#endif
#include <sys/types.h>
#include <sys/stat.h>
-#include <sys/param.h>
#ifdef _WIN32
#include <windows.h>
+/** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
+ * as int64 which is wrong. MSVC doesn't define it at all, so just
+ * don't use it.
+ */
+#define MDB_PID_T int
+#ifdef __GNUC__
+# include <sys/param.h>
+#else
+# define LITTLE_ENDIAN 1234
+# define BIG_ENDIAN 4321
+# define BYTE_ORDER LITTLE_ENDIAN
+# ifndef SSIZE_MAX
+# define SSIZE_MAX INT_MAX
+# endif
+#endif
#else
+#define MDB_PID_T pid_t
+#include <sys/param.h>
#include <sys/uio.h>
#include <sys/mman.h>
#ifdef HAVE_SYS_FILE_H
(((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
/** @} */
- /** A default memory page size.
- * The actual size is platform-dependent, but we use this for
- * boot-strapping. We probably should not be using this any more.
- * The #GET_PAGESIZE() macro is used to get the actual size.
+ /** @brief The maximum size of a database page.
+ *
+ * This is 32k, since it must fit in #MDB_page.#mp_upper.
+ *
+ * LMDB will use database pages < OS pages if needed.
+ * That causes more I/O in write transactions: The OS must
+ * know (read) the whole page before writing a partial page.
*
* Note that we don't currently support Huge pages. On Linux,
* regular data files cannot use Huge pages, and in general
* pressure from other processes is high. So until OSs have
* actual paging support for Huge pages, they're not viable.
*/
-#define MDB_PAGESIZE 4096
+#define MAX_PAGESIZE 0x8000
/** The minimum number of keys required in a database page.
* Setting this to a larger value will place a smaller bound on the
*
* We require that keys all fit onto a regular page. This limit
* could be raised a bit further if needed; to something just
- * under #MDB_PAGESIZE / #MDB_MINKEYS.
+ * under (page size / #MDB_MINKEYS / 3).
*
* Note that data items in an #MDB_DUPSORT database are actually keys
* of a subDB, so they're also limited to this size.
*/
txnid_t mrb_txnid;
/** The process ID of the process owning this reader txn. */
- pid_t mrb_pid;
+ MDB_PID_T mrb_pid;
/** The thread ID of the thread owning this txn. */
pthread_t mrb_tid;
} MDB_rxbody;
txnid_t mm_txnid; /**< txnid that committed this page */
} MDB_meta;
- /** Buffer for a stack-allocated dirty page.
+ /** Buffer for a stack-allocated meta page.
* The members define size and alignment, and silence type
* aliasing warnings. They are not used directly; that could
* mean incorrectly using several union members in parallel.
*/
-typedef union MDB_pagebuf {
- char mb_raw[MDB_PAGESIZE];
+typedef union MDB_metabuf {
MDB_page mb_page;
struct {
char mm_pad[PAGEHDRSZ];
MDB_meta mm_meta;
} mb_metabuf;
-} MDB_pagebuf;
+} MDB_metabuf;
/** Auxiliary DB info.
* The information here is mostly static/read-only. There is
/** Have liveness lock in reader table */
#define MDB_LIVE_READER 0x08000000U
uint32_t me_flags; /**< @ref mdb_env */
- unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
+ unsigned int me_psize; /**< DB page size, inited from me_os_psize */
+ unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
unsigned int me_maxreaders; /**< size of the reader table */
unsigned int me_numreaders; /**< max numreaders set by this env */
MDB_dbi me_numdbs; /**< number of DBs opened */
MDB_dbi me_maxdbs; /**< size of the DB table */
- pid_t me_pid; /**< process ID of this env */
+ MDB_PID_T me_pid; /**< process ID of this env */
char *me_path; /**< path to the DB files */
char *me_map; /**< the memory map of the data file */
MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
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 */
size_t me_mapsize; /**< size of the data memory map */
off_t me_size; /**< current file size */
mdb_page_list(MDB_page *mp)
{
MDB_node *node;
- unsigned int i, nkeys, nsize;
+ unsigned int i, nkeys, nsize, total = 0;
MDB_val key;
DKBUF;
node = NODEPTR(mp, i);
key.mv_size = node->mn_ksize;
key.mv_data = node->mn_data;
- nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
+ nsize = NODESIZE + key.mv_size;
if (IS_BRANCH(mp)) {
fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
DKEY(&key));
+ total += nsize;
} else {
if (F_ISSET(node->mn_flags, F_BIGDATA))
nsize += sizeof(pgno_t);
else
nsize += NODEDSZ(node);
+ total += nsize;
+ nsize += sizeof(indx_t);
fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
}
+ total += (total & 1);
}
+ fprintf(stderr, "Total: %d\n", total);
}
void
{
MDB_env *env = txn->mt_env;
MDB_page *ret = env->me_dpages;
- size_t sz = env->me_psize;
+ size_t psize = env->me_psize, sz = psize, off;
+ /* For ! #MDB_NOMEMINIT, psize counts how much to init.
+ * For a single page alloc, we init everything after the page header.
+ * For multi-page, we init the final page; if the caller needed that
+ * many pages they will be filling in at least up to the last page.
+ */
if (num == 1) {
if (ret) {
VGMEMP_ALLOC(env, ret, sz);
env->me_dpages = ret->mp_next;
return ret;
}
+ psize -= off = PAGEHDRSZ;
} else {
sz *= num;
+ off = sz - psize;
}
if ((ret = malloc(sz)) != NULL) {
+ if (!(env->me_flags & MDB_NOMEMINIT)) {
+ memset((char *)ret + off, 0, psize);
+ ret->mp_pad = 0;
+ }
VGMEMP_ALLOC(env, ret, sz);
}
return ret;
* lock on the lockfile, set at an offset equal to the pid.
*/
static int
-mdb_reader_pid(MDB_env *env, enum Pidlock_op op, pid_t pid)
+mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
{
#if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
int ret = 0;
if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
return MDB_BAD_RSLOT;
} else {
- pid_t pid = env->me_pid;
+ MDB_PID_T pid = env->me_pid;
pthread_t tid = pthread_self();
if (!(env->me_flags & MDB_LIVE_READER)) {
int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
txnid_t pglast = 0, head_id = 0;
pgno_t freecnt = 0, *free_pgs, *mop;
- ssize_t head_room = 0, total_room = 0, mop_len;
+ ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
return rc;
}
+ /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
+ clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
+ ? SSIZE_MAX : maxfree_1pg;
+
for (;;) {
/* Come back here after each Put() in case freelist changed */
MDB_val key, data;
+ pgno_t *pgs;
+ ssize_t j;
/* If using records from freeDB which we have not yet
* deleted, delete them and any we reserved for me_pghead.
rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
if (rc)
return rc;
- *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
+ /* IDL is initially empty, zero out at least the length */
+ pgs = (pgno_t *)data.mv_data;
+ j = head_room > clean_limit ? head_room : 0;
+ do {
+ pgs[j] = 0;
+ } while (--j >= 0);
total_room += head_room;
}
static int
mdb_env_read_header(MDB_env *env, MDB_meta *meta)
{
- MDB_pagebuf pbuf;
+ MDB_metabuf pbuf;
MDB_page *p;
MDB_meta *m;
int i, rc, off;
+ enum { Size = sizeof(pbuf) };
/* We don't know the page size yet, so use a minimum value.
* Read both meta pages so we can use the latest one.
OVERLAPPED ov;
memset(&ov, 0, sizeof(ov));
ov.Offset = off;
- rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
+ rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
rc = 0;
#else
- rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
+ rc = pread(env->me_fd, &pbuf, Size, off);
#endif
- if (rc != MDB_PAGESIZE) {
+ if (rc != Size) {
if (rc == 0 && off == 0)
return ENOENT;
rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
mp->mm_last_pg = txn->mt_next_pgno - 1;
mp->mm_txnid = txn->mt_txnid;
if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
+ unsigned meta_size = env->me_psize;
rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
ptr = env->me_map;
- if (toggle)
- ptr += env->me_psize;
- if (MDB_MSYNC(ptr, env->me_psize, rc)) {
+ if (toggle) {
+#ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
+ if (meta_size < env->me_os_psize)
+ meta_size += meta_size;
+ else
+#endif
+ ptr += meta_size;
+ }
+ if (MDB_MSYNC(ptr, meta_size, rc)) {
rc = ErrCode();
goto fail;
}
e->me_wmutex = SEM_FAILED;
#endif
e->me_pid = getpid();
+ GET_PAGESIZE(e->me_os_psize);
VGMEMP_CREATE(e,0,0);
*env = e;
return MDB_SUCCESS;
return i;
DPUTS("new mdbenv");
newenv = 1;
- GET_PAGESIZE(env->me_psize);
+ env->me_psize = env->me_os_psize;
+ if (env->me_psize > MAX_PAGESIZE)
+ env->me_psize = MAX_PAGESIZE;
} else {
env->me_psize = meta.mm_psize;
}
#pragma comment(linker, "/INCLUDE:_tls_used")
#pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
#pragma const_seg(".CRT$XLB")
-extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
+extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
#pragma const_seg()
#else /* WIN32 */
rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
if (size < rsize && *excl > 0) {
#ifdef _WIN32
- if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
+ if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
|| !SetEndOfFile(env->me_lfd))
goto fail_errno;
#else
* at runtime. Changing other flags requires closing the
* environment and re-opening it with the new flags.
