2005-04-17 05:20:36 +07:00
|
|
|
/* calibrate.c: default delay calibration
|
|
|
|
*
|
|
|
|
* Excised from init/main.c
|
|
|
|
* Copyright (C) 1991, 1992 Linus Torvalds
|
|
|
|
*/
|
|
|
|
|
[PATCH] remove many unneeded #includes of sched.h
After Al Viro (finally) succeeded in removing the sched.h #include in module.h
recently, it makes sense again to remove other superfluous sched.h includes.
There are quite a lot of files which include it but don't actually need
anything defined in there. Presumably these includes were once needed for
macros that used to live in sched.h, but moved to other header files in the
course of cleaning it up.
To ease the pain, this time I did not fiddle with any header files and only
removed #includes from .c-files, which tend to cause less trouble.
Compile tested against 2.6.20-rc2 and 2.6.20-rc2-mm2 (with offsets) on alpha,
arm, i386, ia64, mips, powerpc, and x86_64 with allnoconfig, defconfig,
allmodconfig, and allyesconfig as well as a few randconfigs on x86_64 and all
configs in arch/arm/configs on arm. I also checked that no new warnings were
introduced by the patch (actually, some warnings are removed that were emitted
by unnecessarily included header files).
Signed-off-by: Tim Schmielau <tim@physik3.uni-rostock.de>
Acked-by: Russell King <rmk+kernel@arm.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-02-14 15:33:14 +07:00
|
|
|
#include <linux/jiffies.h>
|
2005-04-17 05:20:36 +07:00
|
|
|
#include <linux/delay.h>
|
|
|
|
#include <linux/init.h>
|
2008-02-06 16:36:42 +07:00
|
|
|
#include <linux/timex.h>
|
2008-06-21 05:06:33 +07:00
|
|
|
#include <linux/smp.h>
|
2011-07-26 07:13:29 +07:00
|
|
|
#include <linux/percpu.h>
|
2005-06-23 14:08:13 +07:00
|
|
|
|
2008-06-24 08:21:56 +07:00
|
|
|
unsigned long lpj_fine;
|
2007-10-16 15:23:46 +07:00
|
|
|
unsigned long preset_lpj;
|
2005-04-17 05:20:36 +07:00
|
|
|
static int __init lpj_setup(char *str)
|
|
|
|
{
|
|
|
|
preset_lpj = simple_strtoul(str,NULL,0);
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
__setup("lpj=", lpj_setup);
|
|
|
|
|
2005-06-23 14:08:13 +07:00
|
|
|
#ifdef ARCH_HAS_READ_CURRENT_TIMER
|
|
|
|
|
|
|
|
/* This routine uses the read_current_timer() routine and gets the
|
|
|
|
* loops per jiffy directly, instead of guessing it using delay().
|
|
|
|
* Also, this code tries to handle non-maskable asynchronous events
|
|
|
|
* (like SMIs)
|
|
|
|
*/
|
|
|
|
#define DELAY_CALIBRATION_TICKS ((HZ < 100) ? 1 : (HZ/100))
|
|
|
|
#define MAX_DIRECT_CALIBRATION_RETRIES 5
|
|
|
|
|
2013-06-20 01:53:51 +07:00
|
|
|
static unsigned long calibrate_delay_direct(void)
|
2005-06-23 14:08:13 +07:00
|
|
|
{
|
|
|
|
unsigned long pre_start, start, post_start;
|
|
|
|
unsigned long pre_end, end, post_end;
|
|
|
|
unsigned long start_jiffies;
|
2008-06-24 08:21:56 +07:00
|
|
|
unsigned long timer_rate_min, timer_rate_max;
|
|
|
|
unsigned long good_timer_sum = 0;
|
|
|
|
unsigned long good_timer_count = 0;
|
2011-05-25 07:13:15 +07:00
|
|
|
unsigned long measured_times[MAX_DIRECT_CALIBRATION_RETRIES];
|
|
|
|
int max = -1; /* index of measured_times with max/min values or not set */
|
|
|
|
int min = -1;
|
2005-06-23 14:08:13 +07:00
|
|
|
int i;
|
|
|
|
|
|
|
|
if (read_current_timer(&pre_start) < 0 )
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* A simple loop like
|
|
|
|
* while ( jiffies < start_jiffies+1)
|
|
|
|
* start = read_current_timer();
|
|
|
|
* will not do. As we don't really know whether jiffy switch
|
|
|
|
* happened first or timer_value was read first. And some asynchronous
|
|
|
|
* event can happen between these two events introducing errors in lpj.
|
|
|
|
*
|
|
|
|
* So, we do
|
|
|
|
* 1. pre_start <- When we are sure that jiffy switch hasn't happened
|
|
|
|
* 2. check jiffy switch
|
|
|
|
* 3. start <- timer value before or after jiffy switch
|
|
|
|
* 4. post_start <- When we are sure that jiffy switch has happened
|
|
|
|
*
|
|
|
|
* Note, we don't know anything about order of 2 and 3.
|
|
|
|
* Now, by looking at post_start and pre_start difference, we can
|
|
|
|
* check whether any asynchronous event happened or not
|
|
|
|
*/
|
|
|
|
|
|
|
|
for (i = 0; i < MAX_DIRECT_CALIBRATION_RETRIES; i++) {
|
|
|
|
pre_start = 0;
|
|
|
|
read_current_timer(&start);
|
|
|
|
start_jiffies = jiffies;
|
2011-02-10 15:50:41 +07:00
|
|
|
while (time_before_eq(jiffies, start_jiffies + 1)) {
|
2005-06-23 14:08:13 +07:00
|
|
|
pre_start = start;
|
|
|
|
read_current_timer(&start);
|
|
|
|
}
|
|
|
|
read_current_timer(&post_start);
|
|
|
|
|
|
|
|
pre_end = 0;
|
|
|
|
end = post_start;
|
2011-02-10 15:50:41 +07:00
|
|
|
while (time_before_eq(jiffies, start_jiffies + 1 +
|
|
|
|
DELAY_CALIBRATION_TICKS)) {
|
2005-06-23 14:08:13 +07:00
|
|
|
pre_end = end;
|
|
|
|
read_current_timer(&end);
|
|
|
|
}
|
|
|
|
read_current_timer(&post_end);
|
|
|
|
|
2008-06-24 08:21:56 +07:00
|
|
|
timer_rate_max = (post_end - pre_start) /
|
|
|
|
DELAY_CALIBRATION_TICKS;
|
|
|
|
timer_rate_min = (pre_end - post_start) /
|
|
|
|
DELAY_CALIBRATION_TICKS;
|
2005-06-23 14:08:13 +07:00
|
|
|
|
|
|
|
/*
|
2008-06-24 08:21:56 +07:00
|
|
|
* If the upper limit and lower limit of the timer_rate is
|
2005-06-23 14:08:13 +07:00
|
|
|
* >= 12.5% apart, redo calibration.
