[Xen-devel] [PATCH] Xen: hibernation is x86-only at the moment

[David Vrabel]

On 07/05/14 11:45, Stefano Stabellini wrote:
> On Tue, 6 May 2014, David Vrabel wrote:
>> On 06/05/14 14:35, Stefano Stabellini wrote:
>>> On Wed, 30 Apr 2014, David Vrabel wrote:
>>>> On 29/04/14 21:00, Arnd Bergmann wrote:
>>>>> In commit 603fb42a66499ab "ARM: 8011/1: ARM hibernation / suspend-to-disk",
>>>>> currently in linux-next, the ARM architecture gains support for
>>>>> hibernation (suspend-to-disk). Xen supports this in principle, but only
>>>>> has an architecture specific hypercall defined on x86, which leads
>>>>> to a build error when both hibernation and Xen support are enabled:
>>>>>
>>>>> drivers/xen/manage.c:105:2: error: implicit declaration of function 'HYPERVISOR_suspend' [-Werror=implicit-function-declaration]
>>>>>   si->cancelled = HYPERVISOR_suspend(si->arg);
>>>>>
>>>>> It is probably a good idea to define this hypercall on ARM as well
>>>>> and provide an implementation in the host, but until that is done,
>>>>> this patch helps disable the broken code in the Xen guest by making
>>>>> it depend on CONFIG_X86.
>>>
>>> Thanks Arnd.
>>>
>>>
>>>> Since this isn't a regression in 3.15-rcX can you provide the
>>>> appropriate hypercall on ARM.  This does not require that there is an
>>>> implementation on Xen.
>>>>
>>>> When the Xen implementation exists it will be easier to use if the
>>>> support already exists in Linux.
>>>
>>>
>>> The sched_on hypercall is already implemented on ARM.
>>> However SCHEDOP_shutdown, defined as it is now, is unusable on ARM:
>>>
>>> /*
>>>  * Halt execution of this domain (all VCPUs) and notify the system controller.
>>>  * @arg == pointer to sched_shutdown_t structure.
>>>  *
>>>  * If the sched_shutdown_t reason is SHUTDOWN_suspend then this
>>>  * hypercall takes an additional extra argument which should be the
>>>  * MFN of the guest's start_info_t.
>>>  *
>>>  * In addition, which reason is SHUTDOWN_suspend this hypercall
>>>  * returns 1 if suspend was cancelled or the domain was merely
>>>  * checkpointed, and 0 if it is resuming in a new domain.
>>>  */
>>> #define SCHEDOP_shutdown    2
>>>
>>>
>>> We don't have a start_info, and even if we had, we wouldn't know the
>>> MFN.
>>> I think we should make it available on ARM only if we change the
>>> interface making the third argument x86 only.  Of course that would
>>> impact the linux side implementation too.
>>
>> We do not want a different number of arguments to this hypercall.  Just
>> pass 0 for the MFN.  This hypercall docs should be updated to say this.
>>
>> This doesn't seem any different to x86 HVM where the start_info_mfn
>> parameter is also not relevant.
>>
>> 	if (xen_hvm_domain()) {
>> 		si.arg = 0UL;
>> 		si.pre = NULL;
>> 		si.post = &xen_hvm_post_suspend;
>> 	} else {
>> 		si.arg = virt_to_mfn(xen_start_info);
>> 		si.pre = &xen_pre_suspend;
>> 		si.post = &xen_post_suspend;
>> 	}
>>
>> Changing this to be !xen_pv_domain() would then do the right thing for arm.
> 
> xen_hvm_domain returns true on arm so that's not needed.
> Unfortunately the patch turns out a bit ugly because of the way
> hypercalls are implemented in assembly on arm and arm64.
> Also I needed to provide quite a few empty stubs for other xen arch
> specific functions.

Thanks.  If you want to add a commit message and apply to
devel/for-linus-3.16:

Reviewed-by: David Vrabel <david.vrabel at citrix.com>

> --- a/arch/arm/include/asm/xen/hypercall.h
> +++ b/arch/arm/include/asm/xen/hypercall.h
> @@ -34,6 +34,7 @@
>  #define _ASM_ARM_XEN_HYPERCALL_H
>  
>  #include <xen/interface/xen.h>
> +#include <xen/interface/sched.h>
>  
>  long privcmd_call(unsigned call, unsigned long a1,
>  		unsigned long a2, unsigned long a3,
> @@ -48,6 +49,16 @@ int HYPERVISOR_memory_op(unsigned int cmd, void *arg);
>  int HYPERVISOR_physdev_op(int cmd, void *arg);
>  int HYPERVISOR_vcpu_op(int cmd, int vcpuid, void *extra_args);
>  int HYPERVISOR_tmem_op(void *arg);
> +int HYPERVISOR_sched_op_shutdown(int cmd, void *arg, unsigned long unused);
> +
> +static inline int
> +HYPERVISOR_suspend(unsigned long start_info_mfn)
> +{
> +	struct sched_shutdown r = { .reason = SHUTDOWN_suspend };
> +
> +	/* start_info_mfn is unused on ARM, pass 0 instead */
> +	return HYPERVISOR_sched_op_shutdown(SCHEDOP_shutdown, &r, 0);
> +}
>  
>  static inline void
>  MULTI_update_va_mapping(struct multicall_entry *mcl, unsigned long va,
> diff --git a/arch/arm/xen/enlighten.c b/arch/arm/xen/enlighten.c
> index b96723e..6d88c18 100644
> --- a/arch/arm/xen/enlighten.c
> +++ b/arch/arm/xen/enlighten.c
> @@ -339,6 +339,17 @@ static int __init xen_pm_init(void)
>  }
>  late_initcall(xen_pm_init);
>  
> +
> +/* empty stubs */
> +void xen_arch_pre_suspend(void) { }
> +void xen_arch_post_suspend(int suspend_cancelled) { }
> +void xen_arch_hvm_post_suspend(int suspend_cancelled) { }
> +void xen_mm_pin_all(void) { }
> +void xen_mm_unpin_all(void) { }

I wonder if the calls to xen_mm_pin_all() and xen_mm_unpin_all() could
be moved into the arch_pre_suspend() and arch_post_suspend() calls
instead. So there wouldn't need to be stubs provided for arm.

David