Date:   Fri, 27 Oct 2017 09:31:09 +0200
From:   Juergen Gross <jgross@...e.com>
To:     Dongli Zhang <dongli.zhang@...cle.com>,
        Boris Ostrovsky <boris.ostrovsky@...cle.com>,
        xen-devel@...ts.xenproject.org, linux-kernel@...r.kernel.org
Cc:     joao.m.martins@...cle.com
Subject: Re: [Xen-devel] [PATCH v3 1/1] xen/time: do not decrease steal time
 after live migration on xen

On 27/10/17 09:16, Dongli Zhang wrote:
> Hi Boris,
> 
> On 10/25/2017 11:12 PM, Boris Ostrovsky wrote:
>> On 10/25/2017 02:45 AM, Dongli Zhang wrote:
>>> After guest live migration on xen, steal time in /proc/stat
>>> (cpustat[CPUTIME_STEAL]) might decrease because steal returned by
>>> xen_steal_lock() might be less than this_rq()->prev_steal_time which is
>>> derived from previous return value of xen_steal_clock().
>>>
>>> For instance, steal time of each vcpu is 335 before live migration.
>>>
>>> cpu  198 0 368 200064 1962 0 0 1340 0 0
>>> cpu0 38 0 81 50063 492 0 0 335 0 0
>>> cpu1 65 0 97 49763 634 0 0 335 0 0
>>> cpu2 38 0 81 50098 462 0 0 335 0 0
>>> cpu3 56 0 107 50138 374 0 0 335 0 0
>>>
>>> After live migration, steal time is reduced to 312.
>>>
>>> cpu  200 0 370 200330 1971 0 0 1248 0 0
>>> cpu0 38 0 82 50123 500 0 0 312 0 0
>>> cpu1 65 0 97 49832 634 0 0 312 0 0
>>> cpu2 39 0 82 50167 462 0 0 312 0 0
>>> cpu3 56 0 107 50207 374 0 0 312 0 0
>>>
>>> Since runstate times are cumulative and cleared during xen live migration
>>> by xen hypervisor, the idea of this patch is to accumulate runstate times
>>> to global percpu variables before live migration suspend. Once guest VM is
>>> resumed, xen_get_runstate_snapshot_cpu() would always return the sum of new
>>> runstate times and previously accumulated times stored in global percpu
>>> variables.
>>>
>>> Similar and more severe issue would impact prior linux 4.8-4.10 as
>>> discussed by Michael Las at
>>> https://0xstubs.org/debugging-a-flaky-cpu-steal-time-counter-on-a-paravirtualized-xen-guest,
>>> which would overflow steal time and lead to 100% st usage in top command
>>> for linux 4.8-4.10. A backport of this patch would fix that issue.
>>>
>>> References: https://0xstubs.org/debugging-a-flaky-cpu-steal-time-counter-on-a-paravirtualized-xen-guest
>>> Signed-off-by: Dongli Zhang <dongli.zhang@...cle.com>
>>>
>>> ---
>>> Changed since v1:
>>>   * relocate modification to xen_get_runstate_snapshot_cpu
>>>
>>> Changed since v2:
>>>   * accumulate runstate times before live migration
>>>
>>> ---
>>>  drivers/xen/manage.c  |  1 +
>>>  drivers/xen/time.c    | 19 +++++++++++++++++++
>>>  include/xen/xen-ops.h |  1 +
>>>  3 files changed, 21 insertions(+)
>>>
>>> diff --git a/drivers/xen/manage.c b/drivers/xen/manage.c
>>> index c425d03..9aa2955 100644
>>> --- a/drivers/xen/manage.c
>>> +++ b/drivers/xen/manage.c
>>> @@ -72,6 +72,7 @@ static int xen_suspend(void *data)
>>>  	}
>>>  
>>>  	gnttab_suspend();
>>> +	xen_accumulate_runstate_time();
>>>  	xen_arch_pre_suspend();
>>>  
>>>  	/*
>>> diff --git a/drivers/xen/time.c b/drivers/xen/time.c
>>> index ac5f23f..6df3f82 100644
>>> --- a/drivers/xen/time.c
>>> +++ b/drivers/xen/time.c
>>> @@ -19,6 +19,8 @@
>>>  /* runstate info updated by Xen */
>>>  static DEFINE_PER_CPU(struct vcpu_runstate_info, xen_runstate);
>>>  
>>> +static DEFINE_PER_CPU(u64[4], old_runstate_time);
>>> +
>>>  /* return an consistent snapshot of 64-bit time/counter value */
>>>  static u64 get64(const u64 *p)
>>>  {
>>> @@ -52,6 +54,7 @@ static void xen_get_runstate_snapshot_cpu(struct vcpu_runstate_info *res,
>>>  {
>>>  	u64 state_time;
>>>  	struct vcpu_runstate_info *state;
>>> +	int i;
>>>  
>>>  	BUG_ON(preemptible());
>>>  
>>> @@ -64,6 +67,22 @@ static void xen_get_runstate_snapshot_cpu(struct vcpu_runstate_info *res,
>>>  		rmb();	/* Hypervisor might update data. */
>>>  	} while (get64(&state->state_entry_time) != state_time ||
>>>  		 (state_time & XEN_RUNSTATE_UPDATE));
>>> +
>>> +	for (i = 0; i < 4; i++)
>>> +		res->time[i] += per_cpu(old_runstate_time, cpu)[i];
>>> +}
>>> +
>>> +void xen_accumulate_runstate_time(void)
>>> +{
>>> +	struct vcpu_runstate_info state;
>>> +	int cpu;
>>> +
>>> +	for_each_possible_cpu(cpu) {
>>> +		xen_get_runstate_snapshot_cpu(&state, cpu);
>>> +		memcpy(per_cpu(old_runstate_time, cpu),
>>> +				state.time,
>>> +				4 * sizeof(u64));
>>
>> sizeof(old_runstate_time). (I think this should work for per_cpu variables)
>>
>>> +	}
>>
>> Hmm.. This may not perform as intended if we are merely checkpointing
>> (or pausing) the guest (i.e. if HYPERVISOR_suspend() returns 1). We will
>> double-account for the last interval that the guest has run.
>>
>> I'd rather not have yet another per-cpu variable but I can't think of
>> anything else. Perhaps you or others can come up with something better.
> 
> I have 3 options so far.
> 
> The 1st option is to another per-cpu variable while you do not like it.
> 
> The 2nd option is to borrow from what do_stolen_accounting() used to do. Compute
> the delta of current and previous time and do nothing if delta is less than 0.
> The drawback of this option is guest might wait for the new time to catch up
> with previous time.

This could be a rather long time. I don't think this is the way to go.

> The 3rd option is to check the return value of HYPERVISOR_suspend() to different
> if this is a migration of checkpointing. As we will double-account the runstate
> time for checkpointing, why not just divide it by 2? The drawback of this option
> is the result is not accurate as we divide the incremental (time before and
> after checkpointing) by 2.

And it is wrong if you do multiple migrations.

> Would you please let me know which option we prefer?

Perhaps option 4:

Allocate a buffer at suspend time for the times to add up and do the
correction after suspend and free the buffer again.


Juergen

Powered by blists - more mailing lists