mirror of
https://github.com/MariaDB/server.git
synced 2025-01-22 14:54:20 +01:00
415ff301a5
{{{ svn merge -c28759 https://svn.tokutek.com/tokudb/toku/tokudb.3312 }}} . git-svn-id: file:///svn/toku/tokudb@28761 c7de825b-a66e-492c-adef-691d508d4ae1
612 lines
17 KiB
C
612 lines
17 KiB
C
#include <toku_portability.h>
|
|
#include <stdio.h>
|
|
#include <toku_assert.h>
|
|
#include <stdint.h>
|
|
#include <unistd.h>
|
|
#include <windows.h>
|
|
#include <toku_atomic.h>
|
|
#include <toku_time.h>
|
|
#include <fcntl.h>
|
|
|
|
int64_t
|
|
pread(int fildes, void *buf, size_t nbyte, int64_t offset) {
|
|
HANDLE filehandle;
|
|
OVERLAPPED win_offset = {0};
|
|
filehandle = (HANDLE)_get_osfhandle(fildes);
|
|
int64_t r;
|
|
if (filehandle==INVALID_HANDLE_VALUE) {
|
|
r = errno; assert(r!=0);
|
|
goto cleanup;
|
|
}
|
|
win_offset.Offset = offset % (1LL<<32LL);
|
|
win_offset.OffsetHigh = offset / (1LL<<32LL);
|
|
|
|
DWORD bytes_read;
|
|
r = ReadFile(filehandle, buf, nbyte, &bytes_read, &win_offset);
|
|
if (!r) {
|
|
r = GetLastError();
|
|
if (r==ERROR_HANDLE_EOF) r = bytes_read;
|
|
else {
|
|
errno = r;
|
|
r = -1;
|
|
}
|
|
}
|
|
else r = bytes_read;
|
|
|
|
// printf("%s: %d %p %u %I64d %I64d\n", __FUNCTION__, fildes, buf, nbyte, offset, r); fflush(stdout);
|
|
cleanup:
|
|
return r;
|
|
}
|
|
|
|
int64_t
|
|
pwrite(int fildes, const void *buf, size_t nbyte, int64_t offset) {
|
|
HANDLE filehandle;
|
|
OVERLAPPED win_offset = {0};
|
|
filehandle = (HANDLE)_get_osfhandle(fildes);
|
|
int64_t r;
|
|
if (filehandle==INVALID_HANDLE_VALUE) {
|
|
r = -1;
|
|
assert(errno!=0);
|
|
goto cleanup;
|
|
}
|
|
win_offset.Offset = offset % (1LL<<32LL);
|
|
win_offset.OffsetHigh = offset / (1LL<<32LL);
|
|
|
|
DWORD bytes_written;
|
|
r = WriteFile(filehandle, buf, nbyte, &bytes_written, &win_offset);
|
|
if (!r) {
|
|
errno = GetLastError();
|
|
if (errno == ERROR_HANDLE_DISK_FULL ||
|
|
errno == ERROR_DISK_FULL) {
|
|
errno = ENOSPC;
|
|
}
|
|
r = -1;
|
|
}
|
|
else r = bytes_written;
|
|
|
|
// printf("%s: %d %p %u %I64d %I64d\n", __FUNCTION__, fildes, buf, nbyte, offset, r); fflush(stdout);
|
|
cleanup:
|
|
return r;
|
|
}
|
|
|
|
int
|
|
fsync(int fd) {
|
|
int r = _commit(fd);
|
|
return r;
|
|
}
|
|
|
|
int
|
|
ftruncate(int fd, toku_off_t offset) {
|
|
int r = _chsize_s(fd, offset);
|
|
if (r!=0) {
|
|
r = -1;
|
|
assert(errno!=0);
|
|
}
|
|
return r;
|
|
}
|
|
|
|
int
|
|
truncate(const char *path, toku_off_t length) {
|
|
int r;
|
|
int saved_errno;
|
|
int fd = open(path, _O_BINARY|_O_RDWR, _S_IREAD|_S_IWRITE);
|
|
if (fd<0) {
|
|
r = -1;
|
|
goto done;
|
|
}
|
|
r = ftruncate(fd, length);
|
|
saved_errno = errno;
|
|
if (r!=0) {
|
|
r = -1;
|
|
assert(errno!=0);
|
|
}
|
|
int r2 = close(fd);
|
|
if (r==0) {
|
|
