AsyncFile::~AsyncFile()
{
if (io_queue_release(aioContext))
{
throw AIOException(NATIVE_ERROR_CANT_RELEASE_AIO,"Can't release aio");
}
if (::close(fileHandle))
{
throw AIOException(NATIVE_ERROR_CANT_OPEN_CLOSE_FILE,"Can't close file");
}
free(events);
::pthread_mutex_destroy(&fileMutex);
::pthread_mutex_destroy(&pollerMutex);
}
JNIEXPORT void JNICALL Java_org_apache_activemq_artemis_jlibaio_LibaioContext_deleteContext(JNIEnv* env, jclass clazz, jobject contextPointer) {
int i;
struct io_control * theControl = getIOControl(env, contextPointer);
if (theControl == NULL) {
return;
}
struct iocb * iocb = getIOCB(theControl);
if (iocb == NULL) {
throwIOException(env, "Not enough space in libaio queue");
return;
}
// Submitting a dumb write so the loop finishes
io_prep_pwrite(iocb, dumbWriteHandler, 0, 0, 0);
iocb->data = (void *) -1;
if (!submit(env, theControl, iocb)) {
return;
}
// to make sure the poll has finished
pthread_mutex_lock(&(theControl->pollLock));
pthread_mutex_unlock(&(theControl->pollLock));
// To return any pending IOCBs
int result = io_getevents(theControl->ioContext, 0, 1, theControl->events, 0);
for (i = 0; i < result; i++) {
struct io_event * event = &(theControl->events[i]);
struct iocb * iocbp = event->obj;
putIOCB(theControl, iocbp);
}
io_queue_release(theControl->ioContext);
pthread_mutex_destroy(&(theControl->pollLock));
pthread_mutex_destroy(&(theControl->iocbLock));
// Releasing each individual iocb
for (i = 0; i < theControl->queueSize; i++) {
free(theControl->iocb[i]);
}
(*env)->DeleteGlobalRef(env, theControl->thisObject);
free(theControl->iocb);
free(theControl->events);
free(theControl);
}
int file_async_done(void)
{
unsigned int i;
if (file_io_mode != FILE_IO_MODE_ASYNC)
return 0;
for (i = 0; i < sb_globals.num_threads; i++)
{
io_queue_release(aio_ctxts[i].io_ctxt);
free(aio_ctxts[i].events);
}
free(aio_ctxts);
return 0;
}
/* this is the meat of the state machine. There is a list of
* active operations structs, and as each one finishes the required
* io it is moved to a list of finished operations. Once they have
* all finished whatever stage they were in, they are given the chance
* to restart and pick a different stage (read/write/random read etc)
*
* various timings are printed in between the stages, along with
* thread synchronization if there are more than one threads.
*/
int worker(struct thread_info *t)
{
struct io_oper *oper;
char *this_stage = NULL;
struct timeval stage_time;
int status = 0;
int iteration = 0;
int cnt;
aio_setup(&t->io_ctx, 512);
restart:
if (num_threads > 1) {
pthread_mutex_lock(&stage_mutex);
threads_starting++;
if (threads_starting == num_threads) {
threads_ending = 0;
gettimeofday(&global_stage_start_time, NULL);
pthread_cond_broadcast(&stage_cond);
}
while (threads_starting != num_threads)
pthread_cond_wait(&stage_cond, &stage_mutex);
pthread_mutex_unlock(&stage_mutex);
}
if (t->active_opers) {
this_stage = stage_name(t->active_opers->rw);
gettimeofday(&stage_time, NULL);
t->stage_mb_trans = 0;
}
cnt = 0;
/* first we send everything through aio */
while(t->active_opers && (cnt < iterations || iterations == RUN_FOREVER)) {
if (stonewall && threads_ending) {
oper = t->active_opers;
oper->stonewalled = 1;
oper_list_del(oper, &t->active_opers);
oper_list_add(oper, &t->finished_opers);
} else {
run_active_list(t, io_iter, max_io_submit);
}
cnt++;
}
if (latency_stats)
print_latency(t);
if (completion_latency_stats)
print_completion_latency(t);
/* then we wait for all the operations to finish */
oper = t->finished_opers;
do {
if (!oper)
break;
io_oper_wait(t, oper);
oper = oper->next;
} while(oper != t->finished_opers);
/* then we do an fsync to get the timing for any future operations
* right, and check to see if any of these need to get restarted
*/
oper = t->finished_opers;
while(oper) {
if (fsync_stages)
fsync(oper->fd);
t->stage_mb_trans += oper_mb_trans(oper);
if (restart_oper(oper)) {
oper_list_del(oper, &t->finished_opers);
oper_list_add(oper, &t->active_opers);
oper = t->finished_opers;
continue;
}
oper = oper->next;
if (oper == t->finished_opers)
break;
}
if (t->stage_mb_trans && t->num_files > 0) {
double seconds = time_since_now(&stage_time);
fprintf(stderr, "thread %llu %s totals (%.2f MB/s) %.2f MB in %.2fs\n",
(unsigned long long)(t - global_thread_info), this_stage,
t->stage_mb_trans/seconds, t->stage_mb_trans, seconds);
}
if (num_threads > 1) {
pthread_mutex_lock(&stage_mutex);
threads_ending++;
if (threads_ending == num_threads) {
threads_starting = 0;
pthread_cond_broadcast(&stage_cond);
global_thread_throughput(t, this_stage);
}
while(threads_ending != num_threads)
pthread_cond_wait(&stage_cond, &stage_mutex);
pthread_mutex_unlock(&stage_mutex);
}
/* someone got restarted, go back to the beginning */
if (t->active_opers && (cnt < iterations || iterations == RUN_FOREVER)) {
iteration++;
goto restart;
}
/* finally, free all the ram */
while(t->finished_opers) {
oper = t->finished_opers;
oper_list_del(oper, &t->finished_opers);
status = finish_oper(t, oper);
}
if (t->num_global_pending) {
fprintf(stderr, "global num pending is %d\n", t->num_global_pending);
}
io_queue_release(t->io_ctx);
return status;
}
/* 一次提交多个io请求的操作 */
int io_tio(char *pathname, int flag, int n, int operation)
{
int res, fd = 0, i = 0;
void *bufptr = NULL;
off_t offset = 0;
struct timespec timeout;
io_context_t myctx;
struct iocb iocb_array[AIO_MAXIO];
struct iocb *iocbps[AIO_MAXIO];
fd = open(pathname, flag, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
if (fd <= 0) {
printf("open for %s failed: %s\n", pathname, strerror(errno));
return -1;
}
res = io_queue_init(n, &myctx);
//printf (" res = %d \n", res);
for (i = 0; i < AIO_MAXIO; i++) {
switch (operation) {
case IO_CMD_PWRITE:
if (posix_memalign(&bufptr, alignment, AIO_BLKSIZE)) {
perror(" posix_memalign failed ");
return -1;
}
memset(bufptr, 0, AIO_BLKSIZE);
io_prep_pwrite(&iocb_array[i], fd, bufptr,
AIO_BLKSIZE, offset);
io_set_callback(&iocb_array[i], work_done);
iocbps[i] = &iocb_array[i];
offset += AIO_BLKSIZE;
break;
case IO_CMD_PREAD:
if (posix_memalign(&bufptr, alignment, AIO_BLKSIZE)) {
perror(" posix_memalign failed ");
return -1;
}
memset(bufptr, 0, AIO_BLKSIZE);
io_prep_pread(&iocb_array[i], fd, bufptr,
AIO_BLKSIZE, offset);
io_set_callback(&iocb_array[i], work_done);
iocbps[i] = &iocb_array[i];
offset += AIO_BLKSIZE;
break;
case IO_CMD_POLL:
case IO_CMD_NOOP:
break;
default:
tst_resm(TFAIL,
"Command failed; opcode returned: %d\n",
operation);
return -1;
break;
}
}
do {
res = io_submit(myctx, AIO_MAXIO, iocbps);
} while (res == -EAGAIN);
if (res < 0) {
io_error("io_submit tio", res);
}
/*
* We have submitted all the i/o requests. Wait for at least one to complete
* and call the callbacks.
