mars/mars_aio.c

1133 lines
26 KiB
C
Raw Normal View History

2010-11-26 13:45:10 +00:00
// (c) 2010 Thomas Schoebel-Theuer / 1&1 Internet AG
//#define BRICK_DEBUGGING
//#define MARS_DEBUGGING
2011-02-23 20:48:06 +00:00
//#define IO_DEBUGGING
2010-11-26 13:45:10 +00:00
#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/string.h>
#include <linux/list.h>
#include <linux/types.h>
#include <linux/blkdev.h>
#include <linux/spinlock.h>
#include <linux/wait.h>
#include <linux/file.h>
#include "mars.h"
2012-10-12 12:17:34 +00:00
#include "lib_timing.h"
2010-11-26 13:45:10 +00:00
#include "mars_aio.h"
2010-11-26 13:45:10 +00:00
#define MARS_MAX_AIO 1024
#define MARS_MAX_AIO_READ 32
static struct timing_stats timings[3] = {};
2011-03-07 10:27:38 +00:00
struct threshold aio_submit_threshold = {
.thr_ban = &mars_global_ban,
.thr_limit = AIO_SUBMIT_MAX_LATENCY,
.thr_factor = 10,
.thr_plus = 10000,
};
EXPORT_SYMBOL_GPL(aio_submit_threshold);
struct threshold aio_io_threshold[2] = {
[0] = {
.thr_ban = &mars_global_ban,
.thr_limit = AIO_IO_R_MAX_LATENCY,
.thr_factor = 100,
.thr_plus = 0,
},
[1] = {
.thr_ban = &mars_global_ban,
.thr_limit = AIO_IO_W_MAX_LATENCY,
.thr_factor = 100,
.thr_plus = 0,
},
};
EXPORT_SYMBOL_GPL(aio_io_threshold);
2011-05-13 11:19:28 +00:00
struct threshold aio_sync_threshold = {
.thr_ban = &mars_global_ban,
.thr_limit = AIO_SYNC_MAX_LATENCY,
.thr_factor = 100,
.thr_plus = 0,
};
EXPORT_SYMBOL_GPL(aio_sync_threshold);
2011-05-13 11:19:28 +00:00
///////////////////////// own type definitions ////////////////////////
2011-05-13 11:19:28 +00:00
2011-03-07 10:27:38 +00:00
////////////////// some helpers //////////////////
2011-03-10 11:40:06 +00:00
static inline
void _enqueue(struct aio_threadinfo *tinfo, struct aio_mref_aspect *mref_a, int prio, bool at_end)
2011-03-07 10:27:38 +00:00
{
unsigned long flags;
2011-05-19 11:36:00 +00:00
#if 1
prio++;
if (unlikely(prio < 0)) {
2011-05-19 11:36:00 +00:00
prio = 0;
} else if (unlikely(prio >= MARS_PRIO_NR)) {
prio = MARS_PRIO_NR - 1;
2011-05-19 11:36:00 +00:00
}
#else
prio = 0;
#endif
2011-03-07 10:27:38 +00:00
mref_a->enqueue_stamp = cpu_clock(raw_smp_processor_id());
2011-03-07 10:27:38 +00:00
traced_lock(&tinfo->lock, flags);
2011-03-10 11:40:06 +00:00
if (at_end) {
list_add_tail(&mref_a->io_head, &tinfo->mref_list[prio]);
} else {
list_add(&mref_a->io_head, &tinfo->mref_list[prio]);
}
tinfo->queued[prio]++;
tinfo->queued_sum++;
2011-03-07 10:27:38 +00:00
traced_unlock(&tinfo->lock, flags);
2011-05-06 10:25:52 +00:00
atomic_inc(&tinfo->total_enqueue_count);
wake_up_interruptible_all(&tinfo->event);
2011-03-07 10:27:38 +00:00
}
2011-03-10 11:40:06 +00:00
static inline
struct aio_mref_aspect *_dequeue(struct aio_threadinfo *tinfo)
2011-03-07 10:27:38 +00:00
{
2011-03-10 11:40:06 +00:00
struct aio_mref_aspect *mref_a = NULL;
int prio;
unsigned long flags = 0;
2011-03-07 10:27:38 +00:00
traced_lock(&tinfo->lock, flags);
2011-03-07 10:27:38 +00:00
2011-05-19 11:36:00 +00:00
for (prio = 0; prio < MARS_PRIO_NR; prio++) {
struct list_head *start = &tinfo->mref_list[prio];
struct list_head *tmp = start->next;
if (tmp != start) {
list_del_init(tmp);
tinfo->queued[prio]--;
tinfo->queued_sum--;
2011-03-10 11:40:06 +00:00
mref_a = container_of(tmp, struct aio_mref_aspect, io_head);
goto done;
}
2011-03-07 10:27:38 +00:00
}
2011-03-10 11:40:06 +00:00
done:
traced_unlock(&tinfo->lock, flags);
if (likely(mref_a && mref_a->object)) {
unsigned long long latency;
latency = cpu_clock(raw_smp_processor_id()) - mref_a->enqueue_stamp;
threshold_check(&aio_io_threshold[mref_a->object->ref_rw & 1], latency);
}
2011-03-08 16:45:52 +00:00
return mref_a;
}
2010-11-26 13:45:10 +00:00
////////////////// own brick / input / output operations //////////////////
2011-03-07 10:27:38 +00:00
static int aio_ref_get(struct aio_output *output, struct mref_object *mref)
2010-11-26 13:45:10 +00:00
{
2011-03-07 10:27:38 +00:00
struct file *file = output->filp;
if (mref->ref_initialized) {
_mref_get(mref);
2011-04-14 14:21:26 +00:00
return mref->ref_len;
}
2011-03-07 10:27:38 +00:00
if (file) {
2011-03-24 16:05:46 +00:00
loff_t total_size = i_size_read(file->f_mapping->host);
mref->ref_total_size = total_size;
/* Only check reads.
* Writes behind EOF are always allowed (sparse files)
*/
if (!mref->ref_may_write) {
loff_t len = total_size - mref->ref_pos;
if (unlikely(len <= 0)) {
2011-06-10 13:57:52 +00:00
/* Special case: allow reads starting _exactly_ at EOF when a timeout is specified.
2011-03-24 16:05:46 +00:00
*/
if (len < 0 || mref->ref_timeout <= 0) {
MARS_DBG("ENODATA %lld\n", len);
return -ENODATA;
}
}
// Shorten below EOF, but allow special case
if (mref->ref_len > len && len > 0) {
mref->ref_len = len;
}
}
2011-03-07 10:27:38 +00:00
}
2011-03-24 16:05:46 +00:00
/* Buffered IO.
2010-11-26 13:45:10 +00:00
*/
2010-12-10 17:40:20 +00:00
if (!mref->ref_data) {
struct aio_mref_aspect *mref_a = aio_mref_get_aspect(output->brick, mref);
2010-12-10 17:40:20 +00:00
if (!mref_a)
return -EILSEQ;
2011-07-28 11:41:06 +00:00
if (mref->ref_len <= 0) {
MARS_ERR("bad ref_len = %d\n", mref->ref_len);
return -ENOMEM;
}
2011-08-12 11:09:48 +00:00
mref->ref_data = brick_block_alloc(mref->ref_pos, (mref_a->alloc_len = mref->ref_len));
2011-03-24 16:05:46 +00:00
if (!mref->ref_data) {
2011-07-15 10:12:06 +00:00
MARS_ERR("ENOMEM %d bytes\n", mref->ref_len);
2010-12-10 17:40:20 +00:00
return -ENOMEM;
2011-03-24 16:05:46 +00:00
}
2011-03-18 13:15:40 +00:00
#if 0 // ???
2010-12-10 17:40:20 +00:00
mref->ref_flags = 0;
#endif
2011-03-18 13:15:40 +00:00
mref_a->do_dealloc = true;
2011-04-12 15:31:08 +00:00
atomic_inc(&output->total_alloc_count);
atomic_inc(&output->alloc_count);
2010-12-10 17:40:20 +00:00
}
2010-11-26 13:45:10 +00:00
_mref_get_first(mref);
2011-04-14 14:21:26 +00:00
return mref->ref_len;
2010-11-26 13:45:10 +00:00
}
2011-03-07 10:27:38 +00:00
static void aio_ref_put(struct aio_output *output, struct mref_object *mref)
2010-11-26 13:45:10 +00:00
{
2011-03-07 10:27:38 +00:00
struct file *file = output->filp;
struct aio_mref_aspect *mref_a;
if (!_mref_put(mref)) {
2011-03-07 10:27:38 +00:00
goto done;
}
2011-03-18 13:15:40 +00:00
if (file) {
mref->ref_total_size = i_size_read(file->f_mapping->host);
}
mref_a = aio_mref_get_aspect(output->brick, mref);
2010-12-10 17:40:20 +00:00
if (mref_a && mref_a->do_dealloc) {
2011-08-12 11:09:48 +00:00
brick_block_free(mref->ref_data, mref_a->alloc_len);
2011-04-12 15:31:08 +00:00
atomic_dec(&output->alloc_count);
2010-12-10 17:40:20 +00:00
}
2011-03-07 10:27:38 +00:00
aio_free_mref(mref);
done:;
2010-11-26 13:45:10 +00:00
}
2011-03-07 10:27:38 +00:00
static
void _complete(struct aio_output *output, struct mref_object *mref, int err)
{
_mref_check(mref);
2011-03-27 15:18:38 +00:00
mars_trace(mref, "aio_endio");
2011-03-07 10:27:38 +00:00
if (err < 0) {
2011-07-15 10:12:06 +00:00
MARS_ERR("IO error %d at pos=%lld len=%d (mref=%p ref_data=%p)\n", err, mref->ref_pos, mref->ref_len, mref, mref->ref_data);
2011-03-07 10:27:38 +00:00
} else {
2012-08-01 10:09:49 +00:00
mref_checksum(mref);
2011-03-07 10:27:38 +00:00
mref->ref_flags |= MREF_UPTODATE;
}
2011-04-12 15:31:08 +00:00
CHECKED_CALLBACK(mref, err, err_found);
2011-04-12 15:31:08 +00:00
2011-05-19 11:36:00 +00:00
done:
2011-04-12 15:31:08 +00:00
if (mref->ref_rw) {
atomic_dec(&output->write_count);
} else {
atomic_dec(&output->read_count);
}
2011-03-07 10:27:38 +00:00
aio_ref_put(output, mref);
atomic_dec(&mars_global_io_flying);
2011-05-19 11:36:00 +00:00
return;
err_found:
MARS_FAT("giving up...\n");
goto done;
2011-03-07 10:27:38 +00:00
}
2011-07-01 14:07:56 +00:00
static
void _complete_all(struct list_head *tmp_list, struct aio_output *output, int err)
{
while (!list_empty(tmp_list)) {
struct list_head *tmp = tmp_list->next;
struct aio_mref_aspect *mref_a = container_of(tmp, struct aio_mref_aspect, io_head);
list_del_init(tmp);
_complete(output, mref_a->object, err);
}
}
2011-03-07 10:27:38 +00:00
static void aio_ref_io(struct aio_output *output, struct mref_object *mref)
2010-11-26 13:45:10 +00:00
{
struct aio_threadinfo *tinfo = &output->tinfo[0];
2011-03-07 10:27:38 +00:00
