mirror of https://gitee.com/openkylin/linux.git
[PARISC] Further updates to timer_interrupt()
This version (relative to the current tree): o eliminates "while (ticks_elapsed)" loop. It's not needed. o drop "ticks_elapsed" completely from timer_interrupt(). o Estimates elapsed cycles (based on HZ) to see which kind of math we want to use to calculate "cycles_remainder". o Fixes a bug where we would loose a tick if we decided we wanted to skip one interrupt. Signed-off-by: Grant Grundler <grundler@parisc-linux.org> Signed-off-by: Kyle McMartin <kyle@parisc-linux.org>
This commit is contained in:
parent
6b799d9222
commit
6e5dc42b5a
|
@ -43,12 +43,11 @@ irqreturn_t timer_interrupt(int irq, void *dev_id, struct pt_regs *regs)
|
||||||
unsigned long now;
|
unsigned long now;
|
||||||
unsigned long next_tick;
|
unsigned long next_tick;
|
||||||
unsigned long cycles_elapsed;
|
unsigned long cycles_elapsed;
|
||||||
unsigned long cycles_remainder;
|
unsigned long cycles_remainder;
|
||||||
unsigned long ticks_elapsed = 1; /* at least one elapsed */
|
unsigned int cpu = smp_processor_id();
|
||||||
int cpu = smp_processor_id();
|
|
||||||
|
|
||||||
/* gcc can optimize for "read-only" case with a local clocktick */
|
/* gcc can optimize for "read-only" case with a local clocktick */
|
||||||
unsigned long local_ct = clocktick;
|
unsigned long cpt = clocktick;
|
||||||
|
|
||||||
profile_tick(CPU_PROFILING, regs);
|
profile_tick(CPU_PROFILING, regs);
|
||||||
|
|
||||||
|
@ -63,28 +62,16 @@ irqreturn_t timer_interrupt(int irq, void *dev_id, struct pt_regs *regs)
|
||||||
|
|
||||||
cycles_elapsed = now - next_tick;
|
cycles_elapsed = now - next_tick;
|
||||||
|
|
||||||
/* Determine how much time elapsed. */
|
if ((cycles_elapsed >> 5) < cpt) {
|
||||||
if (now < next_tick) {
|
/* use "cheap" math (add/subtract) instead
|
||||||
/* Scenario 2: CR16 wrapped after clock tick.
|
* of the more expensive div/mul method
|
||||||
* 1's complement will give us the "elapse cycles".
|
|
||||||
*
|
|
||||||
* This "cr16 wrapped" cruft is primarily for 32-bit kernels.
|
|
||||||
* So think "unsigned long is u32" when reading the code.
|
|
||||||
* And yes, of course 64-bit will someday wrap, but only
|
|
||||||
* every 198841 days on a 1GHz machine.
|
|
||||||
*/
|
*/
|
||||||
cycles_elapsed = ~cycles_elapsed; /* off by one cycle - don't care */
|
|
||||||
}
|
|
||||||
|
|
||||||
if (likely(cycles_elapsed < local_ct)) {
|
|
||||||
/* ticks_elapsed = 1 -- We already assumed one tick elapsed. */
|
|
||||||
cycles_remainder = cycles_elapsed;
|
cycles_remainder = cycles_elapsed;
|
||||||
|
while (cycles_remainder > cpt) {
|
||||||
|
cycles_remainder -= cpt;
|
||||||
|
}
|
||||||
} else {
|
} else {
|
||||||
/* more than one tick elapsed. Do "expensive" math. */
|
cycles_remainder = cycles_elapsed % cpt;
|
||||||
ticks_elapsed += cycles_elapsed / local_ct;
|
|
||||||
|
|
||||||
/* Faster version of "remainder = elapsed % clocktick" */
|
|
||||||
cycles_remainder = cycles_elapsed - (ticks_elapsed * local_ct);
|
|
||||||
}
|
}
|
||||||
|
|
||||||
/* Can we differentiate between "early CR16" (aka Scenario 1) and
|
/* Can we differentiate between "early CR16" (aka Scenario 1) and
|
||||||
|
@ -94,51 +81,65 @@ irqreturn_t timer_interrupt(int irq, void *dev_id, struct pt_regs *regs)
|
||||||
* cycles after the IT fires. But it's arbitrary how much time passes
|
* cycles after the IT fires. But it's arbitrary how much time passes
|
||||||
* before we call it "late". I've picked one second.
|
* before we call it "late". I've picked one second.
