Re: [PATCH] 2.6, 2.4, Nforce2, Experimental idle halt workaroundinstead of apic ack delay.

[Arjen Verweij]

Lately I am trying to find out why my CPU runs so much hotter in W2K
(50ish Celsius) than it does with Linux (36 Celsius) when idle. I am
wondering if Microsoft secretly disables the disconnect :)

On the linux front I am playing with 2.6.3-mm patched aditionally with
nforce2-timer and apic-tack. There is no need for Maciej's patch, because
Andrew was nice enough to include this in -mm.

Personally, I would like to figure out a way of keeping the C1 disconnect,
because I like a silent system and there isn't much cooling fannoise
required when my cpu is at 36 deg. than when it is at 50+ idle.

I gather that not everyone has these big differences, so it may not be
interesting to relentlessly pursue C1 like I am :P

Something else that strikes me as odd, is that 2.6.3-mm like I built it,
immediately locks up solid if you omit apic_tack=2, but 2.4 with acpi and
apic enabled, without Ross' patches would not react this vigorously.


Regards,

Arjen


On Wed, 25 Feb 2004, Ross Dickson wrote:

> On Monday 23 February 2004 11:37, Prakash K. Cheemplavam wrote:
> > Oh before I forget, I had to resolve a reject by hand, but I *think* I
> > did it right. (And yes, I used your corrected versionof the patch.)
> > Well, maybe you take a look over the patch and rediff. :-)
> >
> > Prakash
> >
> >
> >
>
> Hi Prakash,
> Patches attached rediffed for 2.6.3 and 2.6.3-mm3.
>
> Late comers see start of this thread for more detail.
> Also Arjen's page at http://atlas.et.tudelft.nl/verwei90/nforce2/index.html
>
> Note that the ioapic patch is not essential for 2.6.3-mm3, only if you want
> to use nmi_debug=1.
> 2.6.3-mm3 sets up 8254 timer OK in a virtual wire mode without this patch.
>
> It will be interesting to see if there is any clock skew associated with these
> patches, I get no skew on 2.4.24.
>
> The kernel arg to invoke the idle halt workaround is still "idle=C1halt".
> I have increased slightly the ndelay time from 500ns to 600ns - I had
> some lockups on startup and shutdown with certain removable IDE drives,
> this seemed to help. Feel free to adjust the value if you want to.
>
> Also if you want snappier performance you can try different values instead of 6
> in this line:
> +			(apic_read(APIC_TMCCT) > (apic_read(APIC_TMICT)>>6))) {
>
> The 6 yields a divide by 2^6 = 64 for the 1.6% (1/64 = 1.6%)
>
> If you try say 5 for 2^5 = 32 for 3.1% or 4 for 6% or 3 for 12.5%
> then if less than this percentage of time remains in the slice the system won't
> go into hlt disconnect thus avoiding the reconnect - recovery time associated
> with that cycle that would have been.
>
> I have found a need for less ndelay time with the more the percentage
> (lower number)- there appears to be a tradeoff between the two.
>
> 3 for 12.5% or 2 for 25% might even work with no ndelay time i.e. the ndelay
> line removed or commented from the code, for some.
>
> Of course expect a bit more heat under moderate loads with an increase in
> the percentage but some may prefer it for their application.
>
> Happy hacking,
> Ross.
>
>
> For 2.6.3 ioapic
> ---CUT HERE---
> --- linux-2.6.3/arch/i386/kernel/io_apic.c.original	2004-02-18 13:57:22.000000000 +1000
> +++ linux-2.6.3/arch/i386/kernel/io_apic.c	2004-02-24 11:57:04.000000000 +1000
> @@ -2188,12 +2188,56 @@ static inline void check_timer(void)