struct aio_mref_aspect *mref_a;
int err = -EINVAL;
2010-11-26 13:45:10 +00:00
_mref_get(mref);
atomic_inc(&mars_global_io_flying);
2010-11-26 13:45:10 +00:00
2011-04-12 15:31:08 +00:00
// statistics
if (mref->ref_rw) {
atomic_inc(&output->total_write_count);
atomic_inc(&output->write_count);
} else {
atomic_inc(&output->total_read_count);
atomic_inc(&output->read_count);
}
2010-11-26 13:45:10 +00:00
if (unlikely(!output->filp)) {
goto done;
}
2011-02-23 20:48:06 +00:00
MARS_IO("AIO rw=%d pos=%lld len=%d data=%p\n", mref->ref_rw, mref->ref_pos, mref->ref_len, mref->ref_data);
2010-11-26 13:45:10 +00:00
mref_a = aio_mref_get_aspect(output->brick, mref);
2011-03-07 10:27:38 +00:00
if (unlikely(!mref_a)) {
goto done;
}
2011-03-10 11:40:06 +00:00
_enqueue(tinfo, mref_a, mref->ref_prio, true);
2010-11-26 13:45:10 +00:00
return;
done:
2011-03-07 10:27:38 +00:00
_complete(output, mref, err);
2010-11-26 13:45:10 +00:00
}
2011-03-07 10:27:38 +00:00
static int aio_submit(struct aio_output *output, struct aio_mref_aspect *mref_a, bool use_fdsync)
2010-11-26 13:45:10 +00:00
{
2010-12-15 12:13:18 +00:00
struct mref_object *mref = mref_a->object;
2010-11-26 13:45:10 +00:00
mm_segment_t oldfs;
int res;
struct iocb iocb = {
.aio_data = (__u64)mref_a,
2010-12-10 17:40:20 +00:00
.aio_lio_opcode = use_fdsync ? IOCB_CMD_FDSYNC : (mref->ref_rw != 0 ? IOCB_CMD_PWRITE : IOCB_CMD_PREAD),
2010-11-26 13:45:10 +00:00
.aio_fildes = output->fd,
.aio_buf = (unsigned long)mref->ref_data,
.aio_nbytes = mref->ref_len,
.aio_offset = mref->ref_pos,
2011-03-24 16:05:46 +00:00
// .aio_reqprio = something(mref->ref_prio) field exists, but not yet implemented in kernelspace :(
2010-11-26 13:45:10 +00:00
};
struct iocb *iocbp = &iocb;
unsigned long long latency;
2010-11-26 13:45:10 +00:00
2011-03-27 15:18:38 +00:00
mars_trace(mref, "aio_submit");
2010-11-26 13:45:10 +00:00
oldfs = get_fs();
set_fs(get_ds());
latency = TIME_STATS(&timings[mref->ref_rw & 1], res = sys_io_submit(output->ctxp, 1, &iocbp));
2010-11-26 13:45:10 +00:00
set_fs(oldfs);
threshold_check(&aio_submit_threshold, latency);
2010-11-26 13:45:10 +00:00
if (res < 0 && res != -EAGAIN)
MARS_ERR("error = %d\n", res);
2010-11-26 13:45:10 +00:00
return res;
}
2011-03-07 10:27:38 +00:00
static int aio_submit_dummy(struct aio_output *output)
2011-02-23 20:48:06 +00:00
{
mm_segment_t oldfs;
int res;
2011-08-25 10:16:32 +00:00
int dummy;
2011-02-23 20:48:06 +00:00
struct iocb iocb = {
2011-08-25 10:16:32 +00:00
.aio_buf = (__u64)&dummy,
2011-02-23 20:48:06 +00:00
};
struct iocb *iocbp = &iocb;
oldfs = get_fs();
set_fs(get_ds());
res = sys_io_submit(output->ctxp, 1, &iocbp);
set_fs(oldfs);
return res;
}
2011-08-26 10:31:56 +00:00
static
int aio_start_thread(
struct aio_output *output,
struct aio_threadinfo *tinfo,
int(*fn)(void*),
char class)
2010-11-26 13:45:10 +00:00
{
2011-08-26 10:31:56 +00:00
int j;
for (j = 0; j < MARS_PRIO_NR; j++) {
INIT_LIST_HEAD(&tinfo->mref_list[j]);
}
tinfo->output = output;
spin_lock_init(&tinfo->lock);
init_waitqueue_head(&tinfo->event);
init_waitqueue_head(&tinfo->terminate_event);
2011-08-26 10:31:56 +00:00
tinfo->terminated = false;
tinfo->thread = brick_thread_create(fn, tinfo, "mars_aio_%c%d", class, output->index);
if (unlikely(!tinfo->thread)) {
2011-08-26 10:31:56 +00:00
MARS_ERR("cannot create thread\n");
return -ENOENT;
2010-11-26 13:45:10 +00:00
}
return 0;
}
2011-08-26 10:31:56 +00:00
static
void aio_stop_thread(struct aio_output *output, int i, bool do_submit_dummy)
2010-11-26 13:45:10 +00:00
{
2011-08-26 10:31:56 +00:00
struct aio_threadinfo *tinfo = &output->tinfo[i];
2010-11-26 13:45:10 +00:00
2011-08-26 10:31:56 +00:00
if (tinfo->thread) {
MARS_INF("stopping thread %d ...\n", i);
brick_thread_stop_nowait(tinfo->thread);
2011-03-02 09:30:56 +00:00
2011-08-26 10:31:56 +00:00
// workaround for waking up the receiver thread. TODO: check whether signal handlong could do better.
