[RFC PATCH] arm64/efi: use id mapping for Runtime Services

[Ard Biesheuvel]

On 6 August 2014 16:36, Will Deacon <will.deacon at arm.com> wrote:
> On Thu, Jul 31, 2014 at 03:11:49PM +0100, Ard Biesheuvel wrote:
>> There are 2 interesting pieces of information in the UEFI spec section 2.3.6
>> regarding the mapping of runtime regions:
>> (a) the firmware should not request a virtual mapping for configuration tables,
>>     even though they are marked as EfiRuntimeServicesData;
>> (b) calling SetVirtualAddressMap() is optional, and it is equally appropriate to
>>     call Runtime Services using an identity mapping.
>>
>> So we can eliminate some of the complexity around UEFI Runtime Services by not
>> using a virtual mapping at all, and calling the services at their physical
>> address. This is especially useful under kexec, as SetVirtualAddressMap() may
>> only be called once, and there is no guarantee that mappings are stable between
>> different kexec'd kernels.
>>
>> The fallout for other in-kernel users of UEFI data structures should be
>> negligible, as they cannot legally access those data structures through
>> pre-existing virtual mappings anyway (point (a) above)
>>
>> It should also be noted that, as the kernel side of the address space (TTBR1) is
>> retained, the stack and pointer function arguments remain accessible to the
>> runtime service while the id mapping is active.
>>
>> Signed-off-by: Ard Biesheuvel <ard.biesheuvel at linaro.org>
>> ---
>>  arch/arm64/include/asm/efi.h |  24 ++++++++--
>>  arch/arm64/kernel/efi.c      | 106 ++-----------------------------------------
>>  2 files changed, 23 insertions(+), 107 deletions(-)
>>
>> diff --git a/arch/arm64/include/asm/efi.h b/arch/arm64/include/asm/efi.h
>> index a34fd3b12e2b..d42a21e79b39 100644
>> --- a/arch/arm64/include/asm/efi.h
>> +++ b/arch/arm64/include/asm/efi.h
>> @@ -1,8 +1,10 @@
>>  #ifndef _ASM_EFI_H
>>  #define _ASM_EFI_H
>>
>> +#include <asm/cacheflush.h>
>>  #include <asm/io.h>
>>  #include <asm/neon.h>
>> +#include <asm/tlbflush.h>
>>
>>  #ifdef CONFIG_EFI
>>  extern void efi_init(void);
>> @@ -12,23 +14,37 @@ extern void efi_idmap_init(void);
>>  #define efi_idmap_init()
>>  #endif
>>
>> +static inline void switch_pgd(pgd_t *pgd, struct mm_struct *mm)
>> +{
>> +     cpu_switch_mm(pgd, mm);
>> +     flush_tlb_all();
>> +     if (icache_is_aivivt())
>> +             __flush_icache_all();
>> +}
>> +
>>  #define efi_call_virt(f, ...)                                                \
>>  ({                                                                   \
>> -     efi_##f##_t *__f = efi.systab->runtime->f;                      \
>> +     efi_##f##_t *__f;                                               \
>>       efi_status_t __s;                                               \
>>                                                                       \
>> -     kernel_neon_begin();                                            \
>> +     kernel_neon_begin(); /* disables preemption */                  \
>> +     switch_pgd(idmap_pg_dir, &init_mm);                             \
>> +     __f =  efi.systab->runtime->f;                                  \
>>       __s = __f(__VA_ARGS__);                                         \
>> +     switch_pgd(current->active_mm->pgd, current->active_mm);        \
>>       kernel_neon_end();                                              \
>>       __s;                                                            \
>>  })
>
> This scares the bejesus out of me, but I can't put my finger on exactly why.
> I think it does what you intend and I can't break it myself, so it would be
> really good if the EFI folks could confirm that this looks good to them.
>

There is something similar in the x86 code (arch/x86/platform/efi/efi.c)
"""
 * The new method does a pagetable switch in a preemption-safe manner
 * so that we're in a different address space when calling a runtime
 * function. For function arguments passing we do copy the PGDs of the
 * kernel page table into ->trampoline_pgd prior to each call.
"""

How exactly this will turn out for arm64 (and ARM) is still under
discussion, though. My position is that if you are going to switch
pgd's anyway, why not just use the id mapping? And even if you feel it
is mandatory to install a virtual address mapping into UEFI (which I
think is /not/ the case), you could install an id mapping as well,
which means all the related machinery still gets invoked.
The alternative to using a TTBR0 mapping would be to reserve a slice
of kernel virtual memory so that the Runtime Services are guaranteed
to live at the same virtual address after a kexec, ideally the same
region on 4k and 64k pages ...

We are planning to discuss this further at Linaro Connect next month.

-- 
Ard.