2008-06-12 00:50:36 +04:00
|
|
|
/*
|
|
|
|
* Copyright (C) 2007 Oracle. All rights reserved.
|
|
|
|
*
|
|
|
|
* This program is free software; you can redistribute it and/or
|
|
|
|
* modify it under the terms of the GNU General Public
|
|
|
|
* License v2 as published by the Free Software Foundation.
|
|
|
|
*
|
|
|
|
* This program is distributed in the hope that it will be useful,
|
|
|
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
|
* General Public License for more details.
|
|
|
|
*
|
|
|
|
* You should have received a copy of the GNU General Public
|
|
|
|
* License along with this program; if not, write to the
|
|
|
|
* Free Software Foundation, Inc., 59 Temple Place - Suite 330,
|
|
|
|
* Boston, MA 021110-1307, USA.
|
|
|
|
*/
|
|
|
|
|
2008-06-12 01:09:48 +04:00
|
|
|
#include <linux/version.h>
|
2008-06-12 00:50:36 +04:00
|
|
|
#include <linux/kthread.h>
|
|
|
|
#include <linux/list.h>
|
|
|
|
#include <linux/spinlock.h>
|
2008-06-12 01:09:48 +04:00
|
|
|
# include <linux/freezer.h>
|
2008-06-12 00:50:36 +04:00
|
|
|
#include "async-thread.h"
|
|
|
|
|
Btrfs: Add ordered async work queues
Btrfs uses kernel threads to create async work queues for cpu intensive
operations such as checksumming and decompression. These work well,
but they make it difficult to keep IO order intact.
A single writepages call from pdflush or fsync will turn into a number
of bios, and each bio is checksummed in parallel. Once the checksum is
computed, the bio is sent down to the disk, and since we don't control
the order in which the parallel operations happen, they might go down to
the disk in almost any order.
The code deals with this somewhat by having deep work queues for a single
kernel thread, making it very likely that a single thread will process all
the bios for a single inode.
This patch introduces an explicitly ordered work queue. As work structs
are placed into the queue they are put onto the tail of a list. They have
three callbacks:
->func (cpu intensive processing here)
->ordered_func (order sensitive processing here)
->ordered_free (free the work struct, all processing is done)
The work struct has three callbacks. The func callback does the cpu intensive
work, and when it completes the work struct is marked as done.
Every time a work struct completes, the list is checked to see if the head
is marked as done. If so the ordered_func callback is used to do the
order sensitive processing and the ordered_free callback is used to do
any cleanup. Then we loop back and check the head of the list again.
This patch also changes the checksumming code to use the ordered workqueues.
One a 4 drive array, it increases streaming writes from 280MB/s to 350MB/s.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-11-07 06:03:00 +03:00
|
|
|
#define WORK_QUEUED_BIT 0
|
|
|
|
#define WORK_DONE_BIT 1
|
|
|
|
#define WORK_ORDER_DONE_BIT 2
|
|
|
|
|
2008-06-12 00:50:36 +04:00
|
|
|
/*
|
|
|
|
* container for the kthread task pointer and the list of pending work
|
|
|
|
* One of these is allocated per thread.
|
|
|
|
*/
|
|
|
|
struct btrfs_worker_thread {
|
2008-06-12 04:21:24 +04:00
|
|
|
/* pool we belong to */
|
|
|
|
struct btrfs_workers *workers;
|
|
|
|
|
2008-06-12 00:50:36 +04:00
|
|
|
/* list of struct btrfs_work that are waiting for service */
|
|
|
|
struct list_head pending;
|
|
|
|
|
|
|
|
/* list of worker threads from struct btrfs_workers */
|
|
|
|
struct list_head worker_list;
|
|
|
|
|
|
|
|
/* kthread */
|
|
|
|
struct task_struct *task;
|
|
|
|
|
|
|
|
/* number of things on the pending list */
|
|
|
|
atomic_t num_pending;
|
2008-08-15 23:34:18 +04:00
|
|
|
|
2008-08-15 23:34:17 +04:00
|
|
|
unsigned long sequence;
|
2008-06-12 00:50:36 +04:00
|
|
|
|
|
|
|
/* protects the pending list. */
|
|
|
|
spinlock_t lock;
|
|
|
|
|
|
|
|
/* set to non-zero when this thread is already awake and kicking */
|
|
|
|
int working;
|
2008-06-12 04:21:24 +04:00
|
|
|
|
|
|
|
/* are we currently idle */
|
|
|
|
int idle;
|
2008-06-12 00:50:36 +04:00
|
|
|
};
|
|
|
|
|
2008-06-12 04:21:24 +04:00
|
|
|
/*
|
|
|
|
* helper function to move a thread onto the idle list after it
|
|
|
|
* has finished some requests.
