Re: [PATCH 3/3] memblock: cleanup memblock_free interface

[Mike Rapoport]

On Thu, Sep 23, 2021 at 11:47:48AM +0200, Christophe Leroy wrote:
> 
> 
> Le 23/09/2021 à 09:43, Mike Rapoport a écrit :
> > From: Mike Rapoport <rppt@xxxxxxxxxxxxx>
> > 
> > For ages memblock_free() interface dealt with physical addresses even
> > despite the existence of memblock_alloc_xx() functions that return a
> > virtual pointer.
> > 
> > Introduce memblock_phys_free() for freeing physical ranges and repurpose
> > memblock_free() to free virtual pointers to make the following pairing
> > abundantly clear:
> > 
> > 	int memblock_phys_free(phys_addr_t base, phys_addr_t size);
> > 	phys_addr_t memblock_phys_alloc(phys_addr_t base, phys_addr_t size);
> > 
> > 	void *memblock_alloc(phys_addr_t size, phys_addr_t align);
> > 	void memblock_free(void *ptr, size_t size);
> > 
> > Replace intermediate memblock_free_ptr() with memblock_free() and drop
> > unnecessary aliases memblock_free_early() and memblock_free_early_nid().
> > 
> > Suggested-by: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
> > Signed-off-by: Mike Rapoport <rppt@xxxxxxxxxxxxx>
> > ---
> 
> > diff --git a/arch/s390/kernel/smp.c b/arch/s390/kernel/smp.c
> > index 1a04e5bdf655..37826d8c4f74 100644
> > --- a/arch/s390/kernel/smp.c
> > +++ b/arch/s390/kernel/smp.c
> > @@ -723,7 +723,7 @@ void __init smp_save_dump_cpus(void)
> >   			/* Get the CPU registers */
> >   			smp_save_cpu_regs(sa, addr, is_boot_cpu, page);
> >   	}
> > -	memblock_free(page, PAGE_SIZE);
> > +	memblock_phys_free(page, PAGE_SIZE);
> >   	diag_amode31_ops.diag308_reset();
> >   	pcpu_set_smt(0);
> >   }
> > @@ -880,7 +880,7 @@ void __init smp_detect_cpus(void)
> >   	/* Add CPUs present at boot */
> >   	__smp_rescan_cpus(info, true);
> > -	memblock_free_early((unsigned long)info, sizeof(*info));
> > +	memblock_free(info, sizeof(*info));
> >   }
> >   /*
> 
> I'm a bit lost. IIUC memblock_free_early() and memblock_free() where
> identical.

Yes, they were, but all calls to memblock_free_early() were using
__pa(vaddr) because they had a virtual address at hand.

> In the first hunk memblock_free() gets replaced by memblock_phys_free()
> In the second hunk memblock_free_early() gets replaced by memblock_free()

In the first hunk the memory is allocated with memblock_phys_alloc() and we
have a physical range to free. In the second hunk the memory is allocated
with memblock_alloc() and we are freeing a virtual pointer.
 
> I think it would be easier to follow if you could split it in several
> patches:

It was an explicit request from Linus to make it a single commit:

  but the actual commit can and should be just a single commit that just
  fixes 'memblock_free()' to have sane interfaces.

I don't feel strongly about splitting it (except my laziness really
objects), but I don't think doing the conversion in several steps worth the
churn.

> - First patch: Create memblock_phys_free() and change all relevant
> memblock_free() to memblock_phys_free() - Or change memblock_free() to
> memblock_phys_free() and make memblock_free() an alias of it.
> - Second patch: Make memblock_free_ptr() become memblock_free() and change
> all remaining callers to the new semantics (IIUC memblock_free(__pa(ptr))
> becomes memblock_free(ptr) and make memblock_free_ptr() an alias of
> memblock_free()
> - Fourth patch: Replace and drop memblock_free_ptr()
> - Fifth patch: Drop memblock_free_early() and memblock_free_early_nid() (All
> users should have been upgraded to memblock_free_phys() in patch 1 or
> memblock_free() in patch 2)
> 
> Christophe

-- 
Sincerely yours,
Mike.