>  				check_nmi_watchdog();
>  			}
>  			return;
>  		}
>  		clear_IO_APIC_pin(0, pin1);
> -		printk(KERN_ERR "..MP-BIOS bug: 8254 timer not connected to IO-APIC\n");
> +		printk(KERN_ERR "..MP-BIOS bug: 8254 timer not connected to IO-APIC INTIN%d\n",pin1);
> +	}
> +
> +#if defined(CONFIG_ACPI_BOOT) && defined(CONFIG_X86_UP_IOAPIC)
> +	/* for nforce2 try vector 0 on pin0
> +	 * Note 8259a is already masked, also by default
> +	 * the io_apic_set_pci_routing call disables the 8259 irq 0
> +	 * so we must be connected directly to the 8254 timer if this works
> +	 * Note2: this violates the above comment re Subtle but works!
> +	 */
> +	printk(KERN_INFO "..TIMER: Is timer irq0 connected to IO-APIC INTIN0? ...\n");
> +	if (pin1 != -1) {
> +		extern spinlock_t i8259A_lock;
> +		unsigned long flags;
> +		int tok, saved_timer_ack = timer_ack;
> +		/*
> +		 * Ok, does IRQ0 through the IOAPIC work?
> +		 */
> +		io_apic_set_pci_routing ( 0, 0, 0, 0, 0); /* connect pin */
> +		unmask_IO_APIC_irq(0);
> +		timer_ack = 0;
> +
> +		/*
> +		 * Ok, does IRQ0 through the IOAPIC work?
> +		 */
> +		spin_lock_irqsave(&i8259A_lock, flags);
> +		Dprintk("..TIMER 8259A ints disabled?, imr1:%02x, imr2:%02x\n", inb(0x21), inb(0xA1));
> +		tok = timer_irq_works();
> +		spin_unlock_irqrestore(&i8259A_lock, flags);
> +		if (tok) {
> +			if (nmi_watchdog == NMI_IO_APIC) {
> +				disable_8259A_irq(0);
> +				setup_nmi();
> +				enable_8259A_irq(0);
> +				check_nmi_watchdog();
> +			}
> +			printk(KERN_INFO "..TIMER: works OK on IO-APIC INTIN0 irq0\n" );
> +			return;
> +		}
> +		/* failed */
> +		timer_ack = saved_timer_ack;
> +		clear_IO_APIC_pin(0, 0);
> +		io_apic_set_pci_routing ( 0, pin1, 0, 0, 0);
> +		printk(KERN_ERR "..MP-BIOS: 8254 timer not connected to IO-APIC INTIN0\n");
>  	}
> +#endif
>
>  	printk(KERN_INFO "...trying to set up timer (IRQ0) through the 8259A ... ");
>  	if (pin2 != -1) {
>  		printk("\n..... (found pin %d) ...", pin2);
>  		/*
> ---CUT HERE---
>
>
> For 2.6.3 idleC1halt
> ---CUT HERE---
> --- linux-2.6.3/arch/i386/kernel/process.c.original	2004-02-18 13:57:11.000000000 +1000
> +++ linux-2.6.3/arch/i386/kernel/process.c	2004-02-24 11:25:56.000000000 +1000
> @@ -48,10 +48,11 @@
>  #include <asm/irq.h>
>  #include <asm/desc.h>
>  #ifdef CONFIG_MATH_EMULATION
>  #include <asm/math_emu.h>
>  #endif
> +#include <asm/apic.h>
>
>  #include <linux/irq.h>
>  #include <linux/err.h>
>
>  asmlinkage void ret_from_fork(void) __asm__("ret_from_fork");
> @@ -87,11 +88,11 @@ EXPORT_SYMBOL(enable_hlt);
>
>  /*
>   * We use this if we don't have any better
>   * idle routine..
>   */
> -void default_idle(void)
> +static void default_idle(void)
>  {
>  	if (!hlt_counter && current_cpu_data.hlt_works_ok) {
>  		local_irq_disable();
>  		if (!need_resched())
>  			safe_halt();
> @@ -99,10 +100,29 @@ void default_idle(void)
>  			local_irq_enable();
>  	}
>  }
>
>  /*
> + * We use this to avoid nforce2 lockups
> + * Reduces frequency of C1 disconnects
> + */
> +static void c1halt_idle(void)
> +{
> +	if (!hlt_counter && current_cpu_data.hlt_works_ok) {
> +		local_irq_disable();
> +		/* only hlt disconnect if more than 1.6% of apic interval remains */
> +		if( (!need_resched()) &&
> +			(apic_read(APIC_TMCCT) > (apic_read(APIC_TMICT)>>6))) {
> +			ndelay(600); /* helps nforce2 but adds 0.6us hard int latency */
> +			safe_halt();  /* nothing better to do until we wake up */
> +		}
> +		else
> +			local_irq_enable();
> +	}
> +}
> +
> +/*
>   * On SMP it's slightly faster (but much more power-consuming!)
>   * to poll the ->work.need_resched flag instead of waiting for the
>   * cross-CPU IPI to arrive. Use this option with caution.
>   */
>  static void poll_idle (void)
> @@ -136,20 +156,18 @@ static void poll_idle (void)
>   * The idle thread. There's no useful work to be
>   * done, so just try to conserve power and have a
>   * low exit latency (ie sit in a loop waiting for
>   * somebody to say that they'd like to reschedule)
>   */
> +static void (*idle)(void);
>  void cpu_idle (void)
>  {
>  	/* endless idle loop with no priority at all */
>  	while (1) {
>  		while (!need_resched()) {
> -			void (*idle)(void) = pm_idle;
> -
>  			if (!idle)
> -				idle = default_idle;
> -
> +				idle = pm_idle ? pm_idle : default_idle;
>  			irq_stat[smp_processor_id()].idle_timestamp = jiffies;
>  			idle();
>  		}
>  		schedule();
>  	}
> @@ -200,16 +218,18 @@ void __init select_idle_routine(const st
>
>  static int __init idle_setup (char *str)
>  {
>  	if (!strncmp(str, "poll", 4)) {
>  		printk("using polling idle threads.\n");
> -		pm_idle = poll_idle;
> +		idle = poll_idle;
>  	} else if (!strncmp(str, "halt", 4)) {
>  		printk("using halt in idle threads.\n");
> -		pm_idle = default_idle;
> +		idle = default_idle;
> +	} else if (!strncmp(str, "C1halt", 6)) {
> +		printk("using C1 halt disconnect friendly idle threads.\n");
> +		idle = c1halt_idle;
>  	}
> -
>  	return 1;
>  }
>
>  __setup("idle=", idle_setup);
>
> ---CUT HERE---
>
> For 2.6.3-mm3 ioapic
> ---CUT HERE---
> --- linux-2.6.3-mm3/arch/i386/kernel/io_apic.c.original	2004-02-24 12:26:32.000000000 +1000
> +++ linux-2.6.3-mm3/arch/i386/kernel/io_apic.c	2004-02-24 16:40:56.000000000 +1000
> @@ -2206,12 +2206,54 @@ static inline void check_timer(void)
>  					timer_ack = !cpu_has_tsc;
>  			}
>  			return;
>  		}
>  		clear_IO_APIC_pin(0, pin1);
> -		printk(KERN_ERR "..MP-BIOS bug: 8254 timer not connected to IO-APIC\n");
> +		printk(KERN_ERR "..MP-BIOS bug: 8254 timer not connected to IO-APIC INTIN%d\n",pin1);
> +	}
> +
> +#if defined(CONFIG_ACPI_BOOT) && defined(CONFIG_X86_UP_IOAPIC)
> +	/* for nforce2 try vector 0 on pin0
> +	 * Note 8259a is already masked, also by default
> +	 * the io_apic_set_pci_routing call disables the 8259 irq 0
> +	 * so we must be connected directly to the 8254 timer if this works
> +	 */
> +	printk(KERN_INFO "..TIMER: Is timer irq0 connected to IO-APIC INTIN0? ...\n");
> +	if (pin1 != -1) {
> +		extern spinlock_t i8259A_lock;
> +		unsigned long flags;
> +		int tok;
> +		/*
> +		 * Ok, does IRQ0 through the IOAPIC work?
> +		 */
> +		io_apic_set_pci_routing ( 0, 0, 0, 0, 0); /* connect pin */
> +		unmask_IO_APIC_irq(0);
> +
> +		/*
> +		 * Ok, does IRQ0 through the IOAPIC work?
> +		 */
> +		spin_lock_irqsave(&i8259A_lock, flags);
> +		Dprintk("..TIMER 8259A ints disabled?, imr1:%02x, imr2:%02x\n", inb(0x21), inb(0xA1));
> +		tok = timer_irq_works();
> +		spin_unlock_irqrestore(&i8259A_lock, flags);
> +		if (tok) {
> +			if (nmi_watchdog == NMI_IO_APIC) {
> +				disable_8259A_irq(0);
> +				setup_nmi();
> +				enable_8259A_irq(0);
> +				if (check_nmi_watchdog() < 0);
> +					timer_ack = !cpu_has_tsc;
> +			}
> +			printk(KERN_INFO "..TIMER: works OK on IO-APIC INTIN0 irq0\n" );
> +			return;
> +		}
> +		/* failed */
> +		clear_IO_APIC_pin(0, 0);
> +		io_apic_set_pci_routing ( 0, pin1, 0, 0, 0);
