Re: Suppress pte soft-dirty bit with UFFDIO_COPY?

On Tue, May 13, 2025 at 6:25 AM Peter Xu <[email protected]> wrote:
>
> On Mon, May 12, 2025 at 10:16:12AM -0700, Kyle Huey wrote:
> > Personally I don't think it's a real issue to have to create a
> > sacrificial fd once at process initialization to see what features are
> > available. I wouldn't have even said anything if the man page hadn't
> > explicitly told me there was another way.
>
> Yes, that's indeed the part that could be confusing and needs fixing.  Just
> to keep a record (I have you copied), I sent the man-pages changes here:
>
> https://lore.kernel.org/r/[email protected]

Agreed, at a high level I think this is the right fix. I believe I
just forgot the probing required a separate FD when I wrote that
version of the man page. :)

>
> We can stick with the sacrificial fd until there's a solid clue showing
> that we should introduce a new way to probe.

For what it's worth, I'm still convinced the whole handshake / probing
thing is overcomplicated, and it would be simpler to just do:

1. Userspace asks for the features it wants (UFFDIO_API)
2. Kernel responds (fills in the struct) with the (possibly subset) of
features it supports
3. Userspace can react as it sees fit if it gets a subset (fail with
error, gracefully degrade, ...)

But, based on previous discussion of that I believe I'm in the minority. :)

If we are sticking with the handshake approach, I agree needing a
second uffd is no big deal. We could add an ioctl to just probe
without configuring, but that would purely be for convenience, and I
don't think it saves many lines of code in userspace. So, on balance /
considering the small benefit I would probably prefer keeping the
kernel simpler.

>
> Thanks,
>
> --
> Peter Xu
>

On Tue, May 13, 2025 at 6:25 AM Peter Xu <[email protected]> wrote:
>
> On Mon, May 12, 2025 at 10:16:12AM -0700, Kyle Huey wrote:
> > Personally I don't think it's a real issue to have to create a
> > sacrificial fd once at process initialization to see what features are
> > available. I wouldn't have even said anything if the man page hadn't
> > explicitly told me there was another way.
>
> Yes, that's indeed the part that could be confusing and needs fixing.  Just
> to keep a record (I have you copied), I sent the man-pages changes here:
>
> https://lore.kernel.org/r/[email protected]
>
> We can stick with the sacrificial fd until there's a solid clue showing
> that we should introduce a new way to probe.
>
> Thanks,
>
> --
> Peter Xu
>