+#ifndef MDB_WBUF
+#define MDB_WBUF (1024*1024)
+#endif
+
+ /** State needed for a compacting copy. */
+typedef struct mdb_copy {
+ pthread_mutex_t mc_mutex;
+ pthread_cond_t mc_cond;
+ char *mc_wbuf[2];
+ char *mc_over[2];
+ MDB_env *mc_env;
+ MDB_txn *mc_txn;
+ int mc_wlen[2];
+ int mc_olen[2];
+ pgno_t mc_next_pgno;
+ HANDLE mc_fd;
+ int mc_status;
+ volatile int mc_new;
+ int mc_toggle;
+
+} mdb_copy;
+
+ /** Dedicated writer thread for compacting copy. */
+static THREAD_RET ESECT
+mdb_env_copythr(void *arg)
+{
+ mdb_copy *my = arg;
+ char *ptr;
+ int toggle = 0, wsize, rc;
+#ifdef _WIN32
+ DWORD len;
+#define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
+#else
+ int len;
+#define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
+#endif
+
+ pthread_mutex_lock(&my->mc_mutex);
+ my->mc_new = 0;
+ pthread_cond_signal(&my->mc_cond);
+ for(;;) {
+ while (!my->mc_new)
+ pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
+ if (my->mc_new < 0) {
+ my->mc_new = 0;
+ break;
+ }
+ my->mc_new = 0;
+ wsize = my->mc_wlen[toggle];
+ ptr = my->mc_wbuf[toggle];
+again:
+ while (wsize > 0) {
+ DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
+ if (!rc) {
+ rc = ErrCode();
+ break;
+ } else if (len > 0) {
+ rc = MDB_SUCCESS;
+ ptr += len;
+ wsize -= len;
+ continue;
+ } else {
+ rc = EIO;
+ break;
+ }
+ }
+ if (rc) {
+ my->mc_status = rc;
+ break;
+ }
+ /* If there's an overflow page tail, write it too */
+ if (my->mc_olen[toggle]) {
+ wsize = my->mc_olen[toggle];
+ ptr = my->mc_over[toggle];
+ my->mc_olen[toggle] = 0;
+ goto again;
+ }
+ my->mc_wlen[toggle] = 0;
+ toggle ^= 1;
+ pthread_cond_signal(&my->mc_cond);
+ }
+ pthread_cond_signal(&my->mc_cond);
+ pthread_mutex_unlock(&my->mc_mutex);
+ return (THREAD_RET)0;
+#undef DO_WRITE
+}
+
+ /** Tell the writer thread there's a buffer ready to write */
+static int ESECT
+mdb_env_cthr_toggle(mdb_copy *my, int st)
+{
+ int toggle = my->mc_toggle ^ 1;
+ pthread_mutex_lock(&my->mc_mutex);
+ if (my->mc_status) {
+ pthread_mutex_unlock(&my->mc_mutex);
+ return my->mc_status;
+ }
+ while (my->mc_new == 1)
+ pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
+ my->mc_new = st;
+ my->mc_toggle = toggle;
+ pthread_cond_signal(&my->mc_cond);
+ pthread_mutex_unlock(&my->mc_mutex);
+ return 0;
+}
+
+ /** Depth-first tree traversal for compacting copy. */
+static int ESECT
+mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
+{
+ MDB_cursor mc;
+ MDB_txn *txn = my->mc_txn;
+ MDB_node *ni;
+ MDB_page *mo, *mp, *leaf;
+ char *buf, *ptr;
+ int rc, toggle;
+ unsigned int i;
+
+ /* Empty DB, nothing to do */
+ if (*pg == P_INVALID)
+ return MDB_SUCCESS;
+
+ mc.mc_snum = 1;
+ mc.mc_top = 0;
+ mc.mc_txn = txn;
+
+ rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
+ if (rc)
+ return rc;
+ rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
+ if (rc)
+ return rc;
+
+ /* Make cursor pages writable */
+ buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
+ if (buf == NULL)
+ return ENOMEM;
+
+ for (i=0; i<mc.mc_top; i++) {
+ mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
+ mc.mc_pg[i] = (MDB_page *)ptr;
+ ptr += my->mc_env->me_psize;
+ }
+
+ /* This is writable space for a leaf page. Usually not needed. */
+ leaf = (MDB_page *)ptr;
+
+ toggle = my->mc_toggle;
+ while (mc.mc_snum > 0) {
+ unsigned n;
+ mp = mc.mc_pg[mc.mc_top];
+ n = NUMKEYS(mp);
+
+ if (IS_LEAF(mp)) {
+ if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
+ for (i=0; i<n; i++) {
+ ni = NODEPTR(mp, i);
+ if (ni->mn_flags & F_BIGDATA) {
+ MDB_page *omp;