*/
-#define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
-#define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
+#define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
+#define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
+ MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
int
mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
DPRINTF(("opened dbenv %p", (void *) env));
if (excl > 0) {
rc = mdb_env_share_locks(env, &excl);
+ if (rc)
+ goto leave;
}
+ if (!((flags & MDB_RDONLY) ||
+ (env->me_pbuf = calloc(1, env->me_psize))))
+ rc = ENOMEM;
}
leave:
for (i = env->me_maxdbs; --i > MAIN_DBI; )
free(env->me_dbxs[i].md_name.mv_data);
+ free(env->me_pbuf);
free(env->me_dbflags);
free(env->me_dbxs);
free(env->me_path);
if (env->me_fd != INVALID_HANDLE_VALUE)
(void) close(env->me_fd);
if (env->me_txns) {
- pid_t pid = env->me_pid;
+ MDB_PID_T pid = env->me_pid;
/* Clearing readers is done in this function because
* me_txkey with its destructor must be disabled first.
*/
unsigned int mcount = 0, dcount = 0, nospill;
size_t nsize;
int rc, rc2;
- MDB_pagebuf pbuf;
char dbuf[MDB_MAXKEYSIZE+1];
unsigned int nflags;
DKBUF;
/* DB has dups? */
if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
- mp = fp = xdata.mv_data = &pbuf;
+ mp = fp = xdata.mv_data = env->me_pbuf;
mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
/* Was a single item before, must convert now */
dkey.mv_data = dbuf;
fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
fp->mp_lower = PAGEHDRSZ;
- fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
+ xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
if (mc->mc_db->md_flags & MDB_DUPFIXED) {
fp->mp_flags |= P_LEAF2;
fp->mp_pad = data->mv_size;
- fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
+ xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
} else {
- fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
+ xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
(dkey.mv_size & 1) + (data->mv_size & 1);
}
- xdata.mv_size = fp->mp_upper;
+ fp->mp_upper = xdata.mv_size;
} else if (leaf->mn_flags & F_SUBDATA) {
/* Data is on sub-DB, just store it */
flags |= F_DUPDATA|F_SUBDATA;
if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + xdata.mv_size
>= env->me_nodemax) {
/* yes, convert it */
- dummy.md_flags = 0;
if (mc->mc_db->md_flags & MDB_DUPFIXED) {
dummy.md_pad = fp->mp_pad;
dummy.md_flags = MDB_DUPFIXED;
if (mc->mc_db->md_flags & MDB_INTEGERDUP)
dummy.md_flags |= MDB_INTEGERKEY;
+ } else {
+ dummy.md_pad = 0;
+ dummy.md_flags = 0;
}
dummy.md_depth = 1;
dummy.md_branch_pages = 0;
MDB_node *node;
char *base;
size_t len;
- int delta, delta0;
+ int delta, ksize, oksize;
indx_t ptr, i, numkeys, indx;
DKBUF;
}
#endif
- delta0 = delta = key->mv_size - node->mn_ksize;
+ ksize = key->mv_size;
+ ksize += (ksize & 1);
+ oksize = node->mn_ksize;
+ oksize += (oksize & 1);
+ delta = ksize - oksize;
/* Must be 2-byte aligned. If new key is
* shorter by 1, the shift will be skipped.
*/
- delta += (delta & 1);
if (delta) {
if (delta > 0 && SIZELEFT(mp) < delta) {
pgno_t pgno;
}
/* But even if no shift was needed, update ksize */
- if (delta0)
+ if (node->mn_ksize != key->mv_size)
node->mn_ksize = key->mv_size;
if (key->mv_size)
flags = 0;
} else {
srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
- assert(!((long)srcnode&1));
+ assert(!((size_t)srcnode&1));
srcpg = NODEPGNO(srcnode);
flags = srcnode->mn_flags;
if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
/** Insert pid into list if not already present.
* return -1 if already present.
*/
-static int mdb_pid_insert(pid_t *ids, pid_t pid)
+static int mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
{
/* binary search of pid in list */
unsigned base = 0;
{
unsigned int i, j, rdrs;
MDB_reader *mr;
- pid_t *pids, pid;
+ MDB_PID_T *pids, pid;
int count = 0;
if (!env)
if (!env->me_txns)
return MDB_SUCCESS;
rdrs = env->me_txns->mti_numreaders;
- pids = malloc((rdrs+1) * sizeof(pid_t));
+ pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
if (!pids)
return ENOMEM;
pids[0] = 0;
projects / openldap / blobdiff