Re: [PATCH] f2fs: fix the discard thread sleep timeout under high utilization

[Chao Yu]

Hi Sahitya,

On 2021/3/15 15:46, Sahitya Tummala wrote:
> Hi Chao,
>
> On Mon, Mar 15, 2021 at 02:12:44PM +0800, Chao Yu wrote:
> > Sahitya,
> >
> > On 2021/3/15 12:56, Sahitya Tummala wrote:
> > > When f2fs is heavily utilized over 80%, the current discard policy
> > > sets the max sleep timeout of discard thread as 50ms
> > > (DEF_MIN_DISCARD_ISSUE_TIME). But this is set even when there are
> > > no pending discard commands to be issued. This results into
> > > unnecessary frequent and periodic wake ups of the discard thread.
> > > This patch adds check for pending  discard commands in addition
> > > to heavy utilization condition to prevent those wake ups.
> >
> > Could this commit fix your issue?
> >
> > https://git.kernel.org/pub/scm/linux/kernel/git/jaegeuk/f2fs.git/commit/?h=dev&id=43f8c47ea7d59c7b2270835f1d7c4618a1ea27b6
> I don't think it will help because we are changing the max timeout of the
> dpolicy itself in __init_discard_policy() when util > 80% as below -
>
> dpolicy->max_interval = DEF_MIN_DISCARD_ISSUE_TIME;
>
> And issue_discard_thread() uses this value as wait_ms, when there
> are no more pending discard commands to be issued.
> <snip>
>                  } else {
>                          wait_ms = dpolicy.max_interval;
>                  }
> <snip>
>
> The new patch posted above is not changing anything related to the  max_interval.
> Hence, I think it won't help the uncessary wakeup problem I am trying to solve
> for this condition - util > 80% and when there are no pending discards.
>
> Please let me know if i am missing something.

Copied, thanks for the explanation.

But there is another case which can cause this issue in the case of
disk util < 80%.

wait_ms = DEF_MIN_DISCARD_ISSUE_TIME;

do {
	wait_event_interruptible_timeout(, wait_ms);

	...

	if (!atomic_read(&dcc->discard_cmd_cnt))
[1] new statement
		continue;

} while();

Then the loop will wakeup whenever 50ms timeout.

So, to avoid this case, shouldn't we reset wait_ms to dpolicy.max_interval
at [1]?

Meanwhile, how about relocating discard_cmd_cnt check after
__init_discard_policy(DPOLICY_FORCE)? and olny set .max_interval to
DEF_MAX_DISCARD_ISSUE_TIME if there is no discard command, and keep
.granularity to 1?

Thanks,

> Thanks,
> Sahitya.
>
> > Thanks,
> >
> > > Signed-off-by: Sahitya Tummala <stummala@xxxxxxxxxxxxxx>
> > > ---
> > >   fs/f2fs/segment.c | 5 ++++-
> > >   1 file changed, 4 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
> > > index dced46c..df30220 100644
> > > --- a/fs/f2fs/segment.c
> > > +++ b/fs/f2fs/segment.c
> > > @@ -1112,6 +1112,8 @@ static void __init_discard_policy(struct f2fs_sb_info *sbi,
> > >   				struct discard_policy *dpolicy,
> > >   				int discard_type, unsigned int granularity)
> > >   {
> > > +	struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
> > > +
> > >   	/* common policy */
> > >   	dpolicy->type = discard_type;
> > >   	dpolicy->sync = true;
> > > @@ -1129,7 +1131,8 @@ static void __init_discard_policy(struct f2fs_sb_info *sbi,
> > >   		dpolicy->io_aware = true;
> > >   		dpolicy->sync = false;
> > >   		dpolicy->ordered = true;
> > > -		if (utilization(sbi) > DEF_DISCARD_URGENT_UTIL) {
> > > +		if (utilization(sbi) > DEF_DISCARD_URGENT_UTIL &&
> > > +				atomic_read(&dcc->discard_cmd_cnt)) {
> > >   			dpolicy->granularity = 1;
> > >   			dpolicy->max_interval = DEF_MIN_DISCARD_ISSUE_TIME;
> > >   		}