Re: [PATCH v2 13/18] x86/asm/64: Use a percpu trampoline stack for IDT entries

[Andy Lutomirski]

On Thu, Nov 23, 2017 at 3:44 PM, Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> On Tue, 21 Nov 2017, Andy Lutomirski wrote:
>> The asm isn't exactly beautiful,
>
> Delightful euphemism :)
>
>> but I think that fully refactoring
>> it can wait.
>
>> @@ -560,6 +560,14 @@ END(irq_entries_start)
>>       .macro interrupt func
>>       cld
>>       ALLOC_PT_GPREGS_ON_STACK
>> +
>> +     testb   $3, CS(%rsp)
>> +     jz      1f
>> +     SWAPGS
>> +     call    switch_to_thread_stack
>> +     SWAPGS
>
> I'm surely missing something subtle, but the register saving does really
> not care on which GS it is. This swapgs orgy looks odd.

You're mostly right.  switch_to_thread_stack uses
PER_CPU_VAR(cpu_current_top_of_stack), which definitely cares about
which GS it's on, but there's still no legitimate reason for the
SWAPGS orgy.  I'll fix it.

>
>> +1:
>> +
>>       SAVE_C_REGS
>>       SAVE_EXTRA_REGS
>>       ENCODE_FRAME_POINTER
>> @@ -827,6 +835,33 @@ apicinterrupt IRQ_WORK_VECTOR                    irq_work_interrupt              smp_irq_work_interrupt
>>   */
>>  #define CPU_TSS_IST(x) PER_CPU_VAR(cpu_tss) + (TSS_ist + ((x) - 1) * 8)
>>
>> +/*
>> + * Switch to the thread stack.  This is called with the IRET frame and
>> + * orig_ax in pt_regs and the rest of pt_regs allocated, but with all GPRs
>> + * in the CPU registers.
>
> That took several attempts to grok why you left ALLOC_PT_GPRES_ON_STACK in
> place in the interrupts macro above.
>
> In theory it would be sufficient to push %rdi on the entry stack and
> operate from there, but it spares only the 'addq %rsp'. Not worth the
> trouble of dealing with different register offsets.

Hrm.  There wasn't actually a good reason for that.  I got rid of it.