Re: [Xen-devel] what's inside hypercall page?

[Juergen Gross <jgross@xxxxxxxx>]

On 03/03/16 08:09, quizyjones wrote:
> What I want to do is predict how many instructions a hypercall entry of
> hypercall page (not hypercall handler) would execute before it finishes.
> Take HYPERVISOR_iret as an example, it precisely executes five
> instructions then call the hypercall handler, and it doesn't return so
> it just finish. But other hypercall entries expect a return value of
> syscall and might be interrupted during the execution, just as what you
> said about do_sched_up and do_xen_version, so it is hard to predict. Is
> there any solution to this problem? Just like syscall table, how can I
> predict how many instructions the syscall entries would execute before
> it actually go to the handler, thus accurately set traps to the syscall?

Just look at the code? It's not magic, just x86 instructions. In case of
interrupts
the interrupted code is normally resumed where the interrupt occurred.
So the
number of instructions executed in the hypercall page per hypercall is
just the
number of instructions you can see for that specific hypercall. Only
exception
is the case where the call into the hypervisor itself (syscall) may be
repeated
on request of the hypervisor (do_multicall_op).

Juergen

BTW: Please don't top-post.

> 
>> Subject: Re: [Xen-devel] what's inside hypercall page?
>> To: quizy_jones@xxxxxxxxxxx; xen-devel@xxxxxxxxxxxxxxxxxxxx
>> From: jgross@xxxxxxxx
>> Date: Thu, 3 Mar 2016 06:28:06 +0100
>>
>> On 03/03/16 01:56, quizyjones wrote:
>> >> do_sched_op is self explaining: it is used for scheduling of the vcpu.
>> >> A vcpu going to idle is using this hypercall. So any interrupt waking
>> >> the vcpu up will seem to occur very near to the hypercall.
>> >
>> >> do_xen_version is often used as a very fast way to execute the check
>> >> for pending events in the hypervisor (kind of polling).
>> >
>> >> do_multicall might run for a long time. So the hypervisor returns to
>> >> the caller from time to time setting IP to the hypercall. The caller
>> >> has the chance to react to interrupts and will then continue the
>> >> hypercall.
>> >>
>> >>
>> >> HTH, Juergen
>> >
>> >
>> > Thanks for the replying. Does that mean we cannot predict when will
>> > these two hypercalls finish? I want to set up an interval to monitor the
>> > instructions (one time monitor per hypercall), so as to reduce the
>> > performance cost. This requires an accurate prediction of instructions'
>> > execution so as to avoid missing hypercalls. Is that possible? The main
>> > problem is the execution of syscall (0x050f), as each hypercall behaves
>> > different, how can I predict where will it go after the syscall returns?
>>
>> You can't predict how long a hypercall will run, as this depends on
> multiple
>> factors, like the overall load of the host, values of parameters, ...
>>
>> A hypercall is by it's nature much more complicated than e.g. a simple
>> arithmetic operation.
>>
>> What exactly do you want to achieve?
>>
>>
>> Juergen
>>
> 翻译朗读复制正在查询，请稍候……重试朗读复制复制朗读复制via 译


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel