x86/hvm/pmtimer: improving scalability of virtual time update

Mitigate the heavy contention on handle_pmt_io when running a HVM
configured with many cores (e.g., 32 cores). As the virtual time of a
domain must be fresh, there should be someone updating it,
periodically. But it is not necessary to let a VCPU update the virtual
time when another one has been updating it. Thus the update can be
skipped when the VCPU finds someone else is updating the virtual time.
So every time a VCPU invoke handle_pmt_io to update the current
domain's virtual time, it will first try to acquire the pmtimer lock.
If it succeeds, it will update the virtual time. Otherwise, it can
skip the update, waits for the pmtimer lock holder to finish updating
the virtual time and returns the updated time.

Signed-off-by: Xiang Song <xiangsong@fudan.edu.cn>
Signed-off-by: Keir Fraser <keir@xen.org>

diff --git a/xen/arch/x86/hvm/pmtimer.c b/xen/arch/x86/hvm/pmtimer.c
index 2070525abe281b585354e7076eae385329ffedf0..01ae31d38dde1b9b6e321bb6cf58cd78d559d1a4 100644 (file)
--- a/xen/arch/x86/hvm/pmtimer.c
+++ b/xen/arch/x86/hvm/pmtimer.c
@@ -88,7 +88,7 @@ void hvm_acpi_sleep_button(struct domain *d)
 static void pmt_update_time(PMTState *s)
 {
     uint64_t curr_gtime, tmp;
-    uint32_t msb = s->pm.tmr_val & TMR_VAL_MSB;
+    uint32_t tmr_val = s->pm.tmr_val, msb = tmr_val & TMR_VAL_MSB;
     
     ASSERT(spin_is_locked(&s->lock));
 
@@ -96,12 +96,15 @@ static void pmt_update_time(PMTState *s)
     curr_gtime = hvm_get_guest_time(s->vcpu);
     tmp = ((curr_gtime - s->last_gtime) * s->scale) + s->not_accounted;
     s->not_accounted = (uint32_t)tmp;
-    s->pm.tmr_val += tmp >> 32;
-    s->pm.tmr_val &= TMR_VAL_MASK;
+    tmr_val += tmp >> 32;
+    tmr_val &= TMR_VAL_MASK;
     s->last_gtime = curr_gtime;
-    
+
+    /* Update timer value atomically wrt lock-free reads in handle_pmt_io(). */
+    *(volatile uint32_t *)&s->pm.tmr_val = tmr_val;
+
     /* If the counter's MSB has changed, set the status bit */
-    if ( (s->pm.tmr_val & TMR_VAL_MSB) != msb )
+    if ( (tmr_val & TMR_VAL_MSB) != msb )
     {
         s->pm.pm1a_sts |= TMR_STS;
         pmt_update_sci(s);
@@ -215,10 +218,23 @@ static int handle_pmt_io(
     
     if ( dir == IOREQ_READ )
     {
-        spin_lock(&s->lock);
-        pmt_update_time(s);
-        *val = s->pm.tmr_val;
-        spin_unlock(&s->lock);
+        if ( spin_trylock(&s->lock) )
+        {
+            /* We hold the lock: update timer value and return it. */
+            pmt_update_time(s);
+            *val = s->pm.tmr_val;
+            spin_unlock(&s->lock);
+        }
+        else
+        {
+            /*
+             * Someone else is updating the timer: rather than do the work
+             * again ourselves, wait for them to finish and then steal their
+             * updated value with a lock-free atomic read.
+             */
+            spin_barrier(&s->lock);
+            *val = *(volatile uint32_t *)&s->pm.tmr_val;
+        }
         return X86EMUL_OKAY;
     }