// cpu-energy-hold.c: read IOReport "CPU Energy" without sudo and turn it into watts
// two ways: the naive per-interval delta, and a rate held by the driver timestamp.
// Build: clang -O2 -Wall -o cpu-energy-hold cpu-energy-hold.c -framework CoreFoundation -lIOReport
// Run:   ./cpu-energy-hold [interval_ms=1000] [samples=30]
// Tested on macOS 27.0 (Mac17,8, M5 Pro). libIOReport is private: these prototypes
// match what macmon and other open-source samplers declare, not an Apple header.
#include <CoreFoundation/CoreFoundation.h>
#include <mach/mach_time.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>

typedef struct IOReportSubscription *IOReportSubscriptionRef;
CFDictionaryRef IOReportCopyChannelsInGroup(CFStringRef, CFStringRef, uint64_t, uint64_t, uint64_t);
IOReportSubscriptionRef IOReportCreateSubscription(void *, CFMutableDictionaryRef,
                                                   CFMutableDictionaryRef *, uint64_t, CFTypeRef);
CFDictionaryRef IOReportCreateSamples(IOReportSubscriptionRef, CFMutableDictionaryRef, CFTypeRef);
CFStringRef IOReportChannelGetChannelName(CFDictionaryRef);
CFStringRef IOReportChannelGetUnitLabel(CFDictionaryRef);
int64_t IOReportSimpleGetIntegerValue(CFDictionaryRef, int32_t);

static double joules_per_unit(CFStringRef unit) {
    if (unit && CFStringCompare(unit, CFSTR("mJ"), 0) == 0) return 1e-3;
    if (unit && CFStringCompare(unit, CFSTR("nJ"), 0) == 0) return 1e-9;
    return 1e-6;  // "uJ" / "µJ"
}

// One raw sample of "CPU Energy": running total in joules plus the driver's timestamp.
static int read_cpu_energy(IOReportSubscriptionRef sub, CFMutableDictionaryRef subbed,
                           double *joules, uint64_t *driver_ts) {
    CFDictionaryRef s = IOReportCreateSamples(sub, subbed, NULL);
    if (!s) return 0;
    CFArrayRef chans = CFDictionaryGetValue(s, CFSTR("IOReportChannels"));
    int found = 0;
    for (CFIndex i = 0; chans && i < CFArrayGetCount(chans); i++) {
        CFDictionaryRef ch = CFArrayGetValueAtIndex(chans, i);
        CFStringRef name = IOReportChannelGetChannelName(ch);
        if (!name || CFStringCompare(name, CFSTR("CPU Energy"), 0) != 0) continue;
        *joules = (double)IOReportSimpleGetIntegerValue(ch, 0) *
                  joules_per_unit(IOReportChannelGetUnitLabel(ch));
        // RawElements holds an IOReportElement: u64 provider_id, u64 channel_id,
        // 8-byte channel type, then u64 timestamp (mach absolute ticks) at byte 24.
        CFDataRef raw = CFDictionaryGetValue(ch, CFSTR("RawElements"));
        *driver_ts = 0;
        if (raw && CFDataGetLength(raw) >= 32) memcpy(driver_ts, CFDataGetBytePtr(raw) + 24, 8);
        found = 1;
        break;
    }
    CFRelease(s);
    return found;
}

int main(int argc, char **argv) {
    int ms = argc > 1 ? atoi(argv[1]) : 1000, n = argc > 2 ? atoi(argv[2]) : 30;
    mach_timebase_info_data_t tb;
    mach_timebase_info(&tb);
    double sec_per_tick = (double)tb.numer / tb.denom / 1e9;

    CFDictionaryRef group = IOReportCopyChannelsInGroup(CFSTR("Energy Model"), NULL, 0, 0, 0);
    if (!group) return 1;
    CFMutableDictionaryRef desired = CFDictionaryCreateMutableCopy(NULL, 0, group);
    CFMutableDictionaryRef subbed = NULL;
    IOReportSubscriptionRef sub = IOReportCreateSubscription(NULL, desired, &subbed, 0, NULL);
    if (!sub) return 1;

    double j_prev = 0, j_pub = 0, held_w = -1;
    uint64_t ts_pub = 0;
    if (!read_cpu_energy(sub, subbed, &j_prev, &ts_pub)) return 1;
    j_pub = j_prev;
    int pubs = 0;
    printf("%4s  %12s  %12s  %9s\n", "t_s", "naive_W", "held_W", "age_s");
    for (int i = 1; i <= n; i++) {
        usleep((useconds_t)ms * 1000);
        double j;
        uint64_t ts;
        if (!read_cpu_energy(sub, subbed, &j, &ts)) return 1;
        double naive_w = (j - j_prev) / (ms / 1e3);  // what a per-interval sampler shows
        j_prev = j;
        if (ts != ts_pub) {  // a new publication: rate over the driver's own window
            double span = (double)(ts - ts_pub) * sec_per_tick;
            if (ts > ts_pub && span >= 0.01 && j >= j_pub) {
                held_w = (j - j_pub) / span;
                pubs++;
            } else {
                held_w = -1;  // counter reset or clock went backwards: start over
            }
            ts_pub = ts;
            j_pub = j;
        }
        double age = (double)(mach_absolute_time() - ts_pub) * sec_per_tick;
        if (held_w >= 0 && age <= 10.0)
            printf("%4d  %12.2f  %12.2f  %9.1f\n", i * ms / 1000, naive_w, held_w, age);
        else  // never a fake zero: no publication yet, or the last one is stale
            printf("%4d  %12.2f  %12s  %9.1f\n", i * ms / 1000, naive_w, "-", age);
    }
    fprintf(stderr, "%d publications in %d samples\n", pubs, n);
    return 0;
}