|
|
|
|
*/
|
2011-05-25 07:13:15 +07:00
|
|
|
if (start >= post_end)
|
|
|
|
printk(KERN_NOTICE "calibrate_delay_direct() ignoring "
|
|
|
|
"timer_rate as we had a TSC wrap around"
|
|
|
|
" start=%lu >=post_end=%lu\n",
|
|
|
|
start, post_end);
|
|
|
|
if (start < post_end && pre_start != 0 && pre_end != 0 &&
|
2008-06-24 08:21:56 +07:00
|
|
|
(timer_rate_max - timer_rate_min) < (timer_rate_max >> 3)) {
|
|
|
|
good_timer_count++;
|
|
|
|
good_timer_sum += timer_rate_max;
|
2011-05-25 07:13:15 +07:00
|
|
|
measured_times[i] = timer_rate_max;
|
|
|
|
if (max < 0 || timer_rate_max > measured_times[max])
|
|
|
|
max = i;
|
|
|
|
if (min < 0 || timer_rate_max < measured_times[min])
|
|
|
|
min = i;
|
|
|
|
} else
|
|
|
|
measured_times[i] = 0;
|
|
|
|
|
2005-06-23 14:08:13 +07:00
|
|
|
}
|
|
|
|
|
2011-05-25 07:13:15 +07:00
|
|
|
/*
|
|
|
|
* Find the maximum & minimum - if they differ too much throw out the
|
|
|
|
* one with the largest difference from the mean and try again...
|
|
|
|
*/
|
|
|
|
while (good_timer_count > 1) {
|
|
|
|
unsigned long estimate;
|
|
|
|
unsigned long maxdiff;
|
|
|
|
|
|
|
|
/* compute the estimate */
|
|
|
|
estimate = (good_timer_sum/good_timer_count);
|
|
|
|
maxdiff = estimate >> 3;
|
|
|
|
|
|
|
|
/* if range is within 12% let's take it */
|
|
|
|
if ((measured_times[max] - measured_times[min]) < maxdiff)
|
|
|
|
return estimate;
|
|
|
|
|
|
|
|
/* ok - drop the worse value and try again... */
|
|
|
|
good_timer_sum = 0;
|
|
|
|
good_timer_count = 0;
|
|
|
|
if ((measured_times[max] - estimate) <
|
|
|
|
(estimate - measured_times[min])) {
|
|
|
|
printk(KERN_NOTICE "calibrate_delay_direct() dropping "
|
|
|
|
"min bogoMips estimate %d = %lu\n",
|
|
|
|
min, measured_times[min]);
|
|
|
|
measured_times[min] = 0;
|
|
|
|
min = max;
|
|
|
|
} else {
|
|
|
|
printk(KERN_NOTICE "calibrate_delay_direct() dropping "
|
|
|
|
"max bogoMips estimate %d = %lu\n",
|
|
|
|
max, measured_times[max]);
|
|
|
|
measured_times[max] = 0;
|
|
|
|
max = min;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (i = 0; i < MAX_DIRECT_CALIBRATION_RETRIES; i++) {
|
|
|
|
if (measured_times[i] == 0)
|
|
|
|
continue;
|
|
|
|
good_timer_count++;
|
|
|
|
good_timer_sum += measured_times[i];
|
|
|
|
if (measured_times[i] < measured_times[min])
|
|
|
|
min = i;
|
|
|
|
if (measured_times[i] > measured_times[max])
|
|
|
|
max = i;
|
|
|
|
}
|
|
|
|
|
|
|
|
}
|
2005-06-23 14:08:13 +07:00
|
|
|
|
2011-05-25 07:13:15 +07:00
|
|
|
printk(KERN_NOTICE "calibrate_delay_direct() failed to get a good "
|
|
|
|
"estimate for loops_per_jiffy.\nProbably due to long platform "
|
|
|
|
"interrupts. Consider using \"lpj=\" boot option.\n");
|
2005-06-23 14:08:13 +07:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
#else
|
2013-06-20 01:53:51 +07:00
|
|
|
static unsigned long calibrate_delay_direct(void)
|
|
|
|
{
|
|
|
|
return 0;
|
|
|
|
}
|
2005-06-23 14:08:13 +07:00
|
|
|
#endif
|
|
|
|
|
2005-04-17 05:20:36 +07:00
|
|
|
/*
|
|
|
|
* This is the number of bits of precision for the loops_per_jiffy. Each
|
calibrate: home in on correct lpj value more quickly
Binary chop with a jiffy-resync on each step to find an upper bound is
slow, so just race in a tight-ish loop to find an underestimate.
If done with lots of individual steps, sometimes several hundreds of
iterations would be required, which would impose a significant overhead,
and make the initial estimate very low. By taking slowly increasing steps
there will be less overhead.
E.g. an x86_64 2.67GHz could have fitted in 613 individual small delays,
but in reality should have been able to fit in a single delay 644 times
longer, so underestimated by 31 steps. To reach the equivalent of 644
small delays with the accelerating scheme now requires about 130
iterations, so has <1/4th of the overhead, and can therefore be expected
to underestimate by only 7 steps.
As now we have a better initial estimate we can binary chop over a smaller
range. With the loop overhead in the initial estimate kept low, and the
step sizes moderate, we won't have under-estimated by much, so chose as
tight a range as we can.
