#define CACHEFLUSH(addr, bytes, cache)
#endif
+
#include <errno.h>
#include <limits.h>
#include <stddef.h>
#include <inttypes.h>
#include <stdio.h>
#include <stdlib.h>
-#include <malloc.h>
#include <string.h>
#include <time.h>
#include <unistd.h>
+#if defined(__sun)
+/* Most platforms have posix_memalign, older may only have memalign */
+#define HAVE_MEMALIGN 1
+#include <malloc.h>
+#endif
+
#if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
#include <netinet/in.h>
#include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
void *me_pbuf; /**< scratch area for DUPSORT put() */
MDB_txn *me_txn; /**< current write transaction */
+ MDB_txn *me_txn0; /**< prealloc'd write transaction */
size_t me_mapsize; /**< size of the data memory map */
off_t me_size; /**< current file size */
pgno_t me_maxpg; /**< me_mapsize / me_psize */
uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
pthread_key_t me_txkey; /**< thread-key for readers */
+ txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
# define me_pglast me_pgstate.mf_pglast
# define me_pghead me_pgstate.mf_pghead
{
int loose = 0;
pgno_t pgno = mp->mp_pgno;
+ MDB_txn *txn = mc->mc_txn;
if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
- if (mc->mc_txn->mt_parent) {
- MDB_ID2 *dl = mc->mc_txn->mt_u.dirty_list;
+ if (txn->mt_parent) {
+ MDB_ID2 *dl = txn->mt_u.dirty_list;
/* If txn has a parent, make sure the page is in our
* dirty list.
*/
if (x <= dl[0].mid && dl[x].mid == pgno) {
if (mp != dl[x].mptr) { /* bad cursor? */
mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
- mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
+ txn->mt_flags |= MDB_TXN_ERROR;
return MDB_CORRUPTED;
}
/* ok, it's ours */
if (loose) {
DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
mp->mp_pgno));
- NEXT_LOOSE_PAGE(mp) = mc->mc_txn->mt_loose_pgs;
- mc->mc_txn->mt_loose_pgs = mp;
- mc->mc_txn->mt_loose_count++;
+ NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
+ txn->mt_loose_pgs = mp;
+ txn->mt_loose_count++;
mp->mp_flags |= P_LOOSE;
} else {
- int rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, pgno);
+ int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
if (rc)
return rc;
}
#else
enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
#endif
- int rc, retry = num * 20;
+ int rc, retry = num * 60;
MDB_txn *txn = mc->mc_txn;
MDB_env *env = txn->mt_env;
pgno_t pgno, *mop = env->me_pghead;
txnid_t oldest = 0, last;
MDB_cursor_op op;
MDB_cursor m2;
+ int found_old = 0;
/* If there are any loose pages, just use them */
if (num == 1 && txn->mt_loose_pgs) {
if (op == MDB_FIRST) { /* 1st iteration */
/* Prepare to fetch more and coalesce */
- oldest = mdb_find_oldest(txn);
last = env->me_pglast;
+ oldest = env->me_pgoldest;
mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
if (last) {
op = MDB_SET_RANGE;
last++;
/* Do not fetch more if the record will be too recent */
- if (oldest <= last)
- break;
+ if (oldest <= last) {
+ if (!found_old) {
+ oldest = mdb_find_oldest(txn);
+ env->me_pgoldest = oldest;
+ found_old = 1;
+ }
+ if (oldest <= last)
+ break;
+ }
rc = mdb_cursor_get(&m2, &key, NULL, op);
if (rc) {
if (rc == MDB_NOTFOUND)
goto fail;
}
last = *(txnid_t*)key.mv_data;
- if (oldest <= last)
- break;
+ if (oldest <= last) {
+ if (!found_old) {
+ oldest = mdb_find_oldest(txn);
+ env->me_pgoldest = oldest;
+ found_old = 1;
+ }
+ if (oldest <= last)
+ break;
+ }
np = m2.mc_pg[m2.mc_top];
leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
}
size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
if (!(flags & MDB_RDONLY)) {
+ if (!parent) {
+ txn = env->me_txn0;
+ txn->mt_flags = 0;
+ goto ok;
+ }
size += env->me_maxdbs * sizeof(MDB_cursor *);
/* child txns use parent's dbiseqs */
if (!parent)
}
txn->mt_env = env;
+ok:
if (parent) {
unsigned int i;
txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
} else {
rc = mdb_txn_renew0(txn);
}
- if (rc)
- free(txn);
- else {
+ if (rc) {
+ if (txn != env->me_txn0)
+ free(txn);
+ } else {
*ret = txn;
DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
txn->mt_u.reader->mr_pid = 0;
- free(txn);
+ if (txn != txn->mt_env->me_txn0)
+ free(txn);
}
/** Save the freelist as of this transaction to the freeDB.
return rc;
}
- if (!env->me_pghead) {
+ if (!env->me_pghead && txn->mt_loose_pgs) {
/* Put loose page numbers in mt_free_pgs, since
* we may be unable to return them to me_pghead.
*/
if (env->me_txns)
UNLOCK_MUTEX_W(env);
- free(txn);
+ if (txn != env->me_txn0)
+ free(txn);
return MDB_SUCCESS;
env->me_flags |= MDB_FATAL_ERROR;
return rc;
}
-done:
/* MIPS has cache coherency issues, this is a no-op everywhere else */
- if (!(env->me_flags & MDB_WRITEMAP)) {
- CACHEFLUSH(env->me_map + off, len, DCACHE);
- }
+ CACHEFLUSH(env->me_map + off, len, DCACHE);
+done:
/* Memory ordering issues are irrelevant; since the entire writer
* is wrapped by wmutex, all of these changes will become visible
* after the wmutex is unlocked. Since the DB is multi-version,
if (!((flags & MDB_RDONLY) ||
(env->me_pbuf = calloc(1, env->me_psize))))
rc = ENOMEM;
+ if (!(flags & MDB_RDONLY)) {
+ MDB_txn *txn;
+ int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
+ (sizeof(MDB_db)+sizeof(MDB_cursor)+sizeof(unsigned int)+1);
+ txn = calloc(1, size);
+ if (txn) {
+ txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
+ txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
+ txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
+ txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
+ txn->mt_env = env;
+ env->me_txn0 = txn;
+ } else {
+ rc = ENOMEM;
+ }
+ }
}
leave:
free(env->me_dbxs);
free(env->me_path);
free(env->me_dirty_list);
+ free(env->me_txn0);
mdb_midl_free(env->me_free_pgs);
if (env->me_flags & MDB_ENV_TXKEY) {
MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
MDB_GET_KEY(leaf, key);
- if (data)
- rc = mdb_node_read(mc->mc_txn, leaf, data);
+ rc = mdb_node_read(mc->mc_txn, leaf, data);
break;
}
}
#else
pthread_mutex_init(&my.mc_mutex, NULL);
pthread_cond_init(&my.mc_cond, NULL);
+#ifdef HAVE_MEMALIGN
my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
if (my.mc_wbuf[0] == NULL)
return errno;
+#else
+ rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
+ if (rc)
+ return rc;
+#endif
#endif
memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;