1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
|
Without the first part of this patch, any waits for periods shorter than a
single tick return immediately leading to a lot of unnecessary spinning. For
example, I observe that my guest's idle loop does a lot of sleeps with periods
slightly shorter than 1 ms (1/hz), e.g. 900us. All that waiting turns into pure
spinning and VirtualBox eats 100% of a core.
The clamping improves the situation significantly. Also, it (approximately)
follows what tvtohz does. The rest of the patch just chases an upstream
KPI change.
Submitted by: Andriy Gapon <avg@FreeBSD.org>
--- src/VBox/Runtime/r0drv/freebsd/sleepqueue-r0drv-freebsd.h.orig 2025-08-13 19:51:51 UTC
+++ src/VBox/Runtime/r0drv/freebsd/sleepqueue-r0drv-freebsd.h
@@ -94,6 +94,8 @@ DECLINLINE(uint32_t) rtR0SemBsdWaitUpdateTimeout(PRTR0
uint64_t cTicks = ASMMultU64ByU32DivByU32(uTimeout, hz, UINT32_C(1000000000));
if (cTicks >= INT_MAX)
return RTSEMWAIT_FLAGS_INDEFINITE;
+ else if (cTicks == 0 && uTimeout > 0)
+ pWait->iTimeout = 1;
else
pWait->iTimeout = (int)cTicks;
#endif
@@ -310,10 +312,16 @@ DECLINLINE(void) rtR0SemBsdSignal(void *pvWaitChan)
DECLINLINE(void) rtR0SemBsdSignal(void *pvWaitChan)
{
sleepq_lock(pvWaitChan);
+#if __FreeBSD_version < 1500022
int fWakeupSwapProc = sleepq_signal(pvWaitChan, SLEEPQ_CONDVAR, 0, 0);
+#else
+ sleepq_signal(pvWaitChan, SLEEPQ_CONDVAR, 0, 0);
+#endif
sleepq_release(pvWaitChan);
+#if __FreeBSD_version < 1500022
if (fWakeupSwapProc)
kick_proc0();
+#endif
}
/**
|
