Re: [PATCH 1/4] mm/gup: Add FOLL_INTERRUPTIBLE

From: John Hubbard
Date: Wed Jun 29 2022 - 21:53:46 EST


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