[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <4ca4a5af-accb-4424-a5fb-e1e561900348@redhat.com>
Date: Mon, 27 Nov 2023 16:42:25 +0100
From: David Hildenbrand <david@...hat.com>
To: Ryan Roberts <ryan.roberts@....com>,
Barry Song <21cnbao@...il.com>,
Steven Price <steven.price@....com>
Cc: akpm@...ux-foundation.org, catalin.marinas@....com,
will@...nel.org, linux-kernel@...r.kernel.org, linux-mm@...ck.org,
mhocko@...e.com, shy828301@...il.com, v-songbaohua@...o.com,
wangkefeng.wang@...wei.com, willy@...radead.org, xiang@...nel.org,
ying.huang@...el.com, yuzhao@...gle.com
Subject: Re: [RFC V3 PATCH] arm64: mm: swap: save and restore mte tags for
large folios
>>> Are you referring to Steven's suggestion of reading the tag to see if it's
>>> zeros? I think that demonstrates my point that this has to be done per-page and
>>
>> Yes.
>
> OK I'm obviously being thick, because checking a page's tag to see if its zero
> seems logically equivalent to checking the existing per-page flag, just more
> expensive. Yes we could make that change but I don't see how it helps solve the
> real problem at hand. Unless you are also deliberately trying to remove the
> per-page flag at the same time, as per Matthew's master plan?
I think a per-folio interface is cleaner and more future-proof, and
removing the per-page flag might be a nice side product of that.
Anyhow, just some thoughts from my side if it could be easily/cleanly done.
At least the "easy" part does not seem to be the case, so I'm fine with
deferring anything like that for now.
[...]
>>
>> We can identify in 1) the subpages by reading the tag from HW,
>
> I don't think this actually works; I'm pretty sure the optimization to clear the
> tag at the same time as the page clearing only happens for small pages. I don't
> think this will be done when allocating a THP today. Obviously that could change.
>
Could be, absolutely no expert. I was primarily on the "It would be
possible to reverse this scheme - we could drop the page
flag and just look at the actual tag storage. If it's all zeros then
obviously there's no point in storing it." comment from Steven.
>> and on 2) by
>> checking the datastructure. For 3), there is nothing to check.
>>
>> On swapout of a large folio:
>>
>> * For 3) we don't do anything
>> * For 2) we don't do anything
>> * For 1) we store all tags that are non-zero (reading all tags) and
>> transition to 2).
>
> Given a tag architecturally exists for every page even when unused, and we think
> a folio being partially mte-tagged is the corner case, could you simplify this
> further and just write out all the tags for the folio and not care if some are
> not in use?
Likely this could be simplified, yes.
--
Cheers,
David / dhildenb
Powered by blists - more mailing lists