Re: [PATCH v2] mm: introduce validity check on vm dirtiness settings

From: Jan Kara
Date: Wed Sep 20 2017 - 11:34:03 EST


On Tue 19-09-17 19:48:00, Yafang Shao wrote:
> 2017-09-19 16:35 GMT+08:00 Jan Kara <jack@xxxxxxx>:
> > On Tue 19-09-17 06:53:00, Yafang Shao wrote:
> >> + if (vm_dirty_bytes == 0 && vm_dirty_ratio == 0 &&
> >> + (dirty_background_bytes != 0 || dirty_background_ratio != 0))
> >> + ret = false;
> >
> > Hum, why not just:
> > if ((vm_dirty_bytes == 0 && vm_dirty_ratio == 0) ||
> > (dirty_background_bytes == 0 && dirty_background_ratio == 0))
> > ret = false;
> >
> > IMHO setting either tunable to 0 is just wrong and actively dangerous...
> >
>
> Because these four variables all could be set to 0 before, and I'm not
> sure if this
> is needed under some certain conditions, although I think this is
> dangerous but I have
> to keep it as before.
>
> If you think that is wrong, then I will modified it as you suggested.

OK, I see but see below.

> >> int dirty_background_ratio_handler(struct ctl_table *table, int write,
> >> void __user *buffer, size_t *lenp,
> >> loff_t *ppos)
> >> {
> >> int ret;
> >> + int old_ratio = dirty_background_ratio;
> >>
> >> ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
> >> - if (ret == 0 && write)
> >> - dirty_background_bytes = 0;
> >> + if (ret == 0 && write) {
> >> + if (dirty_background_ratio != old_ratio &&
> >> + !vm_dirty_settings_valid()) {
> >
> > Why do you check whether new ratio is different here? If it is really
> > needed, it would deserve a comment.
> >
>
> There're two reseaons,
> 1. if you set a value same with the old value, it's needn't to do this check.
> 2. there's another behavior that I'm not sure whether it is reaonable. i.e.
> if the old value is,
> vm.dirty_background_bytes = 0;
> vm.dirty_background_ratio=10;
> then I execute the bellow command,
> sysctl -w vm.dirty_background_bytes=0
> at the end these two values will be,
> vm.dirty_background_bytes = 0;
> vm.dirty_background_ratio=0;
> I'm not sure if this is needed under some certain conditons, So I have
> to keep it as before.

OK, this is somewhat the problem of the switching logic between _bytes and
_ratio bytes and also the fact that '0' has a special meaning in these
files. I think the cleanest would be to just refuse writing of '0' into any
of these files which would deal with the problem as well.

Honza
--
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR