unsigned char mx_dbflag;
} MDB_xcursor;
+ /** Check if there is an inited xcursor, so #XCURSOR_REFRESH() is proper */
+#define XCURSOR_INITED(mc) \
+ ((mc)->mc_xcursor && ((mc)->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
+
+ /** Update sub-page pointer, if any, in \b mc->mc_xcursor. Needed
+ * when the node which contains the sub-page may have moved. Called
+ * with \b mp = mc->mc_pg[mc->mc_top], \b ki = mc->mc_ki[mc->mc_top].
+ */
+#define XCURSOR_REFRESH(mc, mp, ki) do { \
+ MDB_page *xr_pg = (mp); \
+ MDB_node *xr_node = NODEPTR(xr_pg, ki); \
+ if ((xr_node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) \
+ (mc)->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(xr_node); \
+} while (0)
+
/** State of FreeDB old pages, stored in the MDB_env */
typedef struct MDB_pgstate {
pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
struct MDB_env {
HANDLE me_fd; /**< The main data file */
HANDLE me_lfd; /**< The lock file */
- HANDLE me_mfd; /**< just for writing the meta pages */
+ HANDLE me_mfd; /**< For writing and syncing the meta pages */
#if defined(MDB_VL32) && defined(_WIN32)
HANDLE me_fmh; /**< File Mapping handle */
#endif
}
if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
printf("ack!\n");
- if (mc->mc_xcursor && (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
+ if (XCURSOR_INITED(mc)) {
node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
if (((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) &&
mc->mc_xcursor->mx_cursor.mc_pg[0] != NODEDATA(node)) {
if (m2 == mc) continue;
if (m2->mc_pg[mc->mc_top] == mp) {
m2->mc_pg[mc->mc_top] = np;
- if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
- IS_LEAF(np) &&
- (m2->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
- {
- MDB_node *leaf = NODEPTR(np, m2->mc_ki[mc->mc_top]);
- if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
- m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
- }
+ if (XCURSOR_INITED(m2) && IS_LEAF(np))
+ XCURSOR_REFRESH(m2, np, m2->mc_ki[mc->mc_top]);
}
}
}
len = sizeof(MDB_meta) - off;
off += (char *)mp - env->me_map;
- /* Write to the SYNC fd */
+ /* Write to the SYNC fd unless MDB_NOSYNC/MDB_NOMETASYNC.
+ * (me_mfd goes to the same file as me_fd, but writing to it
+ * also syncs to disk. Avoids a separate fdatasync() call.)
+ */
mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
#ifdef _WIN32
{
/* A comment in mdb_fopen() explains some O_* flag choices. */
MDB_O_RDONLY= O_RDONLY, /**< for RDONLY me_fd */
MDB_O_RDWR = O_RDWR |O_CREAT, /**< for me_fd */
- MDB_O_META = O_RDWR |MDB_DSYNC, /**< for me_mfd */
- MDB_O_COPY = O_WRONLY|O_CREAT|O_EXCL, /**< for #mdb_env_copy() */
+ MDB_O_META = O_WRONLY|MDB_DSYNC |MDB_CLOEXEC, /**< for me_mfd */
+ MDB_O_COPY = O_WRONLY|O_CREAT|O_EXCL|MDB_CLOEXEC, /**< for #mdb_env_copy() */
/** Bitmask for open() flags in enum #mdb_fopen_type. The other bits
* distinguish otherwise-equal MDB_O_* constants from each other.
*/
*
* The lockfile needs FD_CLOEXEC (close file descriptor on exec*())
* to avoid the flock() issues noted under Caveats in lmdb.h.
+ * Also set it for other filehandles which the user cannot get at
+ * and close himself, which he may need after fork(). I.e. all but
+ * me_fd, which programs do use via mdb_env_get_fd().
*/
#ifdef _WIN32
disp = OPEN_EXISTING;
break;
case MDB_O_META: /* for writing metapages */
+ acc = GENERIC_WRITE;
disp = OPEN_EXISTING;
attrs = FILE_ATTRIBUTE_NORMAL|FILE_FLAG_WRITE_THROUGH;
break;
rc = ErrCode();
#ifndef _WIN32
else {
- if (which == MDB_O_LOCKS) {
+ if (which != MDB_O_RDONLY && which != MDB_O_RDWR) {
/* Set CLOEXEC if we could not pass it to open() */
if (!MDB_CLOEXEC && (flags = fcntl(fd, F_GETFD)) != -1)
(void) fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
{
MDB_reader *reader = ptr;
- reader->mr_pid = 0;
+#ifndef _WIN32
+ if (reader->mr_pid == getpid()) /* catch pthread_exit() in child process */
+#endif
+ /* We omit the mutex, so do this atomically (i.e. skip mr_txnid) */
+ reader->mr_pid = 0;
}
#ifdef _WIN32
env->me_wmutex->semnum = 1;
env->me_rmutex->locked = &env->me_txns->mti_rlocked;
env->me_wmutex->locked = &env->me_txns->mti_wlocked;
-#endif
-#ifdef MDB_VL32
-#ifdef _WIN32
- env->me_rpmutex = CreateMutex(NULL, FALSE, NULL);
-#else
- pthread_mutex_init(&env->me_rpmutex, NULL);
-#endif
#endif
return MDB_SUCCESS;
if (rc)
return rc;
+#ifdef MDB_VL32
+#ifdef _WIN32
+ env->me_rpmutex = CreateMutex(NULL, FALSE, NULL);
+ if (!env->me_rpmutex) {
+ rc = ErrCode();
+ goto leave;
+ }
+#else
+ rc = pthread_mutex_init(&env->me_rpmutex, NULL);
+ if (rc)
+ goto leave;
+#endif
+#endif
+ flags |= MDB_ENV_ACTIVE; /* tell mdb_env_close0() to clean up */
+
if (flags & MDB_RDONLY) {
/* silently ignore WRITEMAP when we're only getting read access */
flags &= ~MDB_WRITEMAP;
(env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
rc = ENOMEM;
}
+
+ env->me_flags = flags;
+ if (rc)
+ goto leave;
+
#ifdef MDB_VL32
- if (!rc) {
+ {
env->me_rpages = malloc(MDB_ERPAGE_SIZE * sizeof(MDB_ID3));
if (!env->me_rpages) {
rc = ENOMEM;
env->me_rpcheck = MDB_ERPAGE_SIZE/2;
}
#endif
- env->me_flags = flags |= MDB_ENV_ACTIVE;
- if (rc)
- goto leave;
env->me_path = strdup(path);
env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
}
if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
- if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
- env->me_mfd = env->me_fd;
- } else {
+ if (!(flags & (MDB_RDONLY|MDB_WRITEMAP))) {
/* Synchronous fd for meta writes. Needed even with
* MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
*/
#ifdef MDB_VL32
if (env->me_txn0 && env->me_txn0->mt_rpages)
free(env->me_txn0->mt_rpages);
- { unsigned int x;
- for (x=1; x<=env->me_rpages[0].mid; x++)
- munmap(env->me_rpages[x].mptr, env->me_rpages[x].mcnt * env->me_psize);
+ if (env->me_rpages) {
+ MDB_ID3L el = env->me_rpages;
+ unsigned int x;
+ for (x=1; x<=el[0].mid; x++)
+ munmap(el[x].mptr, el[x].mcnt * env->me_psize);
+ free(el);
}
- free(env->me_rpages);
#endif
free(env->me_txn0);
mdb_midl_free(env->me_free_pgs);
munmap(env->me_map, env->me_mapsize);
#endif
}
- if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
+ if (env->me_mfd != INVALID_HANDLE_VALUE)
(void) close(env->me_mfd);
if (env->me_fd != INVALID_HANDLE_VALUE)
(void) close(env->me_fd);
{
#ifdef MDB_VL32
int rc = mdb_rpage_get(txn, pgno, &p);
- if (rc)
+ if (rc) {
+ txn->mt_flags |= MDB_TXN_ERROR;
return rc;
+ }
#else
MDB_env *env = txn->mt_env;
p = (MDB_page *)(env->me_map + env->me_psize * pgno);
if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) {
m3->mc_ki[i]++;
}
- if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
- MDB_node *n2 = NODEPTR(mp, m3->mc_ki[i]);
- if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
- m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
- }
+ if (XCURSOR_INITED(m3))
+ XCURSOR_REFRESH(m3, mp, m3->mc_ki[i]);
}
}
}
if (m2->mc_ki[i] == mc->mc_ki[i]) {
mdb_xcursor_init2(m2, mx, new_dupdata);
} else if (!insert_key && m2->mc_ki[i] < nkeys) {
- MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
- if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
- m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
