Re: [PATCH] mm/page-writeback.c: fix bug caused by disable periodic writeback

[Yafang Shao]

2017-10-10 17:33 GMT+08:00 Jan Kara <jack@xxxxxxx>:
> On Tue 10-10-17 17:14:48, Yafang Shao wrote:
>> 2017-10-10 16:48 GMT+08:00 Jan Kara <jack@xxxxxxx>:
>> > On Tue 10-10-17 16:00:29, Yafang Shao wrote:
>> >> 2017-10-10 6:42 GMT+08:00 Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>:
>> >> > On Sat,  7 Oct 2017 06:58:04 +0800 Yafang Shao <laoar.shao@xxxxxxxxx> wrote:
>> >> >
>> >> >> After disable periodic writeback by writing 0 to
>> >> >> dirty_writeback_centisecs, the handler wb_workfn() will not be
>> >> >> entered again until the dirty background limit reaches or
>> >> >> sync syscall is executed or no enough free memory available or
>> >> >> vmscan is triggered.
>> >> >> So the periodic writeback can't be enabled by writing a non-zero
>> >> >> value to dirty_writeback_centisecs
>> >> >> As it can be disabled by sysctl, it should be able to enable by
>> >> >> sysctl as well.
>> >> >>
>> >> >> ...
>> >> >>
>> >> >> --- a/mm/page-writeback.c
>> >> >> +++ b/mm/page-writeback.c
>> >> >> @@ -1972,7 +1972,13 @@ bool wb_over_bg_thresh(struct bdi_writeback *wb)
>> >> >>  int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
>> >> >>       void __user *buffer, size_t *length, loff_t *ppos)
>> >> >>  {
>> >> >> -     proc_dointvec(table, write, buffer, length, ppos);
>> >> >> +     unsigned int old_interval = dirty_writeback_interval;
>> >> >> +     int ret;
>> >> >> +
>> >> >> +     ret = proc_dointvec(table, write, buffer, length, ppos);
>> >> >> +     if (!ret && !old_interval && dirty_writeback_interval)
>> >> >> +             wakeup_flusher_threads(0, WB_REASON_PERIODIC);
>> >> >> +
>> >> >>       return 0;
>> >> >
>> >> > We could do with a code comment here, explaining why this code exists.
>> >> >
>> >>
>> >> OK. I will comment here.
>> >>
>> >> > And...  I'm not sure it works correctly?  For example, if a device
>> >> > doesn't presently have bdi_has_dirty_io() then wakeup_flusher_threads()
>> >> > will skip it and the periodic writeback still won't be started?
>> >> >
>> >>
>> >> That's an issue.
>> >> The periodic writeback won't be started.
>> >>
>> >> Maybe we'd better call  wb_wakeup_delayed(wb) here to bypass the
>> >> bdi_has_dirty_io() check ?
>> >
>> > Well, wb_wakeup_delayed() would be more appropriate but you'd then have to
>> > iterate over all bdis and wbs to be able to call it which IMO isn't worth
>> > the pain for a special case like this. But the decision is worth mentioning
>> > in the comment. Also wakeup_flusher_threads() does in principle what you
>> > need - see my reply to Andrew for details.
>> >
>> >                                                                 Honza
>>
>> Thanks for your explaination. I understood.
>> I will mention it in the comment.
>>
>> Should we do the wakeup whenever dirty_writeback_interval changes ?
>> If we still use wakeup_flusher_threads(), it will wakeup the flusher
>> threads immediately after we make the change.
>
> Yes, I think we should wakeup for every change of dirty_writeback_interval.
> And immediate wakeup is not a problem IMO.
>

Got it!


Thanks
Yafang