Re: [PATCH] arm64: Change elfcore for_each_mte_vma() to use VMA iterator

[Catalin Marinas]

On Tue, Feb 22, 2022 at 06:54:38PM +0000, Liam Howlett wrote:
> * Catalin Marinas <catalin.marinas@xxxxxxx> [220222 12:26]:
> > On Tue, Feb 22, 2022 at 04:20:16PM +0000, Will Deacon wrote:
> > > On Tue, Feb 22, 2022 at 02:26:03PM +0000, Liam Howlett wrote:
> > > > The vma iterator uses the maple tree, so this patch would resolve the
> > > > conflict but both branches are needed.
> > > 
> > > I'm not really sure what to do here, then. I think the conflict is nasty
> > > enough that we should resolve it before the trees reach Linus, but there
> > > doesn't seem to be a way forward other than one of us merging the other
> > > branch. I'd like to avoid having MTE coredump support depend on the maple
> > > tree work.
> > > 
> > > Is there some way you could provide a branch which implements
> > > for_each_vma() using the old vma list, and then the maple tree series
> > > could switch that over to the maple tree without breaking things?
> > 
> > Without a branch, we could apply something like below on top of Liam's
> > patch and revert it once the maple tree is upstream:
> > 
> > diff --git a/arch/arm64/kernel/elfcore.c b/arch/arm64/kernel/elfcore.c
> > index 930a0bc4cac4..400ec7a902df 100644
> > --- a/arch/arm64/kernel/elfcore.c
> > +++ b/arch/arm64/kernel/elfcore.c
> > @@ -8,6 +8,13 @@
> >  #include <asm/cpufeature.h>
> >  #include <asm/mte.h>
> > 
> > +#ifndef VMA_ITERATOR
> > +#define VMA_ITERATOR(name, mm, addr) \
> > +	struct mm_struct *name = mm
> > +#define for_each_vma(vmi, vma) \
> > +	for (vma = vmi->mmap; vma; vma = vma->vm_next)
> > +#endif
> > +
> >  #define for_each_mte_vma(vmi, vma)					\
> >  	if (system_supports_mte())					\
> >  		for_each_vma(vmi, vma)					\
> 
> Note that the current VMA_ITERATOR takes a new type and not the mm.

Well, in you proposed fix, it does take current->mm.

> Since I am removing the linked list (mm->mmap and vma->vm_next), then
> the build will fail if this patch and the maple tree branch exist
> together.  The iterator may also not start at the start of the list (but
> usually does) and may not run through the entire list; see
> vma_for_each_range() in the patch set.

My hack above is only temporary to allow building the arm64 tree on its
own (no maple tree branch) and with your patch on top. In -next, when
merged with the maple tree branch, the VMA_ITERATOR macro is already
defined and the above hack skipped. We'll revert this hack around -rc1.

Note that the hack above is only in the arm64 elfcore.c, not a generic
API solution.

-- 
Catalin