Signed-off-by: Phil Carmody <ext-phil.2.carmody@nokia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Tested-by: Stephen Boyd <sboyd@codeaurora.org>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 06:34:13 +07:00
|
|
|
* time we refine our estimate after the first takes 1.5/HZ seconds, so try
|
|
|
|
* to start with a good estimate.
|
2008-06-21 05:06:33 +07:00
|
|
|
* For the boot cpu we can skip the delay calibration and assign it a value
|
2008-06-24 08:21:56 +07:00
|
|
|
* calculated based on the timer frequency.
|
|
|
|
* For the rest of the CPUs we cannot assume that the timer frequency is same as
|
2008-06-21 05:06:33 +07:00
|
|
|
* the cpu frequency, hence do the calibration for those.
|
2005-04-17 05:20:36 +07:00
|
|
|
*/
|
|
|
|
#define LPS_PREC 8
|
|
|
|
|
2013-06-20 01:53:51 +07:00
|
|
|
static unsigned long calibrate_delay_converge(void)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
calibrate: home in on correct lpj value more quickly
Binary chop with a jiffy-resync on each step to find an upper bound is
slow, so just race in a tight-ish loop to find an underestimate.
If done with lots of individual steps, sometimes several hundreds of
iterations would be required, which would impose a significant overhead,
and make the initial estimate very low. By taking slowly increasing steps
there will be less overhead.
E.g. an x86_64 2.67GHz could have fitted in 613 individual small delays,
but in reality should have been able to fit in a single delay 644 times
longer, so underestimated by 31 steps. To reach the equivalent of 644
small delays with the accelerating scheme now requires about 130
iterations, so has <1/4th of the overhead, and can therefore be expected
to underestimate by only 7 steps.
As now we have a better initial estimate we can binary chop over a smaller
range. With the loop overhead in the initial estimate kept low, and the
step sizes moderate, we won't have under-estimated by much, so chose as
tight a range as we can.
Signed-off-by: Phil Carmody <ext-phil.2.carmody@nokia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Tested-by: Stephen Boyd <sboyd@codeaurora.org>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 06:34:13 +07:00
|
|
|
/* First stage - slowly accelerate to find initial bounds */
|
2011-03-23 06:34:15 +07:00
|
|
|
unsigned long lpj, lpj_base, ticks, loopadd, loopadd_base, chop_limit;
|
calibrate: home in on correct lpj value more quickly
Binary chop with a jiffy-resync on each step to find an upper bound is
slow, so just race in a tight-ish loop to find an underestimate.
If done with lots of individual steps, sometimes several hundreds of
iterations would be required, which would impose a significant overhead,
and make the initial estimate very low. By taking slowly increasing steps
there will be less overhead.
E.g. an x86_64 2.67GHz could have fitted in 613 individual small delays,
but in reality should have been able to fit in a single delay 644 times
longer, so underestimated by 31 steps. To reach the equivalent of 644
small delays with the accelerating scheme now requires about 130
iterations, so has <1/4th of the overhead, and can therefore be expected
to underestimate by only 7 steps.
As now we have a better initial estimate we can binary chop over a smaller
range. With the loop overhead in the initial estimate kept low, and the
step sizes moderate, we won't have under-estimated by much, so chose as
tight a range as we can.
Signed-off-by: Phil Carmody <ext-phil.2.carmody@nokia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Tested-by: Stephen Boyd <sboyd@codeaurora.org>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 06:34:13 +07:00
|
|
|
int trials = 0, band = 0, trial_in_band = 0;
|
calibrate: extract fall-back calculation into own helper
The motivation for this patch series is that currently our OMAP calibrates
itself using the trial-and-error binary chop fallback that some other
architectures no longer need to perform. This is a lengthy process,
taking 0.2s in an environment where boot time is of great interest.
Patch 2/4 has two optimisations. Firstly, it replaces the initial
repeated- doubling to find the relevant power of 2 with a tight loop that
just does as much as it can in a jiffy. Secondly, it doesn't binary chop
over an entire power of 2 range, it choses a much smaller range based on
how much it squeezed in, and failed to squeeze in, during the first stage.
Both are significant optimisations, and bring our calibration down from
23 jiffies to 5, and, in the process, often arrive at a more accurate lpj
value.
The 'bands' and 'sub-logarithmic' growth may look over-engineered, but
they only cost a small level of inaccuracy in the initial guess (for all
architectures) in order to avoid the very large inaccuracies that appeared
during testing (on x86_64 architectures, and presumably others with less
metronomic operation). Note that due to the existence of the TSC and
other timers, the x86_64 will not typically use this fallback routine, but
I wanted to code defensively, able to cope with all kinds of processor
behaviours and kernel command line options.
Patch 3/4 is an additional trap for the nightmare scenario where the
initial estimate is very inaccurate, possibly due to things like SMIs.
It simply retries with a larger bound.
Stephen said:
I tried this patch set out on an MSM7630.
:
: Before:
:
: Calibrating delay loop... 681.57 BogoMIPS (lpj=3407872)
:
: After:
:
: Calibrating delay loop... 680.75 BogoMIPS (lpj=3403776)
:
: But the really good news is calibration time dropped from ~247ms to ~56ms.
: Sadly we won't be able to benefit from this should my udelay patches make
: it into ARM because we would be using calibrate_delay_direct() instead (at
: least on machines who choose to). Can we somehow reapply the logic behind
: this to calibrate_delay_direct()? That would be even better, but this is
: definitely a boot time improvement.
:
: Or maybe we could just replace calibrate_delay_direct() with this fallback
: calculation? If __delay() is a thin wrapper around read_current_timer()
: it should work just as well (plus patch 3 makes it handle SMIs). I'll try
: that out.
This patch:
... so that it can be modified more clinically.
This is almost entirely cosmetic. The only change to the operation
is that the global variable is only set once after the estimation is
completed, rather than taking on all the intermediate values. However,
there are no readers of that variable, so this change is unimportant.
Signed-off-by: Phil Carmody <ext-phil.2.carmody@nokia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Tested-by: Stephen Boyd <sboyd@codeaurora.org>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 06:34:12 +07:00
|
|
|
|
|
|
|
lpj = (1<<12);
|
calibrate: home in on correct lpj value more quickly
Binary chop with a jiffy-resync on each step to find an upper bound is
slow, so just race in a tight-ish loop to find an underestimate.
If done with lots of individual steps, sometimes several hundreds of
iterations would be required, which would impose a significant overhead,
and make the initial estimate very low. By taking slowly increasing steps
there will be less overhead.
E.g. an x86_64 2.67GHz could have fitted in 613 individual small delays,
but in reality should have been able to fit in a single delay 644 times
longer, so underestimated by 31 steps. To reach the equivalent of 644
small delays with the accelerating scheme now requires about 130
iterations, so has <1/4th of the overhead, and can therefore be expected
to underestimate by only 7 steps.
As now we have a better initial estimate we can binary chop over a smaller
range. With the loop overhead in the initial estimate kept low, and the
step sizes moderate, we won't have under-estimated by much, so chose as
tight a range as we can.
Signed-off-by: Phil Carmody <ext-phil.2.carmody@nokia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Tested-by: Stephen Boyd <sboyd@codeaurora.org>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 06:34:13 +07:00
|
|
|
|
|
|
|
/* wait for "start of" clock tick */
|
|
|
|
ticks = jiffies;
|
|
|
|
while (ticks == jiffies)
|
|
|
|
; /* nothing */
|
|
|
|
/* Go .. */
|
|
|
|
ticks = jiffies;
|
|
|
|
do {
|
|
|
|
if (++trial_in_band == (1<<band)) {
|
|
|
|
++band;
|
|
|
|
trial_in_band = 0;
|
|
|
|
}
|
|
|
|
__delay(lpj * band);
|
|
|
|
trials += band;
|
|
|
|
} while (ticks == jiffies);
|
|
|
|
/*
|
|
|
|
* We overshot, so retreat to a clear underestimate. Then estimate
|
|
|
|
* the largest likely undershoot. This defines our chop bounds.
|
|
|
|
*/
|
|
|
|
trials -= band;
|
2011-03-23 06:34:15 +07:00
|
|
|
loopadd_base = lpj * band;
|
|
|
|
lpj_base = lpj * trials;
|
|
|
|
|
|
|
|
recalibrate:
|
|
|
|
lpj = lpj_base;
|
|
|
|
loopadd = loopadd_base;
|
calibrate: extract fall-back calculation into own helper
The motivation for this patch series is that currently our OMAP calibrates
itself using the trial-and-error binary chop fallback that some other
architectures no longer need to perform. This is a lengthy process,
taking 0.2s in an environment where boot time is of great interest.
Patch 2/4 has two optimisations. Firstly, it replaces the initial
repeated- doubling to find the relevant power of 2 with a tight loop that
just does as much as it can in a jiffy. Secondly, it doesn't binary chop
over an entire power of 2 range, it choses a much smaller range based on
how much it squeezed in, and failed to squeeze in, during the first stage.
Both are significant optimisations, and bring our calibration down from
23 jiffies to 5, and, in the process, often arrive at a more accurate lpj
value.
The 'bands' and 'sub-logarithmic' growth may look over-engineered, but
they only cost a small level of inaccuracy in the initial guess (for all
architectures) in order to avoid the very large inaccuracies that appeared
during testing (on x86_64 architectures, and presumably others with less
metronomic operation). Note that due to the existence of the TSC and
other timers, the x86_64 will not typically use this fallback routine, but
I wanted to code defensively, able to cope with all kinds of processor
behaviours and kernel command line options.
Patch 3/4 is an additional trap for the nightmare scenario where the
initial estimate is very inaccurate, possibly due to things like SMIs.
It simply retries with a larger bound.
Stephen said:
I tried this patch set out on an MSM7630.
:
: Before:
:
: Calibrating delay loop... 681.57 BogoMIPS (lpj=3407872)
:
: After:
:
: Calibrating delay loop... 680.75 BogoMIPS (lpj=3403776)
:
: But the really good news is calibration time dropped from ~247ms to ~56ms.
: Sadly we won't be able to benefit from this should my udelay patches make
: it into ARM because we would be using calibrate_delay_direct() instead (at
: least on machines who choose to). Can we somehow reapply the logic behind
: this to calibrate_delay_direct()? That would be even better, but this is
: definitely a boot time improvement.
:
: Or maybe we could just replace calibrate_delay_direct() with this fallback
: calculation? If __delay() is a thin wrapper around read_current_timer()
: it should work just as well (plus patch 3 makes it handle SMIs). I'll try
: that out.
This patch:
... so that it can be modified more clinically.
This is almost entirely cosmetic. The only change to the operation
is that the global variable is only set once after the estimation is
completed, rather than taking on all the intermediate values. However,
there are no readers of that variable, so this change is unimportant.
Signed-off-by: Phil Carmody <ext-phil.2.carmody@nokia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Tested-by: Stephen Boyd <sboyd@codeaurora.org>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 06:34:12 +07:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Do a binary approximation to get lpj set to
|
calibrate: home in on correct lpj value more quickly
Binary chop with a jiffy-resync on each step to find an upper bound is
slow, so just race in a tight-ish loop to find an underestimate.
If done with lots of individual steps, sometimes several hundreds of
iterations would be required, which would impose a significant overhead,
and make the initial estimate very low. By taking slowly increasing steps
there will be less overhead.
E.g. an x86_64 2.67GHz could have fitted in 613 individual small delays,
but in reality should have been able to fit in a single delay 644 times
longer, so underestimated by 31 steps. To reach the equivalent of 644
small delays with the accelerating scheme now requires about 130
iterations, so has <1/4th of the overhead, and can therefore be expected
to underestimate by only 7 steps.
As now we have a better initial estimate we can binary chop over a smaller
range. With the loop overhead in the initial estimate kept low, and the
step sizes moderate, we won't have under-estimated by much, so chose as
tight a range as we can.
Signed-off-by: Phil Carmody <ext-phil.2.carmody@nokia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Tested-by: Stephen Boyd <sboyd@codeaurora.org>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 06:34:13 +07:00
|
|
|
* equal one clock (up to LPS_PREC bits)
|
calibrate: extract fall-back calculation into own helper
The motivation for this patch series is that currently our OMAP calibrates
itself using the trial-and-error binary chop fallback that some other
architectures no longer need to perform. This is a lengthy process,
taking 0.2s in an environment where boot time is of great interest.
Patch 2/4 has two optimisations. Firstly, it replaces the initial
repeated- doubling to find the relevant power of 2 with a tight loop that
just does as much as it can in a jiffy. Secondly, it doesn't binary chop
over an entire power of 2 range, it choses a much smaller range based on
how much it squeezed in, and failed to squeeze in, during the first stage.
Both are significant optimisations, and bring our calibration down from
23 jiffies to 5, and, in the process, often arrive at a more accurate lpj
value.
The 'bands' and 'sub-logarithmic' growth may look over-engineered, but
they only cost a small level of inaccuracy in the initial guess (for all
architectures) in order to avoid the very large inaccuracies that appeared
during testing (on x86_64 architectures, and presumably others with less
metronomic operation). Note that due to the existence of the TSC and
other timers, the x86_64 will not typically use this fallback routine, but
I wanted to code defensively, able to cope with all kinds of processor
behaviours and kernel command line options.
Patch 3/4 is an additional trap for the nightmare scenario where the
initial estimate is very inaccurate, possibly due to things like SMIs.
It simply retries with a larger bound.
Stephen said:
I tried this patch set out on an MSM7630.
:
: Before:
:
: Calibrating delay loop... 681.57 BogoMIPS (lpj=3407872)
:
: After:
:
: Calibrating delay loop... 680.75 BogoMIPS (lpj=3403776)
:
: But the really good news is calibration time dropped from ~247ms to ~56ms.
: Sadly we won't be able to benefit from this should my udelay patches make
: it into ARM because we would be using calibrate_delay_direct() instead (at
: least on machines who choose to). Can we somehow reapply the logic behind
: this to calibrate_delay_direct()? That would be even better, but this is
: definitely a boot time improvement.
:
: Or maybe we could just replace calibrate_delay_direct() with this fallback
: calculation? If __delay() is a thin wrapper around read_current_timer()
: it should work just as well (plus patch 3 makes it handle SMIs). I'll try
: that out.
This patch:
... so that it can be modified more clinically.
This is almost entirely cosmetic. The only change to the operation
is that the global variable is only set once after the estimation is
completed, rather than taking on all the intermediate values. However,
there are no readers of that variable, so this change is unimportant.
Signed-off-by: Phil Carmody <ext-phil.2.carmody@nokia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Tested-by: Stephen Boyd <sboyd@codeaurora.org>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 06:34:12 +07:00
|
|
|
*/
|
2011-03-23 06:34:15 +07:00
|
|
|
chop_limit = lpj >> LPS_PREC;
|
calibrate: home in on correct lpj value more quickly
Binary chop with a jiffy-resync on each step to find an upper bound is
slow, so just race in a tight-ish loop to find an underestimate.
If done with lots of individual steps, sometimes several hundreds of
iterations would be required, which would impose a significant overhead,
and make the initial estimate very low. By taking slowly increasing steps
there will be less overhead.
E.g. an x86_64 2.67GHz could have fitted in 613 individual small delays,
but in reality should have been able to fit in a single delay 644 times
longer, so underestimated by 31 steps. To reach the equivalent of 644
small delays with the accelerating scheme now requires about 130
iterations, so has <1/4th of the overhead, and can therefore be expected
to underestimate by only 7 steps.
As now we have a better initial estimate we can binary chop over a smaller
range. With the loop overhead in the initial estimate kept low, and the
step sizes moderate, we won't have under-estimated by much, so chose as
tight a range as we can.
Signed-off-by: Phil Carmody <ext-phil.2.carmody@nokia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Tested-by: Stephen Boyd <sboyd@codeaurora.org>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 06:34:13 +07:00
|
|
|
while (loopadd > chop_limit) {
|
|
|
|
lpj += loopadd;
|
calibrate: extract fall-back calculation into own helper
The motivation for this patch series is that currently our OMAP calibrates
itself using the trial-and-error binary chop fallback that some other
architectures no longer need to perform. This is a lengthy process,
taking 0.2s in an environment where boot time is of great interest.
Patch 2/4 has two optimisations. Firstly, it replaces the initial
repeated- doubling to find the relevant power of 2 with a tight loop that
just does as much as it can in a jiffy. Secondly, it doesn't binary chop
over an entire power of 2 range, it choses a much smaller range based on
how much it squeezed in, and failed to squeeze in, during the first stage.
Both are significant optimisations, and bring our calibration down from
23 jiffies to 5, and, in the process, often arrive at a more accurate lpj
value.
The 'bands' and 'sub-logarithmic' growth may look over-engineered, but
they only cost a small level of inaccuracy in the initial guess (for all
architectures) in order to avoid the very large inaccuracies that appeared
during testing (on x86_64 architectures, and presumably others with less
metronomic operation). Note that due to the existence of the TSC and
other timers, the x86_64 will not typically use this fallback routine, but
I wanted to code defensively, able to cope with all kinds of processor
behaviours and kernel command line options.
Patch 3/4 is an additional trap for the nightmare scenario where the
initial estimate is very inaccurate, possibly due to things like SMIs.
It simply retries with a larger bound.
Stephen said:
I tried this patch set out on an MSM7630.
:
: Before:
:
: Calibrating delay loop... 681.57 BogoMIPS (lpj=3407872)
:
: After:
:
: Calibrating delay loop... 680.75 BogoMIPS (lpj=3403776)
:
: But the really good news is calibration time dropped from ~247ms to ~56ms.
: Sadly we won't be able to benefit from this should my udelay patches make
: it into ARM because we would be using calibrate_delay_direct() instead (at
: least on machines who choose to). Can we somehow reapply the logic behind
: this to calibrate_delay_direct()? That would be even better, but this is
: definitely a boot time improvement.
:
: Or maybe we could just replace calibrate_delay_direct() with this fallback
: calculation? If __delay() is a thin wrapper around read_current_timer()
: it should work just as well (plus patch 3 makes it handle SMIs). I'll try
: that out.
This patch:
... so that it can be modified more clinically.
This is almost entirely cosmetic. The only change to the operation
is that the global variable is only set once after the estimation is
completed, rather than taking on all the intermediate values. However,
there are no readers of that variable, so this change is unimportant.
