Re: [f2fs-dev] [PATCH] f2fs: remove a waiter for checkpoint completion

[Jaegeuk Kim]

On 08/04, Chao Yu wrote:
> On 2020/8/4 9:04, Jaegeuk Kim wrote:
> > On 08/04, Chao Yu wrote:
> > > On 2020/8/4 1:28, Jaegeuk Kim wrote:
> > > > It doesn't need to wait for checkpoint being completed triggered by end_io.
> > > > 
> > > > [   20.157753] ------------[ cut here ]------------
> > > > [   20.158393] do not call blocking ops when !TASK_RUNNING; state=2 set at [<0000000096354225>] prepare_to_wait+0xcd/0x430
> > > > [   20.159858] WARNING: CPU: 1 PID: 1152 at kernel/sched/core.c:7142 __might_sleep+0x149/0x1a0
> > > > ...
> > > > [   20.176110]  __submit_merged_write_cond+0x191/0x310
> > > > [   20.176739]  f2fs_submit_merged_write+0x18/0x20
> > > > [   20.177323]  f2fs_wait_on_all_pages+0x269/0x2d0
> > > > [   20.177899]  ? block_operations+0x980/0x980
> > > > [   20.178441]  ? __kasan_check_read+0x11/0x20
> > > > [   20.178975]  ? finish_wait+0x260/0x260
> > > > [   20.179488]  ? percpu_counter_set+0x147/0x230
> > > > [   20.180049]  do_checkpoint+0x1757/0x2a50
> > > > [   20.180558]  f2fs_write_checkpoint+0x840/0xaf0
> > > > [   20.181126]  f2fs_sync_fs+0x287/0x4a0
> > > > 
> > > > Reported-by: Eric Biggers <ebiggers@xxxxxxxxxx>
> > > > Signed-off-by: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> > > > ---
> > > >    fs/f2fs/checkpoint.c | 6 +-----
> > > >    fs/f2fs/data.c       | 4 ----
> > > >    fs/f2fs/f2fs.h       | 1 -
> > > >    fs/f2fs/super.c      | 1 -
> > > >    4 files changed, 1 insertion(+), 11 deletions(-)
> > > > 
> > > > diff --git a/fs/f2fs/checkpoint.c b/fs/f2fs/checkpoint.c
> > > > index 99c8061da55b9..2bdddc725e677 100644
> > > > --- a/fs/f2fs/checkpoint.c
> > > > +++ b/fs/f2fs/checkpoint.c
> > > > @@ -1255,11 +1255,7 @@ static void unblock_operations(struct f2fs_sb_info *sbi)
> > > >    void f2fs_wait_on_all_pages(struct f2fs_sb_info *sbi, int type)
> > > >    {
> > > > -	DEFINE_WAIT(wait);
> > > > -
> > > >    	for (;;) {
> > > > -		prepare_to_wait(&sbi->cp_wait, &wait, TASK_UNINTERRUPTIBLE);
> > > 
> > > Wouldn't that case high cpu usage before io end?
> > 
> > This is a critical path to wait for IO completion in checkpoint, which would be
> > better to wait for it to avoid long latency to continue filesystem operations.
> 
> Yup, in previous implementation, last end_io wakes up checkpoint() waiter, we
> didn't waste any time there.
> 
> > Moreover, I expect io_schedule_timeout() can mitigate such the CPU overhead and
> > actually we don't need to make there-in context switches as well.
> 
> Then io_schedule_timeout() in this loop may give CPU time slice to other thread
> until scheduler reselect checkpoint(), that would cause longer latency?

Hmm, how about this then?