r = r2;
|
|
}
|
|
else {
|
|
errno = saved_errno;
|
|
}
|
|
done:
|
|
return r;
|
|
}
|
|
|
|
|
|
static ssize_t (*t_write)(int, const void *, size_t) = NULL;
|
|
static ssize_t (*t_full_write)(int, const void *, size_t) = NULL;
|
|
static ssize_t (*t_pwrite)(int, const void *, size_t, toku_off_t) = NULL;
|
|
static ssize_t (*t_full_pwrite)(int, const void *, size_t, toku_off_t) = NULL;
|
|
static FILE * (*t_fdopen)(int, const char *) = NULL;
|
|
static FILE * (*t_fopen)(const char *, const char *) = NULL;
|
|
static int (*t_open)(const char *, int, int) = NULL; // no implementation of variadic form until needed
|
|
static int (*t_fclose)(FILE *) = NULL;
|
|
|
|
int
|
|
toku_set_func_write (ssize_t (*write_fun)(int, const void *, size_t)) {
|
|
t_write = write_fun;
|
|
return 0;
|
|
}
|
|
|
|
int
|
|
toku_set_func_full_write (ssize_t (*write_fun)(int, const void *, size_t)) {
|
|
t_full_write = write_fun;
|
|
return 0;
|
|
}
|
|
|
|
int
|
|
toku_set_func_pwrite (ssize_t (*pwrite_fun)(int, const void *, size_t, toku_off_t)) {
|
|
t_pwrite = pwrite_fun;
|
|
return 0;
|
|
}
|
|
|
|
int
|
|
toku_set_func_full_pwrite (ssize_t (*pwrite_fun)(int, const void *, size_t, toku_off_t)) {
|
|
t_full_pwrite = pwrite_fun;
|
|
return 0;
|
|
}
|
|
|
|
int
|
|
toku_set_func_fdopen(FILE * (*fdopen_fun)(int, const char *)) {
|
|
t_fdopen = fdopen_fun;
|
|
return 0;
|
|
}
|
|
|
|
|
|
int
|
|
toku_set_func_fopen(FILE * (*fopen_fun)(const char *, const char *)) {
|
|
t_fopen = fopen_fun;
|
|
return 0;
|
|
}
|
|
|
|
|
|
int
|
|
toku_set_func_open(int (*open_fun)(const char *, int, int)) {
|
|
t_open = open_fun;
|
|
return 0;
|
|
}
|
|
|
|
int
|
|
toku_set_func_fclose(int (*fclose_fun)(FILE*)) {
|
|
t_fclose = fclose_fun;
|
|
return 0;
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
static int toku_assert_on_write_enospc = 0;
|
|
static const int toku_write_enospc_sleep = 1;
|
|
static uint64_t toku_write_enospc_last_report; // timestamp of most recent report to error log
|
|
static time_t toku_write_enospc_last_time; // timestamp of most recent ENOSPC
|
|
static uint32_t toku_write_enospc_current; // number of threads currently blocked on ENOSPC
|
|
static uint64_t toku_write_enospc_total; // total number of times ENOSPC was returned from an attempt to write
|
|
|
|
void toku_set_assert_on_write_enospc(int do_assert) {
|
|
toku_assert_on_write_enospc = do_assert;
|
|
}
|
|
|
|
void
|
|
toku_fs_get_write_info(time_t *enospc_last_time, uint64_t *enospc_current, uint64_t *enospc_total) {
|
|
*enospc_last_time = toku_write_enospc_last_time;
|
|
*enospc_current = toku_write_enospc_current;
|
|
*enospc_total = toku_write_enospc_total;
|
|
}
|
|
|
|
|
|
//Print any necessary errors
|
|
//Return whether we should try the write again.
