Re: [PATCH v2] mm: page_alloc: dump migrate-failed pages

[Michal Hocko <mhocko@suse.com>]

On Wed 10-03-21 13:26:23, Matthew Wilcox wrote:
> On Tue, Mar 09, 2021 at 10:32:51AM +0100, Michal Hocko wrote:
> > Apart from the above, do we have to warn for something that is a
> > debugging aid? A similar concern wrt dump_page which uses pr_warn and
> > page owner is using even pr_alert.
> > Would it make sense to add a loglevel parameter both into __dump_page
> > and dump_page_owner?
> 
> No.  What would make sense is turning __dump_page() inside-out.
> Something like printk("%pP\n");
> 
> In lib/vsprintf.c, there's a big switch statement in the function
> pointer() that handles printing things like IPv6 addresses, dentries,
> and function symbols.
> 
> Then we can do whatever we want around the new %pP, including choosing
> the log level, adding additional information, choosing to dump the page
> to a sysfs file, etc, etc.

Hmm, __dump_page has grown quite some heavy lifting over time and I am
not sure this is a good candidate to put into printk proper (e.g. is it
safe/reasonable to call get_kernel_nofault from printk - aka arbitrary
context)?.

But you've got a point that such a printk format wouldn't need to be 1:1
with the existing __dump_page. There is quite a lot to infer from page
count, map count, flags, page type already. Then a question would be
what is an actual advantage of %pP over dump_page_info(loglvl, p). One I
can see is that %pP would allow to dump the state into a string and so
it would be more versatile but I am not aware of a usecase for that
(maybe tracing?).
-- 
Michal Hocko
SUSE Labs