Re: [PATCH 1/4] mm: Export flush_vm_area() to sync the PTEs upon construction

[Chris Wilson]

Quoting Joerg Roedel (2020-08-21 11:22:04)
> On Fri, Aug 21, 2020 at 10:54:22AM +0100, Chris Wilson wrote:
> > Ok. I thought it had to be after assigning the *ptep. If we apply the
> > sync first, do not have to worry about PGTBL_PTE_MODIFIED from the
> > *ptep?
> 
> Hmm, if I understand the code correctly, you are re-implementing some
> generic ioremap/vmalloc mapping logic in the i915 driver. I don't know
> the reason, but if it is valid you need to manually call
> arch_sync_kernel_mappings() from your driver like this to be correct:
> 
>         if (ARCH_PAGE_TABLE_SYNC_MASK & PGTBL_PTE_MODIFIED)
>                 arch_sync_kernel_mappings();
> 
> In practice this is a no-op, because nobody sets PGTBL_PTE_MODIFIED in
> ARCH_PAGE_TABLE_SYNC_MASK, so the above code would be optimized away.
> 
> But what you really care about is the tracking in apply_to_page_range(),
> as that allocates the !pte levels of your page-table, which needs
> synchronization on x86-32.
> 
> Btw, what are the reasons you can't use generic vmalloc/ioremap
> interfaces to map the range?

ioremap_prot and ioremap_page_range assume a contiguous IO address. So
we needed to allocate the vmalloc area [and would then need to iterate
over the discontiguous iomem chunks with ioremap_page_range], and since
alloc_vm_area returned the ptep, it looked clearer to then assign those
according to whether we wanted ioremapping or a plain page. So we ended
up with one call to the core to return us a vm_struct and a pte array
that worked for either backing store.
-Chris