Re: FOLL_LONGTERM vs FOLL_EPHEMERAL Re: [PATCH] tee: add FOLL_LONGTERM for CMA case when alloc shm

[John Hubbard <jhubbard@nvidia.com>]

On 5/18/23 06:56, David Hildenbrand wrote:
> On 18.05.23 08:08, Sumit Garg wrote:
>> On Thu, 18 May 2023 at 09:51, Christoph Hellwig <hch@infradead.org> wrote:
>>>
>>> On Wed, May 17, 2023 at 08:23:33PM +0200, David Hildenbrand wrote:
>>>> In general: if user space controls it -> possibly forever -> long-term. Even
>>>> if in most cases it's a short delay: there is no trusting on user space.
>>>>
>>>> For example, iouring fixed buffers keep pages pinned until user space
>>>> decides to unregistered the buffers -> long-term.
>>>>
>>>> Short-term is, for example, something like O_DIRECT where we pin -> DMA ->
>>>> unpin in essentially one operation.
>>>
>>> Btw, one thing that's been on my mind is that I think we got the
>>> polarity on FOLL_LONGTERM wrong.  Instead of opting into the long term
>>> behavior it really should be the default, with a FOLL_EPHEMERAL flag
>>> to opt out of it.  And every users of this flag is required to have
>>> a comment explaining the life time rules for the pin..

I see maybe 10 or 20 call sites today. So it is definitely feasible to add
documentation at each, explaining the why it wants a long term pin.

>>
>> It does look like a better approach to me given the very nature of
>> user space pages.
> 
> Yeah, there is a lot of historical baggage. For example, FOLL_GET should be inaccessible to kernel modules completely at one point, to be only used by selected core-mm pieces.

Yes. When I first mass-converted call sites from gup to pup, I just
preserved FOLL_GET behavior in order to keep from changing too much at
once. But I agree that that it would be nice to make FOLL_GET an
mm internal-only flag like FOLL_PIN.

> 
> Maybe we should even disallow passing in FOLL_LONGTERM as a flag and only provide functions like pin_user_pages() vs. pin_user_pages_longterm(). Then, discussions about conditional flag-setting are no more :)
> 
> ... or even use pin_user_pages_shortterm() vs. pin_user_pages() ... to make the default be longterm.
> 

Yes, it is true that having most gup flags be internal to mm does tend
to avoid some bugs. But it's also a lot of churn. I'm still on the fence
as to whether it's really a good move to do this for FOLL_LONGTERM or
not. But it's really easy to push me off of fences. :)

thanks,
-- 
John Hubbard
NVIDIA