*/
#define DKEY(x) mdb_dkey(x, kbuf)
#else
-#define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
+#define DKBUF
#define DKEY(x) 0
#endif
*/
#define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
- /** Set the \b node's key into \b key, if requested. */
-#define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
- (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
+ /** Set the \b node's key into \b keyptr, if requested. */
+#define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
+ (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
+
+ /** Set the \b node's key into \b key. */
+#define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
/** Information about a single database in the environment. */
typedef struct MDB_db {
dl[0].mid = 0;
}
-/* Set or clear P_KEEP in non-overflow, non-sub pages in this txn's cursors.
+/* Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
* @param[in] mc A cursor handle for the current operation.
* @param[in] pflags Flags of the pages to update:
* P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
+ * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
+ * @return 0 on success, non-zero on failure.
*/
-static void
-mdb_cursorpages_mark(MDB_cursor *mc, unsigned pflags)
+static int
+mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
{
MDB_txn *txn = mc->mc_txn;
MDB_cursor *m3;
MDB_xcursor *mx;
+ MDB_page *dp;
unsigned i, j;
+ int rc = MDB_SUCCESS, level;
+ /* Mark pages seen by cursors */
if (mc->mc_flags & C_UNTRACK)
mc = NULL; /* will find mc in mt_cursors */
for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
if (i == 0)
break;
}
+
+ if (all) {
+ /* Mark dirty root pages */
+ for (i=0; i<txn->mt_numdbs; i++) {
+ if (txn->mt_dbflags[i] & DB_DIRTY) {
+ pgno_t pgno = txn->mt_dbs[i].md_root;
+ if (pgno == P_INVALID)
+ continue;
+ if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
+ break;
+ if ((dp->mp_flags & (P_DIRTY|P_KEEP)) == pflags && level <= 1)
+ dp->mp_flags ^= P_KEEP;
+ }
+ }
+ }
+
+ return rc;
}
-static int mdb_page_flush(MDB_txn *txn);
+static int mdb_page_flush(MDB_txn *txn, int keep);
/** Spill pages from the dirty list back to disk.
* This is intended to prevent running into #MDB_TXN_FULL situations,
MDB_txn *txn = m0->mc_txn;
MDB_page *dp;
MDB_ID2L dl = txn->mt_u.dirty_list;
- unsigned int i, j;
- int rc, level;
+ unsigned int i, j, need;
+ int rc;
if (m0->mc_flags & C_SUB)
return MDB_SUCCESS;
if (key)
i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
i += i; /* double it for good measure */
+ need = i;
if (txn->mt_dirty_room > i)
return MDB_SUCCESS;
return ENOMEM;
}
- /* Mark all the dirty root pages we want to preserve */
- for (i=0; i<txn->mt_numdbs; i++) {
- if (txn->mt_dbflags[i] & DB_DIRTY) {
- pgno_t pgno = txn->mt_dbs[i].md_root;
- if (pgno == P_INVALID)
- continue;
- if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
- goto done;
- if ((dp->mp_flags & P_DIRTY) && level <= 1)
- dp->mp_flags |= P_KEEP;
- }
- }
+ /* Preserve pages which may soon be dirtied again */
+ if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
+ goto done;
- /* Preserve pages used by cursors */
- mdb_cursorpages_mark(m0, P_DIRTY);
+ /* Less aggressive spill - we originally spilled the entire dirty list,
+ * with a few exceptions for cursor pages and DB root pages. But this
+ * turns out to be a lot of wasted effort because in a large txn many
+ * of those pages will need to be used again. So now we spill only 1/8th
+ * of the dirty pages. Testing revealed this to be a good tradeoff,
+ * better than 1/2, 1/4, or 1/10.
