Hi Gu,
> Hi Yuan,
> On 10/08/2013 07:30 PM, Yuan Zhong wrote:
>
>> Hi Gu,
>>
>>> Hi Yuan,
>>> On 10/08/2013 04:30 PM, Yuan Zhong wrote:
>>
>>>> Previously, do_checkpoint() will call congestion_wait() for waiting the pages (previous submitted node/meta/data pages) to be written back.
>>>> Because congestion_wait() will set a regular period (e.g. HZ / 50 ) for waiting.
>>>> For this reason, there is a situation that after the pages have been written back,
>>>> but the checkpoint thread still wait for congestion_wait to exit.
>>
>>> How do you confirm this issue?
>>
>> I traced the execution path.
>> In f2fs_end_io_write, dec_page_count(p->sbi, F2FS_WRITEBACK) will be called.
>> And I found that, when pages of F2FS_WRITEBACK has been zero, but
>> checkpoint thread still congestion_wait for pages of F2FS_WRITEBACK to be zero.
>
>Yes, it maybe. Congestion_wait add the task to a global wait queue which related to
>all back devices, so if F2FS_WRITEBACK has been zero, but other io may be still going on.
>Anyway, using a private wait queue to hold is a better choose.:)
>
>
>> So, I think this point could be improved.
>> And I wrote a simple test case and tested on Micro-SD card, the steps as following:
>> (a) create a fixed-size file (4KB)
>> (b) go on to sync the file
>> (c) go back to step #a (fixed numbers of cycling:1024)
>> The results indicated that the execution time is reduced greatly by using this patch.
>
>Yes, the change is an improvement if the issue is existent.
>
>
>>
>>
>>> I suspect that the block-core does not have a wake-up mechanism
>>> when the back device is uncongested.
>>
>>
>> Yes, you are right.
>> So I wake up the checkpoint thread by myself, when pages of F2FS_WRITEBACK to be zero.
>> In f2fs_end_io_write, f2fs_writeback_wake is called.
>> you cloud find this code in my patch.
>
>Saw it.:)
>But one problem is that the checkpoint routine always is singleton, so the wait queue just
>services only one body, it seems not very worthy. How about just schedule and wake up it
>directly? See the following one.
Yes, your point is right.
My reason for using wait queue is that I am influenced by congestion_wait function.
The inner function of congesiton_wait is also using wait_queue.
And, I think, your patch is also a more efficient method.
>
>Signed-off-by: Gu Zheng <guz.fnst@xxxxxxxxxxxxxx>
>---
> fs/f2fs/checkpoint.c | 11 +++++++++--
> fs/f2fs/f2fs.h | 1 +
> fs/f2fs/segment.c | 4 ++++
> 3 files changed, 14 insertions(+), 2 deletions(-)
>
>diff --git a/fs/f2fs/checkpoint.c b/fs/f2fs/checkpoint.c
>index d808827..2a5999d 100644
>--- a/fs/f2fs/checkpoint.c
>+++ b/fs/f2fs/checkpoint.c
>@@ -757,8 +757,15 @@ static void do_checkpoint(struct f2fs_sb_info *sbi, bool is_umount)
> f2fs_put_page(cp_page, 1);
>
> /* wait for previous submitted node/meta pages writeback */
>- while (get_pages(sbi, F2FS_WRITEBACK))
>- congestion_wait(BLK_RW_ASYNC, HZ / 50);
>+ sbi->cp_task = current;
>+ while (get_pages(sbi, F2FS_WRITEBACK)) {
>+ set_current_state(TASK_UNINTERRUPTIBLE);
>+ if (!get_pages(sbi, F2FS_WRITEBACK))
>+ break;
>+ io_schedule();
>+ }
>+ __set_current_state(TASK_RUNNING);
>+ sbi->cp_task = NULL;
>
> filemap_fdatawait_range(sbi->node_inode->i_mapping, 0, LONG_MAX);
> filemap_fdatawait_range(sbi->meta_inode->i_mapping, 0, LONG_MAX);
>diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
>index a955a59..408ace7 100644
>--- a/fs/f2fs/f2fs.h
>+++ b/fs/f2fs/f2fs.h
>@@ -365,6 +365,7 @@ struct f2fs_sb_info {
> struct mutex writepages; /* mutex for writepages() */
> int por_doing; /* recovery is doing or not */
> int on_build_free_nids; /* build_free_nids is doing */
>+ struct task_struct *cp_task; /* checkpoint task */
>
> /* for orphan inode management */
> struct list_head orphan_inode_list; /* orphan inode list */
>diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
>index bd79bbe..3b20359 100644
>--- a/fs/f2fs/segment.c
>+++ b/fs/f2fs/segment.c
>@@ -597,6 +597,10 @@ static void f2fs_end_io_write(struct bio *bio, int err)
>
> if (p->is_sync)
> complete(p->wait);
>+
>+ if (!get_pages(p->sbi, F2FS_WRITEBACK) && p->sbi->cp_task)
>+ wake_up_process(p->sbi->cp_task);
>+
> kfree(p);
> bio_put(bio);
> }
>--
>1.7.7
>
>Regards,
>Gu
>
Regards,
Yuan
>>
>>
>>>> This is a problem here, especially, when sync a large number of small files or dirs.
>>>> In order to avoid this, a wait_list is introduced,
>>>> the checkpoint thread will be dropped into the wait_list if the pages have not been written back,
>>>> and will be waked up by contrast.
>>
>>> Please pay some attention to the mail form, this mail is out of format in my mail client.
