Re: [RFC PATCH 1/2] kexec: do not place the next kernel in a fixed-address CMA area

[Pandey, Radhey Shyam]

On 10/30/2026 5:33 PM, Mike Rapoport wrote:
> On Fri, Oct 30, 2026 at 01:22:50PM +0530, Pandey, Radhey Shyam wrote:
> > On 10/29/2026 6:41 PM, Pratyush Yadav wrote:
> > > On Thu, Oct 29 2026, Mike Rapoport wrote:
> > > > On Wed, Oct 28, 2026 at 03:03:16PM +0100, Pratyush Yadav wrote:
> > > > > On Wed, Oct 14 2026, Radhey Shyam Pandey wrote:
> > > > > > @@ -304,6 +317,7 @@ int __init cma_init_reserved_mem(phys_addr_t base, phys_addr_t size,
> > > > > >    	cma->ranges[0].count = cma->count;
> > > > > >    	cma->nranges = 1;
> > > > > >    	cma->nid = NUMA_NO_NODE;
> > > > > > +	set_bit(CMA_FIXED_BASE, &cma->flags);
> > > > > >    	*res_cma = cma;
> > > > > > @@ -517,6 +531,8 @@ static int __init __cma_declare_contiguous_nid(phys_addr_t *basep,
> > > > > >    	}
> > > > > >    	(*res_cma)->nid = nid;
> > > > > > +	if (!fixed)
> > > > > > +		clear_bit(CMA_FIXED_BASE, &(*res_cma)->flags);
> > > > >
> > > > > Nit: This reads a bit odd. cma_init_reserved_mem() sets the fixed base
> > > > > and then you clear it later. Wouldn't it be better to just pass the
> > > > > fixed boolean to cma_init_reserved_mem()?
> > > >
> > > > Actually, we shouldn't expose cma_init_reserved_mem() at all.
> > > >
> > > > The current callers of cma_init_reserved_mem() should be switched to
> > > > cma_declare_contiguous_nid() that does sanity checks before calling
> > > > cma_init_reserved_mem().
> >
> > Thanks Pratyush and Mike for the review.
> >
> > The switch from cma_init_reserved_mem() to cma_declare_contiguous_nid()
> > looks like a broader CMA refactoring and requires additional changes to work
> > correctly.
>
> Hmm, you are right.
>
> How about we invert the flag meaning?
> Than cma_declare_contiguous_* will set it where appropriate and we won't
> have the weirdness Pratyush pointed out.

That's a good point. I will invert the flag meaning - something like.

__cma_declare_contiguous_nid()

<snip>
+       if (!fixed)
+               set_bit(CMA_DYNAMIC_BASE, &(*res_cma)->flags);


So the v2 series would look like:
mm/cma: drop the duplicate alignment check in rmem_cma_setup()
of: reserved_mem: keep the reservation when node_init returns -EBUSY
mm/cma: reject CMA areas that overlap the kernel image
kexec: do not place the next kernel in a fixed-address CMA area

> And Radhey, please, make your LLM to spill more concise and to the point
> text and teach it to trim emails ;-)

Noted. Will take care of it in next version.

Thanks,
Radhey