Re: [PATCH v2] mm: Fix memblock_free_late() when using deferred struct page

[Mike Rapoport <rppt@kernel.org>]

On Fri, Feb 20, 2026 at 09:46:50AM +1100, Benjamin Herrenschmidt wrote:
> On Thu, 2026-02-19 at 12:16 +0200, Mike Rapoport wrote:
> > 
> > Let's split it. EFI does weird things with memory already, like mremapping
> > normal memory for example.
> 
> Yup.
> 
> > Here's my take on the split. Lightly tested on qemu and recovered ~45M of
> > ram with the OVMF version I have :)
> 
> Nice :-) I'll test this here.
> 
> > > 
> > +struct efi_freeable_range {
> > +	u64 start;
> > +	u64 end;
> > +};
> > 
> 
> Haha, you went the blunt way :-) I was trying to avoid creating yet-
> another structure with "start/end" :-) 

Well, seems to me the easiest and the most efficient :)
I could have used "struct range", but I don't like it's semantics with
excluding the end. It would mean adding/subtracting 1 everywhere, seems
error prone to me.
 
> > +
> > +static struct efi_freeable_range *ranges_to_free;
> > +
> >  void __init efi_free_boot_services(void)
> >  {
> 
> I was going to call it efi_unmap_boot_services() to avoid having two
> things with almost the same name.

I wanted to minimize churn, but in the end it's not that much to change and
efi_unmap_boot_services() is a better name.
 
> >  	struct efi_memory_map_data data = { 0 };
> >  	efi_memory_desc_t *md;
> >  	int num_entries = 0;
> > +	int idx = 0;
> >  	void *new, *new_md;
> >  
> >  	/* Keep all regions for /sys/kernel/debug/efi */
> >  	if (efi_enabled(EFI_DBG))
> >  		return;
> >  
> > +	ranges_to_free = kzalloc(sizeof(*ranges_to_free) * efi.memmap.nr_map,
> > +				 GFP_KERNEL);
> > +	if (!ranges_to_free) {
> > +		pr_err("Failed to allocate storage for freeable EFI regions\n");
> > +		return;
> > +	}
> 
> Do we still want to do the whole unmap dance in that case ? I mean, OOM
> here means the system is pretty much a goner at that stage but ...

There is another potential OOM in that function. If it happens, we just
skip remapping and return. So return here is consistent :) 
 
> >  	for_each_efi_memory_desc(md) {
> >  		unsigned long long start = md->phys_addr;
> >  		unsigned long long size = md->num_pages << EFI_PAGE_SHIFT;
> > @@ -471,7 +486,15 @@ void __init efi_free_boot_services(void)
> >  			start = SZ_1M;
> >  		}
> >  
> > -		memblock_free_late(start, size);
> > +		/*
> > +		 * With CONFIG_DEFERRED_STRUCT_PAGE_INIT parts of the memory
> > +		 * map are still not initialized and we can't reliably free
> > +		 * memory here.
> > +		 * Queue the ranges to free at a later point.
> > +		 */
> > +		ranges_to_free[idx].start = start;
> > +		ranges_to_free[idx].end = start + size;
> > +		idx++;
> 
> Do we want to make this conditional to CONFIG_DEFERRED_STRUCT_PAGE_INIT
> or we don't care ?

I think it'll add ugliness for no good reason. If we want to keep systems
with CONFIG_DEFERRED_STRUCT_PAGE_INIT=n behave the same way as now, we need
several more if (CONFIG_DEFERRED_STRUCT_PAGE_INIT) and it becomes hairy.

And the change is quite small IMHO to just make it for everything.

> >  	}
> >  
> >  	if (!num_entries)
> > @@ -512,6 +535,23 @@ void __init efi_free_boot_services(void)
> >  	}
> >  }
> >  
> > +static int __init efi_free_boot_services_memory(void)
> > +{
> > +	struct efi_freeable_range *range = ranges_to_free;
> > +
> > +	while (range->start) {
> > +		void *start = phys_to_virt(range->start);
> > +		void *end = phys_to_virt(range->end);
> > +
> > +		free_reserved_area(start, end, -1, NULL);
> 
> I assume here too the total_ram_page_inc stuff is taken care of ? I
> haven't really looked. This feels like a fragile counter.

This is a fragile counter :)
free_reserved_area() -> free_reserved_page() take care of it.
 
> > +		range++;
> > +	}
> > +	kfree(ranges_to_free);
> > +
> > +	return 0;
> > +}
> > +late_initcall(efi_free_boot_services_memory);
> > +
> >  /*
> >   * A number of config table entries get remapped to virtual addresses
> >   * after entering EFI virtual mode. However, the kexec kernel requires
> > 
> > base-commit: 05f7e89ab9731565d8a62e3b5d1ec206485eeb0b
> > -- 
> > 2.51.0
> > 
> >  
> > > Cheers,
> > > Ben.
> > > 
> > 
> 

-- 
Sincerely yours,
Mike.