|
|
|
|
*/
|
|
|
|
static void check_idle_worker(struct btrfs_worker_thread *worker)
|
|
|
|
{
|
|
|
|
if (!worker->idle && atomic_read(&worker->num_pending) <
|
|
|
|
worker->workers->idle_thresh / 2) {
|
|
|
|
unsigned long flags;
|
|
|
|
spin_lock_irqsave(&worker->workers->lock, flags);
|
|
|
|
worker->idle = 1;
|
|
|
|
list_move(&worker->worker_list, &worker->workers->idle_list);
|
|
|
|
spin_unlock_irqrestore(&worker->workers->lock, flags);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* helper function to move a thread off the idle list after new
|
|
|
|
* pending work is added.
|
|
|
|
*/
|
|
|
|
static void check_busy_worker(struct btrfs_worker_thread *worker)
|
|
|
|
{
|
|
|
|
if (worker->idle && atomic_read(&worker->num_pending) >=
|
|
|
|
worker->workers->idle_thresh) {
|
|
|
|
unsigned long flags;
|
|
|
|
spin_lock_irqsave(&worker->workers->lock, flags);
|
|
|
|
worker->idle = 0;
|
|
|
|
list_move_tail(&worker->worker_list,
|
|
|
|
&worker->workers->worker_list);
|
|
|
|
spin_unlock_irqrestore(&worker->workers->lock, flags);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Btrfs: Add ordered async work queues
Btrfs uses kernel threads to create async work queues for cpu intensive
operations such as checksumming and decompression. These work well,
but they make it difficult to keep IO order intact.
A single writepages call from pdflush or fsync will turn into a number
of bios, and each bio is checksummed in parallel. Once the checksum is
computed, the bio is sent down to the disk, and since we don't control
the order in which the parallel operations happen, they might go down to
the disk in almost any order.
The code deals with this somewhat by having deep work queues for a single
kernel thread, making it very likely that a single thread will process all
the bios for a single inode.
This patch introduces an explicitly ordered work queue. As work structs
are placed into the queue they are put onto the tail of a list. They have
three callbacks:
->func (cpu intensive processing here)
->ordered_func (order sensitive processing here)
->ordered_free (free the work struct, all processing is done)
The work struct has three callbacks. The func callback does the cpu intensive
work, and when it completes the work struct is marked as done.
Every time a work struct completes, the list is checked to see if the head
is marked as done. If so the ordered_func callback is used to do the
order sensitive processing and the ordered_free callback is used to do
any cleanup. Then we loop back and check the head of the list again.
This patch also changes the checksumming code to use the ordered workqueues.
One a 4 drive array, it increases streaming writes from 280MB/s to 350MB/s.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-11-07 06:03:00 +03:00
|
|
|
static noinline int run_ordered_completions(struct btrfs_workers *workers,
|
|
|
|
struct btrfs_work *work)
|
|
|
|
{
|
|
|
|
unsigned long flags;
|
|
|
|
|
|
|
|
if (!workers->ordered)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
set_bit(WORK_DONE_BIT, &work->flags);
|
|
|
|
|
|
|
|
spin_lock_irqsave(&workers->lock, flags);
|
|
|
|
|
|
|
|
while(!list_empty(&workers->order_list)) {
|
|
|
|
work = list_entry(workers->order_list.next,
|
|
|
|
struct btrfs_work, order_list);
|
|
|
|
|
|
|
|
if (!test_bit(WORK_DONE_BIT, &work->flags))
|
|
|
|
break;
|
|
|
|
|
|
|
|
/* we are going to call the ordered done function, but
|
|
|
|
* we leave the work item on the list as a barrier so
|
|
|
|
* that later work items that are done don't have their
|
|
|
|
* functions called before this one returns
|
|
|
|
*/
|
|
|
|
if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
|
|
|
|
break;
|
|
|
|
|
|
|
|
spin_unlock_irqrestore(&workers->lock, flags);
|
|
|
|
|
|
|
|
work->ordered_func(work);
|
|
|
|
|
|
|
|
/* now take the lock again and call the freeing code */
|
|
|
|
spin_lock_irqsave(&workers->lock, flags);
|
|
|
|
list_del(&work->order_list);
|
|
|
|
work->ordered_free(work);
|
|
|
|
}
|
|
|
|
|
|
|
|
spin_unlock_irqrestore(&workers->lock, flags);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2008-06-12 00:50:36 +04:00
|
|
|
/*
|
|
|
|
* main loop for servicing work items
|
|
|
|
*/
|
|
|
|
static int worker_loop(void *arg)
|
|
|
|
{
|
|
|
|
struct btrfs_worker_thread *worker = arg;
|
|
|
|
struct list_head *cur;
|
|
|
|
struct btrfs_work *work;
|
|
|
|
do {
|
|
|
|
spin_lock_irq(&worker->lock);
|
|
|
|
while(!list_empty(&worker->pending)) {
|
|
|
|
cur = worker->pending.next;
|
|
|
|
work = list_entry(cur, struct btrfs_work, list);
|
|
|
|
list_del(&work->list);
|
Btrfs: Add ordered async work queues
Btrfs uses kernel threads to create async work queues for cpu intensive
operations such as checksumming and decompression. These work well,
but they make it difficult to keep IO order intact.
