Date:   Mon, 11 Jun 2018 13:53:49 +0200
From:   David Hildenbrand <david@...hat.com>
To:     Michal Hocko <mhocko@...nel.org>
Cc:     linux-mm@...ck.org, linux-kernel@...r.kernel.org,
        Alexander Potapenko <glider@...gle.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Andrey Ryabinin <aryabinin@...tuozzo.com>,
        Balbir Singh <bsingharora@...il.com>,
        Baoquan He <bhe@...hat.com>,
        Benjamin Herrenschmidt <benh@...nel.crashing.org>,
        Boris Ostrovsky <boris.ostrovsky@...cle.com>,
        Dan Williams <dan.j.williams@...el.com>,
        Dave Young <dyoung@...hat.com>,
        Dmitry Vyukov <dvyukov@...gle.com>,
        Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
        Hari Bathini <hbathini@...ux.vnet.ibm.com>,
        Huang Ying <ying.huang@...el.com>,
        Hugh Dickins <hughd@...gle.com>,
        Ingo Molnar <mingo@...nel.org>,
        Jaewon Kim <jaewon31.kim@...sung.com>, Jan Kara <jack@...e.cz>,
        Jérôme Glisse <jglisse@...hat.com>,
        Joonsoo Kim <iamjoonsoo.kim@....com>,
        Juergen Gross <jgross@...e.com>,
        Kate Stewart <kstewart@...uxfoundation.org>,
        "Kirill A. Shutemov" <kirill.shutemov@...ux.intel.com>,
        Matthew Wilcox <mawilcox@...rosoft.com>,
        Mel Gorman <mgorman@...e.de>,
        Michael Ellerman <mpe@...erman.id.au>,
        Miles Chen <miles.chen@...iatek.com>,
        Oscar Salvador <osalvador@...hadventures.net>,
        Paul Mackerras <paulus@...ba.org>,
        Pavel Tatashin <pasha.tatashin@...cle.com>,
        Philippe Ombredanne <pombredanne@...b.com>,
        Rashmica Gupta <rashmica.g@...il.com>,
        Reza Arbab <arbab@...ux.vnet.ibm.com>,
        Souptick Joarder <jrdr.linux@...il.com>,
        Tetsuo Handa <penguin-kernel@...ove.SAKURA.ne.jp>,
        Thomas Gleixner <tglx@...utronix.de>,
        Vlastimil Babka <vbabka@...e.cz>
Subject: Re: [PATCH v1 00/10] mm: online/offline 4MB chunks controlled by
 device driver

On 24.05.2018 23:07, David Hildenbrand wrote:
> On 24.05.2018 16:22, Michal Hocko wrote:
>> I will go over the rest of the email later I just wanted to make this
>> point clear because I suspect we are talking past each other.
> 
> It sounds like we are now talking about how to solve the problem. I like
> that :)
> 

Hi Michal,

did you have time to think about the details of your proposed idea?
(especially the questions I had as response below to make it work at all?)

Personally, I still think that using Pg_reserved is wrong and that your
proposal will be significantly more complicated.

Thanks!

>>
>> On Thu 24-05-18 16:04:38, David Hildenbrand wrote:
>> [...]
>>> The point I was making is: I cannot allocate 8MB/128MB using the buddy
>>> allocator. All I want to do is manage the memory a virtio-mem device
>>> provides as flexible as possible.
>>
>> I didn't mean to use the page allocator to isolate pages from it. We do
>> have other means. Have a look at the page isolation framework and have a
>> look how the current memory hotplug (ab)uses it. In short you mark the
>> desired physical memory range as isolated (nobody can allocate from it)
>> and then simply remove it from the page allocator. And you are done with
>> it. Your particular range is gone, nobody will ever use it. If you mark
>> those struct pages reserved then pfn walkers should already ignore them.
>> If you keep those pages with ref count 0 then even hotplug should work
>> seemlessly (I would have to double check).
>>
>> So all I am arguing is that whatever your driver wants to do can be
>> handled without touching the hotplug code much. You would still need
>> to add new ranges in the mem section units and manage on top of that.
>> You need to do that anyway to keep track of what parts are in use or
>> offlined anyway right? Now the mem sections. You have to do that anyway
>> for memmaps. Our sparse memory model simply works in those units. Even
>> if you make a part of that range unavailable then the section will still
>> be there.
>>
>> Do I make at least some sense or I am completely missing your point?
>>
> 
> I think we're heading somewhere. I understand that you want to separate
> this "semi" offline part from the general offlining code. If so, we
> should definitely enforce segment alignment for online_pages/offline_pages.
> 
> Importantly, what I need is:
> 
> 1. Indicate and prepare memory sections to be used for adding memory
>    chunks (right now add_memory())
> 2. Make memory chunks of a section available to the system (right now
>    online_pages())
> 3. Remove memory chunks of a section from the system (right now
>    offline_pages())
> 4. Remove memory sections from the system (right now remove_memory())
> 5. Hinder dumping tools from reading memory chunks that are logically
>    offline (right now PageOffline())
> 6. For 3. find removable memory chunks in a certain memory range with a
>    variable size.
> 
> In an ideal world, 2. would never fail (in contrast to online_pages()
> right now). This might make some further developments I have in mind
> easier :) So if we can come up with an approach that can guarantee that,
> extra points.
> 
> So what I think you are talking about is the following.
> 
> For 1. Use add_memory() followed by online_pages(). Don't actually
>        online the pages, keep them reserved (like XEN balloon). Fixup
>        stats.
> For 2. Expose reserved pages to Buddy allocator. Clear reserved bit.
>        Fixup stats. This can never fail. (yay)
> For 3. Isolate pages, try to move everything away (basically but not
>        comletely offlining code). Set reserved flag. Fixup flags.
> For 4. offline_pages() followed by remove_memory().
>        -> Q: How to distinguish reserved offline from other reserved
>              pages? offline_pages() has to be able to deal with that
> For 5. I don't think we can use reserved flag here.
>        -> Q: What else to use?
> For 6. Scan for movable ranges. The use
> 
> 
> "You need to do that anyway to keep track of what parts are in use or
>  offlined anyway right?"
> 
> I would manually track which chunks of a section is logically offline (I
> do that right now already).
> 
> Is that what you had in mind? If not, where does your idea differ.
> How could we solve 4/5. Of course, PageOffline() is again an option.
> 
> Thanks!
> 


-- 

Thanks,

David / dhildenb

Powered by blists - more mailing lists