+ */
+ if (need < MDB_IDL_UM_MAX / 8)
+ need = MDB_IDL_UM_MAX / 8;
/* Save the page IDs of all the pages we're flushing */
- for (i=1; i<=dl[0].mid; i++) {
+ /* flush from the tail forward, this saves a lot of shifting later on. */
+ for (i=dl[0].mid; i && need; i--) {
dp = dl[i].mptr;
if (dp->mp_flags & P_KEEP)
continue;
}
if ((rc = mdb_midl_append(&txn->mt_spill_pgs, dl[i].mid)))
goto done;
+ need--;
}
mdb_midl_sort(txn->mt_spill_pgs);
- rc = mdb_page_flush(txn);
+ /* Flush the spilled part of dirty list */
+ if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
+ goto done;
- mdb_cursorpages_mark(m0, P_DIRTY|P_KEEP);
+ /* Reset any dirty pages we kept that page_flush didn't see */
+ rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
done:
if (rc == 0) {
return rc;
}
-/** Flush dirty pages to the map, after clearing their dirty flag.
+/** Flush (some) dirty pages to the map, after clearing their dirty flag.
+ * @param[in] txn the transaction that's being committed
+ * @param[in] keep number of initial pages in dirty_list to keep dirty.
+ * @return 0 on success, non-zero on failure.
*/
static int
-mdb_page_flush(MDB_txn *txn)
+mdb_page_flush(MDB_txn *txn, int keep)
{
MDB_env *env = txn->mt_env;
MDB_ID2L dl = txn->mt_u.dirty_list;
int n = 0;
#endif
- j = 0;
+ j = i = keep;
+
if (env->me_flags & MDB_WRITEMAP) {
/* Clear dirty flags */
- for (i=1; i<=pagecount; i++) {
+ while (++i <= pagecount) {
dp = dl[i].mptr;
/* Don't flush this page yet */
if (dp->mp_flags & P_KEEP) {
}
/* Write the pages */
- for (i = 1;; i++) {
- if (i <= pagecount) {
+ for (;;) {
+ if (++i <= pagecount) {
dp = dl[i].mptr;
/* Don't flush this page yet */
if (dp->mp_flags & P_KEEP) {
#endif /* _WIN32 */
}
- j = 0;
- for (i=1; i<=pagecount; i++) {
+ for (i = keep; ++i <= pagecount; ) {
dp = dl[i].mptr;
/* This is a page we skipped above */
if (!dl[i].mid) {
mdb_audit(txn);
#endif
- if ((rc = mdb_page_flush(txn)) ||
+ if ((rc = mdb_page_flush(txn, 0)) ||
(rc = mdb_env_sync(env, 0)) ||
(rc = mdb_env_write_meta(txn)))
goto fail;
WriteFile(env->me_fd, ptr, len, NULL, &ov);
#else
r2 = pwrite(env->me_fd, ptr, len, off);
+ (void)r2; /* Silence warnings. We don't care about pwrite's return value */
#endif
fail:
env->me_flags |= MDB_FATAL_ERROR;
mdb_hash_enc(MDB_val *val, char *encbuf)
{
mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
- unsigned long *l = (unsigned long *)&h;
- mdb_pack85(l[0], encbuf);
- mdb_pack85(l[1], encbuf+5);
+ mdb_pack85(h, encbuf);
+ mdb_pack85(h>>32, encbuf+5);
encbuf[10] = '\0';
}
#endif
mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
{
MDB_page *mp = mc->mc_pg[mc->mc_top];
- DKBUF;
int rc;
-
+ DKBUF;
while (IS_BRANCH(mp)) {
MDB_node *node;
nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
} else {
leaf = NODEPTR(mp, 0);
- MDB_GET_KEY(leaf, &nodekey);
+ MDB_GET_KEY2(leaf, nodekey);
}
rc = mc->mc_dbx->md_cmp(key, &nodekey);
if (rc == 0) {
nkeys-1, nodekey.mv_size);
} else {
leaf = NODEPTR(mp, nkeys-1);
- MDB_GET_KEY(leaf, &nodekey);
+ MDB_GET_KEY2(leaf, nodekey);
}
rc = mc->mc_dbx->md_cmp(key, &nodekey);
if (rc == 0) {
mc->mc_ki[mc->mc_top], nodekey.mv_size);
} else {
leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
- MDB_GET_KEY(leaf, &nodekey);
+ MDB_GET_KEY2(leaf, nodekey);
}
rc = mc->mc_dbx->md_cmp(key, &nodekey);
if (rc == 0) {
projects / openldap / blobdiff