A single writepages call from pdflush or fsync will turn into a number
of bios, and each bio is checksummed in parallel. Once the checksum is
computed, the bio is sent down to the disk, and since we don't control
the order in which the parallel operations happen, they might go down to
the disk in almost any order.
The code deals with this somewhat by having deep work queues for a single
kernel thread, making it very likely that a single thread will process all
the bios for a single inode.
This patch introduces an explicitly ordered work queue. As work structs
are placed into the queue they are put onto the tail of a list. They have
three callbacks:
->func (cpu intensive processing here)
->ordered_func (order sensitive processing here)
->ordered_free (free the work struct, all processing is done)
The work struct has three callbacks. The func callback does the cpu intensive
work, and when it completes the work struct is marked as done.
Every time a work struct completes, the list is checked to see if the head
is marked as done. If so the ordered_func callback is used to do the
order sensitive processing and the ordered_free callback is used to do
any cleanup. Then we loop back and check the head of the list again.
This patch also changes the checksumming code to use the ordered workqueues.
One a 4 drive array, it increases streaming writes from 280MB/s to 350MB/s.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-11-07 06:03:00 +03:00
|
|
|
clear_bit(WORK_QUEUED_BIT, &work->flags);
|
2008-06-12 00:50:36 +04:00
|
|
|
|
|
|
|
work->worker = worker;
|
|
|
|
spin_unlock_irq(&worker->lock);
|
|
|
|
|
|
|
|
work->func(work);
|
|
|
|
|
|
|
|
atomic_dec(&worker->num_pending);
|
Btrfs: Add ordered async work queues
Btrfs uses kernel threads to create async work queues for cpu intensive
operations such as checksumming and decompression. These work well,
but they make it difficult to keep IO order intact.
A single writepages call from pdflush or fsync will turn into a number
of bios, and each bio is checksummed in parallel. Once the checksum is
computed, the bio is sent down to the disk, and since we don't control
the order in which the parallel operations happen, they might go down to
the disk in almost any order.
The code deals with this somewhat by having deep work queues for a single
kernel thread, making it very likely that a single thread will process all
the bios for a single inode.
This patch introduces an explicitly ordered work queue. As work structs
are placed into the queue they are put onto the tail of a list. They have
three callbacks:
->func (cpu intensive processing here)
->ordered_func (order sensitive processing here)
->ordered_free (free the work struct, all processing is done)
The work struct has three callbacks. The func callback does the cpu intensive
work, and when it completes the work struct is marked as done.
Every time a work struct completes, the list is checked to see if the head
is marked as done. If so the ordered_func callback is used to do the
order sensitive processing and the ordered_free callback is used to do
any cleanup. Then we loop back and check the head of the list again.
This patch also changes the checksumming code to use the ordered workqueues.
One a 4 drive array, it increases streaming writes from 280MB/s to 350MB/s.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-11-07 06:03:00 +03:00
|
|
|
/*
|
|
|
|
* unless this is an ordered work queue,
|
|
|
|
* 'work' was probably freed by func above.
|
|
|
|
*/
|
|
|
|
run_ordered_completions(worker->workers, work);
|
|
|
|
|
2008-06-12 00:50:36 +04:00
|
|
|
spin_lock_irq(&worker->lock);
|
2008-06-12 04:21:24 +04:00
|
|
|
check_idle_worker(worker);
|
Btrfs: Add ordered async work queues
Btrfs uses kernel threads to create async work queues for cpu intensive
operations such as checksumming and decompression. These work well,
but they make it difficult to keep IO order intact.
