Re: [PATCH resend] drop_caches: add some documentation and infomessage

[Michal Hocko]

On Sat 03-08-13 16:16:58, KOSAKI Motohiro wrote:
> >>> You missed the "!".  I'm proposing that setting the new bit 2 will
> >>> permit people to prevent the new printk if it is causing them problems.
> >>
> >> No I don't. I'm sure almost all abuse users think our usage is correct. Then,
> >> I can imagine all crazy applications start to use this flag eventually.
> > 
> > I guess we do not care about those. If somebody wants to shoot his feet
> > then we cannot do much about it. The primary motivation was to find out
> > those that think this is right and they are willing to change the setup
> > once they know this is not the right way to do things.
> > 
> > I think that giving a way to suppress the warning is a good step. Log
> > level might be to coarse and sysctl would be an overkill.
> 
> When Dave Hansen reported this issue originally, he explained a lot of userland
> developer misuse /proc/drop_caches because they don't understand what
> drop_caches do.
> So, if they never understand the fact, why can we trust them? I have no
> idea.

Well, most of that usage I have come across was legacy scripts which
happened to work at a certain point in time because we sucked.
Thinks have changed but such scripts happen to survive a long time.
We are primarily interested in those.

> Or, if you have different motivation w/ Dave, please let me know it.

We have seen reports where users complained about performance drop down
when in fact the real culprit turned out to be such a clever script
which dropped caches on the background thinking it will help to free
some memory. Such cases are tedious to reveal.

> While the purpose is to shoot misuse, I don't think we can trust
> userland app.  If "If somebody wants to shoot his feet then we cannot
> do much about it." is true, this patch is useless. OK, we still catch
> the right user.

I do not think it is useless. It will print a message for all those
users initially. It is a matter of user how to deal with it.

> But we never want to know who is the right users, right?

Well, those that are curious about a new message in the lock and come
back to us asking what is going on are those we are primarily interested
in.
-- 
Michal Hocko
SUSE Labs
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/