Re: [PATCH v7] RISC-V: enable XIP

[Vitaly Wool]

On Mon, Apr 12, 2021 at 7:12 AM Alex Ghiti <alex@xxxxxxxx> wrote:
>
> Le 4/9/21 à 10:42 AM, Vitaly Wool a écrit :
> > On Fri, Apr 9, 2021 at 3:59 PM Mike Rapoport <rppt@xxxxxxxxxxxxx> wrote:
> >>
> >> On Fri, Apr 09, 2021 at 02:46:17PM +0200, David Hildenbrand wrote:
> >>>>>> Also, will that memory properly be exposed in the resource tree as
> >>>>>> System RAM (e.g., /proc/iomem) ? Otherwise some things (/proc/kcore)
> >>>>>> won't work as expected - the kernel won't be included in a dump.
> >>>> Do we really need a XIP kernel to included in kdump?
> >>>> And does not it sound weird to expose flash as System RAM in /proc/iomem? ;-)
> >>>
> >>> See my other mail, maybe we actually want something different.
> >>>
> >>>>
> >>>>> I have just checked and it does not appear in /proc/iomem.
> >>>>>
> >>>>> Ok your conclusion would be to have struct page, I'm going to implement this
> >>>>> version then using memblock as you described.
> >>>>
> >>>> I'm not sure this is required. With XIP kernel text never gets into RAM, so
> >>>> it does not seem to require struct page.
> >>>>
> >>>> XIP by definition has some limitations relatively to "normal" operation,
> >>>> so lack of kdump could be one of them.
> >>>
> >>> I agree.
> >>>
> >>>>
> >>>> I might be wrong, but IMHO, artificially creating a memory map for part of
> >>>> flash would cause more problems in the long run.
> >>>
> >>> Can you elaborate?
> >>
> >> Nothing particular, just a gut feeling. Usually, when you force something
> >> it comes out the wrong way later.
> >
> > It's possible still that MTD_XIP is implemented allowing to write to
> > the flash used for XIP. While flash is being written, memory map
> > doesn't make sense at all. I can't come up with a real life example
> > when it can actually lead to problems but it is indeed weird when
> > System RAM suddenly becomes unreadable. I really don't think exposing
> > it in /proc/iomem is a good idea.
> >
> >>>> BTW, how does XIP account the kernel text on other architectures that
> >>>> implement it?
> >>>
> >>> Interesting point, I thought XIP would be something new on RISC-V (well, at
> >>> least to me :) ). If that concept exists already, we better mimic what
> >>> existing implementations do.
> >>
> >> I had quick glance at ARM, it seems that kernel text does not have memory
> >> map and does not show up in System RAM.
> >
> > Exactly, and I believe ARM64 won't do that too when it gets its own
> > XIP support (which is underway).
> >
>
>
> memmap does not seem necessary and ARM/ARM64 do not use it.
>
> But if someone tries to get a struct page from a physical address that
> lies in flash, as mentioned by David, that could lead to silent
> corruptions if something exists at the address where the struct page
> should be. And it is hard to know which features in the kernel depends
> on that.
>
> Regarding SPARSEMEM, the vmemmap lies in its own region so that's
> unlikely to happen, so we will catch those invalid accesses (and that's
> what I observed on riscv).
>
> But for FLATMEM, memmap is in the linear mapping, then that could very
> likely happen silently.
>
> Could a simple solution be to force SPARSEMEM for those XIP kernels ?
> Then wrong things could happen, but we would see those and avoid
> spending hours to debug :)
>
> I will at least send a v8 to remove the pfn_valid modifications for
> FLATMEM that now returns true to pfn in flash.

That sounds good to me. I am not very keen on spending 200K on struct
pages for flash (we can think of this as of an option but I would
definitely like to have the option to compile it out in the end), so
let's remove pfn_valid and fix things that will eventually break, if
some.

Best regards,
   Vitaly