+ pgno_t pg;
+
+ /* Need writable leaf */
+ if (mp != leaf) {
+ mc.mc_pg[mc.mc_top] = leaf;
+ mdb_page_copy(leaf, mp, my->mc_env->me_psize);
+ mp = leaf;
+ ni = NODEPTR(mp, i);
+ }
+
+ memcpy(&pg, NODEDATA(ni), sizeof(pg));
+ rc = mdb_page_get(txn, pg, &omp, NULL);
+ if (rc)
+ goto done;
+ if (my->mc_wlen[toggle] >= MDB_WBUF) {
+ rc = mdb_env_cthr_toggle(my, 1);
+ if (rc)
+ goto done;
+ toggle = my->mc_toggle;
+ }
+ mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
+ memcpy(mo, omp, my->mc_env->me_psize);
+ mo->mp_pgno = my->mc_next_pgno;
+ my->mc_next_pgno += omp->mp_pages;
+ my->mc_wlen[toggle] += my->mc_env->me_psize;
+ if (omp->mp_pages > 1) {
+ my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
+ my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
+ rc = mdb_env_cthr_toggle(my, 1);
+ if (rc)
+ goto done;
+ toggle = my->mc_toggle;
+ }
+ memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
+ } else if (ni->mn_flags & F_SUBDATA) {
+ MDB_db db;
+
+ /* Need writable leaf */
+ if (mp != leaf) {
+ mc.mc_pg[mc.mc_top] = leaf;
+ mdb_page_copy(leaf, mp, my->mc_env->me_psize);
+ mp = leaf;
+ ni = NODEPTR(mp, i);
+ }
+
+ memcpy(&db, NODEDATA(ni), sizeof(db));
+ my->mc_toggle = toggle;
+ rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
+ if (rc)
+ goto done;
+ toggle = my->mc_toggle;
+ memcpy(NODEDATA(ni), &db, sizeof(db));
+ }
+ }
+ }
+ } else {
+ mc.mc_ki[mc.mc_top]++;
+ if (mc.mc_ki[mc.mc_top] < n) {
+ pgno_t pg;
+again:
+ ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
+ pg = NODEPGNO(ni);
+ rc = mdb_page_get(txn, pg, &mp, NULL);
+ if (rc)
+ goto done;
+ mc.mc_top++;
+ mc.mc_snum++;
+ mc.mc_ki[mc.mc_top] = 0;
+ if (IS_BRANCH(mp)) {
+ /* Whenever we advance to a sibling branch page,
+ * we must proceed all the way down to its first leaf.
+ */
+ mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
+ goto again;
+ } else
+ mc.mc_pg[mc.mc_top] = mp;
+ continue;
+ }
+ }
+ if (my->mc_wlen[toggle] >= MDB_WBUF) {
+ rc = mdb_env_cthr_toggle(my, 1);
+ if (rc)
+ goto done;
+ toggle = my->mc_toggle;
+ }
+ mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
+ mdb_page_copy(mo, mp, my->mc_env->me_psize);
+ mo->mp_pgno = my->mc_next_pgno++;
+ my->mc_wlen[toggle] += my->mc_env->me_psize;
+ if (mc.mc_top) {
+ /* Update parent if there is one */
+ ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
+ SETPGNO(ni, mo->mp_pgno);
+ mdb_cursor_pop(&mc);
+ } else {
+ /* Otherwise we're done */
+ *pg = mo->mp_pgno;
+ break;
+ }
+ }
+done:
+ free(buf);
+ return rc;
+}
+
+ /** Copy environment with compaction. */
+static int ESECT
+mdb_env_copyfd1(MDB_env *env, HANDLE fd)
+{
+ MDB_meta *mm;
+ MDB_page *mp;
+ mdb_copy my;
+ MDB_txn *txn = NULL;
+ pthread_t thr;
+ int rc;
+
+#ifdef _WIN32
+ my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
+ my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
+ my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_psize);
+ if (my.mc_wbuf[0] == NULL)
+ return errno;
+#else
+ pthread_mutex_init(&my.mc_mutex, NULL);
+ pthread_cond_init(&my.mc_cond, NULL);
+ rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_psize, MDB_WBUF*2);
+ if (rc)
+ return rc;
+#endif
+ my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
+ my.mc_wlen[0] = 0;
+ my.mc_wlen[1] = 0;
+ my.mc_olen[0] = 0;
+ my.mc_olen[1] = 0;
+ my.mc_next_pgno = 2;
+ my.mc_status = 0;
+ my.mc_new = 1;
+ my.mc_toggle = 0;
+ my.mc_env = env;
+ my.mc_fd = fd;
+
+ /* Do the lock/unlock of the reader mutex before starting the
+ * write txn. Otherwise other read txns could block writers.
