Message-ID: <8e2f3e24-e2b9-42ee-a401-6c4b681b9ad3@redhat.com>
Date:   Mon, 20 Nov 2023 15:49:22 +0100
From:   David Hildenbrand <david@...hat.com>
To:     Gerald Schaefer <gerald.schaefer@...ux.ibm.com>
Cc:     Sumanth Korikkar <sumanthk@...ux.ibm.com>,
        linux-mm <linux-mm@...ck.org>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Oscar Salvador <osalvador@...e.de>,
        Michal Hocko <mhocko@...e.com>,
        "Aneesh Kumar K.V" <aneesh.kumar@...ux.ibm.com>,
        Anshuman Khandual <anshuman.khandual@....com>,
        Alexander Gordeev <agordeev@...ux.ibm.com>,
        Heiko Carstens <hca@...ux.ibm.com>,
        Vasily Gorbik <gor@...ux.ibm.com>,
        linux-s390 <linux-s390@...r.kernel.org>,
        LKML <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH 6/8] s390/mm: implement MEM_PHYS_ONLINE MEM_PHYS_OFFLINE
 memory notifiers

[catching up on mails]

>>> This new approach has the advantage that we do not need to allocate any
>>> additional memory during online phase, neither for direct mapping page
>>> tables nor struct pages, so that memory hotplug can never fail.
>>
>> Right, just like any other architecture that (triggered by whatever
>> mechanism) ends up calling add_memory() and friends.
> 
> Just for better understanding, are page tables for identity and also
> vmemmap mapping not allocated from system memory on other archs? I.e.
> no altmap support for that, only for struct pages (so far)?

Yes, only the actual "memmap ("struct page")" comes from altmap space, 
everything else comes from the buddy during memory hotplug.

> 
>>
>>>
>>> The old approach (without altmap) is already a hack, because we add
>>> the memmap / struct pages, but for memory that is not really accessible.
>>
>> Yes, it's disgusting. And you still allocate other things like memory
>> block devices or the identify map.
> 
> I would say it is special :-). And again, for understanding, all other

:)

> things apart from struct pages, still would need to be allocated from
> system memory on other archs?

Yes!

> 
> Of course, struct pages would be by far the biggest part, so having
> altmap support for that helps a lot. Doing the other allocations also
> via altmap would feel natural, but it is not possible yet, or did we
> miss something?

The tricky part is making sure ahead of time that that we set aside the 
required number of pageblocks, to properly control during memory 
onlining what to set aside and what to expose to the buddy.

See mhp_supports_memmap_on_memory() / 
memory_block_memmap_on_memory_pages() for the dirty details :)

> 
>>
>>> And with all the disadvantage of pre-allocating struct pages from system
>>> memory.
>>
>> Jep. It never should have been done like that.
> 
> At that time, it gave the benefit over all others, that we do not need
> to allocate struct pages from system memory, at the time of memory online,
> when memory pressure might be high and such allocations might fail.

Agreed. Having the memmap already around can be helpful. But not for all 
standby memory, that's just pure waste.

... but as memory onlining is triggered by user space, it's likely that 
that user space cannot even make progress (e.g., start a process to set 
memory online) to actually trigger memory onlining in serious low-memory 
situations.

> 
> I guess you can say that it should have been done "right" at that time,
> e.g. by already adding something like altmap support, instead of our own
> hack.

Probably yes. IMHO, relying on the existing memory block interface was 
the low hanging fruit. Now, s390x is just special.

> 
>>
>>>
>>> The new approach allows to better integrate s390 to the existing
>>> interface, and also make use of altmap support, which would eliminate
>>> the major disadvantage of the old behaviour. So from s390 perspective,
>>> this new mechanism would be preferred, provided that there is no
>>> functional issue with the "added memory blocks without a memmap"
>>> approach.
>>
>> It achieves that by s390x specific hacks in common code :) Instead of
>> everybody else that simply uses add_memory() and friends.
>>
>>>
>>> Do you see any functional issues, e.g. conflict with common
>>> code?
>>
>> I don't see functional issues right now, just the way it is done to
>> implement support for a new feature is a hack IMHO. Replacing hack #1 by
>> hack #2 is not really something reasonable. Let's try to remove hacks.
> 
> Ok, sounds reasonable, let's try that. Introducing some new s390-specific
> interface also feels a bit hacky, or ugly, but we'll see if we find a
> way so that it is only "special" :-)

As proposed in my other mail, I think there are ways to make s390x happy 
first and look into a cleaner approach long-term.

> Reminds me a bit of that "probe" attribute, that also was an arch-specific
> hack initially, IIRC, and is now to be deprecated...

Yeah, that was for interfaces where the kernel has absolutely no clue 
where/what/how memory gets hotplugged. ARM64 without ACPI.

s390x is completely different though: you know exactly which standby 
memory exists, where it resides, in which granularity in can be 
enabled/disabled, ... you don't have to play dangerous "I'm pretty sure 
there is memory out there although nobody can check so I crash the 
kernel" games.

-- 
Cheers,

David / dhildenb

Powered by blists - more mailing lists