*/
wait_count = AIO_MAXIO;
timeout.tv_sec = 30;
timeout.tv_nsec = 0;
switch (operation) {
case IO_CMD_PREAD:
case IO_CMD_PWRITE:
{
while (wait_count) {
res = io_wait_run(myctx, &timeout);
if (res < 0)
io_error("io_wait_run", res);
}
}
break;
}
close(fd);
for (i = 0; i < AIO_MAXIO; i++) {
if (iocb_array[i].u.c.buf != NULL) {
free(iocb_array[i].u.c.buf);
}
}
io_queue_release(myctx);
return 0;
}
void loop_aio(int fd, void *buf)
{
int res;
unsigned int elapsed;
/* i/o context initialization */
io_context_t myctx;
memset(&myctx, 0, sizeof(myctx));
if ((res = io_queue_init(maxinflight, &myctx)))
io_error("io_queue_init", res);
copied = 0;
inflight = 0;
if (opt_verbose)
printf("[run %d] start\n", runid);
while (copied < mycount)
{
struct iocb *ioq[maxsubmit];
int tosubmit = 0;
unsigned long long int index;
struct iocb *iocb;
struct timeval tv1, tv2;
/* filling a context with io queries */
while (copied + inflight + tosubmit < mycount &&
inflight + tosubmit < maxinflight &&
tosubmit < maxsubmit)
{
/* Simultaneous asynchronous operations using the same iocb produce undefined results. */
iocb = calloc(1, sizeof(struct iocb));
if (mode_rnd)
index = (mydevsize / myiosize) * random() / RAND_MAX;
else
index = copied + inflight + tosubmit;
if (mode_write)
io_prep_pwrite(iocb, fd, buf, myiosize, index * myiosize);
else
io_prep_pread(iocb, fd, buf, myiosize, index * myiosize);
io_set_callback(iocb, io_done);
ioq[tosubmit] = iocb;
tosubmit += 1;
}
/* if there are available slots for submitting queries, do it */
if (tosubmit)
{
/* submit io and check elapsed time */
gettimeofday(&tv1, NULL);
if ((res = io_submit(myctx, tosubmit, ioq)) != tosubmit)
{
printf("only %d io submitted\n", res);
io_error("io_submit write", res);
}
gettimeofday(&tv2, NULL);
elapsed = (tv2.tv_sec - tv1.tv_sec) * 1000 + (tv2.tv_usec - tv1.tv_usec) / 1000;
if (elapsed > 200)
printf("warning: io_submit() took %d ms, this is suspicious, maybe nr_request is too low.\n", elapsed);
/* inflight io += newly submitted io */
inflight += tosubmit;
}
/* handle completed io events */
if ((res = io_queue_run(myctx)) < 0)
io_error("io_queue_run", res);
if (inflight == maxinflight ||
(inflight && copied + inflight == mycount))
{
struct io_event event;
if ((res = io_getevents(myctx, 1, 1, &event, NULL)) < 0)
io_error("io_getevents", res);
if (res != 1)
errx(1, "no events?");
((io_callback_t)event.obj->data)(myctx, event.obj, event.res, event.res2);
}
}
io_queue_release(myctx);
}
static int io_tio(char *pathname, int flag, int operation)
{
int res, fd = 0, i = 0;
void *bufptr = NULL;
off_t offset = 0;
struct timespec timeout;
struct stat fi_stat;
size_t alignment;
io_context_t myctx;
struct iocb iocb_array[AIO_MAXIO];
struct iocb *iocbps[AIO_MAXIO];
fd = SAFE_OPEN(pathname, flag, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
/* determine the alignment from the blksize of the underlying device */
SAFE_FSTAT(fd, &fi_stat);
alignment = fi_stat.st_blksize;
res = io_queue_init(AIO_MAXIO, &myctx);
for (i = 0; i < AIO_MAXIO; i++) {
switch (operation) {
case IO_CMD_PWRITE:
if (posix_memalign(&bufptr, alignment, AIO_BLKSIZE)) {
tst_brk(TBROK | TERRNO, "posix_memalign failed");
return -1;
}
memset(bufptr, 0, AIO_BLKSIZE);
io_prep_pwrite(&iocb_array[i], fd, bufptr,
AIO_BLKSIZE, offset);
io_set_callback(&iocb_array[i], work_done);
iocbps[i] = &iocb_array[i];
offset += AIO_BLKSIZE;
break;
case IO_CMD_PREAD:
if (posix_memalign(&bufptr, alignment, AIO_BLKSIZE)) {
tst_brk(TBROK | TERRNO, "posix_memalign failed");
return -1;
}
memset(bufptr, 0, AIO_BLKSIZE);
io_prep_pread(&iocb_array[i], fd, bufptr,
AIO_BLKSIZE, offset);
io_set_callback(&iocb_array[i], work_done);
iocbps[i] = &iocb_array[i];
offset += AIO_BLKSIZE;
break;
default:
tst_res(TFAIL, "Command failed; opcode returned: %d\n", operation);
return -1;
break;
}
}
do {
res = io_submit(myctx, AIO_MAXIO, iocbps);
} while (res == -EAGAIN);
if (res < 0)
io_error("io_submit tio", res);
/*
* We have submitted all the i/o requests. Wait for them to complete and
* call the callbacks.
*/
wait_count = AIO_MAXIO;
timeout.tv_sec = 30;
timeout.tv_nsec = 0;
switch (operation) {
case IO_CMD_PREAD:
case IO_CMD_PWRITE:
{
while (wait_count) {
res = io_wait_run(myctx, &timeout);
if (res < 0)
io_error("io_wait_run", res);
}
}
break;
}
SAFE_CLOSE(fd);
for (i = 0; i < AIO_MAXIO; i++)
if (iocb_array[i].u.c.buf != NULL)
free(iocb_array[i].u.c.buf);
io_queue_release(myctx);
return 0;
}