+ */
+ rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
+ if (rc)
+ return rc;
+
+ if (env->me_txns) {
+ /* We must start the actual read txn after blocking writers */
+ mdb_txn_reset0(txn, "reset-stage1");
+
+ /* Temporarily block writers until we snapshot the meta pages */
+ LOCK_MUTEX_W(env);
+
+ rc = mdb_txn_renew0(txn);
+ if (rc) {
+ UNLOCK_MUTEX_W(env);
+ goto leave;
+ }
+ }
+
+ THREAD_CREATE(thr, mdb_env_copythr, &my);
+ mp = (MDB_page *)my.mc_wbuf[0];
+ memset(mp, 0, 2*env->me_psize);
+ mp->mp_pgno = 0;
+ mp->mp_flags = P_META;
+ mm = (MDB_meta *)METADATA(mp);
+ mdb_env_init_meta0(env, mm);
+ mm->mm_address = env->me_metas[0]->mm_address;
+
+ mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
+ mp->mp_pgno = 1;
+ mp->mp_flags = P_META;
+ *(MDB_meta *)METADATA(mp) = *mm;
+ mm = (MDB_meta *)METADATA(mp);
+
+ /* Count the number of free pages, subtract from lastpg to find
+ * number of active pages
+ */
+ {
+ MDB_ID freecount = 0;
+ MDB_cursor mc;
+ MDB_val key, data;
+ mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
+ while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
+ freecount += *(MDB_ID *)data.mv_data;
+ freecount += txn->mt_dbs[0].md_branch_pages +
+ txn->mt_dbs[0].md_leaf_pages +
+ txn->mt_dbs[0].md_overflow_pages;
+
+ /* Set metapage 1 */
+ mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
+ mm->mm_dbs[1] = txn->mt_dbs[1];
+ mm->mm_dbs[1].md_root = mm->mm_last_pg;
+ mm->mm_txnid = 1;
+ }
+ my.mc_wlen[0] = env->me_psize * 2;
+ my.mc_txn = txn;
+ pthread_mutex_lock(&my.mc_mutex);
+ while(my.mc_new)
+ pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
+ pthread_mutex_unlock(&my.mc_mutex);
+ rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
+ if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
+ rc = mdb_env_cthr_toggle(&my, 1);
+ mdb_env_cthr_toggle(&my, -1);
+ pthread_mutex_lock(&my.mc_mutex);
+ while(my.mc_new)
+ pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
+ pthread_mutex_unlock(&my.mc_mutex);
+ THREAD_FINISH(thr);
+leave:
+ mdb_txn_abort(txn);
+#ifdef _WIN32
+ CloseHandle(my.mc_cond);
+ CloseHandle(my.mc_mutex);
+ _aligned_free(my.mc_wbuf[0]);
+#else
+ pthread_cond_destroy(&my.mc_cond);
+ pthread_mutex_destroy(&my.mc_mutex);
+ free(my.mc_wbuf[0]);
+#endif
+ return rc;
+}
+
+ /** Copy environment as-is. */
+static int ESECT
+mdb_env_copyfd0(MDB_env *env, HANDLE fd)
+{
+ MDB_txn *txn = NULL;
+ int rc;
+ size_t wsize;
+ char *ptr;
+#ifdef _WIN32
+ DWORD len, w2;
+#define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
+#else
+ ssize_t len;
+ size_t w2;
+#define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
+#endif
+
+ /* Do the lock/unlock of the reader mutex before starting the
+ * write txn. Otherwise other read txns could block writers.
