kernel_optimize_test/init/calibrate.c
Andrew Worsley d2b463135f init/calibrate.c: fix for critical bogoMIPS intermittent calculation failure
A fix to the TSC (Time Stamp Counter) based bogoMIPS calculation used on
secondary CPUs which has two faults:

1: Not handling wrapping of the lower 32 bits of the TSC counter on
   32bit kernel - perhaps TSC is not reset by a warm reset?

2: TSC and Jiffies are no incrementing together properly.  Either
   jiffies increment too quickly or Time Stamp Counter isn't incremented
   in during an SMI but the real time clock is and jiffies are
   incremented.

Case 1 can result in a factor of 16 too large a value which makes udelay()
values too small and can cause mysterious driver errors.  Case 2 appears
to give smaller 10-15% errors after averaging but enough to cause
occasional failures on my own board

I have tested this code on my own branch and attach patch suitable for
current kernel code.  See below for examples of the failures and how the
fix handles these situations now.

I reported this issue earlier here:
     Intermittent problem with BogoMIPs calculation on Intel AP CPUs -
http://marc.info/?l=linux-kernel&m=129947246316875&w=4

I suspect this issue has been seen by others but as it is intermittent and
bogoMIPS for secondary CPUs are no longer printed out it might have been
difficult to identify this as the cause.  Perhaps these unresolved issues,
although quite old, might be relevant as possibly this fault has been
around for a while.  In particular Case 1 may only be relevant to 32bit
kernels on newer HW (most people run 64bit kernels?).  Case 2 is less
dramatic since the earlier fix in this area and also intermittent.

   Re: bogomips discrepancy on Intel Core2 Quad CPU -
http://marc.info/?l=linux-kernel&m=118929277524298&w=4
   slow system and bogus bogomips  -
http://marc.info/?l=linux-kernel&m=116791286716107&w=4
   Re: Re: [RFC-PATCH] clocksource: update lpj if clocksource has -
http://marc.info/?l=linux-kernel&m=128952775819467&w=4

This issue is masked a little by commit feae3203d7 ("timers, init:
Limit the number of per cpu calibration bootup messages") which only
prints out the first bogoMIPS value making it much harder to notice other
values differing.  Perhaps it should be changed to only suppress them when
they are similar values?

Here are some outputs showing faults occurring and the new code handling
them properly.  See my earlier message for examples of the original
failure.

    Case 1:   A Time Stamp Counter wrap:
...
Calibrating delay loop (skipped), value calculated using timer
frequency.. 6332.70 BogoMIPS (lpj=31663540)
....
calibrate_delay_direct() timer_rate_max=31666493
timer_rate_min=31666151 pre_start=4170369255 pre_end=4202035539
calibrate_delay_direct() timer_rate_max=2425955274
timer_rate_min=2425954941 pre_start=4265368533 pre_end=2396356387
calibrate_delay_direct() ignoring timer_rate as we had a TSC wrap
around start=4265368581 >=post_end=2396356511
calibrate_delay_direct() timer_rate_max=31666274
timer_rate_min=31665942 pre_start=2440373374 pre_end=2472039515
calibrate_delay_direct() timer_rate_max=31666492
timer_rate_min=31666160 pre_start=2535372139 pre_end=2567038422
calibrate_delay_direct() timer_rate_max=31666455
timer_rate_min=31666207 pre_start=2630371084 pre_end=2662037415
Calibrating delay using timer specific routine.. 6333.28 BogoMIPS (lpj=31666428)
Total of 2 processors activated (12665.99 BogoMIPS).
....

    Case 2:  Some thing (presumably the SMM interrupt?) causing the
very low increase in TSC counter for the DELAY_CALIBRATION_TICKS
increase in jiffies
...
Calibrating delay loop (skipped), value calculated using timer
frequency.. 6333.25 BogoMIPS (lpj=31666270)
...
calibrate_delay_direct() timer_rate_max=31666483
timer_rate_min=31666074 pre_start=4199536526 pre_end=4231202809
calibrate_delay_direct() timer_rate_max=864348 timer_rate_min=864016
pre_start=2405343672 pre_end=2406207897
calibrate_delay_direct() timer_rate_max=31666483
timer_rate_min=31666179 pre_start=2469540464 pre_end=2501206823
calibrate_delay_direct() timer_rate_max=31666511
timer_rate_min=31666122 pre_start=2564539400 pre_end=2596205712
calibrate_delay_direct() timer_rate_max=31666084
timer_rate_min=31665685 pre_start=2659538782 pre_end=2691204657
calibrate_delay_direct() dropping min bogoMips estimate 1 = 864348
Calibrating delay using timer specific routine.. 6333.27 BogoMIPS (lpj=31666390)
Total of 2 processors activated (12666.53 BogoMIPS).
...