if (do_submit_dummy) {
MARS_INF("submitting dummy for wakeup %d...\n", i);
2011-08-26 10:31:56 +00:00
aio_submit_dummy(output);
2011-03-02 09:30:56 +00:00
}
2011-03-07 10:27:38 +00:00
2011-08-26 10:31:56 +00:00
// wait for termination
MARS_INF("waiting for thread %d ...\n", i);
wait_event_interruptible_timeout(
tinfo->terminate_event,
2011-08-26 10:31:56 +00:00
tinfo->terminated,
(60 - i * 2) * HZ);
if (likely(tinfo->terminated)) {
brick_thread_stop(tinfo->thread);
2011-08-26 10:31:56 +00:00
} else {
MARS_ERR("thread %d did not terminate - leaving a zombie\n", i);
2010-11-26 13:45:10 +00:00
}
}
}
2011-07-01 14:07:56 +00:00
static
int aio_sync(struct file *file)
{
int err;
2012-09-18 06:13:33 +00:00
err = filemap_write_and_wait_range(file->f_mapping, 0, LLONG_MAX);
2011-07-01 14:07:56 +00:00
return err;
}
static
void aio_sync_all(struct aio_output *output, struct list_head *tmp_list)
{
unsigned long long latency;
2011-07-01 14:07:56 +00:00
int err;
output->fdsync_active = true;
atomic_inc(&output->total_fdsync_count);
latency = TIME_STATS(
&timings[2],
err = aio_sync(output->filp)
);
2011-07-01 14:07:56 +00:00
threshold_check(&aio_sync_threshold, latency);
2011-07-01 14:07:56 +00:00
output->fdsync_active = false;
wake_up_interruptible_all(&output->fdsync_event);
if (err < 0) {
MARS_ERR("FDSYNC error %d\n", err);
}
/* Signal completion for the whole list.
* No locking needed, it's on the stack.
*/
_complete_all(tmp_list, output, err);
}
#ifdef USE_CLEVER_SYNC
static
int sync_cmp(struct pairing_heap_sync *_a, struct pairing_heap_sync *_b)
{
struct aio_mref_aspect *a = container_of(_a, struct aio_mref_aspect, heap_head);
struct aio_mref_aspect *b = container_of(_b, struct aio_mref_aspect, heap_head);
struct mref_object *ao = a->object;
struct mref_object *bo = b->object;
if (unlikely(!ao || !bo)) {
MARS_ERR("bad object pointers\n");
return 0;
}
if (ao->ref_pos < bo->ref_pos)
return -1;
if (ao->ref_pos > bo->ref_pos)
return 1;
return 0;
}
_PAIRING_HEAP_FUNCTIONS(static,sync,sync_cmp);
static
void aio_clever_move(struct list_head *tmp_list, int prio, struct q_sync *q_sync)
{
while (!list_empty(tmp_list)) {
struct list_head *tmp = tmp_list->next;
struct aio_mref_aspect *mref_a = container_of(tmp, struct aio_mref_aspect, io_head);
list_del_init(tmp);
ph_insert_sync(&q_sync->heap[prio], &mref_a->heap_head);
}
}
2011-07-05 07:07:06 +00:00
2011-07-01 14:07:56 +00:00
static
void aio_clever_sync(struct aio_output *output, struct q_sync *q_sync)
{
2011-07-05 07:07:06 +00:00
int i;
int max = 64;
for (i = 0; i < MARS_PRIO_NR; i++) {
struct pairing_heap_sync **heap = &q_sync->heap[i];
if (*heap) {
return;
}
}
2011-07-01 14:07:56 +00:00
}
#endif
2010-11-26 13:45:10 +00:00
/* Workaround for non-implemented aio_fsync()
*/
2011-07-01 14:07:56 +00:00
static
int aio_sync_thread(void *data)
2010-11-26 13:45:10 +00:00
{
struct aio_threadinfo *tinfo = data;
2011-03-07 10:27:38 +00:00
struct aio_output *output = tinfo->output;
2011-07-01 14:07:56 +00:00
#ifdef USE_CLEVER_SYNC
struct q_sync q_sync = {};
#endif
2010-11-26 13:45:10 +00:00
MARS_INF("sync thread has started on '%s'.\n", output->brick->brick_path);
2010-11-26 13:45:10 +00:00
//set_user_nice(current, -20);
while (!brick_thread_should_stop() || tinfo->queued_sum > 0) {
2010-11-26 13:45:10 +00:00
LIST_HEAD(tmp_list);
unsigned long flags;
2011-03-10 11:40:06 +00:00
int i;
2010-11-26 13:45:10 +00:00
2011-04-29 09:36:10 +00:00
output->fdsync_active = false;
2011-07-01 14:07:56 +00:00
wake_up_interruptible_all(&output->fdsync_event);
2011-04-29 09:36:10 +00:00
2010-11-26 13:45:10 +00:00
wait_event_interruptible_timeout(
tinfo->event,
tinfo->queued_sum > 0,
HZ / 4);
2010-11-26 13:45:10 +00:00
traced_lock(&tinfo->lock, flags);
2011-05-19 11:36:00 +00:00
for (i = 0; i < MARS_PRIO_NR; i++) {
struct list_head *start = &tinfo->mref_list[i];
if (!list_empty(start)) {
2011-03-10 11:40:06 +00:00
// move over the whole list
2011-05-19 11:36:00 +00:00
list_replace_init(start, &tmp_list);
tinfo->queued_sum -= tinfo->queued[i];
tinfo->queued[i] = 0;
2011-03-10 11:40:06 +00:00
break;
}
2010-11-26 13:45:10 +00:00
}
traced_unlock(&tinfo->lock, flags);
2011-07-01 14:07:56 +00:00
if (!list_empty(&tmp_list)) {
#ifdef USE_CLEVER_SYNC
aio_clever_move(&tmp_list, i, &q_sync);
2011-05-13 11:19:28 +00:00
#else
2011-07-01 14:07:56 +00:00
aio_sync_all(output, &tmp_list);
2011-05-13 11:19:28 +00:00
#endif
2010-11-26 13:45:10 +00:00
}
2011-07-01 14:07:56 +00:00
#ifdef USE_CLEVER_SYNC
aio_clever_sync(output, &q_sync);
#endif
2010-11-26 13:45:10 +00:00
}
MARS_INF("sync thread has stopped.\n");
2011-02-23 20:48:06 +00:00
tinfo->terminated = true;
wake_up_interruptible_all(&tinfo->terminate_event);
2010-11-26 13:45:10 +00:00
return 0;
}
2011-08-26 10:31:56 +00:00
static int aio_event_thread(void *data)
{
struct aio_threadinfo *tinfo = data;
struct aio_output *output = tinfo->output;
struct aio_threadinfo *other = &output->tinfo[2];
int err = -ENOMEM;
MARS_INF("event thread has started.\n");
2011-08-26 10:31:56 +00:00
//set_user_nice(current, -20);
use_fake_mm();
if (!current->mm)
goto err;
err = aio_start_thread(output, &output->tinfo[2], aio_sync_thread, 'y');
2011-08-26 10:31:56 +00:00
if (unlikely(err < 0))
goto err;
while (!brick_thread_should_stop() || tinfo->queued_sum > 0) {
2011-08-26 10:31:56 +00:00
mm_segment_t oldfs;
int count;
int i;
struct timespec timeout = {
.tv_sec = 1,
2011-08-26 10:31:56 +00:00
};
struct io_event events[MARS_MAX_AIO_READ];
oldfs = get_fs();
set_fs(get_ds());
/* TODO: don't timeout upon termination.
* Probably we should submit a dummy request.
*/
count = sys_io_getevents(output->ctxp, 1, MARS_MAX_AIO_READ, events, &timeout);
set_fs(oldfs);
//MARS_INF("count = %d\n", count);
for (i = 0; i < count; i++) {
struct aio_mref_aspect *mref_a = (void*)events[i].data;
struct mref_object *mref;
int err = events[i].res;
if (!mref_a) {
continue; // this was a dummy request
}
mref = mref_a->object;
MARS_IO("AIO done %p pos = %lld len = %d rw = %d\n", mref, mref->ref_pos, mref->ref_len, mref->ref_rw);
if (output->brick->o_fdsync
&& err >= 0
&& mref->ref_rw != READ
&& !mref->ref_skip_sync
&& !mref_a->resubmit++) {
// workaround for non-implemented AIO FSYNC operation
if (!output->filp->f_op->aio_fsync) {
mars_trace(mref, "aio_fsync");
_enqueue(other, mref_a, mref->ref_prio, true);
continue;
}
err = aio_submit(output, mref_a, true);
if (likely(err >= 0))
continue;
}
_complete(output, mref, err);
}
}
err = 0;
err:
MARS_INF("event thread has stopped, err = %d\n", err);
2011-08-26 10:31:56 +00:00
aio_stop_thread(output, 2, false);
unuse_fake_mm();
tinfo->terminated = true;
wake_up_interruptible_all(&tinfo->terminate_event);
2011-08-26 10:31:56 +00:00
return err;
}
2011-08-31 11:42:04 +00:00
#if 1
/* This should go to fs/open.c (as long as vfs_submit() is not implemented)
*/
#include <linux/fdtable.h>
void fd_uninstall(unsigned int fd)
{
struct files_struct *files = current->files;
struct fdtable *fdt;
2011-09-27 08:46:14 +00:00
MARS_INF("fd = %d\n", fd);
2011-08-31 11:42:04 +00:00
spin_lock(&files->file_lock);
fdt = files_fdtable(files);
rcu_assign_pointer(fdt->fd[fd], NULL);
spin_unlock(&files->file_lock);
}
EXPORT_SYMBOL(fd_uninstall);
#endif
2012-02-20 09:34:42 +00:00
static
void _mapfree_pages(struct aio_output *output, bool force)
{
struct aio_brick *brick = output->brick;
struct address_space *mapping;
pgoff_t start;
pgoff_t end;
pgoff_t gap;
if (brick->linear_cache_size <= 0)
goto done;
if (!force && ++output->rounds < brick->linear_cache_rounds)
goto done;
if (unlikely(!output->filp || !(mapping = output->filp->f_mapping)))
goto done;
if (unlikely(brick->linear_cache_rounds <= 0))
brick->linear_cache_rounds = 1024;
if (force) {
start = 0;
end = -1;
} else {
gap = brick->linear_cache_size * (1024 * 1024 / PAGE_SIZE);
end = output->min_pos / PAGE_SIZE - gap;
2012-02-20 11:20:55 +00:00
output->min_pos = 0;
2012-02-20 09:34:42 +00:00
if (end <= 0)
goto done;
start = end - gap * 4;
if (start < 0)
start = 0;
}
output->rounds = 0;
atomic_inc(&output->total_mapfree_count);
invalidate_mapping_pages(mapping, start, end);
done:;
}
2011-08-26 10:31:56 +00:00
static int aio_submit_thread(void *data)
{
struct aio_threadinfo *tinfo = data;
struct aio_output *output = tinfo->output;
struct file *file = output->filp;
mm_segment_t oldfs;
int err = -EINVAL;
CHECK_PTR_NULL(file, done);
2011-08-31 11:42:04 +00:00
/* TODO: this is provisionary. We only need it for sys_io_submit()
* which uses userspace concepts like file handles.
* This should be accompanied by a future kernelsapce vfs_submit() or
2011-08-26 10:31:56 +00:00
* do_submit() which currently does not exist :(
* FIXME: corresponding cleanup NYI
*/
err = get_unused_fd();
MARS_INF("fd = %d\n", err);
if (unlikely(err < 0))
goto done;
output->fd = err;
fd_install(err, file);
2011-08-26 10:31:56 +00:00
MARS_INF("submit thread has started.\n");
2011-08-26 10:31:56 +00:00
//set_user_nice(current, -20);
use_fake_mm();
err = -ENOMEM;
if (unlikely(!current->mm))
goto cleanup_fd;
2011-08-26 10:31:56 +00:00
oldfs = get_fs();
set_fs(get_ds());
err = sys_io_setup(MARS_MAX_AIO, &output->ctxp);
set_fs(oldfs);
if (unlikely(err < 0))
goto cleanup_mm;
err = aio_start_thread(output, &output->tinfo[1], aio_event_thread, 'e');
2011-08-26 10:31:56 +00:00
if (unlikely(err < 0))
goto cleanup_ctxp;
2011-08-26 10:31:56 +00:00
while (!brick_thread_should_stop() || atomic_read(&output->read_count) + atomic_read(&output->write_count) + tinfo->queued_sum > 0) {
2011-08-26 10:31:56 +00:00
struct aio_mref_aspect *mref_a;
struct mref_object *mref;
int sleeptime;
int err;
2012-02-20 09:34:42 +00:00
_mapfree_pages(output, false);
2011-08-26 10:31:56 +00:00
wait_event_interruptible_timeout(
tinfo->event,
tinfo->queued_sum > 0,
HZ / 4);
2011-08-26 10:31:56 +00:00
mref_a = _dequeue(tinfo);
2011-08-26 10:31:56 +00:00
if (!mref_a) {
continue;
}
mref = mref_a->object;
2012-02-20 09:34:42 +00:00
err = -EINVAL;
CHECK_PTR(mref, err);
if (!output->min_pos || mref->ref_pos < output->min_pos)
output->min_pos = mref->ref_pos;
// check for reads exactly at EOF (special case)
2011-08-26 10:31:56 +00:00
if (mref->ref_pos == mref->ref_total_size &&
!mref->ref_rw &&
mref->ref_timeout > 0) {
loff_t total_size = i_size_read(file->f_mapping->host);
loff_t len = total_size - mref->ref_pos;
if (len > 0) {
mref->ref_total_size = total_size;
mref->ref_len = len;
} else {
if (!mref_a->start_jiffies) {
mref_a->start_jiffies = jiffies;
}
if ((long long)jiffies - mref_a->start_jiffies <= mref->ref_timeout) {
if (!tinfo->queued_sum) {
2011-08-26 10:31:56 +00:00
atomic_inc(&output->total_msleep_count);
brick_msleep(1000 * 4 / HZ);
2011-08-26 10:31:56 +00:00
}
_enqueue(tinfo, mref_a, MARS_PRIO_LOW, true);
continue;
}
MARS_DBG("ENODATA %lld\n", len);
_complete(output, mref, -ENODATA);
continue;
}
}
sleeptime = 1;
2011-08-26 10:31:56 +00:00
for (;;) {
/* This is just a test. Don't enable it for performance reasons.
2011-08-26 10:31:56 +00:00
*/
if (output->brick->wait_during_fdsync && mref->ref_rw != READ) {
if (output->fdsync_active) {
long long delay = 60 * HZ;
atomic_inc(&output->total_fdsync_wait_count);
__wait_event_interruptible_timeout(
output->fdsync_event,
!output->fdsync_active,
2011-08-26 10:31:56 +00:00
delay);
}
}
/* Now really do the work
*/
err = aio_submit(output, mref_a, false);
if (likely(err != -EAGAIN)) {
break;
}
atomic_inc(&output->total_delay_count);
brick_msleep(sleeptime);
2011-08-26 10:31:56 +00:00
if (sleeptime < 100) {
sleeptime++;
2011-08-26 10:31:56 +00:00
}
}
2012-02-20 09:34:42 +00:00
err:
2011-08-26 10:31:56 +00:00
if (unlikely(err < 0)) {
_complete(output, mref, err);
}
}
MARS_INF("submit thread has stopped.\n");
2011-08-26 10:31:56 +00:00
aio_stop_thread(output, 1, true);
cleanup_ctxp:
MARS_DBG("destroying ioctx.....\n");
oldfs = get_fs();
set_fs(get_ds());
sys_io_destroy(output->ctxp);
set_fs(oldfs);
output->ctxp = 0;
cleanup_mm:
unuse_fake_mm();
cleanup_fd:
MARS_DBG("destroying fd %d\n", output->fd);
2011-08-31 11:42:04 +00:00
fd_uninstall(output->fd);
2011-08-26 10:31:56 +00:00
put_unused_fd(output->fd);
2011-08-26 10:31:56 +00:00
err = 0;
done:
MARS_DBG("status = %d\n", err);
2011-08-26 10:31:56 +00:00
tinfo->terminated = true;
wake_up_interruptible_all(&tinfo->terminate_event);
2011-08-26 10:31:56 +00:00
return err;
}
2011-03-07 10:27:38 +00:00
static int aio_get_info(struct aio_output *output, struct mars_info *info)
2010-11-26 13:45:10 +00:00
{
struct file *file = output->filp;
2011-02-23 20:48:06 +00:00
if (unlikely(!file || !file->f_mapping || !file->f_mapping->host))
return -EINVAL;
2010-11-26 13:45:10 +00:00
info->current_size = i_size_read(file->f_mapping->host);
2011-02-23 20:48:06 +00:00
MARS_DBG("determined file size = %lld\n", info->current_size);
2010-11-26 13:45:10 +00:00
info->backing_file = file;
return 0;
}
2011-04-12 15:31:08 +00:00
//////////////// informational / statistics ///////////////
static noinline
char *aio_statistics(struct aio_brick *brick, int verbose)
{
struct aio_output *output = brick->outputs[0];
2012-10-12 12:17:34 +00:00
char *res = brick_string_alloc(4096);
2011-05-13 11:19:28 +00:00
char *sync = NULL;
2012-10-12 12:17:34 +00:00
int pos = 0;
2011-04-12 15:31:08 +00:00
if (!res)
return NULL;
2012-10-12 12:17:34 +00:00
pos += report_timing(&timings[0], res + pos, 4096 - pos);
pos += report_timing(&timings[1], res + pos, 4096 - pos);
pos += report_timing(&timings[2], res + pos, 4096 - pos);
2011-04-12 15:31:08 +00:00
2012-10-12 12:17:34 +00:00
snprintf(res + pos, 4096 - pos,
"total "
"reads = %d "
2012-02-12 11:19:57 +00:00
"writes = %d "
"allocs = %d "
"delays = %d "
"msleeps = %d "
"fdsyncs = %d "
2012-02-20 09:34:42 +00:00
"fdsync_waits = %d "
"map_free = %d | "
2012-02-20 10:06:50 +00:00
"linear_cache_size = %d "
"linear_cache_rounds = %d "
"min_pos = %lld | "
2012-02-12 11:19:57 +00:00
"flying reads = %d "
"writes = %d "
"allocs = %d "
"q0 = %d "
"q1 = %d "
"q2 = %d "
"| total "
"q0 = %d "
"q1 = %d "
"q2 = %d "
"%s\n",
2012-02-12 11:19:57 +00:00
atomic_read(&output->total_read_count),
atomic_read(&output->total_write_count),
atomic_read(&output->total_alloc_count),
atomic_read(&output->total_delay_count),
atomic_read(&output->total_msleep_count),
atomic_read(&output->total_fdsync_count),
atomic_read(&output->total_fdsync_wait_count),
2012-02-20 09:34:42 +00:00
atomic_read(&output->total_mapfree_count),
2012-02-20 10:06:50 +00:00
brick->linear_cache_size,
brick->linear_cache_rounds,
output->min_pos,
2012-02-12 11:19:57 +00:00
atomic_read(&output->read_count),
atomic_read(&output->write_count),
atomic_read(&output->alloc_count),
output->tinfo[0].queued_sum,
output->tinfo[1].queued_sum,
output->tinfo[2].queued_sum,
atomic_read(&output->tinfo[0].total_enqueue_count),
atomic_read(&output->tinfo[1].total_enqueue_count),
atomic_read(&output->tinfo[2].total_enqueue_count),
2012-02-12 11:19:57 +00:00
sync ? sync : "");
2011-05-13 11:19:28 +00:00
if (sync)
2011-08-12 11:09:48 +00:00
brick_string_free(sync);
2011-04-12 15:31:08 +00:00
return res;
}
static noinline
void aio_reset_statistics(struct aio_brick *brick)
{
struct aio_output *output = brick->outputs[0];
2011-05-06 10:25:52 +00:00
int i;
2011-04-12 15:31:08 +00:00
atomic_set(&output->total_read_count, 0);
atomic_set(&output->total_write_count, 0);
atomic_set(&output->total_alloc_count, 0);
2011-04-29 09:36:10 +00:00
atomic_set(&output->total_delay_count, 0);
2011-05-06 10:25:52 +00:00
atomic_set(&output->total_msleep_count, 0);
atomic_set(&output->total_fdsync_count, 0);
2011-04-29 09:36:10 +00:00
atomic_set(&output->total_fdsync_wait_count, 0);
2012-02-20 09:34:42 +00:00
atomic_set(&output->total_mapfree_count, 0);
2011-05-06 10:25:52 +00:00
for (i = 0; i < 3; i++) {
struct aio_threadinfo *tinfo = &output->tinfo[i];
atomic_set(&tinfo->total_enqueue_count, 0);
}
2011-04-12 15:31:08 +00:00
}
2010-11-26 13:45:10 +00:00
//////////////// object / aspect constructors / destructors ///////////////
static int aio_mref_aspect_init_fn(struct generic_aspect *_ini)
2010-11-26 13:45:10 +00:00
{
2011-03-07 10:27:38 +00:00
struct aio_mref_aspect *ini = (void*)_ini;
2010-11-26 13:45:10 +00:00
INIT_LIST_HEAD(&ini->io_head);
return 0;
}
static void aio_mref_aspect_exit_fn(struct generic_aspect *_ini)
2010-11-26 13:45:10 +00:00
{
2011-03-07 10:27:38 +00:00
struct aio_mref_aspect *ini = (void*)_ini;
2010-11-26 13:45:10 +00:00
(void)ini;
}
2011-03-07 10:27:38 +00:00
MARS_MAKE_STATICS(aio);
2010-11-26 13:45:10 +00:00
////////////////////// brick constructors / destructors ////////////////////
2011-03-07 10:27:38 +00:00
static int aio_brick_construct(struct aio_brick *brick)
2010-11-26 13:45:10 +00:00
{
return 0;
}
2011-03-07 10:27:38 +00:00
static int aio_switch(struct aio_brick *brick)
2010-11-26 13:45:10 +00:00
{
static int index;
2011-03-07 10:27:38 +00:00
struct aio_output *output = brick->outputs[0];
const char *path = output->brick->brick_path;
2010-11-26 13:45:10 +00:00
int flags = O_CREAT | O_RDWR | O_LARGEFILE;
int prot = 0600;
mm_segment_t oldfs;
int err = 0;
2011-02-23 20:48:06 +00:00
MARS_DBG("power.button = %d\n", brick->power.button);
if (!brick->power.button)
goto cleanup;
if (brick->power.led_on || output->filp)
2011-03-01 18:00:14 +00:00
goto done;
2011-02-23 20:48:06 +00:00
mars_power_led_off((void*)brick, false);
2011-03-11 13:57:54 +00:00
if (brick->o_direct) {
2010-11-26 13:45:10 +00:00
flags |= O_DIRECT;
MARS_INF("using O_DIRECT on %s\n", path);
}
oldfs = get_fs();
set_fs(get_ds());
output->filp = filp_open(path, flags, prot);
set_fs(oldfs);
MARS_DBG("opened file '%s' flags = %d prot = %d filp = %p\n", path, flags, prot, output->filp);
if (unlikely(!output->filp || IS_ERR(output->filp))) {
2010-11-26 13:45:10 +00:00
err = PTR_ERR(output->filp);
MARS_ERR("can't open file '%s' status=%d\n", path, err);
