Re: [PATCH 0/6] Boot-time switching between 4- and 5-level paging for 4.15, Part 1

From: Kirill A. Shutemov
Date: Tue Oct 31 2017 - 08:04:39 EST


On Tue, Oct 31, 2017 at 10:47:27AM +0100, Ingo Molnar wrote:
>
> * Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> wrote:
> > I don't think this design is reasonable.
> >
> > - It introduces memory references where we haven't had them before.
> >
> > At this point all variable would fit a cache line, which is not that
> > bad. But I don't see what would stop the list from growing in the
> > future.
>
> Is any of these actually in a hotpath?

Probably, no. Closest to hotpath I see so far is page_zone_id() in page
allocator.

> Also, note the context: your changes turn some of these into variables. Yes, I
> suggest structuring them all and turning them all into variables, exactly because
> the majority are now dynamic, yet their _naming_ suggests that they are constants.

Another way to put it would be that you suggest significant rework of kernel
machinery based on cosmetic nitpick. :)

> > - We loose ability to optimize out change with static branches
> > (cpu_feature_enabled() instead of pgtable_l5_enabled variable).
> >
> > It's probably, not that big of an issue here, but if we are going to
> > use the same approach for other dynamic macros in the patchset, it
> > might be.
>
> Here too I think the (vast) majority of the uses here are for bootup/setup/init
> purposes, where clarity and maintainability of code matters a lot.

I would argue that it makes maintainability worse.

It makes dependencies between values less obvious. For instance, checking
MAXMEM definition on x86-64 makes it obvious that it depends directly
on MAX_PHYSMEM_BITS.

If we would convert MAXMEM to variable, we would need to check where the
variable is initialized and make sure that nobody changes it afterwards.

Does it sound like a win for maintainability?

> > - AFAICS, it requires changes to all architectures to provide such
> > structures as we now partly in generic code.
> >
> > Or to introduce some kind of compatibility layer, but it would make
> > the kernel as a whole uglier than cleaner. Especially, given that
> > nobody beyond x86 need this.
>
> Yes, all the uses should be harmonized (no compatibility layer) - but as you can
> see it from the histogram I generated it's a few dozen uses, i.e. not too bad.

Without a compatibility layer, I would need to change every architecture.
It's few dozen patches easily. Not fun.

---------------------------------8<------------------------------------

Putting, my disagreement with the design aside, I try to prototype it.
And stumble an issue that I don't see how to solve.

If we are going to convert macros to variable whether they need to be
variable in the configuration we quickly put ourself into corner:

- SECTIONS_SHIFT is dependent on MAX_PHYSMEM_BITS.

- SECTIONS_SHIFT is used to define SECTIONS_WIDTH, but only if
CONFIG_SPARSEMEM_VMEMMAP is not enabled. SECTIONS_WIDTH is zero
otherwise.

At this point we can convert both SECTIONS_SHIFT and SECTIONS_WIDTH to
variables.

But SECTIONS_WIDTH used on preprocessor level to determinate NODES_WIDTH,
which used to determinate if we going to define NODE_NOT_IN_PAGE_FLAGS and
the value of LAST_CPUPID_WIDTH.

Making SECTIONS_WIDTH variable breaks the preprocessor logic. But problems
don't stop there:

- LAST_CPUPID_WIDTH determinate if LAST_CPUPID_NOT_IN_PAGE_FLAGS is defined.

- LAST_CPUPID_NOT_IN_PAGE_FLAGS is used define struct page and therefore
cannot be dynamic (read variable).


In my patchset I made X86_5LEVEL select SPARSEMEM_VMEMMAP. It breaks the
chain and SECTIONS_WIDTH is never dynamic.

But how get it work with the design?

I can only think of hack like making machine.physmem.sections.shift a
constant macro if we don't want it dynamic for the configuration and leave
SECTHION_WITH as a constant in generic code.

To me it's ugly as hell.

--
Kirill A. Shutemov