After 70 boots I saw 2 variations <1% slip through

[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: fix straggly printk mess]
Signed-off-by: Andrew Worsley <amworsley@gmail.com>
Reviewed-by: Phil Carmody <ext-phil.2.carmody@nokia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-05-25 08:39:46 -07:00

278 lines
7.8 KiB
C

/* calibrate.c: default delay calibration
*
* Excised from init/main.c
* Copyright (C) 1991, 1992 Linus Torvalds
*/
#include <linux/jiffies.h>
#include <linux/delay.h>
#include <linux/init.h>
#include <linux/timex.h>
#include <linux/smp.h>
unsigned long lpj_fine;
unsigned long preset_lpj;
static int __init lpj_setup(char *str)
{
preset_lpj = simple_strtoul(str,NULL,0);
return 1;
}
__setup("lpj=", lpj_setup);
#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
static unsigned long __cpuinit calibrate_delay_direct(void)
{
unsigned long pre_start, start, post_start;
unsigned long pre_end, end, post_end;
unsigned long start_jiffies;
unsigned long timer_rate_min, timer_rate_max;
unsigned long good_timer_sum = 0;
unsigned long good_timer_count = 0;
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;
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;
while (time_before_eq(jiffies, start_jiffies + 1)) {
pre_start = start;
read_current_timer(&start);
}
read_current_timer(&post_start);
pre_end = 0;
end = post_start;
while (time_before_eq(jiffies, start_jiffies + 1 +
DELAY_CALIBRATION_TICKS)) {
pre_end = end;
read_current_timer(&end);
}
read_current_timer(&post_end);
timer_rate_max = (post_end - pre_start) /
DELAY_CALIBRATION_TICKS;
timer_rate_min = (pre_end - post_start) /
DELAY_CALIBRATION_TICKS;
/*
* If the upper limit and lower limit of the timer_rate is
* >= 12.5% apart, redo calibration.
*/
printk(KERN_DEBUG "calibrate_delay_direct() timer_rate_max=%lu "
"timer_rate_min=%lu pre_start=%lu pre_end=%lu\n",
timer_rate_max, timer_rate_min, pre_start, pre_end);
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 &&
(timer_rate_max - timer_rate_min) < (timer_rate_max >> 3)) {
good_timer_count++;
good_timer_sum += timer_rate_max;
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;
}
/*
* 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;
}
}
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");
return 0;
}
#else
static unsigned long __cpuinit calibrate_delay_direct(void) {return 0;}
#endif
/*
* This is the number of bits of precision for the loops_per_jiffy. Each
* time we refine our estimate after the first takes 1.5/HZ seconds, so try
* to start with a good estimate.
* For the boot cpu we can skip the delay calibration and assign it a value
* calculated based on the timer frequency.
* For the rest of the CPUs we cannot assume that the timer frequency is same as
* the cpu frequency, hence do the calibration for those.
*/
#define LPS_PREC 8
static unsigned long __cpuinit calibrate_delay_converge(void)
{
/* First stage - slowly accelerate to find initial bounds */
unsigned long lpj, lpj_base, ticks, loopadd, loopadd_base, chop_limit;
int trials = 0, band = 0, trial_in_band = 0;
lpj = (1<<12);
/* 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;
loopadd_base = lpj * band;
lpj_base = lpj * trials;
recalibrate:
lpj = lpj_base;
loopadd = loopadd_base;
/*
* Do a binary approximation to get lpj set to
* equal one clock (up to LPS_PREC bits)
*/
chop_limit = lpj >> LPS_PREC;
while (loopadd > chop_limit) {
lpj += loopadd;
ticks = jiffies;
while (ticks == jiffies)
; /* nothing */
ticks = jiffies;
__delay(lpj);
if (jiffies != ticks) /* longer than 1 tick */
lpj -= loopadd;
loopadd >>= 1;
}
/*
* 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;
}
return lpj;
}
void __cpuinit calibrate_delay(void)
{
static bool printed;
if (preset_lpj) {
loops_per_jiffy = preset_lpj;
if (!printed)
pr_info("Calibrating delay loop (skipped) "
"preset value.. ");
} else if ((!printed) && lpj_fine) {
loops_per_jiffy = lpj_fine;
pr_info("Calibrating delay loop (skipped), "
"value calculated using timer frequency.. ");
} else if ((loops_per_jiffy = calibrate_delay_direct()) != 0) {
if (!printed)
pr_info("Calibrating delay using timer "
"specific routine.. ");
} else {
if (!printed)
pr_info("Calibrating delay loop... ");
loops_per_jiffy = calibrate_delay_converge();
}
if (!printed)
pr_cont("%lu.%02lu BogoMIPS (lpj=%lu)\n",
loops_per_jiffy/(500000/HZ),
(loops_per_jiffy/(5000/HZ)) % 100, loops_per_jiffy);
printed = true;
}