[BUG linux-4.9.x] xen hotplug cpu leads to 100% steal usage

From: Dongli Zhang
Date: Thu Feb 28 2019 - 21:36:04 EST


This issue is only for stable 4.9.x (e.g., 4.9.160), while the root cause is
still in the lasted mainline kernel.

This is obviated by new feature patch set ended with b672592f0221
("sched/cputime: Remove generic asm headers").

After xen guest is up for long time, once we hotplug new vcpu, the corresponding
steal usage might become 100% and the steal time from /proc/stat would increase
abnormally.

As we cannot wait for long time to reproduce the issue, here is how I reproduce
it on purpose by accounting a large initial steal clock for new vcpu 2 and 3.

1. Apply the below patch to guest 4.9.160 to account large initial steal clock
for new vcpu 2 and 3:

diff --git a/drivers/xen/time.c b/drivers/xen/time.c
index ac5f23f..3cf629e 100644
--- a/drivers/xen/time.c
+++ b/drivers/xen/time.c
@@ -85,7 +85,14 @@ u64 xen_steal_clock(int cpu)
struct vcpu_runstate_info state;

xen_get_runstate_snapshot_cpu(&state, cpu);
- return state.time[RUNSTATE_runnable] + state.time[RUNSTATE_offline];
+
+ if (cpu == 2 || cpu == 3)
+ return state.time[RUNSTATE_runnable]
+ + state.time[RUNSTATE_offline]
+ + 0x00071e87e677aa12;
+ else
+ return state.time[RUNSTATE_runnable]
+ + state.time[RUNSTATE_offline];
}

void xen_setup_runstate_info(int cpu)


2. Boot hvm guest with "vcpus=2" and "maxvcpus=4". By default, VM boot with
vcpu 0 and 1.

3. Hotplug vcpu 2 and 3 via "xl vcpu-set <domid> 4" on dom0.

In my env, the steal becomes 100% within 10s after the "xl vcpu-set" command on
dom0.

I can reproduce on kvm with similar method. However, as the initial steal clock
on kvm guest is always 0, I do not think it is easy to hit this issue on kvm.

--------------------------------------------------------

The root cause is that the return type of jiffies_to_usecs() is 'unsigned int',
but not 'unsigned long'. As a result, the leading 32 bits are discarded.

jiffies_to_usecs() is indirectly triggered by cputime_to_nsecs() at line 264.
If guest is already up for long time, the initial steal time for new vcpu might
be large and the leading 32 bits of jiffies_to_usecs() would be discarded.

As a result, the steal at line 259 is always large and the
this_rq()->prev_steal_time at line 264 is always small. The difference at line
260 is always large during each time steal_account_process_time() is involved.
Finally, the steal time in /proc/stat would increase abnormally.

252 static __always_inline cputime_t steal_account_process_time(cputime_t maxtime)
253 {
254 #ifdef CONFIG_PARAVIRT
255 if (static_key_false(&paravirt_steal_enabled)) {
256 cputime_t steal_cputime;
257 u64 steal;
258
259 steal = paravirt_steal_clock(smp_processor_id());
260 steal -= this_rq()->prev_steal_time;
261
262 steal_cputime = min(nsecs_to_cputime(steal), maxtime);
263 account_steal_time(steal_cputime);
264 this_rq()->prev_steal_time += cputime_to_nsecs(steal_cputime);
265
266 return steal_cputime;
267 }
268 #endif
269 return 0;
270 }

--------------------------------------------------------

I have emailed the kernel mailing list about the return type of
jiffies_to_usecs() and jiffies_to_msecs():

https://lkml.org/lkml/2019/2/26/899


So far, I have two solutions:

1. Change the return type from 'unsigned int' to 'unsigned long' as in above
link and I am afraid it would bring side effect. The return type in latest
mainline kernel is still 'unsigned int'.

2. Something like below based on stable 4.9.160:

diff --git a/include/linux/jiffies.h b/include/linux/jiffies.h
index 734377a..9b1fc40 100644
--- a/include/linux/jiffies.h
+++ b/include/linux/jiffies.h
@@ -286,10 +286,11 @@ extern unsigned long preset_lpj;
*/
extern unsigned int jiffies_to_msecs(const unsigned long j);
extern unsigned int jiffies_to_usecs(const unsigned long j);
+extern unsigned long jiffies_to_usecs64(const unsigned long j);

static inline u64 jiffies_to_nsecs(const unsigned long j)
{
- return (u64)jiffies_to_usecs(j) * NSEC_PER_USEC;
+ return (u64)jiffies_to_usecs64(j) * NSEC_PER_USEC;
}

extern u64 jiffies64_to_nsecs(u64 j);
diff --git a/kernel/time/time.c b/kernel/time/time.c
index a5b6d98..256c147 100644
--- a/kernel/time/time.c
+++ b/kernel/time/time.c
@@ -288,6 +288,27 @@ unsigned int jiffies_to_usecs(const unsigned long j)
}
EXPORT_SYMBOL(jiffies_to_usecs);

+unsigned long jiffies_to_usecs64(const unsigned long j)
+{
+ /*
+ * Hz usually doesn't go much further MSEC_PER_SEC.
+ * jiffies_to_usecs() and usecs_to_jiffies() depend on that.
+ */
+ BUILD_BUG_ON(HZ > USEC_PER_SEC);
+
+#if !(USEC_PER_SEC % HZ)
+ return (USEC_PER_SEC / HZ) * j;
+#else
+# if BITS_PER_LONG == 32
+ return (HZ_TO_USEC_MUL32 * j) >> HZ_TO_USEC_SHR32;
+# else
+ return (j * HZ_TO_USEC_NUM) / HZ_TO_USEC_DEN;
+# endif
+#endif
+}
+EXPORT_SYMBOL(jiffies_to_usecs64);
+
+
/**
* timespec_trunc - Truncate timespec to a granularity
* @t: Timespec

People may dislike the 2nd solution.

3. Backport patch set ended with b672592f0221 ("sched/cputime:
Remove generic asm headers").

This is not reasonable for stable branch as the patch set involves
lots of changes.


Would you please let me know if there is any suggestion on this issue?

Thank you very much!

Dongli Zhang