A single writepages call from pdflush or fsync will turn into a number
of bios, and each bio is checksummed in parallel. Once the checksum is
computed, the bio is sent down to the disk, and since we don't control
the order in which the parallel operations happen, they might go down to
the disk in almost any order.
The code deals with this somewhat by having deep work queues for a single
kernel thread, making it very likely that a single thread will process all
the bios for a single inode.
This patch introduces an explicitly ordered work queue. As work structs
are placed into the queue they are put onto the tail of a list. They have
three callbacks:
->func (cpu intensive processing here)
->ordered_func (order sensitive processing here)
->ordered_free (free the work struct, all processing is done)
The work struct has three callbacks. The func callback does the cpu intensive
work, and when it completes the work struct is marked as done.
Every time a work struct completes, the list is checked to see if the head
is marked as done. If so the ordered_func callback is used to do the
order sensitive processing and the ordered_free callback is used to do
any cleanup. Then we loop back and check the head of the list again.
This patch also changes the checksumming code to use the ordered workqueues.
One a 4 drive array, it increases streaming writes from 280MB/s to 350MB/s.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-11-07 06:03:00 +03:00
|
|
|
|
2008-06-12 00:50:36 +04:00
|
|
|
}
|
|
|
|
worker->working = 0;
|
|
|
|
if (freezing(current)) {
|
|
|
|
refrigerator();
|
|
|
|
} else {
|
|
|
|
set_current_state(TASK_INTERRUPTIBLE);
|
|
|
|
spin_unlock_irq(&worker->lock);
|
2008-11-12 22:36:58 +03:00
|
|
|
if (!kthread_should_stop())
|
|
|
|
schedule();
|
2008-06-12 00:50:36 +04:00
|
|
|
__set_current_state(TASK_RUNNING);
|
|
|
|
}
|
|
|
|
} while (!kthread_should_stop());
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* this will wait for all the worker threads to shutdown
|
|
|
|
*/
|
|
|
|
int btrfs_stop_workers(struct btrfs_workers *workers)
|
|
|
|
{
|
|
|
|
struct list_head *cur;
|
|
|
|
struct btrfs_worker_thread *worker;
|
|
|
|
|
2008-06-12 04:21:24 +04:00
|
|
|
list_splice_init(&workers->idle_list, &workers->worker_list);
|
2008-06-12 00:50:36 +04:00
|
|
|
while(!list_empty(&workers->worker_list)) {
|
|
|
|
cur = workers->worker_list.next;
|
|
|
|
worker = list_entry(cur, struct btrfs_worker_thread,
|
|
|
|
worker_list);
|
|
|
|
kthread_stop(worker->task);
|
|
|
|
list_del(&worker->worker_list);
|
|
|
|
kfree(worker);
|
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* simple init on struct btrfs_workers
|
|
|
|
*/
|
2008-08-15 23:34:16 +04:00
|
|
|
void btrfs_init_workers(struct btrfs_workers *workers, char *name, int max)
|
2008-06-12 00:50:36 +04:00
|
|
|
{
|
|
|
|
workers->num_workers = 0;
|
|
|
|
INIT_LIST_HEAD(&workers->worker_list);
|
2008-06-12 04:21:24 +04:00
|
|
|
INIT_LIST_HEAD(&workers->idle_list);
|
Btrfs: Add ordered async work queues
Btrfs uses kernel threads to create async work queues for cpu intensive
operations such as checksumming and decompression. These work well,
but they make it difficult to keep IO order intact.
A single writepages call from pdflush or fsync will turn into a number
of bios, and each bio is checksummed in parallel. Once the checksum is
computed, the bio is sent down to the disk, and since we don't control
the order in which the parallel operations happen, they might go down to
the disk in almost any order.
The code deals with this somewhat by having deep work queues for a single
kernel thread, making it very likely that a single thread will process all
the bios for a single inode.
