On Mon, May 18, 2015 at 02:40:56PM +0300, Andriy Gapon wrote: > On 18/05/2015 14:29, Andriy Gapon wrote: > > On 09/05/2015 15:28, Konstantin Belousov wrote: > >> +void > >> +acpi_cpu_idle_mwait(uint32_t mwait_hint) > >> +{ > >> + int *state; > >> + > >> + state = (int *)PCPU_PTR(monitorbuf); > >> + /* > >> + * XXXKIB. Software coordination mode should be supported, > >> + * but all Intel CPUs provide hardware coordination. > >> + */ > >> + cpu_monitor(state, 0, 0); > >> + cpu_mwait(MWAIT_INTRBREAK, mwait_hint); > >> +} > >> + > > > > Kostik, > > > > it's been a while since I studied this code, so please pardon me if I am > > asking > > something obvious or silly. > > > > I wonder why this function does not set 'state' before monitor + mwait. > > As far as I can see, all other idling functions do that. And > > cpu_idle_wakeup() > > compares the state to STATE_MWAIT before changing it. > > > > So, I am concerned that if the state happens to be anything other than > > STATE_MWAIT when acpi_cpu_idle_mwait() is called, then cpu_idle_wakeup() > > won't > > wake up the idled CPU. It seems that if the state is not STATE_SLEEPING > > then an > > IPI won't be sent either. Actually, that leaves STATE_RUNNING is the only > > problematic case, but that's probably the state that the CPU would have > > before > > idling. > > > > After having written the above I realized what I overlooked: > acpi_cpu_idle_mwait() is called from the ACPI idle method, so the state must > already be set to STATE_SLEEPING. > So, looks like the wake-up would always be done by an IPI... > > Just in case, here's what I had in my old local code: > void > acpi_cpu_mwait_cx(u_int hints) > { > int *state; > > state = (int *)PCPU_PTR(monitorbuf); > KASSERT(*state == STATE_SLEEPING, > ("cpu_mwait_cx: wrong monitorbuf state")); > *state = STATE_MWAIT; > cpu_monitor(state, 0, 0); > if (*state == STATE_MWAIT) > cpu_mwait(MWAIT_INTR_BRK, hints); > > /* > * We should exit on any event that interrupts mwait, > * because that event might be a wanted interrupt. > */ > *state = STATE_RUNNING; > } > > This code also accounted for a time window between the CPU wanting to go idle > and it calling mwait. During that window other CPU could want to wake up the > first CPU.
Right. And there is very similar, but having some small important difference, function cpu_idle_mwait(). My plan was to give the commit some time for wider testing, then axe cpu_idle_mwait and corresponding machdep.idle selector, with the move of the fast wakeup code under acpi_cpu_idle_mwait(). Additional reason to wait was the use of cpu_idle_mwait() as the 'fast' idle method for busy CPU, whatever this means. I think I would commit the patch below shortly, with 'Submitted by: avg'. diff --git a/sys/x86/x86/cpu_machdep.c b/sys/x86/x86/cpu_machdep.c index 7925713..f07b97e 100644 --- a/sys/x86/x86/cpu_machdep.c +++ b/sys/x86/x86/cpu_machdep.c @@ -101,6 +101,10 @@ __FBSDID("$FreeBSD$"); #include <vm/vm_pager.h> #include <vm/vm_param.h> +#define STATE_RUNNING 0x0 +#define STATE_MWAIT 0x1 +#define STATE_SLEEPING 0x2 + /* * Machine dependent boot() routine * @@ -134,13 +138,24 @@ acpi_cpu_idle_mwait(uint32_t mwait_hint) { int *state; - state = (int *)PCPU_PTR(monitorbuf); /* * XXXKIB. Software coordination mode should be supported, * but all Intel CPUs provide hardware coordination. */ + + state = (int *)PCPU_PTR(monitorbuf); + KASSERT(*state == STATE_SLEEPING, + ("cpu_mwait_cx: wrong monitorbuf state")); + *state = STATE_MWAIT; cpu_monitor(state, 0, 0); - cpu_mwait(MWAIT_INTRBREAK, mwait_hint); + if (*state == STATE_MWAIT) + cpu_mwait(MWAIT_INTRBREAK, mwait_hint); + + /* + * We should exit on any event that interrupts mwait, because + * that event might be a wanted interrupt. + */ + *state = STATE_RUNNING; } /* Get current clock frequency for the given cpu id. */ @@ -231,10 +246,6 @@ static int idle_mwait = 1; /* Use MONITOR/MWAIT for short idle. */ SYSCTL_INT(_machdep, OID_AUTO, idle_mwait, CTLFLAG_RWTUN, &idle_mwait, 0, "Use MONITOR/MWAIT for short idle"); -#define STATE_RUNNING 0x0 -#define STATE_MWAIT 0x1 -#define STATE_SLEEPING 0x2 - #ifndef PC98 static void cpu_idle_acpi(sbintime_t sbt) _______________________________________________ svn-src-all@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/svn-src-all To unsubscribe, send any mail to "svn-src-all-unsubscr...@freebsd.org"