+ XCURSOR_REFRESH(m2, mp, m2->mc_ki[i]);
}
}
}
if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
if (!(m2->mc_flags & C_INITIALIZED)) continue;
if (m2->mc_pg[mc->mc_top] == mp) {
- if (m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
- m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
- } else {
- MDB_node *n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
- if (!(n2->mn_flags & F_SUBDATA))
- m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
+ MDB_node *n2 = leaf;
+ if (m2->mc_ki[mc->mc_top] != mc->mc_ki[mc->mc_top]) {
+ n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
+ if (n2->mn_flags & F_SUBDATA) continue;
}
+ m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
}
}
}
m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
m3->mc_ki[csrc->mc_top-1]++;
}
- if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
- IS_LEAF(mps)) {
- MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
- if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
- m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
- }
+ if (XCURSOR_INITED(m3) && IS_LEAF(mps))
+ XCURSOR_REFRESH(m3, m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
}
} else
/* Adding on the right, bump others down */
} else {
m3->mc_ki[csrc->mc_top]--;
}
- if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
- IS_LEAF(mps)) {
- MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
- if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
- m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
- }
+ if (XCURSOR_INITED(m3) && IS_LEAF(mps))
+ XCURSOR_REFRESH(m3, m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
}
}
}
m3->mc_ki[top-1] > csrc->mc_ki[top-1]) {
m3->mc_ki[top-1]--;
}
- if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
- IS_LEAF(psrc)) {
- MDB_node *node = NODEPTR(m3->mc_pg[top], m3->mc_ki[top]);
- if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
- m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
- }
+ if (XCURSOR_INITED(m3) && IS_LEAF(psrc))
+ XCURSOR_REFRESH(m3, m3->mc_pg[top], m3->mc_ki[top]);
}
}
{
if (m3->mc_pg[mc->mc_top] == mp) {
if (m3->mc_ki[mc->mc_top] == ki) {
m3->mc_flags |= C_DEL;
+ if (mc->mc_db->md_flags & MDB_DUPSORT) {
+ /* Sub-cursor referred into dataset which is gone */
+ m3->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
+ }
+ continue;
} else if (m3->mc_ki[mc->mc_top] > ki) {
m3->mc_ki[mc->mc_top]--;
}
- if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
- MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
- if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
- m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
- }
+ if (XCURSOR_INITED(m3))
+ XCURSOR_REFRESH(m3, m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
}
}
}
}
if (mc->mc_db->md_flags & MDB_DUPSORT) {
MDB_node *node = NODEPTR(m3->mc_pg[m3->mc_top], m3->mc_ki[m3->mc_top]);
- if (node->mn_flags & F_DUPDATA) {
- mdb_xcursor_init1(m3, node);
- m3->mc_xcursor->mx_cursor.mc_flags |= C_DEL;
+ /* If this node is a fake page, it needs to be reinited
+ * because its data has moved. But just reset mc_pg[0]
+ * if the xcursor is already live.
+ */
+ if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) {
+ if (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)
+ m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
+ else
+ mdb_xcursor_init1(m3, node);
}
}
}
m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
m3->mc_ki[ptop]++;
}
- if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
- IS_LEAF(mp)) {
- MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
- if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
- m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
- }
+ if (XCURSOR_INITED(m3) && IS_LEAF(mp))
+ XCURSOR_REFRESH(m3, m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
}
}
DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));