On Thu, Aug 26, 2021 at 6:17 AM Will Deacon [off-list ref] wrote:

Please pull this single arm64 fix for 5.14.

Pulled.

But adding Christoph to the cc, since I do think the eventual fix
needs to be in the DMA mapping code:

We received a report this week
that the generic version of pfn_valid(), which we switched to this merge
window in 16c9afc77660 ("arm64/mm: drop HAVE_ARCH_PFN_VALID"), interacts
badly with dma_map_resource() due to the following check:

        /* Don't allow RAM to be mapped */
        if (WARN_ON_ONCE(pfn_valid(PHYS_PFN(phys_addr))))
                return DMA_MAPPING_ERROR;

Since the ongoing saga to determine the semantics of pfn_valid() is
unlikely to be resolved this week (does it indicate valid memory, or
just the presence of a struct page, or whether that struct page has been
initialised?), just revert back to our old version of pfn_valid() for
5.14.

I think that's the right thing for now, but yeah, that condition for
WARN_ON_ONCE() seems very questionable.

"pfn_valid()" is more about whether you can do a "pfn_to_page()" lookup on it.

II get the feeling that the dma-mapping code should allow pages that
are PageReserved() to be mapped - they aren't "ram" in the kernel
sense.

Perhaps also make sure it's not the zero page (which is
PageReserved(), but most definitely RAM).

In a PC world that would be (for example) the legacy PCI space at
0xa0000-0xfffff, but I could easily imagine other platforms having
other situations.

              Linus

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel