Re: your mail

From: Joonsoo Kim
Date: Fri Apr 21 2017 - 00:38:57 EST


On Thu, Apr 20, 2017 at 09:28:20AM +0200, Michal Hocko wrote:
> On Thu 20-04-17 10:27:55, Joonsoo Kim wrote:
> > On Mon, Apr 17, 2017 at 10:15:15AM +0200, Michal Hocko wrote:
> [...]
> > > Which pfn walkers you have in mind?
> >
> > For example, kpagecount_read() in fs/proc/page.c. I searched it by
> > using pfn_valid().
>
> Yeah, I've checked that one and in fact this is a good example of the
> case where you do not really care about holes. It just checks the page
> count which is a valid information under any circumstances.

I don't think so. First, it checks the page *map* count. Is it still valid
even if PageReserved() is set? What I'd like to ask in this example is
that what information is valid if PageReserved() is set. Is there any
design document on this? I think that we need to define/document it first.

And, I hope that all the information in flags field is valid in all
cases if pfn_valid() return true. By the design.

This makes all the exsiting pfn walkers happy since we don't need an
additional check for PageReserved().

>
> > > > The other problem I found is that your change will makes some
> > > > contiguous zones to be considered as non-contiguous. Memory allocated
> > > > by memblock API is also marked as PageResereved. If we consider this as
> > > > a hole, we will set such a zone as non-contiguous.
> > >
> > > Why would that be a problem? We shouldn't touch those pages anyway?
> >
> > Skipping those pages in compaction are valid so no problem in this
> > case.
> >
> > The problem I mentioned above is that adding PageReserved() check in
> > __pageblock_pfn_to_page() invalidates optimization by
> > set_zone_contiguous(). In compaction, we need to get a valid struct
> > page and it requires a lot of work. There is performance problem
> > report due to this so set_zone_contiguous() optimization is added. It
> > checks if the zone is contiguous or not in boot time. If zone is
> > determined as contiguous, we can easily get a valid struct page in
> > runtime without expensive checks.
>
> OK, I see. I've had some vague understading and the clarification helps.
>
> > Your patch try to add PageReserved() to __pageblock_pfn_to_page(). It
> > woule make that zone->contiguous usually returns false since memory
> > used by memblock API is marked as PageReserved() and your patch regard
> > it as a hole. It invalidates set_zone_contiguous() optimization and I
> > worry about it.
>
> OK, fair enough. I did't consider memblock allocations. I will rethink
> this patch but there are essentially 3 options
> - use a different criterion for the offline holes dection. I
> have just realized we might do it by storing the online
> information into the mem sections
> - drop this patch
> - move the PageReferenced check down the chain into
> isolate_freepages_block resp. isolate_migratepages_block
>
> I would prefer 3 over 2 over 1. I definitely want to make this more
> robust so 1 is preferable long term but I do not want this to be a
> roadblock to the rest of the rework. Does that sound acceptable to you?

I like #1 among of above options and I already see your patch for #1.
It's much better than your first attempt but I'm still not happy due
to the semantic of pfn_valid().

> [..]
> > Let me clarify my desire(?) for this issue.
> >
> > 1. If pfn_valid() returns true, struct page has valid information, at
> > least, in flags (zone id, node id, flags, etc...). So, we can use them
> > without checking PageResereved().
>
> This is no longer true after my rework. Pages are associated with the
> zone during _onlining_ rather than when they are physically hotpluged.

If your rework make information valid during _onlining_, my
suggestion is making pfn_valid() return false until onlining.

Caller of pfn_valid() expects that they can get valid information from
the struct page. There is no reason to access the struct page if they
can't get valid information from it. So, passing pfn_valid() should
guarantee that, at least, some kind of information is valid.

If pfn_valid() doesn't guarantee it, most of the pfn walker should
check PageResereved() to make sure that validity of information from
the struct page.

> Basically only the nid is set properly. Strictly speaking this is the
> case also without my rework because the zone might change during online
> phase so you cannot assume it is correct even now. It just happens that
> it more or less works just fine.
>
> > 2. pfn_valid() for offlined holes returns false. This can be easily
> > (?) implemented by manipulating SECTION_MAP_MASK in hotplug code. I
> > guess that there is no reason that pfn_valid() returns true for
> > offlined holes. If there is, please let me know.
>
> There is some code which really expects that pfn_valid returns true iff
> there is a struct page and it doesn't care about the online status.
> E.g. hotplug code itself so no, we cannot change pfn_valid. What we can
> do though is to add pfn_to_online_page which would do the proper check.
> I have already sent [1]. As noted above we can (ab)use the remaining bit
> in SECTION_MAP_MASK to detect offline pages more robustly.

Some pfn_valid() caller in hotplug code look wrong. They want to check
section's validity rather than pfn's validity. Others want to access
the struct page so they fit for my assumption (?) for pfn_valid().
Therefore, we can change that pfn_valid() return false until online.

> > 3. We don't need to check PageReserved() in most of pfn walkers in
> > order to check offline holes.
>
> We still have to distinguish those who care about offline pages from
> those who do not care about it.

Hotplug code can distinguish those by another way by using new section
mask as you did in a new patch. If someone excluding hotplug code do
care about offline pages, it would be just for optimization rather
than correteness. I think that it's okay.

Thanks.