This patch introduces an explicitly ordered work queue. As work structs
are placed into the queue they are put onto the tail of a list. They have
three callbacks:
->func (cpu intensive processing here)
->ordered_func (order sensitive processing here)
->ordered_free (free the work struct, all processing is done)
The work struct has three callbacks. The func callback does the cpu intensive
work, and when it completes the work struct is marked as done.
Every time a work struct completes, the list is checked to see if the head
is marked as done. If so the ordered_func callback is used to do the
order sensitive processing and the ordered_free callback is used to do
any cleanup. Then we loop back and check the head of the list again.
This patch also changes the checksumming code to use the ordered workqueues.
One a 4 drive array, it increases streaming writes from 280MB/s to 350MB/s.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-11-07 06:03:00 +03:00
|
|
|
INIT_LIST_HEAD(&workers->order_list);
|
2008-06-12 00:50:36 +04:00
|
|
|
spin_lock_init(&workers->lock);
|
|
|
|
workers->max_workers = max;
|
2008-07-31 23:42:53 +04:00
|
|
|
workers->idle_thresh = 32;
|
2008-08-15 23:34:16 +04:00
|
|
|
workers->name = name;
|
Btrfs: Add ordered async work queues
Btrfs uses kernel threads to create async work queues for cpu intensive
operations such as checksumming and decompression. These work well,
but they make it difficult to keep IO order intact.
A single writepages call from pdflush or fsync will turn into a number
of bios, and each bio is checksummed in parallel. Once the checksum is
computed, the bio is sent down to the disk, and since we don't control
the order in which the parallel operations happen, they might go down to
the disk in almost any order.
The code deals with this somewhat by having deep work queues for a single
kernel thread, making it very likely that a single thread will process all
the bios for a single inode.
This patch introduces an explicitly ordered work queue. As work structs
are placed into the queue they are put onto the tail of a list. They have
three callbacks:
->func (cpu intensive processing here)
->ordered_func (order sensitive processing here)
->ordered_free (free the work struct, all processing is done)
The work struct has three callbacks. The func callback does the cpu intensive
work, and when it completes the work struct is marked as done.
Every time a work struct completes, the list is checked to see if the head
is marked as done. If so the ordered_func callback is used to do the
order sensitive processing and the ordered_free callback is used to do
any cleanup. Then we loop back and check the head of the list again.
This patch also changes the checksumming code to use the ordered workqueues.
One a 4 drive array, it increases streaming writes from 280MB/s to 350MB/s.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-11-07 06:03:00 +03:00
|
|
|
workers->ordered = 0;
|
2008-06-12 00:50:36 +04:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* starts new worker threads. This does not enforce the max worker
|
|
|
|
* count in case you need to temporarily go past it.
|
|
|
|
*/
|
|
|
|
int btrfs_start_workers(struct btrfs_workers *workers, int num_workers)
|
|
|
|
{
|
|
|
|
struct btrfs_worker_thread *worker;
|
|
|
|
int ret = 0;
|
|
|
|
int i;
|
|
|
|
|
|
|
|
for (i = 0; i < num_workers; i++) {
|
|
|
|
worker = kzalloc(sizeof(*worker), GFP_NOFS);
|
|
|
|
if (!worker) {
|
|
|
|
ret = -ENOMEM;
|
|
|
|
goto fail;
|
|
|
|
}
|
|
|
|
|
|
|
|
INIT_LIST_HEAD(&worker->pending);
|
|
|
|
INIT_LIST_HEAD(&worker->worker_list);
|
|
|
|
spin_lock_init(&worker->lock);
|
|
|
|
atomic_set(&worker->num_pending, 0);
|
2008-08-15 23:34:16 +04:00
|
|
|
worker->task = kthread_run(worker_loop, worker,
|
|
|
|
"btrfs-%s-%d", workers->name,
|
|
|
|
workers->num_workers + i);
|
2008-06-12 04:21:24 +04:00
|
|
|
worker->workers = workers;
|
2008-06-12 00:50:36 +04:00
|
|
|
if (IS_ERR(worker->task)) {
|
2008-07-30 17:24:37 +04:00
|
|
|
kfree(worker);
|
2008-06-12 00:50:36 +04:00
|
|
|
ret = PTR_ERR(worker->task);
|
|
|
|
goto fail;
|
|
|
|
}
|
|
|
|
|
|
|
|
spin_lock_irq(&workers->lock);
|
2008-06-12 04:21:24 +04:00
|
|
|
list_add_tail(&worker->worker_list, &workers->idle_list);
|
2008-08-15 23:34:17 +04:00
|
|
|
worker->idle = 1;
|
2008-06-12 00:50:36 +04:00
|
|
|
workers->num_workers++;
|
|
|
|
spin_unlock_irq(&workers->lock);
|
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
fail:
|
|
|
|
btrfs_stop_workers(workers);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* run through the list and find a worker thread that doesn't have a lot
|
|
|
|
* to do right now. This can return null if we aren't yet at the thread
|
|
|
|
* count limit and all of the threads are busy.
