Re: [PATCH] mm: move mask update out of the atomic context

[Alexander Gordeev <agordeev@linux.ibm.com>]

On Tue, Jun 24, 2025 at 11:40:05AM +0200, David Hildenbrand wrote:
> On 24.06.25 11:37, Alexander Gordeev wrote:
> > On Mon, Jun 23, 2025 at 09:45:34PM +0200, David Hildenbrand wrote:
> > ...
> > > Let's ask the real questions: who checks PGTBL_PTE_MODIFIED?
> > > 
> > > I see
> > > 
> > > if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
> > > 	arch_sync_kernel_mappings(start, start + size);
> > > 
> > > And then
> > > 
> > > arch/arm/include/asm/page.h:#define ARCH_PAGE_TABLE_SYNC_MASK   PGTBL_PMD_MODIFIED
> > > arch/x86/include/asm/pgtable-2level_types.h:#define ARCH_PAGE_TABLE_SYNC_MASK   PGTBL_PMD_MODIFIED
> > > arch/x86/include/asm/pgtable-3level_types.h:#define ARCH_PAGE_TABLE_SYNC_MASK   PGTBL_PMD_MODIFIED
> > > 
> > > 
> > > Which makes me wonder why we need PGTBL_PTE_MODIFIED at all? Is there some other check I am missing?
> > > 
> > > (same question regarding everything excepy PGTBL_PMD_MODIFIED, because that actually seems to be used)
> > 
> > AFAICT it was thought as architecture-specific:
> > 
> > /*
> >   * Architectures can set this mask to a combination of PGTBL_P?D_MODIFIED values
> >   * and let generic vmalloc and ioremap code know when arch_sync_kernel_mappings()
> >   * needs to be called.
> >   */
> > #ifndef ARCH_PAGE_TABLE_SYNC_MASK
> > #define ARCH_PAGE_TABLE_SYNC_MASK 0
> > #endif
> > 
> > Not sure if that needs to be addressed at all.
> 
> Okay, if there are no users of PGTBL_PTE_MODIFIED we could just ... remove
> it. Dead code.

As you noticed, PGTBL_PMD_MODIFIED bit is used only. Thus, all other
bits would have to be removed as well, not just PGTBL_PTE_MODIFIED?

That is more or less revert of at least the below commits and rewriting
it in a PMD-focused manner:

2ba3e6947aed ("mm/vmalloc: track which page-table levels were modified")
d8626138009b ("mm: add functions to track page directory modifications")

That would be a completely different effort, which I am not aming at ;)

> -- 
> Cheers,
> 
> David / dhildenb

Thanks!