Re: linux-next: boot warning after merge of the vfs-brauner tree

[Mike Rapoport <rppt@kernel.org>]

On Mon, Aug 26, 2024 at 01:52:54PM -0700, Luis Chamberlain wrote:
> On Mon, Aug 26, 2024 at 07:43:20PM +0200, Christophe Leroy wrote:
> > 
> > 
> > Le 26/08/2024 à 17:48, Pankaj Raghav (Samsung) a écrit :
> > > On Mon, Aug 26, 2024 at 05:59:31PM +1000, Stephen Rothwell wrote:
> > > > Hi all,
> > > > 
> > > > After merging the vfs-brauner tree, today's linux-next boot test (powerpc
> > > > pseries_le_defconfig) produced this warning:
> > > 
> > > iomap dio calls set_memory_ro() on the page that is used for sub block
> > > zeroing.
> > > 
> > > But looking at powerpc code, they don't support set_memory_ro() for
> > > memory region that belongs to the kernel(LINEAR_MAP_REGION_ID).
> > > 
> > > /*
> > >   * On hash, the linear mapping is not in the Linux page table so
> > >   * apply_to_existing_page_range() will have no effect. If in the future
> > >   * the set_memory_* functions are used on the linear map this will need
> > >   * to be updated.
> > >   */
> > > if (!radix_enabled()) {
> > >          int region = get_region_id(addr);
> > > 
> > >          if (WARN_ON_ONCE(region != VMALLOC_REGION_ID && region != IO_REGION_ID))
> > >                  return -EINVAL;
> > > }
> > > 
> > > We call set_memory_ro() on the zero page as a extra security measure.
> > > I don't know much about powerpc, but looking at the comment, is it just
> > > adding the following to support it in powerpc:
> > > 
> > > diff --git a/arch/powerpc/mm/pageattr.c b/arch/powerpc/mm/pageattr.c
> > > index ac22bf28086fa..e6e0b40ba6db4 100644
> > > --- a/arch/powerpc/mm/pageattr.c
> > > +++ b/arch/powerpc/mm/pageattr.c
> > > @@ -94,7 +94,9 @@ int change_memory_attr(unsigned long addr, int numpages, long action)
> > >          if (!radix_enabled()) {
> > >                  int region = get_region_id(addr);
> > > -               if (WARN_ON_ONCE(region != VMALLOC_REGION_ID && region != IO_REGION_ID))
> > > +               if (WARN_ON_ONCE(region != VMALLOC_REGION_ID &&
> > > +                                region != IO_REGION_ID &&
> > > +                                region != LINEAR_MAP_REGION_ID))
> > >                          return -EINVAL;
> > >          }
> > >   #endif
> > 
> > By doing this you will just hide the fact that it didn't work.
> > 
> > See commit 1f9ad21c3b38 ("powerpc/mm: Implement set_memory() routines") for
> > details. The linear memory region is not mapped using page tables so
> > set_memory_ro() will have no effect on it.
> > 
> > You can either use vmalloc'ed pages, or do a const static allocation at
> > buildtime so that it will be allocated in the kernel static rodata area.
> > 
> > By the way, your code should check the value returned by set_memory_ro(),
> > there is some work in progress to make it mandatory, see
> > https://github.com/KSPP/linux/issues/7
> 
> Our users expect contiguous memory [0] and so we use alloc_pages() here,
> so if we're architecture limitted by this I'd rather we just remove the
> set_memory_ro() only for PPC, I don't see why other have to skip this.
> 
> diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
> index c02b266bba52..aba5cde89e14 100644
> --- a/fs/iomap/direct-io.c
> +++ b/fs/iomap/direct-io.c
> @@ -775,14 +775,22 @@ EXPORT_SYMBOL_GPL(iomap_dio_rw);
>  
>  static int __init iomap_dio_init(void)
>  {
> +	int ret;
> +
>  	zero_page = alloc_pages(GFP_KERNEL | __GFP_ZERO,
>  				IOMAP_ZERO_PAGE_ORDER);
>  
>  	if (!zero_page)
>  		return -ENOMEM;
>  
> -	set_memory_ro((unsigned long)page_address(zero_page),
> -		      1U << IOMAP_ZERO_PAGE_ORDER);
> -	return 0;
> +	if (IS_ENABLED(CONFIG_PPC))
> +		return 0;
> +
> +	ret = set_memory_ro((unsigned long)page_address(zero_page),
> +			    1U << IOMAP_ZERO_PAGE_ORDER);
> +	if (ret)
> +		free_pages((unsigned long) zero_page, IOMAP_ZERO_PAGE_ORDER);

arm64 will return -EINVAL here, their code for changing memory attributes
only works on vmalloc:

	 * Let's restrict ourselves to mappings created by vmalloc (or vmap).
	 * Those are guaranteed to consist entirely of page mappings, and
	 * splitting is never needed.


> +
> +	return ret;
>  }
>  fs_initcall(iomap_dio_init);
> 
> Thoughts?
> 
> [0] https://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs.git/commit/?h=vfs.blocksize&id=d940b3b7b76b409b0550fdf2de6dc2183f01526f
> 
>   Luis

-- 
Sincerely yours,
Mike.