|
|
|
|
*/
|
|
|
|
static struct btrfs_worker_thread *next_worker(struct btrfs_workers *workers)
|
|
|
|
{
|
|
|
|
struct btrfs_worker_thread *worker;
|
|
|
|
struct list_head *next;
|
|
|
|
int enforce_min = workers->num_workers < workers->max_workers;
|
|
|
|
|
|
|
|
/*
|
2008-06-12 04:21:24 +04:00
|
|
|
* if we find an idle thread, don't move it to the end of the
|
|
|
|
* idle list. This improves the chance that the next submission
|
|
|
|
* will reuse the same thread, and maybe catch it while it is still
|
|
|
|
* working
|
2008-06-12 00:50:36 +04:00
|
|
|
*/
|
2008-06-12 04:21:24 +04:00
|
|
|
if (!list_empty(&workers->idle_list)) {
|
|
|
|
next = workers->idle_list.next;
|
2008-06-12 00:50:36 +04:00
|
|
|
worker = list_entry(next, struct btrfs_worker_thread,
|
|
|
|
worker_list);
|
2008-06-12 04:21:24 +04:00
|
|
|
return worker;
|
2008-06-12 00:50:36 +04:00
|
|
|
}
|
2008-06-12 04:21:24 +04:00
|
|
|
if (enforce_min || list_empty(&workers->worker_list))
|
|
|
|
return NULL;
|
|
|
|
|
2008-06-12 00:50:36 +04:00
|
|
|
/*
|
2008-06-12 04:21:24 +04:00
|
|
|
* if we pick a busy task, move the task to the end of the list.
|
2008-09-29 23:18:18 +04:00
|
|
|
* hopefully this will keep things somewhat evenly balanced.
|
|
|
|
* Do the move in batches based on the sequence number. This groups
|
|
|
|
* requests submitted at roughly the same time onto the same worker.
|
2008-06-12 00:50:36 +04:00
|
|
|
*/
|
2008-06-12 04:21:24 +04:00
|
|
|
next = workers->worker_list.next;
|
|
|
|
worker = list_entry(next, struct btrfs_worker_thread, worker_list);
|
2008-08-15 23:34:17 +04:00
|
|
|
atomic_inc(&worker->num_pending);
|
|
|
|
worker->sequence++;
|
2008-09-29 23:18:18 +04:00
|
|
|
|
2008-08-15 23:34:18 +04:00
|
|
|
if (worker->sequence % workers->idle_thresh == 0)
|
2008-08-15 23:34:17 +04:00
|
|
|
list_move_tail(next, &workers->worker_list);
|
2008-06-12 00:50:36 +04:00
|
|
|
return worker;
|
|
|
|
}
|
|
|
|
|
2008-09-29 23:18:18 +04:00
|
|
|
/*
|
|
|
|
* selects a worker thread to take the next job. This will either find
|
|
|
|
* an idle worker, start a new worker up to the max count, or just return
|
|
|
|
* one of the existing busy workers.
