[PATCH v2 1/1] riscv: restore global data pointer in trap handler

[Heinrich Schuchardt]

On 9/27/20 10:51 AM, Bin Meng wrote:
> Hi Heinrich,
>
> On Sun, Sep 27, 2020 at 4:46 PM Heinrich Schuchardt <xypron.glpk at gmx.de> wrote:
>>
>> On 9/27/20 10:40 AM, Bin Meng wrote:
>>> On Sun, Sep 27, 2020 at 4:24 PM Heinrich Schuchardt <xypron.glpk at gmx.de> wrote:
>>>>
>>>> The gp register is used to store U-Boot's global data pointer. We should
>>>> not assume that an UEFI application leaves the gp register unchanged as
>>>> the UEFI specifications does not define who is the owner of the gp and tp
>>>> registers.
>>>>
>>>> So the following sequence should be followed in the trap handler:
>>>>
>>>> * save the caller's gp register
>>>> * restore the global data pointer
>>>> * serve interrupts or print crash dump and reset
>>>> * restore the caller's gp register
>>>>
>>>> Cc: Abner Chang <abner.chang at hpe.com>
>>>> Signed-off-by: Heinrich Schuchardt <xypron.glpk at gmx.de>
>>>> ---
>>>> v2:
>>>>         Saving and restoring the caller's x3 is already handled in mtrap.S.
>>>>         efi_loader.h provides an empty fallback efi_restore_gd() function
>>>>         if CONFIG_EFI_LOADER=n.
>>>> ---
>>>>  arch/riscv/lib/interrupts.c | 3 +++
>>>>  1 file changed, 3 insertions(+)
>>>>
>>>> diff --git a/arch/riscv/lib/interrupts.c b/arch/riscv/lib/interrupts.c
>>>> index cd47e64487..8ff40f0f36 100644
>>>> --- a/arch/riscv/lib/interrupts.c
>>>> +++ b/arch/riscv/lib/interrupts.c
>>>> @@ -111,6 +111,9 @@ ulong handle_trap(ulong cause, ulong epc, ulong tval, struct pt_regs *regs)
>>>>  {
>>>>         ulong is_irq, irq;
>>>>
>>>> +       /* An UEFI application may have changed gd. Restore U-Boot's gd. */
>>>> +       efi_restore_gd();
>>>> +
>>>>         is_irq = (cause & MCAUSE_INT);
>>>>         irq = (cause & ~MCAUSE_INT);
>>>>
>>>
>>> Reviewed-by: Bin Meng <bin.meng at windriver.com>
>>>
>>> Does other arch suffer the same issue?
>>>
>>
>> Hello Bin,
>>
>> x86 does not use a register of gd.
>
> x86 uses fs segment register to point to gd. Does UEFI not use fs?

The UEFI 2.8B specification is available at
https://uefi.org/sites/default/files/resources/UEFI%20Spec%202.8B%20May%202020.pdf.

Chapter 2.3.2 "IA-32 Platform" and chapter 2.3.4 x64 "PlatformsAll" both
have this sentence: "Other general purpose flag registers are undefined."

I could find no provision that segment registers remain unchanged.

Chapter 2.3.4.3 "Enabling Paging or Alternate Translations" and 2.3.4.3
"Enabling Paging or Alternate Translations in an Application" both allow
UEFI payloads to have their own GDT and IDT. But it seems these have to
be restored by the payload before entering an API call: "An application
with translations enabled can restore firmware required mapping before
each UEFI call."

This implies we have to change multiple places in U-Boot. For ARM and
RISC-V we have routines

* efi_save_gd()
* efi_restore_gd()
* efi_set_gd()
* trace_save_gd();
* trace_swap_gd();

So what we need to do is:

* Save our value of fs in efi_save_gd() and trace_save_gd().
* On API entry from the caller save the value of the callers fs and
restore our value.
* Before returning restore the callers fs.

Best regards

Heinrich

>
>> arm has this statement already.
>> No further architectures support UEFI.
>>
>
> Regards,
> Bin
>