Signed-off-by: Phil Carmody <ext-phil.2.carmody@nokia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Tested-by: Stephen Boyd <sboyd@codeaurora.org>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 06:34:12 +07:00
|
|
|
ticks = jiffies;
|
|
|
|
while (ticks == jiffies)
|
calibrate: home in on correct lpj value more quickly
Binary chop with a jiffy-resync on each step to find an upper bound is
slow, so just race in a tight-ish loop to find an underestimate.
If done with lots of individual steps, sometimes several hundreds of
iterations would be required, which would impose a significant overhead,
and make the initial estimate very low. By taking slowly increasing steps
there will be less overhead.
E.g. an x86_64 2.67GHz could have fitted in 613 individual small delays,
but in reality should have been able to fit in a single delay 644 times
longer, so underestimated by 31 steps. To reach the equivalent of 644
small delays with the accelerating scheme now requires about 130
iterations, so has <1/4th of the overhead, and can therefore be expected
to underestimate by only 7 steps.
As now we have a better initial estimate we can binary chop over a smaller
range. With the loop overhead in the initial estimate kept low, and the
step sizes moderate, we won't have under-estimated by much, so chose as
tight a range as we can.
Signed-off-by: Phil Carmody <ext-phil.2.carmody@nokia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Tested-by: Stephen Boyd <sboyd@codeaurora.org>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 06:34:13 +07:00
|
|
|
; /* nothing */
|
calibrate: extract fall-back calculation into own helper
The motivation for this patch series is that currently our OMAP calibrates
itself using the trial-and-error binary chop fallback that some other
architectures no longer need to perform. This is a lengthy process,
taking 0.2s in an environment where boot time is of great interest.
Patch 2/4 has two optimisations. Firstly, it replaces the initial
repeated- doubling to find the relevant power of 2 with a tight loop that
just does as much as it can in a jiffy. Secondly, it doesn't binary chop
over an entire power of 2 range, it choses a much smaller range based on
how much it squeezed in, and failed to squeeze in, during the first stage.
Both are significant optimisations, and bring our calibration down from
23 jiffies to 5, and, in the process, often arrive at a more accurate lpj
value.
The 'bands' and 'sub-logarithmic' growth may look over-engineered, but
they only cost a small level of inaccuracy in the initial guess (for all
architectures) in order to avoid the very large inaccuracies that appeared
during testing (on x86_64 architectures, and presumably others with less
metronomic operation). Note that due to the existence of the TSC and
other timers, the x86_64 will not typically use this fallback routine, but
I wanted to code defensively, able to cope with all kinds of processor
behaviours and kernel command line options.
Patch 3/4 is an additional trap for the nightmare scenario where the
initial estimate is very inaccurate, possibly due to things like SMIs.
It simply retries with a larger bound.
Stephen said:
I tried this patch set out on an MSM7630.
:
: Before:
:
: Calibrating delay loop... 681.57 BogoMIPS (lpj=3407872)
:
: After:
:
: Calibrating delay loop... 680.75 BogoMIPS (lpj=3403776)
:
: But the really good news is calibration time dropped from ~247ms to ~56ms.
: Sadly we won't be able to benefit from this should my udelay patches make
: it into ARM because we would be using calibrate_delay_direct() instead (at
: least on machines who choose to). Can we somehow reapply the logic behind
: this to calibrate_delay_direct()? That would be even better, but this is
: definitely a boot time improvement.
:
: Or maybe we could just replace calibrate_delay_direct() with this fallback
: calculation? If __delay() is a thin wrapper around read_current_timer()
: it should work just as well (plus patch 3 makes it handle SMIs). I'll try
: that out.
This patch:
... so that it can be modified more clinically.
This is almost entirely cosmetic. The only change to the operation
is that the global variable is only set once after the estimation is
completed, rather than taking on all the intermediate values. However,
there are no readers of that variable, so this change is unimportant.
Signed-off-by: Phil Carmody <ext-phil.2.carmody@nokia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Tested-by: Stephen Boyd <sboyd@codeaurora.org>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 06:34:12 +07:00
|
|
|
ticks = jiffies;
|
|
|
|
__delay(lpj);
|
|
|
|
if (jiffies != ticks) /* longer than 1 tick */
|
calibrate: home in on correct lpj value more quickly
Binary chop with a jiffy-resync on each step to find an upper bound is
slow, so just race in a tight-ish loop to find an underestimate.
If done with lots of individual steps, sometimes several hundreds of
iterations would be required, which would impose a significant overhead,
and make the initial estimate very low. By taking slowly increasing steps
there will be less overhead.
E.g. an x86_64 2.67GHz could have fitted in 613 individual small delays,
but in reality should have been able to fit in a single delay 644 times
longer, so underestimated by 31 steps. To reach the equivalent of 644
small delays with the accelerating scheme now requires about 130
iterations, so has <1/4th of the overhead, and can therefore be expected
to underestimate by only 7 steps.
As now we have a better initial estimate we can binary chop over a smaller
range. With the loop overhead in the initial estimate kept low, and the
step sizes moderate, we won't have under-estimated by much, so chose as
tight a range as we can.
Signed-off-by: Phil Carmody <ext-phil.2.carmody@nokia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Tested-by: Stephen Boyd <sboyd@codeaurora.org>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 06:34:13 +07:00
|
|
|
lpj -= loopadd;
|
|
|
|
loopadd >>= 1;
|
calibrate: extract fall-back calculation into own helper
The motivation for this patch series is that currently our OMAP calibrates
itself using the trial-and-error binary chop fallback that some other
architectures no longer need to perform. This is a lengthy process,
taking 0.2s in an environment where boot time is of great interest.
Patch 2/4 has two optimisations. Firstly, it replaces the initial
repeated- doubling to find the relevant power of 2 with a tight loop that
just does as much as it can in a jiffy. Secondly, it doesn't binary chop
over an entire power of 2 range, it choses a much smaller range based on
how much it squeezed in, and failed to squeeze in, during the first stage.
Both are significant optimisations, and bring our calibration down from
23 jiffies to 5, and, in the process, often arrive at a more accurate lpj
value.
The 'bands' and 'sub-logarithmic' growth may look over-engineered, but
they only cost a small level of inaccuracy in the initial guess (for all
architectures) in order to avoid the very large inaccuracies that appeared
during testing (on x86_64 architectures, and presumably others with less
metronomic operation). Note that due to the existence of the TSC and
other timers, the x86_64 will not typically use this fallback routine, but
I wanted to code defensively, able to cope with all kinds of processor
behaviours and kernel command line options.
Patch 3/4 is an additional trap for the nightmare scenario where the
initial estimate is very inaccurate, possibly due to things like SMIs.
It simply retries with a larger bound.
Stephen said:
I tried this patch set out on an MSM7630.
:
: Before:
:
: Calibrating delay loop... 681.57 BogoMIPS (lpj=3407872)
:
: After:
:
: Calibrating delay loop... 680.75 BogoMIPS (lpj=3403776)
:
: But the really good news is calibration time dropped from ~247ms to ~56ms.
: Sadly we won't be able to benefit from this should my udelay patches make
: it into ARM because we would be using calibrate_delay_direct() instead (at
: least on machines who choose to). Can we somehow reapply the logic behind
: this to calibrate_delay_direct()? That would be even better, but this is
: definitely a boot time improvement.
:
: Or maybe we could just replace calibrate_delay_direct() with this fallback
: calculation? If __delay() is a thin wrapper around read_current_timer()
: it should work just as well (plus patch 3 makes it handle SMIs). I'll try
: that out.
This patch:
... so that it can be modified more clinically.
This is almost entirely cosmetic. The only change to the operation
is that the global variable is only set once after the estimation is
completed, rather than taking on all the intermediate values. However,
there are no readers of that variable, so this change is unimportant.
Signed-off-by: Phil Carmody <ext-phil.2.carmody@nokia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Tested-by: Stephen Boyd <sboyd@codeaurora.org>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 06:34:12 +07:00
|
|
|
}
|
2011-03-23 06:34:15 +07:00
|
|
|
/*
|
|
|
|
* If we incremented every single time possible, presume we've
|
|
|
|
* massively underestimated initially, and retry with a higher
|
|
|
|
* start, and larger range. (Only seen on x86_64, due to SMIs)
|
|
|
|
*/
|
|
|
|
if (lpj + loopadd * 2 == lpj_base + loopadd_base * 2) {
|
|
|
|
lpj_base = lpj;
|
|
|
|
loopadd_base <<= 2;
|
|
|
|
goto recalibrate;
|
|
|
|
}
|
calibrate: extract fall-back calculation into own helper
The motivation for this patch series is that currently our OMAP calibrates
itself using the trial-and-error binary chop fallback that some other
architectures no longer need to perform. This is a lengthy process,
taking 0.2s in an environment where boot time is of great interest.
Patch 2/4 has two optimisations. Firstly, it replaces the initial
repeated- doubling to find the relevant power of 2 with a tight loop that
just does as much as it can in a jiffy. Secondly, it doesn't binary chop
over an entire power of 2 range, it choses a much smaller range based on
how much it squeezed in, and failed to squeeze in, during the first stage.
Both are significant optimisations, and bring our calibration down from
23 jiffies to 5, and, in the process, often arrive at a more accurate lpj
value.
The 'bands' and 'sub-logarithmic' growth may look over-engineered, but
they only cost a small level of inaccuracy in the initial guess (for all
architectures) in order to avoid the very large inaccuracies that appeared
during testing (on x86_64 architectures, and presumably others with less
metronomic operation). Note that due to the existence of the TSC and
other timers, the x86_64 will not typically use this fallback routine, but
I wanted to code defensively, able to cope with all kinds of processor
behaviours and kernel command line options.
Patch 3/4 is an additional trap for the nightmare scenario where the
initial estimate is very inaccurate, possibly due to things like SMIs.
It simply retries with a larger bound.
Stephen said:
I tried this patch set out on an MSM7630.
:
: Before:
:
: Calibrating delay loop... 681.57 BogoMIPS (lpj=3407872)
:
: After:
:
: Calibrating delay loop... 680.75 BogoMIPS (lpj=3403776)
:
: But the really good news is calibration time dropped from ~247ms to ~56ms.
: Sadly we won't be able to benefit from this should my udelay patches make
: it into ARM because we would be using calibrate_delay_direct() instead (at
: least on machines who choose to). Can we somehow reapply the logic behind
: this to calibrate_delay_direct()? That would be even better, but this is
: definitely a boot time improvement.
:
: Or maybe we could just replace calibrate_delay_direct() with this fallback
: calculation? If __delay() is a thin wrapper around read_current_timer()
: it should work just as well (plus patch 3 makes it handle SMIs). I'll try
: that out.
This patch:
... so that it can be modified more clinically.
This is almost entirely cosmetic. The only change to the operation
is that the global variable is only set once after the estimation is
completed, rather than taking on all the intermediate values. However,
there are no readers of that variable, so this change is unimportant.
Signed-off-by: Phil Carmody <ext-phil.2.carmody@nokia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Tested-by: Stephen Boyd <sboyd@codeaurora.org>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 06:34:12 +07:00
|
|
|
|
|
|
|
return lpj;
|
|
|
|
}
|
|
|
|
|
2011-07-26 07:13:29 +07:00
|
|
|
static DEFINE_PER_CPU(unsigned long, cpu_loops_per_jiffy) = { 0 };
|
|
|
|
|
2011-11-16 06:33:56 +07:00
|
|
|
/*
|
|
|
|
* Check if cpu calibration delay is already known. For example,
|
|
|
|
* some processors with multi-core sockets may have all cores
|
|
|
|
* with the same calibration delay.
|
|
|
|
*
|
|
|
|
* Architectures should override this function if a faster calibration
|
|
|
|
* method is available.
|
|
|
|
*/
|
2013-06-20 01:53:51 +07:00
|
|
|
unsigned long __attribute__((weak)) calibrate_delay_is_known(void)
|
2011-11-16 06:33:56 +07:00
|
|
|
{
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2014-06-12 22:58:27 +07:00
|
|
|
/*
|
|
|
|
* Indicate the cpu delay calibration is done. This can be used by
|
|
|
|
* architectures to stop accepting delay timer registrations after this point.
|
|
|
|
*/
|
|
|
|
|
|
|
|
void __attribute__((weak)) calibration_delay_done(void)
|
|
|
|
{
|
|
|
|
}
|
|
|
|
|
2013-06-20 01:53:51 +07:00
|
|
|
void calibrate_delay(void)
|
calibrate: extract fall-back calculation into own helper
The motivation for this patch series is that currently our OMAP calibrates
itself using the trial-and-error binary chop fallback that some other
architectures no longer need to perform. This is a lengthy process,
taking 0.2s in an environment where boot time is of great interest.