|
||||||
*/
|
*/
|
||||||
if (ticks_elapsed > HZ) {
|
/* aproximate HZ with shifts. Intended math is "(elapsed/clocktick) > HZ" */
|
||||||
|
#if HZ == 1000
|
||||||
|
if (cycles_elapsed > (cpt << 10) )
|
||||||
|
#elif HZ == 250
|
||||||
|
if (cycles_elapsed > (cpt << 8) )
|
||||||
|
#elif HZ == 100
|
||||||
|
if (cycles_elapsed > (cpt << 7) )
|
||||||
|
#else
|
||||||
|
#warn WTF is HZ set to anyway?
|
||||||
|
if (cycles_elapsed > (HZ * cpt) )
|
||||||
|
#endif
|
||||||
|
{
|
||||||
/* Scenario 3: very long delay? bad in any case */
|
/* Scenario 3: very long delay? bad in any case */
|
||||||
printk (KERN_CRIT "timer_interrupt(CPU %d): delayed!"
|
printk (KERN_CRIT "timer_interrupt(CPU %d): delayed!"
|
||||||
" ticks %ld cycles %lX rem %lX"
|
" cycles %lX rem %lX "
|
||||||
" next/now %lX/%lX\n",
|
" next/now %lX/%lX\n",
|
||||||
cpu,
|
cpu,
|
||||||
ticks_elapsed, cycles_elapsed, cycles_remainder,
|
cycles_elapsed, cycles_remainder,
|
||||||
next_tick, now );
|
next_tick, now );
|
||||||
}
|
}
|
||||||
|
|
||||||
|
/* convert from "division remainder" to "remainder of clock tick" */
|
||||||
|
cycles_remainder = cpt - cycles_remainder;
|
||||||
|
|
||||||
/* Determine when (in CR16 cycles) next IT interrupt will fire.
|
/* Determine when (in CR16 cycles) next IT interrupt will fire.
|
||||||
* We want IT to fire modulo clocktick even if we miss/skip some.
|
* We want IT to fire modulo clocktick even if we miss/skip some.
|
||||||
* But those interrupts don't in fact get delivered that regularly.
|
* But those interrupts don't in fact get delivered that regularly.
|
||||||
*/
|
*/
|
||||||
next_tick = now + (local_ct - cycles_remainder);
|
next_tick = now + cycles_remainder;
|
||||||
|
|
||||||
|
cpu_data[cpu].it_value = next_tick;
|
||||||
|
|
||||||
/* Skip one clocktick on purpose if we are likely to miss next_tick.
|
/* Skip one clocktick on purpose if we are likely to miss next_tick.
|
||||||
* We'll catch what we missed on the tick after that.
|
* We want to avoid the new next_tick being less than CR16.
|
||||||
* We should never need 0x1000 cycles to read CR16, calc the
|
* If that happened, itimer wouldn't fire until CR16 wrapped.
|
||||||
* new next_tick, then write CR16 back. */
|
* We'll catch the tick we missed on the tick after that.
|
||||||
if (!((local_ct - cycles_remainder) >> 12))
|
*/
|
||||||
next_tick += local_ct;
|
if (!(cycles_remainder >> 13))
|
||||||
|
next_tick += cpt;
|
||||||
|
|
||||||
/* Program the IT when to deliver the next interrupt. */
|
/* Program the IT when to deliver the next interrupt. */
|
||||||
/* Only bottom 32-bits of next_tick are written to cr16. */
|
/* Only bottom 32-bits of next_tick are written to cr16. */
|
||||||
cpu_data[cpu].it_value = next_tick;
|
|
||||||
mtctl(next_tick, 16);
|
mtctl(next_tick, 16);
|
||||||
|
|
||||||
/* Now that we are done mucking with unreliable delivery of interrupts,
|
|
||||||
* go do system house keeping.
|
/* Done mucking with unreliable delivery of interrupts.
|
||||||
|
* Go do system house keeping.
|
||||||
*/
|
*/
|
||||||
while (ticks_elapsed--) {
|
|
||||||
#ifdef CONFIG_SMP
|
#ifdef CONFIG_SMP
|
||||||
smp_do_timer(regs);
|
smp_do_timer(regs);
|
||||||
#else
|
#else
|
||||||
update_process_times(user_mode(regs));
|
update_process_times(user_mode(regs));
|
||||||
#endif
|
#endif
|
||||||
if (cpu == 0) {
|
if (cpu == 0) {
|
||||||
write_seqlock(&xtime_lock);
|
write_seqlock(&xtime_lock);
|
||||||
do_timer(1);
|
do_timer(regs);
|
||||||
write_sequnlock(&xtime_lock);
|
write_sequnlock(&xtime_lock);
|
||||||
}
|
|
||||||
}
|
}
|
||||||
|
|
||||||
/* check soft power switch status */
|
/* check soft power switch status */
|
||||||
if (cpu == 0 && !atomic_read(&power_tasklet.count))
|
if (cpu == 0 && !atomic_read(&power_tasklet.count))
|
||||||
tasklet_schedule(&power_tasklet);
|
tasklet_schedule(&power_tasklet);
|
||||||
|
@ -164,14 +165,12 @@ unsigned long profile_pc(struct pt_regs *regs)
|
||||||
EXPORT_SYMBOL(profile_pc);
|
EXPORT_SYMBOL(profile_pc);
|
||||||
|
|
||||||
|
|
||||||
/*** converted from ia64 ***/
|
|
||||||
/*
|
/*
|
||||||
* Return the number of micro-seconds that elapsed since the last
|
* Return the number of micro-seconds that elapsed since the last
|
||||||
* update to wall time (aka xtime). The xtime_lock
|
* update to wall time (aka xtime). The xtime_lock
|
||||||
* must be at least read-locked when calling this routine.