+ */
+ rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
+ if (rc)
+ return rc;
+
+ if (env->me_txns) {
+ /* We must start the actual read txn after blocking writers */
+ mdb_txn_reset0(txn, "reset-stage1");
+
+ /* Temporarily block writers until we snapshot the meta pages */
+ LOCK_MUTEX_W(env);
+
+ rc = mdb_txn_renew0(txn);
+ if (rc) {
+ UNLOCK_MUTEX_W(env);
+ goto leave;
+ }
+ }
+
+ wsize = env->me_psize * 2;
+ ptr = env->me_map;
+ w2 = wsize;
+ while (w2 > 0) {
+ DO_WRITE(rc, fd, ptr, w2, len);
+ if (!rc) {
+ rc = ErrCode();
+ break;
+ } else if (len > 0) {
+ rc = MDB_SUCCESS;
+ ptr += len;
+ w2 -= len;
+ continue;
+ } else {
+ /* Non-blocking or async handles are not supported */
+ rc = EIO;
+ break;
+ }
+ }
+ if (env->me_txns)
+ UNLOCK_MUTEX_W(env);
+
+ if (rc)
+ goto leave;
+
+ w2 = txn->mt_next_pgno * env->me_psize;
+#ifdef WIN32
+ {
+ LARGE_INTEGER fsize;
+ GetFileSizeEx(env->me_fd, &fsize);
+ if (w2 > fsize.QuadPart)
+ w2 = fsize.QuadPart;
+ }
+#else
+ {
+ struct stat st;
+ fstat(env->me_fd, &st);
+ if (w2 > (size_t)st.st_size)
+ w2 = st.st_size;
+ }
+#endif
+ wsize = w2 - wsize;
+ while (wsize > 0) {
+ if (wsize > MAX_WRITE)
+ w2 = MAX_WRITE;
+ else
+ w2 = wsize;
+ DO_WRITE(rc, fd, ptr, w2, len);
+ if (!rc) {
+ rc = ErrCode();
+ break;
+ } else if (len > 0) {
+ rc = MDB_SUCCESS;
+ ptr += len;
+ wsize -= len;
+ continue;
+ } else {
+ rc = EIO;
+ break;
+ }
+ }
+
+leave:
+ mdb_txn_abort(txn);
+ return rc;
+}
+
+int ESECT
+mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
+{
+ if (flags & MDB_CP_COMPACT)
+ return mdb_env_copyfd1(env, fd);
+ else
+ return mdb_env_copyfd0(env, fd);
+}
+
+int ESECT
+mdb_env_copyfd(MDB_env *env, HANDLE fd)
+{
+ return mdb_env_copyfd2(env, fd, 0);
+}
+
+int ESECT
+mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
+{
+ int rc, len;
+ char *lpath;
+ HANDLE newfd = INVALID_HANDLE_VALUE;
+
+ if (env->me_flags & MDB_NOSUBDIR) {
+ lpath = (char *)path;
+ } else {
+ len = strlen(path);
+ len += sizeof(DATANAME);
+ lpath = malloc(len);
+ if (!lpath)
+ return ENOMEM;
+ sprintf(lpath, "%s" DATANAME, path);
+ }
+
+ /* The destination path must exist, but the destination file must not.
+ * We don't want the OS to cache the writes, since the source data is
+ * already in the OS cache.
+ */
+#ifdef _WIN32
+ newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
+ FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
+#else
+ newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
+#endif
+ if (newfd == INVALID_HANDLE_VALUE) {
+ rc = ErrCode();
+ goto leave;
+ }
+
+#ifdef O_DIRECT
+ /* Set O_DIRECT if the file system supports it */
+ if ((rc = fcntl(newfd, F_GETFL)) != -1)
+ (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
+#endif
+#ifdef F_NOCACHE /* __APPLE__ */
+ rc = fcntl(newfd, F_NOCACHE, 1);
+ if (rc) {
+ rc = ErrCode();
+ goto leave;
+ }
+#endif
+
+ rc = mdb_env_copyfd2(env, newfd, flags);
+
+leave:
+ if (!(env->me_flags & MDB_NOSUBDIR))
+ free(lpath);
+ if (newfd != INVALID_HANDLE_VALUE)
+ if (close(newfd) < 0 && rc == MDB_SUCCESS)
+ rc = ErrCode();
+
+ return rc;
+}
+
+int ESECT
+mdb_env_copy(MDB_env *env, const char *path)
+{
+ return mdb_env_copy2(env, path, 0);
+}
+
+int ESECT