#include "db_config.h"
#ifndef lint
static const char revid[] = "$Id: os_rw.c,v 1.1.1.1 2003/02/15 04:56:09 zarzycki Exp $";
#endif
#ifndef NO_SYSTEM_INCLUDES
#include <sys/types.h>
#include <sys/stat.h>
#include <string.h>
#include <unistd.h>
#endif
#include "db_int.h"
#ifdef HAVE_FILESYSTEM_NOTZERO
static int __os_zerofill __P((DB_ENV *, DB_FH *));
#endif
static int __os_physwrite __P((DB_ENV *, DB_FH *, void *, size_t, size_t *));
int
__os_io(dbenv, db_iop, op, niop)
DB_ENV *dbenv;
DB_IO *db_iop;
int op;
size_t *niop;
{
int ret;
DB_ASSERT(F_ISSET(db_iop->fhp, DB_FH_VALID) && db_iop->fhp->fd != -1);
#if defined(HAVE_PREAD) && defined(HAVE_PWRITE)
switch (op) {
case DB_IO_READ:
if (DB_GLOBAL(j_read) != NULL)
goto slow;
*niop = pread(db_iop->fhp->fd, db_iop->buf,
db_iop->bytes, (off_t)db_iop->pgno * db_iop->pagesize);
break;
case DB_IO_WRITE:
if (DB_GLOBAL(j_write) != NULL)
goto slow;
#ifdef HAVE_FILESYSTEM_NOTZERO
if (__os_fs_notzero())
goto slow;
#endif
*niop = pwrite(db_iop->fhp->fd, db_iop->buf,
db_iop->bytes, (off_t)db_iop->pgno * db_iop->pagesize);
break;
}
if (*niop == (size_t)db_iop->bytes)
return (0);
slow:
#endif
MUTEX_THREAD_LOCK(dbenv, db_iop->mutexp);
if ((ret = __os_seek(dbenv, db_iop->fhp,
db_iop->pagesize, db_iop->pgno, 0, 0, DB_OS_SEEK_SET)) != 0)
goto err;
switch (op) {
case DB_IO_READ:
ret = __os_read(dbenv,
db_iop->fhp, db_iop->buf, db_iop->bytes, niop);
break;
case DB_IO_WRITE:
ret = __os_write(dbenv,
db_iop->fhp, db_iop->buf, db_iop->bytes, niop);
break;
}
err: MUTEX_THREAD_UNLOCK(dbenv, db_iop->mutexp);
return (ret);
}
int
__os_read(dbenv, fhp, addr, len, nrp)
DB_ENV *dbenv;
DB_FH *fhp;
void *addr;
size_t len;
size_t *nrp;
{
size_t offset;
ssize_t nr;
int ret;
u_int8_t *taddr;
DB_ASSERT(F_ISSET(fhp, DB_FH_VALID) && fhp->fd != -1);
for (taddr = addr,
offset = 0; offset < len; taddr += nr, offset += nr) {
retry: if ((nr = DB_GLOBAL(j_read) != NULL ?
DB_GLOBAL(j_read)(fhp->fd, taddr, len - offset) :
read(fhp->fd, taddr, len - offset)) < 0) {
if ((ret = __os_get_errno()) == EINTR)
goto retry;
__db_err(dbenv, "read: 0x%x, %lu: %s", taddr,
(u_long)len-offset, strerror(ret));
return (ret);
}
if (nr == 0)
break;
}
*nrp = taddr - (u_int8_t *)addr;
return (0);
}
int
__os_write(dbenv, fhp, addr, len, nwp)
DB_ENV *dbenv;
DB_FH *fhp;
void *addr;
size_t len;
size_t *nwp;
{
DB_ASSERT(F_ISSET(fhp, DB_FH_VALID) && fhp->fd != -1);
#ifdef HAVE_FILESYSTEM_NOTZERO
if (__os_fs_notzero()) {
int ret;
if ((ret = __os_zerofill(dbenv, fhp)) != 0)
return (ret);
}
#endif
return (__os_physwrite(dbenv, fhp, addr, len, nwp));
}
static int
__os_physwrite(dbenv, fhp, addr, len, nwp)
DB_ENV *dbenv;
DB_FH *fhp;
void *addr;
size_t len;
size_t *nwp;
{
size_t offset;
ssize_t nw;
int ret;
u_int8_t *taddr;
#if defined(HAVE_FILESYSTEM_NOTZERO) && defined(DIAGNOSTIC)
if (__os_fs_notzero()) {
struct stat sb;
off_t cur_off;
DB_ASSERT(fstat(fhp->fd, &sb) != -1 &&
(cur_off = lseek(fhp->fd, (off_t)0, SEEK_CUR)) != -1 &&
cur_off <= sb.st_size);
}
#endif
for (taddr = addr,
offset = 0; offset < len; taddr += nw, offset += nw)
retry: if ((nw = DB_GLOBAL(j_write) != NULL ?
DB_GLOBAL(j_write)(fhp->fd, taddr, len - offset) :
write(fhp->fd, taddr, len - offset)) < 0) {
if ((ret = __os_get_errno()) == EINTR)
goto retry;
__db_err(dbenv, "write: 0x%x, %lu: %s", taddr,
(u_long)len-offset, strerror(ret));
return (ret);
}
*nwp = len;
return (0);
}
#ifdef HAVE_FILESYSTEM_NOTZERO
static int
__os_zerofill(dbenv, fhp)
DB_ENV *dbenv;
DB_FH *fhp;
{
off_t stat_offset, write_offset;
size_t blen, nw;
u_int32_t bytes, mbytes;
int group_sync, need_free, ret;
u_int8_t buf[8 * 1024], *bp;
write_offset = (off_t)fhp->pgno * fhp->pgsize + fhp->offset;
if ((ret = __os_ioinfo(dbenv, NULL, fhp, &mbytes, &bytes, NULL)) != 0)
return (ret);
stat_offset = (off_t)mbytes * MEGABYTE + bytes;
if (stat_offset >= write_offset)
return (0);
#undef ZF_LARGE_WRITE
#define ZF_LARGE_WRITE (64 * 1024)
if (write_offset - stat_offset > ZF_LARGE_WRITE) {
if ((ret = __os_calloc(dbenv, 1, ZF_LARGE_WRITE, &bp)) != 0)
return (ret);
blen = ZF_LARGE_WRITE;
need_free = 1;
} else {
bp = buf;
blen = sizeof(buf);
need_free = 0;
memset(buf, 0, sizeof(buf));
}
if ((ret = __os_seek(
dbenv, fhp, MEGABYTE, mbytes, bytes, 0, DB_OS_SEEK_SET)) != 0)
goto err;
for (group_sync = 0; stat_offset < write_offset; group_sync = 1) {
if (write_offset - stat_offset <= blen) {
blen = (size_t)(write_offset - stat_offset);
if (group_sync && (ret = __os_fsync(dbenv, fhp)) != 0)
goto err;
}
if ((ret = __os_physwrite(dbenv, fhp, bp, blen, &nw)) != 0)
goto err;
stat_offset += blen;
}
if ((ret = __os_fsync(dbenv, fhp)) != 0)
goto err;
mbytes = (u_int32_t)(write_offset / MEGABYTE);
bytes = (u_int32_t)(write_offset % MEGABYTE);
ret = __os_seek(dbenv, fhp, MEGABYTE, mbytes, bytes, 0, DB_OS_SEEK_SET);
err: if (need_free)
__os_free(dbenv, bp);
return (ret);
}
#endif