|
* must be at least read-locked when calling this routine.
|
||||||
*/
|
*/
|
||||||
static inline unsigned long
|
static inline unsigned long gettimeoffset (void)
|
||||||
gettimeoffset (void)
|
|
||||||
{
|
{
|
||||||
#ifndef CONFIG_SMP
|
#ifndef CONFIG_SMP
|
||||||
/*
|
/*
|
||||||
|
@ -185,36 +184,40 @@ gettimeoffset (void)
|
||||||
unsigned long elapsed_cycles;
|
unsigned long elapsed_cycles;
|
||||||
unsigned long usec;
|
unsigned long usec;
|
||||||
unsigned long cpuid = smp_processor_id();
|
unsigned long cpuid = smp_processor_id();
|
||||||
unsigned long local_ct = clocktick;
|
unsigned long cpt = clocktick;
|
||||||
|
|
||||||
next_tick = cpu_data[cpuid].it_value;
|
next_tick = cpu_data[cpuid].it_value;
|
||||||
now = mfctl(16); /* Read the hardware interval timer. */
|
now = mfctl(16); /* Read the hardware interval timer. */
|
||||||
|
|
||||||
prev_tick = next_tick - local_ct;
|
prev_tick = next_tick - cpt;
|
||||||
|
|
||||||
/* Assume Scenario 1: "now" is later than prev_tick. */
|
/* Assume Scenario 1: "now" is later than prev_tick. */
|
||||||
elapsed_cycles = now - prev_tick;
|
elapsed_cycles = now - prev_tick;
|
||||||
|
|
||||||
if (now < prev_tick) {
|
/* aproximate HZ with shifts. Intended math is "(elapsed/clocktick) > HZ" */
|
||||||
/* Scenario 2: CR16 wrapped!
|
#if HZ == 1000
|
||||||
* ones complement is off-by-one. Don't care.
|
if (elapsed_cycles > (cpt << 10) )
|
||||||
*/
|
#elif HZ == 250
|
||||||
elapsed_cycles = ~elapsed_cycles;
|
if (elapsed_cycles > (cpt << 8) )
|
||||||
}
|
#elif HZ == 100
|
||||||
|
if (elapsed_cycles > (cpt << 7) )
|
||||||
if (elapsed_cycles > (HZ * local_ct)) {
|
#else
|
||||||
|
#warn WTF is HZ set to anyway?
|
||||||
|
if (elapsed_cycles > (HZ * cpt) )
|
||||||
|
#endif
|
||||||
|
{
|
||||||
/* Scenario 3: clock ticks are missing. */
|
/* Scenario 3: clock ticks are missing. */
|
||||||
printk (KERN_CRIT "gettimeoffset(CPU %d): missing ticks!"
|
printk (KERN_CRIT "gettimeoffset(CPU %ld): missing %ld ticks!"
|
||||||
"cycles %lX prev/now/next %lX/%lX/%lX clock %lX\n",
|
" cycles %lX prev/now/next %lX/%lX/%lX clock %lX\n",
|
||||||
cpuid,
|
cpuid, elapsed_cycles / cpt,
|
||||||
elapsed_cycles, prev_tick, now, next_tick, local_ct);
|
elapsed_cycles, prev_tick, now, next_tick, cpt);
|
||||||
}
|
}
|
||||||
|
|
||||||
/* FIXME: Can we improve the precision? Not with PAGE0. */
|
/* FIXME: Can we improve the precision? Not with PAGE0. */
|
||||||
usec = (elapsed_cycles * 10000) / PAGE0->mem_10msec;
|
usec = (elapsed_cycles * 10000) / PAGE0->mem_10msec;
|
||||||
|
|
||||||
/* add in "lost" jiffies */
|
/* add in "lost" jiffies */
|
||||||
usec += local_ct * (jiffies - wall_jiffies);
|
usec += cpt * (jiffies - wall_jiffies);
|
||||||
return usec;
|
return usec;
|
||||||
#else
|
#else
|
||||||
return 0;
|
return 0;
|
||||||
|
|
Loading…
Reference in New Issue