Linux Virtualization - Aug 2020 - [PATCH v3 4/7] x86/paravirt: remove 32-bit support from PARAVIRT_XXL

On Fri, Aug 07, 2020 at 10:38:23AM +0200, Juergen Gross wrote:
> -# else
> -	const unsigned char	cpu_iret[1];
> -# endif
>  };
>  
>  static const struct patch_xxl patch_data_xxl = {
> @@ -42,7 +38,6 @@ static const struct patch_xxl patch_data_xxl = {
>  	.irq_save_fl		= { 0x9c, 0x58 },	// pushf; pop %[re]ax
>  	.mmu_read_cr2		= { 0x0f, 0x20, 0xd0 },	// mov %cr2, %[re]ax
>  	.mmu_read_cr3		= { 0x0f, 0x20, 0xd8 },	// mov %cr3, %[re]ax
> -# ifdef CONFIG_X86_64
>  	.mmu_write_cr3		= { 0x0f, 0x22, 0xdf },	// mov %rdi, %cr3
>  	.irq_restore_fl		= { 0x57, 0x9d },	// push %rdi; popfq
>  	.cpu_wbinvd		= { 0x0f, 0x09 },	// wbinvd
> @@ -50,19 +45,11 @@ static const struct patch_xxl patch_data_xxl = {
>  				    0x48, 0x0f, 0x07 },	// swapgs; sysretq
>  	.cpu_swapgs		= { 0x0f, 0x01, 0xf8 },	// swapgs
>  	.mov64			= { 0x48, 0x89, 0xf8 },	// mov %rdi, %rax
> -# else
> -	.mmu_write_cr3		= { 0x0f, 0x22, 0xd8 },	// mov %eax, %cr3
> -	.irq_restore_fl		= { 0x50, 0x9d },	// push %eax; popf
> -	.cpu_iret		= { 0xcf },		// iret
> -# endif
I was looking at x86_64 paravirt the other day and found we actually
have pv_ops.cpu.iret users there..

So we want to change the above to also patch iret on x86_64 or do we
need to fix x86_64 to not have pv-iret?

On 07.08.20 11:39, peterz at infradead.org wrote:
> On Fri, Aug 07, 2020 at 10:38:23AM +0200, Juergen Gross wrote:
> 
>> -# else
>> -	const unsigned char	cpu_iret[1];
>> -# endif
>>   };
>>   
>>   static const struct patch_xxl patch_data_xxl = {
>> @@ -42,7 +38,6 @@ static const struct patch_xxl patch_data_xxl = {
>>   	.irq_save_fl		= { 0x9c, 0x58 },	// pushf; pop %[re]ax
>>   	.mmu_read_cr2		= { 0x0f, 0x20, 0xd0 },	// mov %cr2, %[re]ax
>>   	.mmu_read_cr3		= { 0x0f, 0x20, 0xd8 },	// mov %cr3, %[re]ax
>> -# ifdef CONFIG_X86_64
>>   	.mmu_write_cr3		= { 0x0f, 0x22, 0xdf },	// mov %rdi, %cr3
>>   	.irq_restore_fl		= { 0x57, 0x9d },	// push %rdi; popfq
>>   	.cpu_wbinvd		= { 0x0f, 0x09 },	// wbinvd
>> @@ -50,19 +45,11 @@ static const struct patch_xxl patch_data_xxl = {
>>   				    0x48, 0x0f, 0x07 },	// swapgs; sysretq
>>   	.cpu_swapgs		= { 0x0f, 0x01, 0xf8 },	// swapgs
>>   	.mov64			= { 0x48, 0x89, 0xf8 },	// mov %rdi, %rax
>> -# else
>> -	.mmu_write_cr3		= { 0x0f, 0x22, 0xd8 },	// mov %eax, %cr3
>> -	.irq_restore_fl		= { 0x50, 0x9d },	// push %eax; popf
>> -	.cpu_iret		= { 0xcf },		// iret
>> -# endif
> 
> I was looking at x86_64 paravirt the other day and found we actually
> have pv_ops.cpu.iret users there..
On x86_64 we have (without PARAVIRT_XXL):

#define INTERRUPT_RETURN        jmp native_iret

and with PARAVIRT_XXL this is basically a jmp *pv_ops.cpu.iret which
will then be patched to either jmp native_iret or jmp xen_iret.

On x86_32 INTERRUPT_RETURN was just "iret" for the non-paravirt case.
This is the reason for above dropping of the static patch data.
> So we want to change the above to also patch iret on x86_64 or do we
> need to fix x86_64 to not have pv-iret?
We want it to stay how it is. This will let both variants (PARVIRT y/n)
continue to work.


Juergen

On Fri, Aug 07, 2020 at 12:02:59PM +0200, J?rgen Gro?
wrote:
> On 07.08.20 11:39, peterz at infradead.org wrote:
> > On Fri, Aug 07, 2020 at 10:38:23AM +0200, Juergen Gross wrote:
> > 
> > > -# else
> > > -	const unsigned char	cpu_iret[1];
> > > -# endif
> > >   };
> > >   static const struct patch_xxl patch_data_xxl = {
> > > @@ -42,7 +38,6 @@ static const struct patch_xxl patch_data_xxl =
{
> > >   	.irq_save_fl		= { 0x9c, 0x58 },	// pushf; pop %[re]ax
> > >   	.mmu_read_cr2		= { 0x0f, 0x20, 0xd0 },	// mov %cr2, %[re]ax
> > >   	.mmu_read_cr3		= { 0x0f, 0x20, 0xd8 },	// mov %cr3, %[re]ax
> > > -# ifdef CONFIG_X86_64
> > >   	.mmu_write_cr3		= { 0x0f, 0x22, 0xdf },	// mov %rdi, %cr3
> > >   	.irq_restore_fl		= { 0x57, 0x9d },	// push %rdi; popfq
> > >   	.cpu_wbinvd		= { 0x0f, 0x09 },	// wbinvd
> > > @@ -50,19 +45,11 @@ static const struct patch_xxl patch_data_xxl
= {
> > >   				    0x48, 0x0f, 0x07 },	// swapgs; sysretq
> > >   	.cpu_swapgs		= { 0x0f, 0x01, 0xf8 },	// swapgs
> > >   	.mov64			= { 0x48, 0x89, 0xf8 },	// mov %rdi, %rax
> > > -# else
> > > -	.mmu_write_cr3		= { 0x0f, 0x22, 0xd8 },	// mov %eax, %cr3
> > > -	.irq_restore_fl		= { 0x50, 0x9d },	// push %eax; popf
> > > -	.cpu_iret		= { 0xcf },		// iret
> > > -# endif
> > 
> > I was looking at x86_64 paravirt the other day and found we actually
> > have pv_ops.cpu.iret users there..
> 
> On x86_64 we have (without PARAVIRT_XXL):
> 
> #define INTERRUPT_RETURN        jmp native_iret
> 
> and with PARAVIRT_XXL this is basically a jmp *pv_ops.cpu.iret which
> will then be patched to either jmp native_iret or jmp xen_iret.
Ah, okay. Clearly I didn't look hard enough. Thanks!