Patch 2/4 has two optimisations. Firstly, it replaces the initial
repeated- doubling to find the relevant power of 2 with a tight loop that
just does as much as it can in a jiffy. Secondly, it doesn't binary chop
over an entire power of 2 range, it choses a much smaller range based on
how much it squeezed in, and failed to squeeze in, during the first stage.
Both are significant optimisations, and bring our calibration down from
23 jiffies to 5, and, in the process, often arrive at a more accurate lpj
value.
The 'bands' and 'sub-logarithmic' growth may look over-engineered, but
they only cost a small level of inaccuracy in the initial guess (for all
architectures) in order to avoid the very large inaccuracies that appeared
during testing (on x86_64 architectures, and presumably others with less
metronomic operation). Note that due to the existence of the TSC and
other timers, the x86_64 will not typically use this fallback routine, but
I wanted to code defensively, able to cope with all kinds of processor
behaviours and kernel command line options.
Patch 3/4 is an additional trap for the nightmare scenario where the
initial estimate is very inaccurate, possibly due to things like SMIs.
It simply retries with a larger bound.
Stephen said:
I tried this patch set out on an MSM7630.
:
: Before:
:
: Calibrating delay loop... 681.57 BogoMIPS (lpj=3407872)
:
: After:
:
: Calibrating delay loop... 680.75 BogoMIPS (lpj=3403776)
:
: But the really good news is calibration time dropped from ~247ms to ~56ms.
: Sadly we won't be able to benefit from this should my udelay patches make
: it into ARM because we would be using calibrate_delay_direct() instead (at
: least on machines who choose to). Can we somehow reapply the logic behind
: this to calibrate_delay_direct()? That would be even better, but this is
: definitely a boot time improvement.
:
: Or maybe we could just replace calibrate_delay_direct() with this fallback
: calculation? If __delay() is a thin wrapper around read_current_timer()
: it should work just as well (plus patch 3 makes it handle SMIs). I'll try
: that out.
This patch:
... so that it can be modified more clinically.
This is almost entirely cosmetic. The only change to the operation
is that the global variable is only set once after the estimation is
completed, rather than taking on all the intermediate values. However,
there are no readers of that variable, so this change is unimportant.
Signed-off-by: Phil Carmody <ext-phil.2.carmody@nokia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Tested-by: Stephen Boyd <sboyd@codeaurora.org>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 06:34:12 +07:00
|
|
|
{
|
2011-06-22 17:55:50 +07:00
|
|
|
unsigned long lpj;
|
2009-11-18 07:22:13 +07:00
|
|
|
static bool printed;
|
2011-07-26 07:13:29 +07:00
|
|
|
int this_cpu = smp_processor_id();
|
2005-04-17 05:20:36 +07:00
|
|
|
|
2011-07-26 07:13:29 +07:00
|
|
|
if (per_cpu(cpu_loops_per_jiffy, this_cpu)) {
|
|
|
|
lpj = per_cpu(cpu_loops_per_jiffy, this_cpu);
|
2012-03-24 05:02:28 +07:00
|
|
|
if (!printed)
|
|
|
|
pr_info("Calibrating delay loop (skipped) "
|
2011-07-26 07:13:29 +07:00
|
|
|
"already calibrated this CPU");
|
|
|
|
} else if (preset_lpj) {
|
2011-06-22 17:55:50 +07:00
|
|
|
lpj = preset_lpj;
|
2009-11-18 07:22:13 +07:00
|
|
|
if (!printed)
|
|
|
|
pr_info("Calibrating delay loop (skipped) "
|
|
|
|
"preset value.. ");
|
|
|
|
} else if ((!printed) && lpj_fine) {
|
2011-06-22 17:55:50 +07:00
|
|
|
lpj = lpj_fine;
|
2009-11-18 07:22:13 +07:00
|
|
|
pr_info("Calibrating delay loop (skipped), "
|
2008-06-24 08:21:56 +07:00
|
|
|
"value calculated using timer frequency.. ");
|
2011-11-16 06:33:56 +07:00
|
|
|
} else if ((lpj = calibrate_delay_is_known())) {
|
|
|
|
;
|
2011-06-22 17:55:50 +07:00
|
|
|
} else if ((lpj = calibrate_delay_direct()) != 0) {
|
2009-11-18 07:22:13 +07:00
|
|
|
if (!printed)
|
|
|
|
pr_info("Calibrating delay using timer "
|
|
|
|
"specific routine.. ");
|
2005-04-17 05:20:36 +07:00
|
|
|
} else {
|
2009-11-18 07:22:13 +07:00
|
|
|
if (!printed)
|
|
|
|
pr_info("Calibrating delay loop... ");
|
2011-06-22 17:55:50 +07:00
|
|
|
lpj = calibrate_delay_converge();
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
2011-07-26 07:13:29 +07:00
|
|
|
per_cpu(cpu_loops_per_jiffy, this_cpu) = lpj;
|
2009-11-18 07:22:13 +07:00
|
|
|
if (!printed)
|
|
|
|
pr_cont("%lu.%02lu BogoMIPS (lpj=%lu)\n",
|
2011-06-22 17:55:50 +07:00
|
|
|
lpj/(500000/HZ),
|
|
|
|
(lpj/(5000/HZ)) % 100, lpj);
|
2009-11-18 07:22:13 +07:00
|
|
|
|
2011-06-22 17:55:50 +07:00
|
|
|
loops_per_jiffy = lpj;
|
2009-11-18 07:22:13 +07:00
|
|
|
printed = true;
|
2014-06-12 22:58:27 +07:00
|
|
|
|
|
|
|
calibration_delay_done();
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|