> +		printk(KERN_ERR "..MP-BIOS: 8254 timer not connected to IO-APIC INTIN0\n");
>  	}
> +#endif
>
>  	printk(KERN_INFO "...trying to set up timer (IRQ0) through the 8259A ... ");
>  	if (pin2 != -1) {
>  		printk("\n..... (found pin %d) ...", pin2);
>  		/*
> ---CUT HERE---
>
> For 2.6.3-mm3 idleC1halt
> ---CUT HERE---
> --- linux-2.6.3-mm3/arch/i386/kernel/process.c.original	2004-02-24 12:26:32.000000000 +1000
> +++ linux-2.6.3-mm3/arch/i386/kernel/process.c	2004-02-24 15:54:26.000000000 +1000
> @@ -49,10 +49,11 @@
>  #include <asm/desc.h>
>  #include <asm/atomic_kmap.h>
>  #ifdef CONFIG_MATH_EMULATION
>  #include <asm/math_emu.h>
>  #endif
> +#include <asm/apic.h>
>
>  #include <linux/irq.h>
>  #include <linux/err.h>
>
>  asmlinkage void ret_from_fork(void) __asm__("ret_from_fork");
> @@ -88,11 +89,11 @@ EXPORT_SYMBOL(enable_hlt);
>
>  /*
>   * We use this if we don't have any better
>   * idle routine..
>   */
> -void default_idle(void)
> +static void default_idle(void)
>  {
>  	if (!hlt_counter && current_cpu_data.hlt_works_ok) {
>  		local_irq_disable();
>  		if (!need_resched())
>  			safe_halt();
> @@ -100,10 +101,29 @@ void default_idle(void)
>  			local_irq_enable();
>  	}
>  }
>
>  /*
> + * We use this to avoid nforce2 lockups
> + * Reduces frequency of C1 disconnects
> + */
> +static void c1halt_idle(void)
> +{
> +	if (!hlt_counter && current_cpu_data.hlt_works_ok) {
> +		local_irq_disable();
> +		/* only hlt disconnect if more than 1.6% of apic interval remains */
> +		if( (!need_resched()) &&
> +			(apic_read(APIC_TMCCT) > (apic_read(APIC_TMICT)>>6))) {
> +			ndelay(600); /* helps nforce2 but adds 0.6us hard int latency */
> +			safe_halt();  /* nothing better to do until we wake up */
> +		}
> +		else
> +			local_irq_enable();
> +	}
> +}
> +
> +/*
>   * On SMP it's slightly faster (but much more power-consuming!)
>   * to poll the ->work.need_resched flag instead of waiting for the
>   * cross-CPU IPI to arrive. Use this option with caution.
>   */
>  static void poll_idle (void)
> @@ -137,20 +157,18 @@ static void poll_idle (void)
>   * The idle thread. There's no useful work to be
>   * done, so just try to conserve power and have a
>   * low exit latency (ie sit in a loop waiting for
>   * somebody to say that they'd like to reschedule)
>   */
> +static void (*idle)(void);
>  void cpu_idle (void)
>  {
>  	/* endless idle loop with no priority at all */
>  	while (1) {
>  		while (!need_resched()) {
> -			void (*idle)(void) = pm_idle;
> -
>  			if (!idle)
> -				idle = default_idle;
> -
> +				idle = pm_idle ? pm_idle : default_idle;
>  			irq_stat[smp_processor_id()].idle_timestamp = jiffies;
>  			idle();
>  		}
>  		schedule();
>  	}
> @@ -201,16 +219,18 @@ void __init select_idle_routine(const st
>
>  static int __init idle_setup (char *str)
>  {
>  	if (!strncmp(str, "poll", 4)) {
>  		printk("using polling idle threads.\n");
> -		pm_idle = poll_idle;
> +		idle = poll_idle;
>  	} else if (!strncmp(str, "halt", 4)) {
>  		printk("using halt in idle threads.\n");
> -		pm_idle = default_idle;
> +		idle = default_idle;
> +	} else if (!strncmp(str, "C1halt", 6)) {
> +		printk("using C1 halt disconnect friendly idle threads.\n");
> +		idle = c1halt_idle;
>  	}
> -
>  	return 1;
>  }
>
>  __setup("idle=", idle_setup);
>
> ---CUT HERE---
>
> Attached patches also in tarballs if whitespace problems.
>

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/