Message-ID: <17f9eae0-01bb-4793-201e-16ee267c07f2@nvidia.com>
Date:   Wed, 29 Jun 2022 18:53:30 -0700
From:   John Hubbard <jhubbard@...dia.com>
To:     Peter Xu <peterx@...hat.com>
Cc:     kvm@...r.kernel.org, linux-kernel@...r.kernel.org,
        Paolo Bonzini <pbonzini@...hat.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        David Hildenbrand <david@...hat.com>,
        "Dr . David Alan Gilbert" <dgilbert@...hat.com>,
        Andrea Arcangeli <aarcange@...hat.com>,
        Linux MM Mailing List <linux-mm@...ck.org>,
        Sean Christopherson <seanjc@...gle.com>
Subject: Re: [PATCH 1/4] mm/gup: Add FOLL_INTERRUPTIBLE

On 6/29/22 08:47, Peter Xu wrote:
>> It looks like part of this comment is trying to document a pre-existing
>> concept, which is that faultin_page() only ever sets FAULT_FLAG_KILLABLE
>> if locked != NULL.
> 
> I'd say that's not what I wanted to comment.. I wanted to express that
> INTERRUPTIBLE should rely on KILLABLE, that's also why I put the comment to
> be after KILLABLE, not before.  IMHO it makes sense already to have
> "interruptible" only if "killable", no matter what's the pre-requisite for
> KILLABLE (in this case it's having "locked" being non-null).
>

OK, I think I finally understand both the intention of the comment,
and (thanks to your notes, below) the interaction between *locked and
_RETRY, _KILLABLE, and _INTERRUPTIBLE. Really appreciate your leading
me by the nose through that. The pre-existing code is abusing *locked
a bit, by treating it as a flag when really it is a side effect of
flags, but at least now that's clear to me.

Anyway...this leads to finally getting into the comment, which I now
think is not quite what we want: there is no need for a hierarchy of
"_INTERRUPTIBLE should depend upon _KILLABLE". That is: even though an
application allows a fatal signal to get through, it's not clear to me
that that implies that non-fatal signal handling should be prevented.

The code is only vaguely enforcing such a thing, because it just so
happens that both cases require the same basic prerequisites. So the
code looks good, but I don't see a need to claim a hierarchy in the
comments.

So I'd either delete the comment entirely, or go with something that is
doesn't try to talk about hierarchy nor locked/retry either. Does this
look reasonable to you:


	/*
	 * FAULT_FLAG_INTERRUPTIBLE is opt-in: kernel callers must set
	 * FOLL_INTERRUPTIBLE. That's because some callers may not be
	 * prepared to handle early exits caused by non-fatal signals.
	 */

?

>> The problem I am (personally) having is that I don't yet understand why
>> or how those are connected: what is it about having locked non-NULL that
>> means the process is killable? (Can you explain why that is?)
> 
> Firstly RETRY_KILLABLE relies on ALLOW_RETRY, because if we don't allow
> retry at all it means we'll never wait in handle_mm_fault() anyway, then no
> need to worry on being interrupted by any kind of signal (fatal or not).
> 
> Then if we allow retry, we need some way to know "whether mmap_sem is
> released or not" during the process for the caller (because the caller
> cannot see VM_FAULT_RETRY).  That's why we added "locked" parameter, so
> that we can set *locked=false to tell the caller we have released mmap_sem.
> 
> I think that's why we have "locked" defined as "we allow this page fault
> request to retry and wait, during wait we can always allow fatal signals".
> I think that's defined throughout the gup call interfaces too, and
> faultin_page() is the last step to talk to handle_mm_fault().
> 
> To make this whole picture complete, NOWAIT is another thing that relies on
> ALLOW_RETRY but just to tell "oh please never release the mmap_sem at all".
> For example, when we want to make sure no vma will be released after
> faultin_page() returned.
> 

Again, thanks for taking the time to explain that for me. :)

>>
>> If that were clear, I think I could suggest a good comment wording.
> 
> IMHO it's a little bit weird to explain "locked" here, especially after
> KILLABLE is set, that's why I didn't try to mention "locked" in my 2nd
> attempt.  There are some comments for "locked" above the definition of
> faultin_page(), I think that'll be a nicer place to enrich explanations for
> "locked", and it seems even more suitable as a separate patch?
> 

Totally agreed. I didn't intend to ask for that kind of documentation
here.

For that, I'm thinking a combination of cleaning up *locked a little
bit, plus maybe some higher level notes like what you wrote above, added
to either pin_user_pages.rst or a new get_user_pages.rst or some .rst
anyway. Definitely a separately thing.


thanks,
-- 
John Hubbard
NVIDIA

Powered by blists - more mailing lists