[PATCH v2 09/10] arm64/efi: ignore unusable regions instead of reserving them

[Mark Rutland]

On Tue, Nov 11, 2014 at 05:12:09PM +0000, Mark Salter wrote:
> On Tue, 2014-11-11 at 10:42 -0500, Mark Salter wrote:
> > On Mon, 2014-11-10 at 08:31 +0100, Ard Biesheuvel wrote:
> > > On 10 November 2014 05:11, Mark Salter <msalter at redhat.com> wrote:
> > > > On Thu, 2014-11-06 at 15:13 +0100, Ard Biesheuvel wrote:
> > > >> This changes the way memblocks are installed based on the contents
> > > of
> > > >> the UEFI memory map. Formerly, all regions would be
> > > memblock_add()'ed,
> > > >> after which unusable regions would be memblock_reserve()'d as well.
> > > >> To simplify things, but also to allow access to the unusable
> > > regions
> > > >> through mmap(/dev/mem), even with CONFIG_STRICT_DEVMEM set, change
> > > >> this so that only usable regions are memblock_add()'ed in the first
> > > >> place.
> > > >
> > > > This patch is crashing 64K pagesize kernels during boot. I'm not
> > > exactly
> > > > sure why, though. Here is what I get on an APM Mustang box:
> > > >
> > > 
> > > Ah, yes, I meant to mention this patch
> > > 
> > > https://git.kernel.org/cgit/linux/kernel/git/glikely/linux.git/commit/?id=8cccffc52694938fc88f3d90bc7fed8460e27191
> > > 
> > > in the cover letter, which addresses this issue at least for the DT
> > > case.
> > > 
> > 
> > That isn't the problem. In general, with 64K kernel pages, you can't be
> > sure if you leave something you need out of the kernel linear mapping.

Regardless of 64k pages you can never assume that anything will be in
the linear mapping due to the current relationship between the start of
the linear map and the address the kernel was loaded at.

> > If you have Loader Code/Data regions begin and/or end at something other
> > than a 64K boundary and that region is adjacent to a region not being
> > added, then you end up leaving out the unaligned bits from the linear
> > mapping. This could be bits of the initramfs or devicetree.
> > 
> > What I don't get with this failure is that it is an alignment fault
> > which should be masked at EL1 for the kernel. The same unaligned
> > access happens without this patch and it doesn't generate a fault.
> > 
> 
> Ah, but unaligned accesses are not ignored for device memory.
> I have this in include/acpi/acpi_io.h:
> 
> static inline void __iomem *acpi_os_ioremap(acpi_physical_address phys,
> 					    acpi_size size)
> {
> #ifdef CONFIG_ARM64
> 	if (!page_is_ram(phys >> PAGE_SHIFT))
> 		return ioremap(phys, size);
> #endif
> 
>        return ioremap_cache(phys, size);
> }
> 
> Because the table isn't in the linear mapping, it fails the
> page_is_ram() test and it gits mapped with ioremap() leading to
> the alignment fault.
> 
> If I take out the code inside the #ifdef, I get a different
> fault:
> 
> [    0.350057] Unhandled fault: synchronous external abort (0x96000010) at 0xfffffe0000fae6f4
> [    0.358704] pgd = fffffe0001160000
> [    0.362276] [fffffe0000fae6f4] *pgd=0000004001370003, *pud=0000004001370003, *pmd=0000004001370003, *pte=02c00040011a0713
> [    0.373746] Internal error: : 96000010 [#1] SMP
> [    0.378484] Modules linked in:
> [    0.381601] CPU: 1 PID: 1 Comm: swapper/0 Not tainted 3.18.0-rc4+ #15
> [    0.388248] Hardware name: APM X-Gene Mustang board (DT)
> [    0.393738] task: fffffe03dbe10000 ti: fffffe03dbf00000 task.ti: fffffe03dbf00000
> [    0.401503] PC is at acpi_ex_system_memory_space_handler+0x238/0x2e0
> [    0.408160] LR is at acpi_ex_system_memory_space_handler+0x130/0x2e0
> 
> That happens because AML is trying to access a hardware register
> which has been mapped as normal memory.

Which is why removing the check in the ifdef is completely nonsensical.
We already knew we can't map everything cacheable -- regardless of what
AML does the CPU can prefetch from anything mapped cacheable (or simply
executable) at any time.

> So, we need a way to tell a table in ram from an io address in AML.
> And page_is_ram() no longer cuts it if the tables are not in the
> linear mapping.

If the kernel were loaded at an address above the tables, it would fail
similarly. So the page_is_ram was never sufficient to ensure tables
would be mapped cacheable.

As we haven't decoupled the kernel text mapping from the linear mapping,
and that doesn't look to be happening any time soon, we can't fix up
page_is_ram -- we need something else entirely.

Thanks,
Mark.