|
|
static void
|
|
try_again_after_handling_write_error(int fd, size_t len, ssize_t r_write) {
|
|
int try_again = 0;
|
|
|
|
assert(r_write < 0);
|
|
int errno_write = errno;
|
|
assert(errno_write != 0);
|
|
switch (errno_write) {
|
|
case EINTR: { //The call was interrupted by a signal before any data was written; see signal(7).
|
|
char err_msg[sizeof("Write of [] bytes to fd=[] interrupted. Retrying.") + 20+10]; //64 bit is 20 chars, 32 bit is 10 chars
|
|
snprintf(err_msg, sizeof(err_msg), "Write of [%"PRIu64"] bytes to fd=[%d] interrupted. Retrying.", (uint64_t)len, fd);
|
|
perror(err_msg);
|
|
fflush(stderr);
|
|
try_again = 1;
|
|
break;
|
|
}
|
|
case ENOSPC: {
|
|
if (toku_assert_on_write_enospc) {
|
|
char err_msg[sizeof("Failed write of [] bytes to fd=[].") + 20+10]; //64 bit is 20 chars, 32 bit is 10 chars
|
|
snprintf(err_msg, sizeof(err_msg), "Failed write of [%"PRIu64"] bytes to fd=[%d].", (uint64_t)len, fd);
|
|
perror(err_msg);
|
|
fflush(stderr);
|
|
int out_of_disk_space = 1;
|
|
assert(!out_of_disk_space); //Give an error message that might be useful if this is the only one that survives.
|
|
} else {
|
|
toku_sync_fetch_and_increment_uint64(&toku_write_enospc_total);
|
|
toku_sync_fetch_and_increment_uint32(&toku_write_enospc_current);
|
|
|
|
time_t tnow = time(0);
|
|
toku_write_enospc_last_time = tnow;
|
|
if (toku_write_enospc_last_report == 0 || tnow - toku_write_enospc_last_report >= 60) {
|
|
toku_write_enospc_last_report = tnow;
|
|
|
|
const int tstr_length = 26;
|
|
char tstr[tstr_length];
|
|
time_t t = time(0);
|
|
ctime_r(&t, tstr);
|
|
|
|
#if 0
|
|
//TODO: Find out how to get name from fd or handle
|
|
//In vista/server08 and later there exists GetFinalPathNameByHandle()
|
|
//In XP there exists:
|
|
// Example code that requires on kernel-level functions that
|
|
// are not guaranteed to stay around.
|
|
// http://rubyforge.org/pipermail/win32utils-devel/2008-May/001091.html
|
|
// Example code that works for files of length > 0 (according
|
|
// to author). This one appears to not require unsafe apis
|
|
// (not guaranteed to stick around)
|
|
// http://msdn.microsoft.com/en-us/library/aa366789%28VS.85%29.aspx
|
|
// Can we use runtime-checks to determine what version of OS we
|
|
// are using so we can choose which function to use?
|
|
// We CAN do compile-time checks, but then we would need to
|
|
// release multiple versions of binary.
|
|
// Is this important?
|
|
//
|
|
const int MY_MAX_PATH = 256;
|
|
char fname[MY_MAX_PATH], symname[MY_MAX_PATH+1];
|
|
sprintf(fname, "/proc/%d/fd/%d", getpid(), fd);
|
|
ssize_t n = readlink(fname, symname, MY_MAX_PATH);
|
|
if ((int)n == -1)
|
|
fprintf(stderr, "%.24s Tokudb No space when writing %"PRIu64" bytes to fd=%d ", tstr, (uint64_t) len, fd);
|
|
else {
|
|
tstr[n] = 0; // readlink doesn't append a NUL to the end of the buffer.
|
|
fprintf(stderr, "%.24s Tokudb No space when writing %"PRIu64" bytes to %*s ", tstr, (uint64_t) len, (int) n, symname);
|
|
}
|
|
#else
|
|
fprintf(stderr, "%.24s Tokudb No space when writing %"PRIu64" bytes to fd=%d ", tstr, (uint64_t) len, fd);
|
|
#endif
|
|
|
|
fprintf(stderr, "retry in %d second%s\n", toku_write_enospc_sleep, toku_write_enospc_sleep > 1 ? "s" : "");
|
|
fflush(stderr);
|
|
}
|
|
sleep(toku_write_enospc_sleep);
|
|
try_again = 1;
|
|
toku_sync_fetch_and_decrement_uint32(&toku_write_enospc_current);
|
|
break;
|
|
}
|
|
}
|
|
default:
|
|
break;
|
|
}
|
|
assert(try_again);
|
|
errno = errno_write;
|
|
}
|
|
|
|
|
|
void
|
|
toku_os_full_write (int fd, const void *buf, size_t len) {
|
|
const uint8_t *bp = (const uint8_t *) buf;
|
|
while (len > 0) {
|
|
ssize_t r;
|
|
if (t_full_write) {
|
|
r = t_full_write(fd, bp, len);
|
|
} else {
|
|
r = write(fd, bp, len);
|
|
}
|
|
if (r > 0) {
|
|
len -= r;
|
|
bp += r;
|
|
}
|
|
else {
|
|
try_again_after_handling_write_error(fd, len, r);
|
|
}
|
|
}
|
|
assert(len == 0);
|
|
}
|
|
|
|
int
|
|
toku_os_write (int fd, const void *buf, size_t len) {
|
|
const uint8_t *bp = (const uint8_t *) buf;
|
|
int result = 0;
|
|
while (len > 0) {
|
|
ssize_t r;
|
|
if (t_write) {
|
|
r = t_write(fd, bp, len);
|
|
} else {
|
|
r = write(fd, bp, len);
|
|
}
|
|
if (r < 0) {
|
|
result = errno;
|
|
break;
|
|
}
|
|
len -= r;
|
|
bp += r;
|
|
}
|
|
return result;
|
|
}
|
|
|
|
void
|
|
toku_os_full_pwrite (int fd, const void *buf, size_t len, toku_off_t off) {
|
|
const uint8_t *bp = (const uint8_t *) buf;
|
|
while (len > 0) {
|
|
ssize_t r;
|
|
if (t_full_pwrite) {
|
|
r = t_full_pwrite(fd, bp, len, off);
|
|
} else {
|
|
r = pwrite(fd, bp, len, off);
|
|
}
|
|
if (r > 0) {
|
|
len -= r;
|
|
bp += r;
|
|
off += r;
|
|
}
|
|
else {
|
|
try_again_after_handling_write_error(fd, len, r);
|
|
}
|
|
}
|
|
assert(len == 0);
|
|
}
|
|
|
|
ssize_t
|
|
toku_os_pwrite (int fd, const void *buf, size_t len, toku_off_t off) {
|
|
const uint8_t *bp = (const uint8_t *) buf;
|
|
ssize_t result = 0;
|
|
while (len > 0) {
|
|
ssize_t r;
|
|
if (t_pwrite) {
|
|
r = t_pwrite(fd, bp, len, off);
|
|
} else {
|
|
r = pwrite(fd, bp, len, off);
|
|
}
|
|
if (r < 0) {
|
|
result = errno;
|
|
break;
|
|
}
|
|
len -= r;
|
|
bp += r;
|
|
off += r;
|
|
}
|
|
return result;
|
|
}
|
|
|
|
FILE *
|
|
toku_os_fdopen(int fildes, const char *mode) {
|
|
FILE * rval;
|
|
if (t_fdopen)
|
|
rval = t_fdopen(fildes, mode);
|
|
else
|
|
rval = fdopen(fildes, mode);
|
|
return rval;
|
|
}
|
|
|
|
|
|
FILE *
|
|
toku_os_fopen(const char *filename, const char *mode){
|
|
FILE * rval;
|
|
if (t_fopen)
|
|
rval = t_fopen(filename, mode);
|
|
else
|
|
rval = fopen(filename, mode);
|
|
return rval;
|
|
}
|
|
|
|
int
|
|
toku_os_open(const char *path, int oflag, int mode) {
|
|
int rval;
|
|
if (t_open)
|
|
rval = t_open(path, oflag, mode);
|
|
else
|
|
rval = open(path, oflag, mode);
|
|
return rval;
|
|
}
|
|
|
|
int
|
|
toku_os_fclose(FILE * stream) {
|
|
int rval = -1;
|
|
if (t_fclose)
|
|
rval = t_fclose(stream);
|
|
else { // if EINTR, retry until success
|
|
while (rval != 0) {
|
|
rval = fclose(stream);
|
|
if (rval && (errno != EINTR))
|
|
break;
|
|
}
|
|
}
|
|
return rval;
|
|
}
|
|
|
|
int
|
|
toku_os_close (int fd) { // if EINTR, retry until success
|
|
int r = -1;
|
|
while (r != 0) {
|
|
r = close(fd);
|
|
if (r) {
|
|
int rr = errno;
|
|
if (rr!=EINTR) printf("rr=%d (%s)\n", rr, strerror(rr));
|
|
assert(rr==EINTR);
|
|
}
|
|
}
|
|
return r;
|
|
}
|
|
|
|
|
|
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
// fsync logic:
|
|
|
|
|
|
// t_fsync exists for testing purposes only
|
|
static int (*t_fsync)(int) = 0;
|
|
static uint64_t toku_fsync_count;
|
|
static uint64_t toku_fsync_time;
|
|
|
|
static uint64_t sched_fsync_count;
|
|
static uint64_t sched_fsync_time;
|
|
|
|
#if !TOKU_WINDOWS_HAS_ATOMIC_64
|
|
static toku_pthread_mutex_t fsync_lock;
|
|
#endif
|
|
|
|
int
|
|
toku_fsync_init(void) {
|
|
int r = 0;
|
|
#if !TOKU_WINDOWS_HAS_ATOMIC_64
|
|
r = toku_pthread_mutex_init(&fsync_lock, NULL); assert(r == 0);
|
|
#endif
|
|
return r;
|
|
}
|
|
|
|
int
|
|
toku_fsync_destroy(void) {
|
|
int r = 0;
|
|
#if !TOKU_WINDOWS_HAS_ATOMIC_64
|
|
r = toku_pthread_mutex_destroy(&fsync_lock); assert(r == 0);
|
|
#endif
|
|
return r;
|
|
}
|
|
|
|
int
|
|
toku_set_func_fsync(int (*fsync_function)(int)) {
|
|
t_fsync = fsync_function;
|
|
return 0;
|
|
}
|
|
static uint64_t get_tnow(void) {
|
|
struct timeval tv;
|
|
int r = gettimeofday(&tv, NULL); assert(r == 0);
|
|
return tv.tv_sec * 1000000ULL + tv.tv_usec;
|
|
}
|
|
|
|
|
|
// keep trying if fsync fails because of EINTR
|
|
static int
|
|
toku_file_fsync_internal (int fd, uint64_t *duration_p) {
|
|
uint64_t tstart = get_tnow();
|
|
int r = -1;
|
|
while (r != 0) {
|
|
if (t_fsync)
|
|
r = t_fsync(fd);
|
|
else
|
|
r = fsync(fd);
|
|
if (r) {
|
|
int rr = errno;
|
|
if (rr!=EINTR) printf("rr=%d (%s)\n", rr, strerror(rr));
|
|
assert(rr==EINTR);
|
|
}
|
|
}
|
|
uint64_t duration;
|
|
duration = get_tnow() - tstart;
|
|
#if TOKU_WINDOWS_HAS_ATOMIC_64
|
|
toku_sync_fetch_and_increment_uint64(&toku_fsync_count);
|
|
toku_sync_fetch_and_add_uint64(&toku_fsync_time, duration);
|
|
#else
|
|
//These two need to be fully 64 bit and atomic.
|
|
//The windows atomic add 64 bit is not available.
|
|
//toku_sync_fetch_and_add_uint64 (and increment) treat it as 32 bit, and
|
|
//would overflow.
|
|
//Even on 32 bit machines, aligned 64 bit writes/writes are atomic, so we just
|
|
//need to make sure there's only one writer for these two variables.
|
|
//Protect with a mutex. Fsync is rare/slow enough that this should be ok.
|
|
int r_mutex;
|
|
r_mutex = toku_pthread_mutex_lock(&fsync_lock); assert(r_mutex == 0);
|
|
toku_fsync_count++;
|
|
toku_fsync_time += duration;
|
|
r_mutex = toku_pthread_mutex_unlock(&fsync_lock); assert(r_mutex == 0);
|
|
#endif
|
|
if (duration_p) *duration_p = duration;
|
|
return r;
|
|
}
|
|
|
|
// keep trying if fsync fails because of EINTR
|
|
int
|
|
toku_file_fsync_without_accounting (int fd) {
|
|
int r = toku_file_fsync_internal(fd, NULL);
|
|
return r;
|
|
}
|
|
|
|
int
|
|
toku_file_fsync(int fd) {
|
|
uint64_t duration;
|
|
int r = toku_file_fsync_internal(fd, &duration);
|
|
#if TOKU_WINDOWS_HAS_ATOMIC_64
|
|
toku_sync_fetch_and_increment_uint64(&sched_fsync_count);
|
|
toku_sync_fetch_and_add_uint64(&sched_fsync_time, duration);
|
|
#else
|
|
//These two need to be fully 64 bit and atomic.
|
|
//The windows atomic add 64 bit is not available.
|
|
//toku_sync_fetch_and_add_uint64 (and increment) treat it as 32 bit, and
|
|
//would overflow.
|
|
//Even on 32 bit machines, aligned 64 bit writes/writes are atomic, so we just
|
|
//need to make sure there's only one writer for these two variables.
|
|
//Protect with a mutex. Fsync is rare/slow enough that this should be ok.
|
|
int r_mutex;
|
|
r_mutex = toku_pthread_mutex_lock(&fsync_lock); assert(r_mutex == 0);
|
|
sched_fsync_count++;
|
|
sched_fsync_time += duration;
|
|
r_mutex = toku_pthread_mutex_unlock(&fsync_lock); assert(r_mutex == 0);
|
|
#endif
|
|
return r;
|
|
}
|
|
|
|
// for real accounting
|
|
void
|
|
toku_get_fsync_times(uint64_t *fsync_count, uint64_t *fsync_time) {
|
|
*fsync_count = toku_fsync_count;
|
|
*fsync_time = toku_fsync_time;
|
|
}
|
|
|
|
// for scheduling algorithm only
|
|
void
|
|
toku_get_fsync_sched(uint64_t *fsync_count, uint64_t *fsync_time) {
|
|
*fsync_count = sched_fsync_count;
|
|
*fsync_time = sched_fsync_time;
|
|
}
|
|
|
|
static toku_pthread_mutex_t mkstemp_lock;
|
|
|
|
int
|
|
toku_mkstemp_init(void) {
|
|
int r = 0;
|
|
r = toku_pthread_mutex_init(&mkstemp_lock, NULL); assert(r == 0);
|
|
return r;
|
|
}
|
|
|
|
int
|
|
toku_mkstemp_destroy(void) {
|
|
int r = 0;
|
|
r = toku_pthread_mutex_destroy(&mkstemp_lock); assert(r == 0);
|
|
return r;
|
|
}
|
|
|
|
int mkstemp (char * template) {
|
|
int fd;
|
|
int r_mutex;
|
|
r_mutex = toku_pthread_mutex_lock(&mkstemp_lock);
|
|
assert(r_mutex == 0);
|
|
errno_t err = _mktemp_s(template, strlen(template)+1);
|
|
if (err!=0) {
|
|
fd = -1;
|
|
errno = err;
|
|
goto cleanup;
|
|
}
|
|
assert(err==0);
|
|
fd = open(template, _O_BINARY|_O_CREAT|_O_SHORT_LIVED|_O_EXCL|_O_RDWR, _S_IREAD|_S_IWRITE);
|
|
cleanup:
|
|
r_mutex = toku_pthread_mutex_unlock(&mkstemp_lock);
|
|
assert(r_mutex == 0);
|
|
return fd;
|
|
}
|
|
|
|
toku_off_t
|
|
ftello(FILE *stream) {
|
|
toku_off_t offset = _ftelli64(stream);
|
|
return offset;
|
|
}
|
|
|
|
ssize_t
|
|
toku_os_read(int fd, void *buf, size_t count) {
|
|
return read(fd, buf, count);
|
|
}
|
|
|
|
|
|
|
|
|