|
|
|
|
*/
|
2008-06-12 00:50:36 +04:00
|
|
|
static struct btrfs_worker_thread *find_worker(struct btrfs_workers *workers)
|
|
|
|
{
|
|
|
|
struct btrfs_worker_thread *worker;
|
|
|
|
unsigned long flags;
|
|
|
|
|
|
|
|
again:
|
|
|
|
spin_lock_irqsave(&workers->lock, flags);
|
|
|
|
worker = next_worker(workers);
|
|
|
|
spin_unlock_irqrestore(&workers->lock, flags);
|
|
|
|
|
|
|
|
if (!worker) {
|
|
|
|
spin_lock_irqsave(&workers->lock, flags);
|
|
|
|
if (workers->num_workers >= workers->max_workers) {
|
2008-06-12 04:21:24 +04:00
|
|
|
struct list_head *fallback = NULL;
|
2008-06-12 00:50:36 +04:00
|
|
|
/*
|
|
|
|
* we have failed to find any workers, just
|
|
|
|
* return the force one
|
|
|
|
*/
|
2008-06-12 04:21:24 +04:00
|
|
|
if (!list_empty(&workers->worker_list))
|
|
|
|
fallback = workers->worker_list.next;
|
|
|
|
if (!list_empty(&workers->idle_list))
|
|
|
|
fallback = workers->idle_list.next;
|
|
|
|
BUG_ON(!fallback);
|
|
|
|
worker = list_entry(fallback,
|
2008-06-12 00:50:36 +04:00
|
|
|
struct btrfs_worker_thread, worker_list);
|
|
|
|
spin_unlock_irqrestore(&workers->lock, flags);
|
|
|
|
} else {
|
|
|
|
spin_unlock_irqrestore(&workers->lock, flags);
|
|
|
|
/* we're below the limit, start another worker */
|
|
|
|
btrfs_start_workers(workers, 1);
|
|
|
|
goto again;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return worker;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* btrfs_requeue_work just puts the work item back on the tail of the list
|
|
|
|
* it was taken from. It is intended for use with long running work functions
|
|
|
|
* that make some progress and want to give the cpu up for others.
|
|
|
|
*/
|
|
|
|
int btrfs_requeue_work(struct btrfs_work *work)
|
|
|
|
{
|
|
|
|
struct btrfs_worker_thread *worker = work->worker;
|
|
|
|
unsigned long flags;
|
|
|
|
|
Btrfs: Add ordered async work queues
Btrfs uses kernel threads to create async work queues for cpu intensive
operations such as checksumming and decompression. These work well,
but they make it difficult to keep IO order intact.
A single writepages call from pdflush or fsync will turn into a number
of bios, and each bio is checksummed in parallel. Once the checksum is
computed, the bio is sent down to the disk, and since we don't control
the order in which the parallel operations happen, they might go down to
the disk in almost any order.
The code deals with this somewhat by having deep work queues for a single
kernel thread, making it very likely that a single thread will process all
the bios for a single inode.
This patch introduces an explicitly ordered work queue. As work structs
are placed into the queue they are put onto the tail of a list. They have
three callbacks:
->func (cpu intensive processing here)
->ordered_func (order sensitive processing here)
->ordered_free (free the work struct, all processing is done)
The work struct has three callbacks. The func callback does the cpu intensive
work, and when it completes the work struct is marked as done.
Every time a work struct completes, the list is checked to see if the head
is marked as done. If so the ordered_func callback is used to do the
order sensitive processing and the ordered_free callback is used to do
any cleanup. Then we loop back and check the head of the list again.
This patch also changes the checksumming code to use the ordered workqueues.
One a 4 drive array, it increases streaming writes from 280MB/s to 350MB/s.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-11-07 06:03:00 +03:00
|
|
|
if (test_and_set_bit(WORK_QUEUED_BIT, &work->flags))
|
2008-06-12 00:50:36 +04:00
|
|
|
goto out;
|
|
|
|
|
|
|
|
spin_lock_irqsave(&worker->lock, flags);
|
|
|
|
atomic_inc(&worker->num_pending);
|
|
|
|
list_add_tail(&work->list, &worker->pending);
|
2008-10-01 03:24:06 +04:00
|
|
|
|
|
|
|
/* by definition we're busy, take ourselves off the idle
|
|
|
|
* list
|
|
|
|
*/
|
|
|
|
if (worker->idle) {
|
|
|
|
spin_lock_irqsave(&worker->workers->lock, flags);
|
|
|
|
worker->idle = 0;
|
|
|
|
list_move_tail(&worker->worker_list,
|
|
|
|
&worker->workers->worker_list);
|
|
|
|
spin_unlock_irqrestore(&worker->workers->lock, flags);
|
|
|
|
}
|
|
|
|
|
2008-06-12 00:50:36 +04:00
|
|
|
spin_unlock_irqrestore(&worker->lock, flags);
|
2008-10-01 03:24:06 +04:00
|
|
|
|
2008-06-12 00:50:36 +04:00
|
|
|
out:
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* places a struct btrfs_work into the pending queue of one of the kthreads
|
|
|
|
*/
|
|
|
|
int btrfs_queue_worker(struct btrfs_workers *workers, struct btrfs_work *work)
|
|
|
|
{
|
|
|
|
struct btrfs_worker_thread *worker;
|
|
|
|
unsigned long flags;
|
|
|
|
int wake = 0;
|
|
|
|
|
|
|
|
/* don't requeue something already on a list */
|
Btrfs: Add ordered async work queues
Btrfs uses kernel threads to create async work queues for cpu intensive
operations such as checksumming and decompression. These work well,
but they make it difficult to keep IO order intact.
