[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] Big Bug:Time in VM goes slower; foud Solution but demand Judgement! A Interesting Story!



>>> On 10.08.12 at 17:17, tupeng212 <tupeng212@xxxxxxxxx> wrote:
> 2 Bug in xen
> JVM is OK, so left the bug to xen, I have found both the reason and 
> solution. As Jan mentioned avoiding call create_periodic_time, it got much 
> better. so I modified it like this, if the pt timer is created before, 
> setting RegA down is just changing the period value, so I do nothing except 

What you describe doesn't sound accurate (i.e. I'm getting the
impression that you might have suppressed the call in cases
where you shouldn't).

Below/attached a first draft of a patch to fix not only this issue,
but a few more with the RTC emulation. Would you give this a
try?

Keir, Tim, others - the change to xen/arch/x86/hvm/vpt.c really
looks more like a hack than a solution, but I don't see another
way without much more intrusive changes. The point is that we
want the RTC code to decide whether to generate an interrupt
(so that RTC_PF can become set correctly even without RTC_PIE
getting enabled by the guest).

Jan

x86/HVM: assorted RTC emulation adjustments

- don't call rtc_timer_update() on REG_A writes when the value didn't
  change (doing the call always was reported to cause wall clock time
  lagging with the JVM under Windows)
- in the same spirit, don't call rtc_timer_update() or
  alarm_timer_update() on REG_B writes when the respective RTC_xIE bit
  didn't change
- raise the RTC IRQ not only when RTC_UIE gets set while RTC_UF was
  already set, but generalize this to alarm and periodic interrupts as
  well
- properly handle RTC_PF when the guest is not also setting RTC_PIE
- also handle the two other clock bases

--- a/xen/arch/x86/hvm/rtc.c
+++ b/xen/arch/x86/hvm/rtc.c
@@ -50,11 +50,24 @@ static void rtc_set_time(RTCState *s);
 static inline int from_bcd(RTCState *s, int a);
 static inline int convert_hour(RTCState *s, int hour);
 
-static void rtc_periodic_cb(struct vcpu *v, void *opaque)
+static void rtc_toggle_irq(RTCState *s)
+{
+    struct domain *d = vrtc_domain(s);
+
+    ASSERT(spin_is_locked(&s->lock));
+    s->hw.cmos_data[RTC_REG_C] |= RTC_IRQF;
+    hvm_isa_irq_deassert(d, RTC_IRQ);
+    hvm_isa_irq_assert(d, RTC_IRQ);
+}
+
+void rtc_periodic_interrupt(void *opaque)
 {
     RTCState *s = opaque;
+
     spin_lock(&s->lock);
-    s->hw.cmos_data[RTC_REG_C] |= 0xc0;
+    s->hw.cmos_data[RTC_REG_C] |= RTC_PF;
+    if ( s->hw.cmos_data[RTC_REG_B] & RTC_PIE )
+        rtc_toggle_irq(s);
     spin_unlock(&s->lock);
 }
 
@@ -68,19 +81,25 @@ static void rtc_timer_update(RTCState *s
     ASSERT(spin_is_locked(&s->lock));
 
     period_code = s->hw.cmos_data[RTC_REG_A] & RTC_RATE_SELECT;
-    if ( (period_code != 0) && (s->hw.cmos_data[RTC_REG_B] & RTC_PIE) )
+    switch ( s->hw.cmos_data[RTC_REG_A] & RTC_DIV_CTL )
     {
-        if ( period_code <= 2 )
+    case RTC_REF_CLCK_32KHZ:
+        if ( (period_code != 0) && (period_code <= 2) )
             period_code += 7;
-
-        period = 1 << (period_code - 1); /* period in 32 Khz cycles */
-        period = DIV_ROUND((period * 1000000000ULL), 32768); /* period in ns */
-        create_periodic_time(v, &s->pt, period, period, RTC_IRQ,
-                             rtc_periodic_cb, s);
-    }
-    else
-    {
+        /* fall through */
+    case RTC_REF_CLCK_1MHZ:
+    case RTC_REF_CLCK_4MHZ:
+        if ( period_code != 0 )
+        {
+            period = 1 << (period_code - 1); /* period in 32 Khz cycles */
+            period = DIV_ROUND(period * 1000000000ULL, 32768); /* in ns */
+            create_periodic_time(v, &s->pt, period, period, RTC_IRQ, NULL, s);
+            break;
+        }
+        /* fall through */
+    default:
         destroy_periodic_time(&s->pt);
+        break;
     }
 }
 
@@ -144,7 +163,6 @@ static void rtc_update_timer(void *opaqu
 static void rtc_update_timer2(void *opaque)
 {
     RTCState *s = opaque;
-    struct domain *d = vrtc_domain(s);
 
     spin_lock(&s->lock);
     if (!(s->hw.cmos_data[RTC_REG_B] & RTC_SET))
@@ -152,11 +170,7 @@ static void rtc_update_timer2(void *opaq
         s->hw.cmos_data[RTC_REG_C] |= RTC_UF;
         s->hw.cmos_data[RTC_REG_A] &= ~RTC_UIP;
         if ((s->hw.cmos_data[RTC_REG_B] & RTC_UIE))
-        {
-            s->hw.cmos_data[RTC_REG_C] |= RTC_IRQF;
-            hvm_isa_irq_deassert(d, RTC_IRQ);
-            hvm_isa_irq_assert(d, RTC_IRQ);
-        }
+            rtc_toggle_irq(s);
         check_update_timer(s);
     }
     spin_unlock(&s->lock);
@@ -343,7 +357,6 @@ static void alarm_timer_update(RTCState 
 static void rtc_alarm_cb(void *opaque)
 {
     RTCState *s = opaque;
-    struct domain *d = vrtc_domain(s);
 
     spin_lock(&s->lock);
     if (!(s->hw.cmos_data[RTC_REG_B] & RTC_SET))
@@ -351,11 +364,7 @@ static void rtc_alarm_cb(void *opaque)
         s->hw.cmos_data[RTC_REG_C] |= RTC_AF;
         /* alarm interrupt */
         if (s->hw.cmos_data[RTC_REG_B] & RTC_AIE)
-        {
-            s->hw.cmos_data[RTC_REG_C] |= RTC_IRQF;
-            hvm_isa_irq_deassert(d, RTC_IRQ);
-            hvm_isa_irq_assert(d, RTC_IRQ);
-        }
+            rtc_toggle_irq(s);
         alarm_timer_update(s);
     }
     spin_unlock(&s->lock);
@@ -365,6 +374,7 @@ static int rtc_ioport_write(void *opaque
 {
     RTCState *s = opaque;
     struct domain *d = vrtc_domain(s);
+    uint32_t orig, mask;
 
     spin_lock(&s->lock);
 
@@ -382,6 +392,7 @@ static int rtc_ioport_write(void *opaque
         return 0;
     }
 
+    orig = s->hw.cmos_data[s->hw.cmos_index];
     switch ( s->hw.cmos_index )
     {
     case RTC_SECONDS_ALARM:
@@ -405,9 +416,9 @@ static int rtc_ioport_write(void *opaque
         break;
     case RTC_REG_A:
         /* UIP bit is read only */
-        s->hw.cmos_data[RTC_REG_A] = (data & ~RTC_UIP) |
-            (s->hw.cmos_data[RTC_REG_A] & RTC_UIP);
-        rtc_timer_update(s);
+        s->hw.cmos_data[RTC_REG_A] = (data & ~RTC_UIP) | (orig & RTC_UIP);
+        if ( (data ^ orig) & (RTC_RATE_SELECT | RTC_DIV_CTL) )
+            rtc_timer_update(s);
         break;
     case RTC_REG_B:
         if ( data & RTC_SET )
@@ -415,7 +426,7 @@ static int rtc_ioport_write(void *opaque
             /* set mode: reset UIP mode */
             s->hw.cmos_data[RTC_REG_A] &= ~RTC_UIP;
             /* adjust cmos before stopping */
-            if (!(s->hw.cmos_data[RTC_REG_B] & RTC_SET))
+            if (!(orig & RTC_SET))
             {
                 s->current_tm = gmtime(get_localtime(d));
                 rtc_copy_date(s);
@@ -424,21 +435,27 @@ static int rtc_ioport_write(void *opaque
         else
         {
             /* if disabling set mode, update the time */
-            if ( s->hw.cmos_data[RTC_REG_B] & RTC_SET )
+            if ( orig & RTC_SET )
                 rtc_set_time(s);
         }
-        /* if the interrupt is already set when the interrupt become
-         * enabled, raise an interrupt immediately*/
-        if ((data & RTC_UIE) && !(s->hw.cmos_data[RTC_REG_B] & RTC_UIE))
-            if (s->hw.cmos_data[RTC_REG_C] & RTC_UF)
+        /*
+         * If the interrupt is already set when the interrupt becomes
+         * enabled, raise an interrupt immediately.
+         * NB: RTC_{A,P,U}IE == RTC_{A,P,U}F respectively.
+         */
+        for ( mask = RTC_UIE; mask <= RTC_PIE; mask <<= 1 )
+            if ( (data & mask) && !(orig & mask) &&
+                 (s->hw.cmos_data[RTC_REG_C] & mask) )
             {
-                hvm_isa_irq_deassert(d, RTC_IRQ);
-                hvm_isa_irq_assert(d, RTC_IRQ);
+                rtc_toggle_irq(s);
+                break;
             }
         s->hw.cmos_data[RTC_REG_B] = data;
-        rtc_timer_update(s);
+        if ( (data ^ orig) & RTC_PIE )
+            rtc_timer_update(s);
         check_update_timer(s);
-        alarm_timer_update(s);
+        if ( (data ^ orig) & RTC_AIE )
+            alarm_timer_update(s);
         break;
     case RTC_REG_C:
     case RTC_REG_D:
--- a/xen/arch/x86/hvm/vpt.c
+++ b/xen/arch/x86/hvm/vpt.c
@@ -22,6 +22,7 @@
 #include <asm/hvm/vpt.h>
 #include <asm/event.h>
 #include <asm/apic.h>
+#include <asm/mc146818rtc.h>
 
 #define mode_is(d, name) \
     ((d)->arch.hvm_domain.params[HVM_PARAM_TIMER_MODE] == HVMPTM_##name)
@@ -218,6 +219,7 @@ void pt_update_irq(struct vcpu *v)
     struct periodic_time *pt, *temp, *earliest_pt = NULL;
     uint64_t max_lag = -1ULL;
     int irq, is_lapic;
+    void *pt_priv;
 
     spin_lock(&v->arch.hvm_vcpu.tm_lock);
 
@@ -251,13 +253,14 @@ void pt_update_irq(struct vcpu *v)
     earliest_pt->irq_issued = 1;
     irq = earliest_pt->irq;
     is_lapic = (earliest_pt->source == PTSRC_lapic);
+    pt_priv = earliest_pt->priv;
 
     spin_unlock(&v->arch.hvm_vcpu.tm_lock);
 
     if ( is_lapic )
-    {
         vlapic_set_irq(vcpu_vlapic(v), irq, 0);
-    }
+    else if ( irq == RTC_IRQ )
+        rtc_periodic_interrupt(pt_priv);
     else
     {
         hvm_isa_irq_deassert(v->domain, irq);
--- a/xen/include/asm-x86/hvm/vpt.h
+++ b/xen/include/asm-x86/hvm/vpt.h
@@ -181,6 +181,7 @@ void rtc_migrate_timers(struct vcpu *v);
 void rtc_deinit(struct domain *d);
 void rtc_reset(struct domain *d);
 void rtc_update_clock(struct domain *d);
+void rtc_periodic_interrupt(void *);
 
 void pmtimer_init(struct vcpu *v);
 void pmtimer_deinit(struct domain *d);

Attachment: x86-hvm-rtc.patch
Description: Text document

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

 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.