>>
>>> Regards,
>>> Gu
>>
>> Regards,
>> Yuan
>>
>>>>
>>>> Signed-off-by: Yuan Zhong <yuan.mark.zhong@xxxxxxxxxxx>
>>>> ---
>>>> fs/f2fs/checkpoint.c | 3 +--
>>>> fs/f2fs/f2fs.h | 19 +++++++++++++++++++
>>>> fs/f2fs/segment.c | 1 +
>>>> fs/f2fs/super.c | 1 +
>>>> 4 files changed, 22 insertions(+), 2 deletions(-)
>>>>
>>>> diff --git a/fs/f2fs/checkpoint.c b/fs/f2fs/checkpoint.c
>>>> index ca39442..5d69ae0 100644
>>>> --- a/fs/f2fs/checkpoint.c
>>>> +++ b/fs/f2fs/checkpoint.c
>>>> @@ -758,8 +758,7 @@ static void do_checkpoint(struct f2fs_sb_info *sbi, bool is_umount)
>>>> f2fs_put_page(cp_page, 1);
>>>>
>>>> /* wait for previous submitted node/meta pages writeback */
>>>> - while (get_pages(sbi, F2FS_WRITEBACK))
>>>> - congestion_wait(BLK_RW_ASYNC, HZ / 50);
>>>> + f2fs_writeback_wait(sbi);
>>>>
>>>> filemap_fdatawait_range(sbi->node_inode->i_mapping, 0, LONG_MAX);
>>>> filemap_fdatawait_range(sbi->meta_inode->i_mapping, 0, LONG_MAX);
>>>> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
>>>> index 7fd99d8..4b0d70e 100644
>>>> --- a/fs/f2fs/f2fs.h
>>>> +++ b/fs/f2fs/f2fs.h
>>>> @@ -18,6 +18,8 @@
>>>> #include <linux/crc32.h>
>>>> #include <linux/magic.h>
>>>> #include <linux/kobject.h>
>>>> +#include <linux/wait.h>
>>>> +#include <linux/sched.h>
>>>>
>>>> /*
>>>> * For mount options
>>>> @@ -368,6 +370,7 @@ struct f2fs_sb_info {
>>>> struct mutex fs_lock[NR_GLOBAL_LOCKS]; /* blocking FS operations */
>>>> struct mutex node_write; /* locking node writes */
>>>> struct mutex writepages; /* mutex for writepages() */
>>>> + wait_queue_head_t writeback_wqh; /* wait_queue for writeback */
>>>> unsigned char next_lock_num; /* round-robin global locks */
>>>> int por_doing; /* recovery is doing or not */
>>>> int on_build_free_nids; /* build_free_nids is doing */
>>>> @@ -961,6 +964,22 @@ static inline int f2fs_readonly(struct super_block *sb)
>>>> return sb->s_flags & MS_RDONLY;
>>>> }
>>>>
>>>> +static inline void f2fs_writeback_wait(struct f2fs_sb_info *sbi)
>>>> +{
>>>> + DEFINE_WAIT(wait);
>>>> +
>>>> + prepare_to_wait(&sbi->writeback_wqh, &wait, TASK_UNINTERRUPTIBLE);
>>>> + if (get_pages(sbi, F2FS_WRITEBACK))
>>>> + io_schedule();
>>>> + finish_wait(&sbi->writeback_wqh, &wait);
>>>> +}
>>>> +
>>>> +static inline void f2fs_writeback_wake(struct f2fs_sb_info *sbi)
>>>> +{
>>>> + if (!get_pages(sbi, F2FS_WRITEBACK))
>>>> + wake_up_all(&sbi->writeback_wqh);
>>>> +}
>>>> +
>>>> /*
>>>> * file.c
>>>> */
>>>> diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
>>>> index bd79bbe..0708aa9 100644
>>>> --- a/fs/f2fs/segment.c
>>>> +++ b/fs/f2fs/segment.c
>>>> @@ -597,6 +597,7 @@ static void f2fs_end_io_write(struct bio *bio, int err)
>>>>
>>>> if (p->is_sync)
>>>> complete(p->wait);
>>>> + f2fs_writeback_wake(p->sbi);
>>>> kfree(p);
>>>> bio_put(bio);
>>>> }
>>>> diff --git a/fs/f2fs/super.c b/fs/f2fs/super.c
>>>> index 094ccc6..3ac6d85 100644
>>>> --- a/fs/f2fs/super.c
>>>> +++ b/fs/f2fs/super.c
>>>> @@ -835,6 +835,7 @@ static int f2fs_fill_super(struct super_block *sb, void *data, int silent)
>>>> mutex_init(&sbi->gc_mutex);
>>>> mutex_init(&sbi->writepages);
>>>> mutex_init(&sbi->cp_mutex);
>>>> + init_waitqueue_head(&sbi->writeback_wqh);
>>>> for (i = 0; i < NR_GLOBAL_LOCKS; i++)
>>>> mutex_init(&sbi->fs_lock[i]);
>>>> mutex_init(&sbi->node_write);
ÿôèº{.nÇ+‰·Ÿ®‰†+%ŠËÿ±éݶ¥Šwÿº{.nÇ+‰·¥Š{±ýû z÷¥þ)í…æèw* jg¬±¨ ¶‰šŽŠÝ¢jÿ¾«þG«?éÿ¢¸ ¢·¦j:+v‰¨ŠwèjØm¶Ÿÿþø ¯ù ®w¥þŠàþf£¢·hš?â?úÿ†Ù¥
