* BerkeleyDB API, but much simplified.
*/
/*
- * Copyright 2011-2014 Howard Chu, Symas Corp.
+ * Copyright 2011-2015 Howard Chu, Symas Corp.
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
#define CACHEFLUSH(addr, bytes, cache)
#endif
+#if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
+/** fdatasync is broken on ext3/ext4fs on older kernels, see
+ * description in #mdb_env_open2 comments. You can safely
+ * define MDB_FDATASYNC_WORKS if this code will only be run
+ * on kernels 3.6 and newer.
+ */
+#define BROKEN_FDATASYNC
+#endif
#include <errno.h>
#include <limits.h>
* started from so we can avoid overwriting any data used in that
* particular version.
*/
- txnid_t mrb_txnid;
+ volatile txnid_t mrb_txnid;
/** The process ID of the process owning this reader txn. */
- MDB_PID_T mrb_pid;
+ volatile MDB_PID_T mrb_pid;
/** The thread ID of the thread owning this txn. */
- MDB_THR_T mrb_tid;
+ volatile MDB_THR_T mrb_tid;
} MDB_rxbody;
/** The actual reader record, with cacheline padding. */
* This is recorded here only for convenience; the value can always
* be determined by reading the main database meta pages.
*/
- txnid_t mtb_txnid;
+ volatile txnid_t mtb_txnid;
/** The number of slots that have been used in the reader table.
* This always records the maximum count, it is not decremented
* when readers release their slots.
*/
- unsigned mtb_numreaders;
+ volatile unsigned mtb_numreaders;
} MDB_txbody;
/** The actual reader table definition. */
/** Any persistent environment flags. @ref mdb_env */
#define mm_flags mm_dbs[0].md_flags
pgno_t mm_last_pg; /**< last used page in file */
- txnid_t mm_txnid; /**< txnid that committed this page */
+ volatile txnid_t mm_txnid; /**< txnid that committed this page */
} MDB_meta;
/** Buffer for a stack-allocated meta page.
#define MDB_ENV_ACTIVE 0x20000000U
/** me_txkey is set */
#define MDB_ENV_TXKEY 0x10000000U
+ /** fdatasync is unreliable */
+#define MDB_FSYNCONLY 0x08000000U
uint32_t me_flags; /**< @ref mdb_env */
unsigned int me_psize; /**< DB page size, inited from me_os_psize */
unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
mdb_env_sync(MDB_env *env, int force)
{
int rc = 0;
+ if (env->me_flags & MDB_RDONLY)
+ return EACCES;
if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
if (env->me_flags & MDB_WRITEMAP) {
int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
rc = ErrCode();
#endif
} else {
+#ifdef BROKEN_FDATASYNC
+ if (env->me_flags & MDB_FSYNCONLY) {
+ if (fsync(env->me_fd))
+ rc = ErrCode();
+ } else
+#endif
if (MDB_FDATASYNC(env->me_fd))
rc = ErrCode();
}
int rc, new_notls = 0;
if (txn->mt_flags & MDB_TXN_RDONLY) {
+ txn->mt_flags = MDB_TXN_RDONLY;
/* Setup db info */
txn->mt_numdbs = env->me_numdbs;
txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
return rc;
}
}
- txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
+ do /* LY: Retry on a race, ITS#7970. */
+ r->mr_txnid = ti->mti_txnid;
+ while(r->mr_txnid != ti->mti_txnid);
+ txn->mt_txnid = r->mr_txnid;
txn->mt_u.reader = r;
meta = env->me_metas[txn->mt_txnid & 1];
}
/* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
if (n) {
+retry_write:
/* Write previous page(s) */
#ifdef MDB_USE_PWRITEV
wres = pwritev(env->me_fd, iov, n, wpos);
if (n == 1) {
wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
} else {
+retry_seek:
if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
rc = ErrCode();
+ if (rc == EINTR)
+ goto retry_seek;
DPRINTF(("lseek: %s", strerror(rc)));
return rc;
}
if (wres != wsize) {
if (wres < 0) {
rc = ErrCode();
+ if (rc == EINTR)
+ goto retry_write;
DPRINTF(("Write error: %s", strerror(rc)));
} else {
rc = EIO; /* TODO: Use which error code? */
int len;
#define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
len = pwrite(fd, ptr, size, pos); \
- rc = (len >= 0); } while(0)
+ if (len == -1 && ErrCode() == EINTR) continue; \
+ rc = (len >= 0); break; } while(1)
#endif
DPUTS("writing new meta page");
/* Write to the SYNC fd */
mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
env->me_fd : env->me_mfd;
+retry_write:
#ifdef _WIN32
{
memset(&ov, 0, sizeof(ov));
#endif
if (rc != len) {
rc = rc < 0 ? ErrCode() : EIO;
+ if (rc == EINTR)
+ goto retry_write;
DPUTS("write failed, disk error?");
/* On a failure, the pagecache still contains the new data.
