOVERY_HEAD offsetof(struct tdb_header, recovery_start)
#define TDB_SEQNUM_OFS offsetof(struct tdb_header, sequence_number)
#define TDB_PAD_BYTE 0x42
@@ -144,6 +148,7 @@ struct tdb_context {
tdb_len_t map_size; /* how much space has been mapped */
int read_only; /* opened read-only */
int traverse_read; /* read-only traversal */
+ int traverse_write; /* read-write traversal */
struct tdb_lock_type global_lock;
int num_lockrecs;
struct tdb_lock_type *lockrecs; /* only real locks, all with count>0 */
@@ -173,7 +178,6 @@ struct tdb_context {
int tdb_munmap(struct tdb_context *tdb);
void tdb_mmap(struct tdb_context *tdb);
int tdb_lock(struct tdb_context *tdb, int list, int ltype);
-int tdb_lock_nonblock(struct tdb_context *tdb, int list, int ltype);
int tdb_unlock(struct tdb_context *tdb, int list, int ltype);
int tdb_brlock(struct tdb_context *tdb, tdb_off_t offset, int rw_type, int lck_type, int probe, size_t len);
int tdb_transaction_lock(struct tdb_context *tdb, int ltype);
diff --git a/source/lib/tdb/common/transaction.c b/source/lib/tdb/common/transaction.c
index 7eaacf7..2da09fa 100644
--- a/source/lib/tdb/common/transaction.c
+++ b/source/lib/tdb/common/transaction.c
@@ -87,13 +87,6 @@
--
Samba Shared Repository
More information about the samba-cvs
mailing list