mirror of
https://github.com/Plexi09/Nadir-Kernel.git
synced 2026-10-20 15:07:27 +02:00
Nadir 8254 PIT timer driver
This commit is contained in:
parent
47e83cf795
commit
6ee1900c1a
2 changed files with 151 additions and 0 deletions
95
kernel/arch/x86_64/pit.c
Normal file
95
kernel/arch/x86_64/pit.c
Normal file
|
|
@ -0,0 +1,95 @@
|
|||
/* 8254 PIT timer driver (see pit.h).
|
||||
*
|
||||
* Programs channel 0 in mode 3 (square-wave generator) at 100Hz on IRQ0.
|
||||
*
|
||||
* Why 1193182Hz base: the PIT input clock is the IBM-PC crystal frequency
|
||||
* (14.31818MHz / 12); every PC-compatible PIT ticks at this rate, so the
|
||||
* divisor for N Hz is always 1193182 / N.
|
||||
* Why mode 3: square-wave output stays high/low for half the period and
|
||||
* retriggers the IRQ line cleanly each cycle; mode 2 (rate generator)
|
||||
* emits narrow pulses that are fine too, but mode 3 is the conventional
|
||||
* choice for periodic system ticks on channel 0.
|
||||
* Why volatile + no locking: single-core kernel; s_ticks is written only
|
||||
* in IRQ0 context and read in mainline code. `volatile` keeps the compiler
|
||||
* from caching the counter across the hlt loop, and on x86-64 an aligned
|
||||
* 64-bit load/store is atomic, so no spinlock is needed while IF gates the
|
||||
* writer.
|
||||
*/
|
||||
|
||||
#include <stdint.h>
|
||||
|
||||
#include "idt.h"
|
||||
#include "io.h"
|
||||
#include "pit.h"
|
||||
|
||||
#define PIT_CMD 0x43U
|
||||
#define PIT_CH0 0x40U
|
||||
/* Ch0, lobyte/hibyte access, mode 3 square-wave, binary (not BCD). */
|
||||
#define PIT_CMD_CH0_MODE3 0x36U
|
||||
|
||||
/* Compile-time guard: a zero divisor would mean 65536 to the hardware and
|
||||
* silently run at ~18Hz instead of PIT_HZ. Frequency is fixed on purpose. */
|
||||
_Static_assert(PIT_DIVISOR != 0, "PIT divisor must not be zero");
|
||||
_Static_assert(PIT_DIVISOR <= 0xFFFFU, "PIT divisor must fit in 16 bits");
|
||||
|
||||
/* Port I/O is owned by include/io.h (single owner); irq_register_handler
|
||||
* is owned by idt.c via include/idt.h. */
|
||||
|
||||
/* Monotonic tick counter: written in IRQ context. */
|
||||
static volatile uint64_t s_ticks;
|
||||
/* pit_init is idempotent; re-entry after the first programming is a no-op. */
|
||||
static uint8_t s_initialized;
|
||||
|
||||
void pit_init(void)
|
||||
{
|
||||
uint16_t divisor;
|
||||
uint8_t lo;
|
||||
uint8_t hi;
|
||||
|
||||
if (s_initialized != 0) {
|
||||
return;
|
||||
}
|
||||
|
||||
/* No cli/sti here on purpose: latch programming is two plain outb
|
||||
* writes (safe with IF=0 at boot) and interrupt policy belongs to the
|
||||
* caller (kmain integrator), which unmasks IRQ0 and stis afterwards. */
|
||||
divisor = (uint16_t)PIT_DIVISOR;
|
||||
lo = (uint8_t)(divisor & 0xFFU);
|
||||
hi = (uint8_t)((divisor >> 8) & 0xFFU);
|
||||
|
||||
/* Latch order matters: command byte first selects ch0 + lobyte/hibyte
|
||||
* mode, then the divisor follows low byte first, high byte second. */
|
||||
outb(PIT_CMD, PIT_CMD_CH0_MODE3);
|
||||
outb(PIT_CH0, lo);
|
||||
outb(PIT_CH0, hi);
|
||||
|
||||
/* Strong idt.h registration: IRQ0 -> vector 32 wires pit_on_tick. */
|
||||
irq_register_handler(PIT_IRQ, pit_on_tick);
|
||||
|
||||
s_initialized = 1;
|
||||
}
|
||||
|
||||
void pit_on_tick(void)
|
||||
{
|
||||
/* Single increment async-safe, no logging or port I/O here. The PIC
|
||||
* EOI is owned by the IRQ dispatcher not this driver. */
|
||||
s_ticks++;
|
||||
}
|
||||
|
||||
uint64_t pit_ticks(void)
|
||||
{
|
||||
return s_ticks;
|
||||
}
|
||||
|
||||
void pit_sleep(uint64_t ticks)
|
||||
{
|
||||
uint64_t start = s_ticks;
|
||||
|
||||
/* hlt (not a spin loop) so the CPU sleeps until the next interrupt.
|
||||
* Assumes IF=1 and IRQ0 unmasked; with IF=0 this would halt forever,
|
||||
* which is why pit_init leaves interrupt policy to the caller. The
|
||||
* unsigned delta keeps working if s_ticks wraps mid-sleep. */
|
||||
while ((s_ticks - start) < ticks) {
|
||||
__asm__ volatile("hlt");
|
||||
}
|
||||
}
|
||||
Loading…
Add table
Add a link
Reference in a new issue