* Write some old data back, to prevent it from being used.
return MDB_SUCCESS;
}
+#ifdef BROKEN_FDATASYNC
+#include <sys/utsname.h>
+#include <sys/vfs.h>
+#endif
+
/** Further setup required for opening an LMDB environment
*/
static int ESECT
else
env->me_pidquery = PROCESS_QUERY_INFORMATION;
#endif /* _WIN32 */
+#ifdef BROKEN_FDATASYNC
+ /* ext3/ext4 fdatasync is broken on some older Linux kernels.
+ * https://lkml.org/lkml/2012/9/3/83
+ * Kernels after 3.6-rc6 are known good.
+ * https://lkml.org/lkml/2012/9/10/556
+ * See if the DB is on ext3/ext4, then check for new enough kernel
+ * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
+ * to be patched.
+ */
+ {
+ struct statfs st;
+ fstatfs(env->me_fd, &st);
+ while (st.f_type == 0xEF53) {
+ struct utsname uts;
+ int i;
+ uname(&uts);
+ if (uts.release[0] < '3') {
+ if (!strncmp(uts.release, "2.6.32.", 7)) {
+ i = atoi(uts.release+7);
+ if (i >= 60)
+ break; /* 2.6.32.60 and newer is OK */
+ } else if (!strncmp(uts.release, "2.6.34.", 7)) {
+ i = atoi(uts.release+7);
+ if (i >= 15)
+ break; /* 2.6.34.15 and newer is OK */
+ }
+ } else if (uts.release[0] == '3') {
+ i = atoi(uts.release+2);
+ if (i > 5)
+ break; /* 3.6 and newer is OK */
+ if (i == 5) {
+ i = atoi(uts.release+4);
+ if (i >= 4)
+ break; /* 3.5.4 and newer is OK */
+ } else if (i == 2) {
+ i = atoi(uts.release+4);
+ if (i >= 30)
+ break; /* 3.2.30 and newer is OK */
+ }
+ } else { /* 4.x and newer is OK */
+ break;
+ }
+ env->me_flags |= MDB_FSYNCONLY;
+ break;
+ }
+ }
+#endif
memset(&meta, 0, sizeof(meta));
return;
/* Doing this here since me_dbxs may not exist during mdb_env_close */
- for (i = env->me_maxdbs; --i > MAIN_DBI; )
- free(env->me_dbxs[i].md_name.mv_data);
+ if (env->me_dbxs) {
+ for (i = env->me_maxdbs; --i > MAIN_DBI; )
+ free(env->me_dbxs[i].md_name.mv_data);
+ free(env->me_dbxs);
+ }
free(env->me_pbuf);
free(env->me_dbiseqs);
free(env->me_dbflags);
- free(env->me_dbxs);
free(env->me_path);
free(env->me_dirty_list);
free(env->me_txn0);
mc->mc_snum = 0;
mc->mc_top = 0;
mc->mc_pg[0] = 0;
+ mc->mc_ki[0] = 0;
mc->mc_flags = 0;
if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
mdb_tassert(txn, mx != NULL);
cdst->mc_ki[cdst->mc_top] = 0;
rc = mdb_update_key(cdst, &nullkey);
cdst->mc_ki[cdst->mc_top] = ix;
- mdb_cassert(csrc, rc == MDB_SUCCESS);
+ mdb_cassert(cdst, rc == MDB_SUCCESS);
}
}
m3 = m2;
if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
if (m3->mc_pg[0] == mp) {
- m3->mc_snum--;
- m3->mc_top--;
for (i=0; i<m3->mc_snum; i++) {
m3->mc_pg[i] = m3->mc_pg[i+1];
m3->mc_ki[i] = m3->mc_ki[i+1];
}
+ m3->mc_snum--;
+ m3->mc_top--;
}
}
}
if (mc->mc_ki[ptop] == 0) {
rc = mdb_page_merge(&mn, mc);
} else {
+ MDB_cursor dummy;
oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
+ /* We want mdb_rebalance to find mn when doing fixups */
+ if (mc->mc_flags & C_SUB) {
+ dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
+ mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
+ dummy.mc_xcursor = (MDB_xcursor *)&mn;
+ } else {
+ mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
+ mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
+ }
rc = mdb_page_merge(mc, &mn);
+ if (mc->mc_flags & C_SUB)
+ mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
+ else
+ mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
mdb_cursor_copy(&mn, mc);
}
mc->mc_flags &= ~C_EOF;
MDB_cursor *m2, *m3;
MDB_dbi dbi = mc->mc_dbi;
+ /* DB is totally empty now, just bail out.
+ * Other cursors adjustments were already done
+ * by mdb_rebalance and aren't needed here.
+ */
+ if (!mc->mc_snum)
+ return rc;
+
mp = mc->mc_pg[mc->mc_top];
nkeys = NUMKEYS(mp);