Re: [PATCH 6 of 7] x86: use early_ioremap in __acpi_map_table

[Jeremy Fitzhardinge <jeremy@goop.org>]

Ingo Molnar wrote:
> uhm, there's a nasty trap in that route: it can potentially cause a lot 
> of breakage.
>
> It's not robust to assume that the ACPI code is sane wrt. 
> mapping/unmapping, because it currently simply doesnt rely on robust 
> unmapping (in the linear range).
>   

You mean there's code which just assumes that it can keep using a
linear-mapped acpi even after __acpi_map_table() should have implicitly
unmapped it?
> I tried it in the past and i found tons of crappy ACPI code all around 
> that just never unmapped tables. Leaking ACPI maps are hard to find as 
> well, and it can occur anytime during bootup.
>   

__acpi_map_table() is called by acpi_map_table(), which does have a
acpi_unmap_table() counterpart.  But it only calls iounmap() once we're
past the stage of calling early_*().  I could easily make it call
__acpi_unmap_table()->early_iounmap().  But if the concern is that the
early boot callers of acpi_map_table() "know" that they never need to
unmap, then yes, I see the problem.

> As a general principle it might be worth fixing those places, and we've 
> hardened up the early-ioremap code for leaks during the PAT rewrite, 
> still please realize that it can become non-trivial and it might cause a 
> lot of unhappy users.
>
> So i'd suggest a different, more carful approach: keep the new code you 
> wrote, but print a WARN()ing if prev_map is not unmapped yet when the 
> next mapping is acquired. That way the ACPI code can be fixed gradually 
> and without breaking existing functionality.
>   

Yep.

> There's another complication: ACPI might rely on multiple mappings being 
> present at once, so unmapping the previous one might not be safe. But it 
> _should_ be fine most of the time as __acpi_map_table() is only used 
> inearly init code - and we fixed most of these things in the PAT 
> patchset in any case.

And the current behaviour of __acpi_map_table() is to remove the
previous mapping (implicitly, by overwriting the same fixmap slots), so
its only an issue if the callers assume they can keep using
linear-mapped acpi tables after a subsequent call to __acpi_map_table().

    J