Message-ID: <0f7a6b69-5706-4010-ba7a-68a071922c80@redhat.com>
Date: Mon, 21 Oct 2024 23:20:03 +0200
From: David Hildenbrand <david@...hat.com>
To: Lorenzo Stoakes <lorenzo.stoakes@...cle.com>
Cc: Vlastimil Babka <vbabka@...e.cz>,
 Andrew Morton <akpm@...ux-foundation.org>,
 Suren Baghdasaryan <surenb@...gle.com>,
 "Liam R . Howlett" <Liam.Howlett@...cle.com>,
 Matthew Wilcox <willy@...radead.org>, "Paul E . McKenney"
 <paulmck@...nel.org>, Jann Horn <jannh@...gle.com>, linux-mm@...ck.org,
 linux-kernel@...r.kernel.org, Muchun Song <muchun.song@...ux.dev>,
 Richard Henderson <richard.henderson@...aro.org>,
 Ivan Kokshaysky <ink@...assic.park.msu.ru>, Matt Turner
 <mattst88@...il.com>, Thomas Bogendoerfer <tsbogend@...ha.franken.de>,
 "James E . J . Bottomley" <James.Bottomley@...senpartnership.com>,
 Helge Deller <deller@....de>, Chris Zankel <chris@...kel.net>,
 Max Filippov <jcmvbkbc@...il.com>, Arnd Bergmann <arnd@...db.de>,
 linux-alpha@...r.kernel.org, linux-mips@...r.kernel.org,
 linux-parisc@...r.kernel.org, linux-arch@...r.kernel.org,
 Shuah Khan <shuah@...nel.org>, Christian Brauner <brauner@...nel.org>,
 linux-kselftest@...r.kernel.org, Sidhartha Kumar
 <sidhartha.kumar@...cle.com>, Jeff Xu <jeffxu@...omium.org>,
 Christoph Hellwig <hch@...radead.org>, linux-api@...r.kernel.org,
 John Hubbard <jhubbard@...dia.com>
Subject: Re: [PATCH v2 3/5] mm: madvise: implement lightweight guard page
 mechanism

>> Yes, I see from Lorenzo's reply that there is apparently some history to
>> this (maybe it's all nicely summarized in the cover letter / this patch,
>> have to dig further).
>>
>> Not sure yet what the problem is, I would have thought it's all protected by
>> the PTL, and concurrent faults are user space doing something stupid and
>> we'd detect it.
> 
> The looping mechanism is fine for dealing with concurrent faults. There's
> no actual _race_ due to PTL, it's just that a user could repeatedly
> populate stuff stupidly in a range that is meant to have poison markers put
> in.
> 
> It's not likely and would be kind of an abusive of the interface, and it'd
> really be a process just hurting itself.
> 
> In nearly all cases you won't zap at all. The whole point is it's
> optimistic. In 99.99% of others you zap once...

Exactly! And that's why I am questioning whether the kernel should care 
about that. See below.

> 
>>
>> Have to do some more reading on this.
> 
> May I suggest a book on the history of the prodigy?

:D

> 
>>
>>>
>>> I'd normally agree with the KIS principle, but..
>>>
>>>> We can always implement support for that later if
>>>
>>> it would either mean later we change behavior (installing guards on
>>> non-zapped PTEs would have to be an error now but maybe start working later,
>>> which is user observable change thus can break somebody)
>>>
>>>> really required (leave behavior open when documenting).
>>>
>>> and leaving it open when documenting doesn't really mean anything for the
>>> "we don't break userspace" promise vs what the implementation actually does.
>>
>> Not quite I think. You could start return -EEXIST or -EOPNOTSUPP and
>> document that this can change in the future to succeed if there is
>> something. User space can sense support.
> 
> Yeah I mean originally I had a -EAGAIN which was sort of equivalent of this
> but Jann pointed out you're just shifting work to userland who would loop
> and repeat.
> 
> I just don't see why we'd do this.
> 
> In fact I was looking at the series and thinking 'wow it's actually a
> really small delta' and being proud but... still not KIS enough apparently
> ;)

You know, I read a lot of kernel code ... and mfill_atomic_install_pte() 
is what popped in my head: if there is already something, let user space 
handle it, because it is unexpected.

The uffd interface is slightly better, as it gives you the number of 
processed PTEs back, which madvise() is not designed for.

But maybe this (returning how many we already processed) is not required 
due to the nature of guard pages (below).

> 
>>
>> Something failing that at one point starts working is not really breaking
>> user space, unless someone really *wants* to fail if there is already
>> something (e.g., concurrent fault -> bail out instead of hiding it).
>>
>> Of course, a more elegant solution would be GUARD_INSTALL vs.
>> GUARD_FORCE_INSTALL.
>>
>> .. but again, there seems to be more history to this.
> 
> I don't think there's really any value in that. There's just no sensible
> situation in which a user would care about this I don't think.

Making sure nobody touches an area, and wile doing that somebody already 
touched that area? I guess it could be worked around by 
mprotect(PROT_NONE),madvise(GUARD),mprotect(PROT_READ|PROT_WRITE) ... 
which is not particularly nice :)

> 
> And if you're saying 'hey do MADV_DONTNEED if this fails and keep trying!'
> then why not just do that in the kernel?

Heh, no!

If user space doesn't expect there to be something, it should *fail*. 
That's likely going to be the majority of use cases for guard pages 
(happy to be told otherwise). No retry.

And if user space expects there to be something it should zap ahead of 
time (which some allocators maybe already do to free up memory after 
free()) to then install the guard. No retry.

There is this case where user space might be unsure. There, it might 
make sense to retry exactly once.

> 
> Trying to explain to a user 'hey this is for installing guard pages but if
> there's a facing fault it'll fail and that could keep happening and then
> you'll have to zap and maybe in a loop' just... seems like a bloody awful
> interface?

Hope my example above made it clearer. This "retry forever until it 
works" use case doesn't quite make sense to me, but I might just be 
missing something important.

But again, I have to do more reading on the history of the current 
approach ... and it's fairly late here.


-- 
Cheers,

David / dhildenb

Powered by blists - more mailing lists