output->filp = NULL;
if (err >= 0)
err = -ENOENT;
2010-11-26 13:45:10 +00:00
return err;
}
2011-03-11 13:57:54 +00:00
#if 1
{
struct inode *inode = output->filp->f_mapping->host;
if (S_ISBLK(inode->i_mode)) {
MARS_INF("changing readahead from %lu to %d\n", inode->i_bdev->bd_disk->queue->backing_dev_info.ra_pages, brick->readahead);
inode->i_bdev->bd_disk->queue->backing_dev_info.ra_pages = brick->readahead;
}
}
#endif
2010-11-26 13:45:10 +00:00
output->index = ++index;
err = aio_start_thread(output, &output->tinfo[0], aio_submit_thread, 's');
2011-08-26 10:31:56 +00:00
if (err < 0)
goto err;
2010-11-26 13:45:10 +00:00
MARS_INF("opened file '%s'\n", path);
2011-02-23 20:48:06 +00:00
mars_power_led_on((void*)brick, true);
MARS_DBG("successfully switched on.\n");
2011-03-01 18:00:14 +00:00
done:
2010-11-26 13:45:10 +00:00
return 0;
err:
MARS_ERR("status = %d\n", err);
cleanup:
2011-03-01 18:00:14 +00:00
if (brick->power.led_off) {
goto done;
}
2011-02-23 20:48:06 +00:00
mars_power_led_on((void*)brick, false);
2011-08-26 10:31:56 +00:00
aio_stop_thread(output, 0, false);
2011-08-25 10:16:32 +00:00
2011-02-23 20:48:06 +00:00
mars_power_led_off((void*)brick,
(output->tinfo[0].thread == NULL &&
output->tinfo[1].thread == NULL &&
output->tinfo[2].thread == NULL));
2011-08-25 10:16:32 +00:00
2011-02-23 20:48:06 +00:00
if (brick->power.led_off) {
if (output->filp) {
2012-02-20 09:34:42 +00:00
_mapfree_pages(output, true);
2011-02-23 20:48:06 +00:00
filp_close(output->filp, NULL);
output->filp = NULL;
}
2010-11-26 13:45:10 +00:00
}
2011-02-23 20:48:06 +00:00
MARS_DBG("switch off status = %d\n", err);
2010-11-26 13:45:10 +00:00
return err;
}
2011-03-07 10:27:38 +00:00
static int aio_output_construct(struct aio_output *output)
2010-11-26 13:45:10 +00:00
{
2011-04-29 09:36:10 +00:00
init_waitqueue_head(&output->fdsync_event);
2010-11-26 13:45:10 +00:00
return 0;
}
2011-03-07 10:27:38 +00:00
static int aio_output_destruct(struct aio_output *output)
2010-11-26 13:45:10 +00:00
{
2011-03-23 17:58:02 +00:00
return 0;
2010-11-26 13:45:10 +00:00
}
///////////////////////// static structs ////////////////////////
2011-03-07 10:27:38 +00:00
static struct aio_brick_ops aio_brick_ops = {
.brick_switch = aio_switch,
2011-04-12 15:31:08 +00:00
.brick_statistics = aio_statistics,
.reset_statistics = aio_reset_statistics,
2010-11-26 13:45:10 +00:00
};
2011-03-07 10:27:38 +00:00
static struct aio_output_ops aio_output_ops = {
.mref_get = aio_ref_get,
.mref_put = aio_ref_put,
.mref_io = aio_ref_io,
.mars_get_info = aio_get_info,
2010-11-26 13:45:10 +00:00
};
2011-03-23 17:58:02 +00:00
const struct aio_input_type aio_input_type = {
.type_name = "aio_input",
.input_size = sizeof(struct aio_input),
};
static const struct aio_input_type *aio_input_types[] = {
&aio_input_type,
};
2011-03-07 10:27:38 +00:00
const struct aio_output_type aio_output_type = {
.type_name = "aio_output",
.output_size = sizeof(struct aio_output),
.master_ops = &aio_output_ops,
.output_construct = &aio_output_construct,
.output_destruct = &aio_output_destruct,
2010-11-26 13:45:10 +00:00
};
2011-03-07 10:27:38 +00:00
static const struct aio_output_type *aio_output_types[] = {
&aio_output_type,
2010-11-26 13:45:10 +00:00
};
2011-03-07 10:27:38 +00:00
const struct aio_brick_type aio_brick_type = {
.type_name = "aio_brick",
.brick_size = sizeof(struct aio_brick),
2010-11-26 13:45:10 +00:00
.max_inputs = 0,
.max_outputs = 1,
2011-03-07 10:27:38 +00:00
.master_ops = &aio_brick_ops,
.aspect_types = aio_aspect_types,
2011-03-23 17:58:02 +00:00
.default_input_types = aio_input_types,
2011-03-07 10:27:38 +00:00
.default_output_types = aio_output_types,
.brick_construct = &aio_brick_construct,
2010-11-26 13:45:10 +00:00
};
2011-03-07 10:27:38 +00:00
EXPORT_SYMBOL_GPL(aio_brick_type);
2010-11-26 13:45:10 +00:00
////////////////// module init stuff /////////////////////////
2011-08-25 10:16:32 +00:00
int __init init_mars_aio(void)
2010-11-26 13:45:10 +00:00
{
2011-03-07 10:27:38 +00:00
MARS_INF("init_aio()\n");
_aio_brick_type = (void*)&aio_brick_type;
return aio_register_brick_type();
2010-11-26 13:45:10 +00:00
}
2011-08-25 10:16:32 +00:00
void __exit exit_mars_aio(void)
2010-11-26 13:45:10 +00:00
{
2011-03-07 10:27:38 +00:00
MARS_INF("exit_aio()\n");
aio_unregister_brick_type();
2010-11-26 13:45:10 +00:00
}
2011-08-25 10:16:32 +00:00
#ifndef CONFIG_MARS_HAVE_BIGMODULE
2011-03-07 10:27:38 +00:00
MODULE_DESCRIPTION("MARS aio brick");
2010-11-26 13:45:10 +00:00
MODULE_AUTHOR("Thomas Schoebel-Theuer <tst@1und1.de>");
MODULE_LICENSE("GPL");
2011-08-25 10:16:32 +00:00
module_init(init_mars_aio);
module_exit(exit_mars_aio);
#endif