A single writepages call from pdflush or fsync will turn into a number
of bios, and each bio is checksummed in parallel. Once the checksum is
computed, the bio is sent down to the disk, and since we don't control
the order in which the parallel operations happen, they might go down to
the disk in almost any order.
The code deals with this somewhat by having deep work queues for a single
kernel thread, making it very likely that a single thread will process all
the bios for a single inode.
This patch introduces an explicitly ordered work queue. As work structs
are placed into the queue they are put onto the tail of a list. They have
three callbacks:
->func (cpu intensive processing here)
->ordered_func (order sensitive processing here)
->ordered_free (free the work struct, all processing is done)
The work struct has three callbacks. The func callback does the cpu intensive
work, and when it completes the work struct is marked as done.
Every time a work struct completes, the list is checked to see if the head
is marked as done. If so the ordered_func callback is used to do the
order sensitive processing and the ordered_free callback is used to do
any cleanup. Then we loop back and check the head of the list again.
This patch also changes the checksumming code to use the ordered workqueues.
One a 4 drive array, it increases streaming writes from 280MB/s to 350MB/s.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-11-07 06:03:00 +03:00
|
|
|
if (test_and_set_bit(WORK_QUEUED_BIT, &work->flags))
|
2008-06-12 00:50:36 +04:00
|
|
|
goto out;
|
|
|
|
|
|
|
|
worker = find_worker(workers);
|
Btrfs: Add ordered async work queues
Btrfs uses kernel threads to create async work queues for cpu intensive
operations such as checksumming and decompression. These work well,
but they make it difficult to keep IO order intact.
A single writepages call from pdflush or fsync will turn into a number
of bios, and each bio is checksummed in parallel. Once the checksum is
computed, the bio is sent down to the disk, and since we don't control
the order in which the parallel operations happen, they might go down to
the disk in almost any order.
The code deals with this somewhat by having deep work queues for a single
kernel thread, making it very likely that a single thread will process all
the bios for a single inode.
This patch introduces an explicitly ordered work queue. As work structs
are placed into the queue they are put onto the tail of a list. They have
three callbacks:
->func (cpu intensive processing here)
->ordered_func (order sensitive processing here)
->ordered_free (free the work struct, all processing is done)
The work struct has three callbacks. The func callback does the cpu intensive
work, and when it completes the work struct is marked as done.
Every time a work struct completes, the list is checked to see if the head
is marked as done. If so the ordered_func callback is used to do the
order sensitive processing and the ordered_free callback is used to do
any cleanup. Then we loop back and check the head of the list again.
This patch also changes the checksumming code to use the ordered workqueues.
One a 4 drive array, it increases streaming writes from 280MB/s to 350MB/s.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-11-07 06:03:00 +03:00
|
|
|
if (workers->ordered) {
|
|
|
|
spin_lock_irqsave(&workers->lock, flags);
|
|
|
|
list_add_tail(&work->order_list, &workers->order_list);
|
|
|
|
spin_unlock_irqrestore(&workers->lock, flags);
|
|
|
|
} else {
|
|
|
|
INIT_LIST_HEAD(&work->order_list);
|
|
|
|
}
|
2008-06-12 00:50:36 +04:00
|
|
|
|
|
|
|
spin_lock_irqsave(&worker->lock, flags);
|
|
|
|
atomic_inc(&worker->num_pending);
|
2008-06-12 04:21:24 +04:00
|
|
|
check_busy_worker(worker);
|
2008-06-12 00:50:36 +04:00
|
|
|
list_add_tail(&work->list, &worker->pending);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* avoid calling into wake_up_process if this thread has already
|
|
|
|
* been kicked
|
|
|
|
*/
|
|
|
|
if (!worker->working)
|
|
|
|
wake = 1;
|
|
|
|
worker->working = 1;
|
|
|
|
|
|
|
|
spin_unlock_irqrestore(&worker->lock, flags);
|
|
|
|
|
|
|
|
if (wake)
|
|
|
|
wake_up_process(worker->task);
|
|
|
|
out:
|
|
|
|
return 0;
|
|
|
|
}
|