Re: [PATCH] x86, UV: Improve access time to RTC clock in nodecontroller

[Ingo Molnar]

* Jack Steiner <steiner@xxxxxxx> wrote:

> UV has a globally synchronized clock (UV RTC) that is located in the node
> controller chip of each blade. Although any cpu can access it's own RTC
> without making offnode references, access to a register in the node
> controller still requires a few hundred nsec.  This is problematic for
> some workloads.
> 
> This patch changes the clock code so that the RTC value is calculated
> based on the in-core TSC clock. Access to the TSC is much faster than
> access to the RTC.
> 	OLD: 292 nsec
> 	NEW:  18 nsec

Impressive speedup!

> 
> 
> Periodically, a pair of TSC/RTC values is read. New requests to read the
> RTC will read the current TSC, then calculate an RTC value
> based on the TSC delta from the last sync point and the ratio of TSC to
> RTC frequencies.
> 
> Signed-off-by: Jack Steiner <steiner@xxxxxxx>
> Reviewed-by: Robin Holt <holt@xxxxxxx>
> 
> ---
>  arch/x86/kernel/uv_time.c |  494 ++++++++++++++++++++++++++++++++++++++++++++--
>  1 file changed, 473 insertions(+), 21 deletions(-)
> 
> Index: linux/arch/x86/kernel/uv_time.c
> ===================================================================
> --- linux.orig/arch/x86/kernel/uv_time.c	2010-09-09 14:02:09.000000000 -0500
> +++ linux/arch/x86/kernel/uv_time.c	2010-09-09 14:03:32.302037155 -0500
> @@ -15,12 +15,14 @@
>   *  along with this program; if not, write to the Free Software
>   *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
>   *
> - *  Copyright (c) 2009 Silicon Graphics, Inc.  All Rights Reserved.
> + *  Copyright (c) 2009-2010 Silicon Graphics, Inc.  All Rights Reserved.
>   *  Copyright (c) Dimitri Sivanich
>   */
>  #include <linux/clockchips.h>
>  #include <linux/slab.h>
> +#include <linux/sysfs.h>
>  
> +#include <asm/timer.h>
>  #include <asm/uv/uv_mmrs.h>
>  #include <asm/uv/uv_hub.h>
>  #include <asm/uv/bios.h>
> @@ -29,8 +31,137 @@
>  #include <asm/cpu.h>
>  
>  #define RTC_NAME		"sgi_rtc"
> +#define RTC_NAME_CALC		"sgi_rtc_calc"
> +
> +/*
> + * On UV, the latency to read the hardware RTC clock is ~300 nsec.
> + * Numerous workloads show this to be a performance issue.
> + *
> + * The following code changes the RTC access code to read the TSC
> + * instead of the RTC. Based on the current TSC and the ratio
> + * of TSC to RTC frequency, the code calculates the RTC value.
> + *
> + * The following are the requirements for the algorithms that
> + * calculate the RTC value:
> + *	- a cpu will never see it's own RTC value go backward
> + *	- the error in the calculation of the RTC value must be
> + *	  less than the time required to bounce a cacheline between
> + *	  cpus. For example cpu A must not be allowed to
> + *		- read a RTC value
> + *		- write it to memory
> + *		- have cpu B read A's RTC value
> + *		- cpu B read it's own RTC value
> + *		- B's RTC is less than A's
> + *	  Or saying this another way, no pair of cpus can
> + *	  perceive time to go backward.
> + *
> + * Assumptions:
> + *	- RTCs are synchronized across all blades. Any cpu reading
> + *	  it's own RTC will see the same value.
> + *	- TSCs on a blade are synchronized
> + *	- TSCs on different blades are NOT synchronized
> + *	- TSCs on different blades may run at different frequencies
> + *	  (UV supports SSIs with mixed processor types & running
> + *	  at different core frequencies)
> + *	- clock frequencies of TSC vs RTC may drift but at a
> + *	  slow rate. No significant change over a period of a
> + *	  few 10's of seconds.
> + *	- don't do divide instructions in the critical path
> + *
> + * Tables:
> + *
> + *      RTC Blade Control (RBC) - Per blade tables allocated blade local
> + *                      (2 processor sockets on each blade)
> + *
> + *               +-----------+                  +-----------+
> + *               '    RBC    '                  '   RBC     '
> + *               '           '                  '           '
> + *               +-----------+                  +-----------+
> + *                ^    ^   ^                     ^    ^    ^
> + *               /     |    \                   /     |     \
> + *              /      |     \                 /      |      \
> + *             /       |      \               /       |       \
> + *        +-----+   +------  -------      +-----+   +-----+   +-----+
> + *        ' RCC '   ' RCC '  ' RCC '      ' RCC '   ' RCC '   ' RCC '
> + *        +-----+   +-----+  +-----+      +-----+   +-----+   +-----+
> + *
> + *      RTC Cpu Control (RCC) - Per-cpu tables allocated socket-local
> + *
> + *
> + *   RBC contain per-blade state:
> + *	- recent RTC/TSC pair. Current RTC value is
> + *	  calculated by reading the live TSC, then calculating
> + *	  the current RTC value based on the TSC delta.
> + *	- fixed-point (32.32) multiplier:
> + *		rtc_delta = (tsc delta) * fixmult
> + *	- RTC/TSC pair for time when fixmult was last calculated
> + *	- TSC intervals for resyncing the recent RTC/TSC pair
> + *	  or for recalculating the fixmult multiplier
> + *
> + *   RCC contain per-cpu state for estimating the RTC value
> + *	- pointer to the RBC structure for all cpus on the blade
> + *	- last_rtc_value returned on this cpu. This is used to
> + *	  make sure time never goes backward.
> + *
> + *  Implementation Notes:
> + *	- I tried making the RBC table a per-socket table instead of a
> + *	  per-blade table. That did not work & a pair of cpus could see
> + *	  time go backward. Cachelines bounce between sockets faster
> + *	  than the the accuracy of the RTC calculation.
> + *	- the most critical inter-cpu timing is for passing RTC
> + *	  values between HT threads in the same core. Other inter-cpu
> + *	  times are much larger.
> + */
> +
> +/*
> + * The following structures are used to manage the per-node UV RTC
> + * structures used to calculate RTC values from the local RTC. Access
> + * to the local TSC is much faster than a RTC access.
> + *
> + */
> +
> +/* Maximum attempts to read a TSC/RTC pair before giving up */
> +#define MAXREADS	 			10
> +
> +/* Maximum TSC nsec to read a TSC/RTC pair. Try again if exceeded. */
> +#define DEFAULT_MAX_READCLOCKS_TSC_NSEC		10000
> +
> +/* Reread a TSC/RTC pair if it is older than the following. */
> +#define DEFAULT_RESYNC_CLOCKS_MSEC		2000
> +
> +/* Recalculate fixmult if it is older than the following. */
> +#define DEFAULT_RECALC_DRIFT_MSEC		20000
> +
> +static long resync_clocks_msec = DEFAULT_RESYNC_CLOCKS_MSEC;
> +static long recalc_drift_msec = DEFAULT_RECALC_DRIFT_MSEC;
> +static long max_readclocks_tsc_nsec = DEFAULT_MAX_READCLOCKS_TSC_NSEC;
> +
> +static atomic_t resync_clocks_failures;

Do you want all these to end up the same cacheline and bounce around nodes?

If not then judicious application of __read_mostly and 
__cacheline_aligned ought to cure it.

> +
> +#define ZZMAXDELTA	1000		//ZZZ debug
> +atomic_t zzdelta[ZZMAXDELTA + 1];	//ZZZ debug

We dont want any of the ZZZ things upstream, right?

>  /*
> + * Atomic cmpxchg of a 128-bit value (pair of DW)
> + */
> +static inline int uv_cmpxchg128(void *p, unsigned long oldhi, unsigned long oldlo,
> +			unsigned long newhi, unsigned long newlo)
> +{
> +	unsigned char ret;
> +
> +	asm volatile(   "lock; cmpxchg16b (%1)\n\t"
> +			"sete %0\n\t"
> +			: "=qm" (ret)
> +			: "r"(p), "d"(oldhi), "a"(oldlo), "c"(newhi), "b"(newlo)
> +			: "memory", "cc");
> +	return ret;
> +}

That's a rather interesting piece of code that is not UV specific at 
all, so it should move into the appropriate arch/x86/include/asm/ header 
file.

> +
> +

-ETOOMANYNEWLINES

> +/*
> + * Atomic read of 128 bit value (pair of DW)
> + */
> +static inline void uv_read128(unsigned long *wd,
> +				unsigned long *hi, unsigned long *lo)

Should move into a header file too.

> +{
> +	unsigned long t0, t1, t0x;
> +
> +	do {
> +		barrier();
> +		t0 = wd[0];
> +		barrier();
> +		t1 = wd[1];
> +		barrier();
> +		t0x = wd[0];
> +	} while (t0 != t0x);
> +	*hi = t1;
> +	*lo = t0;
> +}
> +
> +/*
> + * Convert a nsec/msec to a TSC delta for the specified cpu
> + */
> +static unsigned long long uv_nsec_2_cycles(unsigned long nsec, int cpu)
> +{
> +	unsigned long long cyc;
> +	cyc = (nsec << CYC2NS_SCALE_FACTOR) / (per_cpu(cyc2ns, cpu));
> +	return cyc;

-ETOOFEWNEWLINES

Also, there doesnt seem to be anything UV specific about this, right?

> +}
> +
> +static unsigned long long uv_msec_2_cycles(unsigned long msec, int cpu)
> +{
> +	return uv_nsec_2_cycles(msec * 1000000, cpu);
> +}
> +
> +/*
> + * Atomically read the RTC & TSC. For accuracy, if it
> + * takes too long to access the RTC, try again. It is imporatnt that
> + * the TSC/RTC values be read as close as possible at the same time.
> + * Under certain pathological workloads, it can take a long time to
> + * read the RTC. Be paranoid - after a few attempts at reading a
> + * RTC/TSC pair, return a failure. Caller can try again later.
> + */
> +static noinline int uv_readclocks(struct uv_rtc_cpu_control *rcc,
> +				unsigned long *tsc, unsigned long *rtc)
> +{
> +	unsigned long z, flags, delta, rtc0, tsc0, tsc1;
> +	unsigned volatile long *myrtcp = rcc->myrtcp;
> +
> +	int cnt = MAXREADS;
> +
> +	do {
> +		local_irq_save(flags);
> +		mb();
> +		rdtscll(tsc0);
> +		mb();
> +		rtc0 = *myrtcp;
> +		mb();

Note, we have rdtsc_barrier() for such purposes. Beyond being faster 
it's even self-documenting. (unlike standalone mb()s)

> +		rdtscll(tsc1);
> +		local_irq_restore(flags);
> +		delta = tsc1 - tsc0;
> +		cnt--;
> +		z = delta / 100;		//ZZZZZZZZZ debug
> +		if (z > ZZMAXDELTA)		//ZZZZZZZZZ debug
> +			z = ZZMAXDELTA;		//ZZZZZZZZZ debug
> +		atomic_inc(&zzdelta[z]);	//ZZZZZZZZZ debug

And we dont want ZZZZZZZZZ's (and C++ comments) upstream either, right?

> +		if (cnt == 0)
> +			return -1;
> +	} while (delta > rcc->max_readclocks_tsc_delta);
> +	*rtc = rtc0;
> +	*tsc = tsc1;
> +	return 0;
> +}
> +
> +/*
> + * Read the current RTC value. RTC access is several hundred nsec &
> + * can be a performance issue. In most cases, the current RTC value
> + * is calculated based on the current TSC value (fast access) and the
> + * TSC delta since a TSC/RTC pair was last read..
> + */
> +static cycle_t uv_read_rtc_calc(struct clocksource *cs)
> +{
> +	struct uv_rtc_cpu_control *rcc = __get_cpu_var(uv_rtc_cpu_control);
> +	struct uv_rtc_blade_control *rbc = rcc->rbc;
> +	unsigned long rtc, tsc, rtcx, tscx, rtc0m, tsc0m, rtc0, tsc0;
> +	unsigned long fixmult;
> +	unsigned long adj;
> +
> +	uv_read128(&rbc->rtc0, &tsc0, &rtc0);
> +	rdtscll(tsc);
> +	fixmult = rbc->fixmult;
> +
> +	if (unlikely((tsc0 + rcc->max_tsc_delta < tsc) || fixmult == 0))
> +		goto resync;
> +
> +	/*
> +	 * Cast to long to get sign extension on the shift  - delta might
> +	 * be negative
> +	 */
> +	adj = ((long)((tsc - tsc0) * fixmult)) >> 32;
> +	rtc = rtc0 + adj;
> +
> +done:
> +	/*
> +	 * Don't return a rtc value less than the previous. Note: optimized
> +	 * for typical case where last_rtc must be updated. Other case
> +	 * rarely happens.
> +	 */
> +	rtc = max(rcc->last_rtc_value, rtc);
> +	rcc->last_rtc_value = rtc;
> +	return rtc;
> +
> +resync:
> +	/*
> +	 * Periodically, reread a TSC/RTC pair. This prevents cumulative
> +	 * clock drift from causing incorect values to be calculated.
> +	 * If unable to read a TSC/RTC pair, disable RTC calculation
> +	 * and read it directly.
> +	 */
> +
> +	if (unlikely(uv_readclocks(rcc, &tsc0, &rtc0) < 0))
> +		goto syncfail;
> +	uv_cmpxchg128(&rbc->rtc0, rbc->tsc0, rbc->rtc0, tsc0, rtc0);
> +	uv_read128(&rbc->rtc0m, &tsc0m, &rtc0m);
> +	if (unlikely(tsc0m + rcc->recalc_drift_tsc_delta < tsc0 ||
> +					tsc0m == 0)) {
> +		uv_cmpxchg128(&rbc->rtc0m, tsc0m, rtc0m, tsc0, rtc0);
> +		if (tsc0m) {
> +			rtcx = rtc0 - rtc0m;
> +			tscx = tsc0 - tsc0m;
> +			while (tscx > 0x7fffffff || rtcx > 0x7fffffff) {
> +				tscx >>= 1;
> +				rtcx >>= 1;
> +			}
> +			rbc->fixmult = (rtcx << 32) / tscx;
> +		}
> +	}
> +	rtc = rtc0;
> +	goto done;
> +
> +syncfail:
> +	/*
> +	 * Unable to read a TSC/RTC pair. Disable dynamic calculation
> +	 * of the RTC value until a pair can be read.
> +	 */
> +	rbc->tsc0m = 0;
> +	rbc->fixmult = 0;
> +	rtc = *rcc->myrtcp;
> +	atomic_inc(&resync_clocks_failures);
> +	goto done;
> +}

-EFUNCTIONTOOLONG

The goto maze is pretty disgusting as well - splitting the function will 
help that problem too.

> +
> +static cycle_t uv_read_rtc(struct clocksource *cs)
> +{
> +	struct uv_rtc_cpu_control *rcc = __get_cpu_var(uv_rtc_cpu_control);
> +
> +	return *rcc->myrtcp;
> +}
> +
> +/*
>   * Setup the RTC timer in oneshot mode
>   */
>  static void uv_rtc_timer_setup(enum clock_event_mode mode,
> @@ -354,6 +675,108 @@ static int __init uv_enable_evt_rtc(char
>  }
>  __setup("uvrtcevt", uv_enable_evt_rtc);
>  
> +
> +#define UVCLK_ATTR(_name) \
> +	static struct kobj_attribute uv_##_name##_attr = \
> +		__ATTR(_name, 0644, uv_##_name##_show, uv_##_name##_store)
> +
> +#define UVCLK_ATTR_RO(_name) \
> +	static struct kobj_attribute uv_##_name##_attr = \
> +		__ATTR(_name, 0444, uv_##_name##_show, NULL)
> +
> +static void uv_update_all_blade_control(void)
> +{
> +	struct uv_rtc_cpu_control *rcc;
> +	int cpu;
> +
> +	for_each_online_cpu(cpu) {
> +		rcc = per_cpu(uv_rtc_cpu_control, cpu);
> +		rcc->recalc_drift_tsc_delta = uv_msec_2_cycles(recalc_drift_msec, cpu);
> +		rcc->max_tsc_delta = uv_msec_2_cycles(resync_clocks_msec, cpu);
> +		rcc->max_readclocks_tsc_delta = uv_nsec_2_cycles(max_readclocks_tsc_nsec, cpu);
> +	}
> +}
> +
> +static ssize_t uv_recalculate_drift_msec_show(struct kobject *kobj,
> +			struct kobj_attribute *attr, char *buf)
> +{
> +	return sprintf(buf, "%lu\n", recalc_drift_msec);
> +}
> +static ssize_t uv_recalculate_drift_msec_store(struct kobject *kobj,
> +		struct kobj_attribute *attr, const char *buf, size_t count)
> +{
> +	unsigned long input;
> +
> +	if (strict_strtoul(buf, 10, &input))
> +		return 0;
> +
> +	recalc_drift_msec = input;
> +	uv_update_all_blade_control();
> +
> +	return count;
> +}
> +UVCLK_ATTR(recalculate_drift_msec);
> +
> +static ssize_t uv_resync_clocks_msec_show(struct kobject *kobj,
> +			struct kobj_attribute *attr, char *buf)
> +{
> +	return sprintf(buf, "%lu\n", resync_clocks_msec);
> +}
> +static ssize_t uv_resync_clocks_msec_store(struct kobject *kobj,
> +		struct kobj_attribute *attr, const char *buf, size_t count)
> +{
> +	unsigned long input;
> +
> +	if (strict_strtoul(buf, 10, &input))
> +		return 0;
> +
> +	resync_clocks_msec = input;
> +	uv_update_all_blade_control();
> +
> +	return count;
> +}
> +UVCLK_ATTR(resync_clocks_msec);
> +
> +static ssize_t uv_max_readclocks_tsc_nsec_show(struct kobject *kobj,
> +			struct kobj_attribute *attr, char *buf)
> +{
> +	return sprintf(buf, "%lu\n", max_readclocks_tsc_nsec);
> +}
> +static ssize_t uv_max_readclocks_tsc_nsec_store(struct kobject *kobj,
> +		struct kobj_attribute *attr, const char *buf, size_t count)
> +{
> +	unsigned long input;
> +
> +	if (strict_strtoul(buf, 10, &input))
> +		return 0;
> +
> +	max_readclocks_tsc_nsec = input;
> +	uv_update_all_blade_control();
> +
> +	return count;
> +}
> +UVCLK_ATTR(max_readclocks_tsc_nsec);
> +
> +static ssize_t uv_resync_clocks_failures_show(struct kobject *kobj,
> +			struct kobj_attribute *attr, char *buf)
> +{
> +	return sprintf(buf, "%u\n", atomic_read(&resync_clocks_failures));
> +}
> +UVCLK_ATTR_RO(resync_clocks_failures);
> +
> +static struct attribute *uv_rtc_attrs[] = {
> +	&uv_recalculate_drift_msec_attr.attr,
> +	&uv_resync_clocks_msec_attr.attr,
> +	&uv_max_readclocks_tsc_nsec_attr.attr,
> +	&uv_resync_clocks_failures_attr.attr,
> +	NULL,
> +};
> +
> +static struct attribute_group uv_rtc_attr_group = {
> +	.attrs = uv_rtc_attrs,
> +	.name = "sgi_rtc",
> +};
> +
>  static __init void uv_rtc_register_clockevents(struct work_struct *dummy)
>  {
>  	struct clock_event_device *ced = &__get_cpu_var(cpu_ced);
> @@ -365,29 +788,57 @@ static __init void uv_rtc_register_clock
>  
>  static __init int uv_rtc_setup_clock(void)
>  {
> -	int rc;
> +	struct uv_rtc_cpu_control *rcc = __get_cpu_var(uv_rtc_cpu_control);
> +	struct uv_rtc_blade_control *rbc, **rbcp;
> +	struct clocksource *cs;
> +	int b, rc, node, cpu, csi, numblades;
>  
>  	if (!is_uv_system())
>  		return -ENODEV;
>  
> -	clocksource_uv.mult = clocksource_hz2mult(sn_rtc_cycles_per_second,
> -				clocksource_uv.shift);
> +	/* Allocate per node structures for RTC */
> +	numblades = uv_num_possible_blades();
> +	rbcp = kzalloc(sizeof(void *) * numblades, GFP_ATOMIC);
> +	for_each_online_node(node) {
> +		b = uv_node_to_blade_id(node);
> +		if (!rbcp[b])
> +			rbcp[b] = kzalloc_node(sizeof(*rbc), GFP_ATOMIC, node);
> +		rbc = rbcp[b];
> +		for_each_online_cpu(cpu)
> +			if (cpu_to_node(cpu) == node) {
> +				rcc = kzalloc_node(sizeof(*rcc), GFP_ATOMIC,
> +									node);

Ignore checkpatch in cases where the solution to a warning is an uglie 
like this one.

> +				rcc->rbc = rbc;
> +				rcc->myrtcp = uv_raw_rtc_address(cpu);
> +				per_cpu(uv_rtc_cpu_control, cpu) = rcc;
> +			}
> +	}
> +	uv_update_all_blade_control();
> +	kfree(rbcp);
>  
> -	/* If single blade, prefer tsc */
> -	if (uv_num_possible_blades() == 1)
> -		clocksource_uv.rating = 250;
> +	for (csi = 0; csi < ARRAY_SIZE(clocksources); csi++) {
> +		cs = clocksources[csi];
> +		cs->mult = clocksource_hz2mult(sn_rtc_cycles_per_second,
> +				cs->shift);
> +
> +		/* If single blade, prefer tsc */
> +		if (uv_num_possible_blades() == 1)
> +			cs->rating = 250;
> +		rc = clocksource_register(cs);
> +		if (rc)
> +			goto error;
> +	}
>  
> -	rc = clocksource_register(&clocksource_uv);
> -	if (rc)
> -		printk(KERN_INFO "UV RTC clocksource failed rc %d\n", rc);
> -	else
> -		printk(KERN_INFO "UV RTC clocksource registered freq %lu MHz\n",
> -			sn_rtc_cycles_per_second/(unsigned long)1E6);
> +	printk(KERN_INFO "UV RTC clocksource registered freq %lu MHz\n",
> +		sn_rtc_cycles_per_second/(unsigned long)1E6);
>  
> -	if (rc || !uv_rtc_evt_enable || x86_platform_ipi_callback)
> -		return rc;
> +	if (sysfs_create_group(kernel_kobj, &uv_rtc_attr_group))
> +		printk(KERN_INFO "UV RTC - unable to create /sys files\n");
>  
>  	/* Setup and register clockevents */
> +	if (!uv_rtc_evt_enable)
> +		return 0;
> +
>  	rc = uv_rtc_allocate_timers();
>  	if (rc)
>  		goto error;
> @@ -415,8 +866,9 @@ static __init int uv_rtc_setup_clock(voi
>  	return 0;
>  
>  error:
> -	clocksource_unregister(&clocksource_uv);
>  	printk(KERN_INFO "UV RTC clockevents failed rc %d\n", rc);
> +	for (csi = csi - 1; csi >= 0; csi--)
> +		clocksource_unregister(clocksources[csi]);
>  
>  	return rc;
>  }

Probably: -EFUNCTIONTOOBIG.

Thanks,

	Ingo
--
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/