/**
 * Copyright 2019 Sims Hill Osborne and Joshua Bakita
 *
 * This header provides facilities by which to separably run and time TACLeBench
 **/
#define _GNU_SOURCE
#include <time.h>
#include <sys/mman.h>
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <signal.h>
#include <limits.h>
#include <fcntl.h>
#include <stdint.h>
#include <sched.h>

// This is only visible if _GNU_SOURCE is defined, and that define does not
// come along to places where this file is included. Address this by manually
// forcing it into the global namespace.
extern int sched_getcpu();

// These constants correspond to the imx6q-sabredb platform
#define LINE_SIZE 32
#define L2_SIZE 16*2048*32

#if __arm__
#include <unistd.h>
#include <sys/syscall.h>
#endif

#define LITMUS 0
#define MC2 0
#define MMDC_PROF 0

#if LITMUS
#include <litmus.h>
#endif

#if MMDC_PROF
#include  "/media/speedy/litmus/tools/mmdc/mmdc.h"
#endif

#if LITMUS
#define SET_UP LOAD_PARAMS SETUP_LITMUS
#else
#define SET_UP LOAD_PARAMS
#endif

#if MMDC_PROF
#define LOAD_PARAMS LOAD_PARAMS_ITRL SETUP_MMDC
#else
#define LOAD_PARAMS LOAD_PARAMS_ITRL
#endif

#define LOAD_PARAMS_ITRL \
    if (argc != 6) { \
        fprintf(stderr, "Usage: %s <name> <loops> <my core> <runID> <save results?>\n", argv[0]);\
        fprintf(stderr, " <loops> integer number of iterations. -1 for infitite.\n");\
        fprintf(stderr, " <save results?> 1 to save results, 0 to discard.\n");\
        exit(1);\
    }\
    char *thisProgram=argv[1];\
    int parsedMaxJobs=atoi(argv[2]);\
    unsigned int thisCore=atoi(argv[3]);\
    thisCore = sched_getcpu();\
    char *runID=argv[4];\
    int output=atoi(argv[5]);\
    if (parsedMaxJobs < 0 && output != 0){\
        fprintf(stderr, "Infinite loops only supported when output is disabled!\n");\
        exit(1);\
    }\
    /* Cheat. -1 is larger than jobsComplete can ever reach. */\
    unsigned int maxJobs = parsedMaxJobs;\
    struct timespec _start, _end;\
    int jobsComplete;\
    int jobs_complete = -1;\
    float *progTime = malloc(sizeof(float)*maxJobs*output);\
    memset(progTime, 0, sizeof(float)*maxJobs*output);\
    char fileName[64];\
    float *mmdc_read = malloc(sizeof(float)*maxJobs*output);\
    float *mmdc_write = malloc(sizeof(float)*maxJobs*output);\
    memset(mmdc_read, 0, sizeof(float)*maxJobs*output);\
    memset(mmdc_write, 0, sizeof(float)*maxJobs*output);\
    if (strlen(runID) + 5 > sizeof(fileName)) {\
        fprintf(stderr, "Run ID is too large! Keep it to less than 60 characters.\n");\
        exit(1);\
    }\
    strcpy(fileName, runID);\
    strcat(fileName, ".txt");\
    mlockall(MCL_CURRENT || MCL_FUTURE);

#define SETUP_MMDC \
    MMDC_PROFILE_RES_t mmdc_res; \
    memset(&mmdc_res, 0, sizeof(MMDC_PROFILE_RES_t));\
    int fd = open("/dev/mem", O_RDWR, 0);\
    if (fd < 0) {\
        perror("Unable to open /dev/mem");\
        exit(1);\
    }\
    pMMDC_t mmdc = mmap(NULL, 0x4000, PROT_READ | PROT_WRITE, MAP_SHARED, fd, MMDC_P0_IPS_BASE_ADDR);\
    if (mmdc == MAP_FAILED) {\
        perror("Unable to map MMDC address space");\
        exit(1);\
    }\
    mmdc->madpcr1 = axi_arm1;\
    msync(&(mmdc->madpcr1),4,MS_SYNC);

#define SETUP_LITMUS \
    unsigned int cpu = atoi(thisCore); \
    unsigned int wait = 0; \
    if (be_migrate_to_domain(cpu) < 0) { \
        perror("Unable to migrate to specified CPU"); \
        exit(1); \
    } \
    struct reservation_config res; \
    res.id = gettid(); \
    res.cpu = cpu; \
    res.priority = LITMUS_HIGHEST_PRIORITY; \
    /* we take over half the CPU time (these are ns) */ \
    res.polling_params.budget = ms2ns(3000); \
    res.polling_params.period = ms2ns(3000); \
    res.polling_params.offset = 0; \
    res.polling_params.relative_deadline = ms2ns(3000); \
    /* Not 100% sure that we should use periodic polling */ \
    if (reservation_create(PERIODIC_POLLING, &res) < 0) { \
        perror("Unable to create reservation"); \
        exit(1); \
    } \
    struct rt_task rt_param; \
    init_rt_task_param(&rt_param); \
    /* Supposedly the next two parameters are irrelevant when reservations are enabled, but I'm leaving them anyway... */ \
    rt_param.exec_cost = ms2ns(999); \
    rt_param.period = ms2ns(1000); \
    rt_param.priority = LITMUS_HIGHEST_PRIORITY; \
    rt_param.cls = RT_CLASS_HARD; \
    rt_param.release_policy = TASK_PERIODIC; \
    rt_param.budget_policy = NO_ENFORCEMENT; \
    rt_param.cpu = cpu; \
    if (set_rt_task_param(gettid(), &rt_param) < 0) { \
        perror("Unable to set real-time parameters"); \
        exit(1); \
    } \
    if (init_litmus() != 0) { \
        perror("init_litmus failed"); \
        exit(1); \
    } \
    MC2_SETUP \
    if (task_mode(LITMUS_RT_TASK) != 0) { \
        perror("Unable to become real-time task"); \
        exit(1); \
    } \
    if (wait && wait_for_ts_release() != 0) { \
        perror("Unable to wait for taskset release"); \
        exit(1); \
    }

#if MC2
#define MC2_SETUP \
    struct mc2_task mc2_param; \
    mc2_param.res_id = gettid(); \
    mc2_param.crit = CRIT_LEVEL_A; \
    if (set_mc2_task_param(gettid(), &mc2_param) < 0) { \
        perror("Unable to set MC^2 task params"); \
        exit(1); \
    } \
    set_page_color(rt_param.cpu);
#else
#define MC2_SETUP
#endif

#define CLEANUP_LITMUS \
    if (task_mode(BACKGROUND_TASK) != 0) { \
        perror("Unable to become a real-time task"); \
        exit(1); \
    } \
    reservation_destroy(gettid(), rt_param.cpu);

#if MMDC_PROF
#define SAVE_RESULTS \
    if(jobs_complete >= maxJobs) {\
        fprintf(stderr, "Max jobs setting too small! Trying to record job #%d when we only have space for %d jobs. Exiting...\n", jobs_complete, maxJobs);\
        exit(1);\
    }\
    if(jobs_complete > -1 && output) {\
        progTime[jobs_complete] = _end.tv_sec - _start.tv_sec;\
        progTime[jobs_complete] *= 1000000000;\
        progTime[jobs_complete] += _end.tv_nsec - _start.tv_nsec;\
        mmdc_read[jobs_complete] = mmdc_res.read_bytes;\
        mmdc_write[jobs_complete] = mmdc_res.write_bytes;\
    }
#else
#define SAVE_RESULTS \
    if(jobs_complete >= maxJobs) {\
        fprintf(stderr, "Max jobs setting too small! Trying to record job #%d when we only have space for %d jobs. Exiting...\n", jobs_complete, maxJobs);\
        exit(1);\
    }\
    if(jobs_complete > -1 && output) {\
        progTime[jobs_complete] = _end.tv_sec - _start.tv_sec;\
        progTime[jobs_complete] *= 1000000000;\
        progTime[jobs_complete] += _end.tv_nsec - _start.tv_nsec;\
    }
#endif


#if LITMUS
#define WRITE_TO_FILE WRITE_TO_FILE_ITRNL CLEANUP_LITMUS
#else
#define WRITE_TO_FILE WRITE_TO_FILE_ITRNL
#endif

#define WRITE_TO_FILE_ITRNL if (output){\
    munlockall();\
    FILE *fp=fopen(fileName, "a");\
    if (fp == NULL) {\
        perror("Unable to open output file");\
        exit(1);\
    }\
    for (int i = 0; i <= jobs_complete; i++){\
        fprintf(fp, "%s none %u none %d %.f %s %d %.f %.f \n",\
        thisProgram, thisCore, maxJobs,\
        progTime[i], runID, i, mmdc_read[i], mmdc_write[i]);\
    }\
    fclose(fp);\
}

#if __arm__
// On ARM, manually flush the cache
#define FLUSH_CACHES \
    volatile uint8_t buffer[L2_SIZE * 4]; \
    for (uint32_t j = 0; j < 4; j++) \
        for (uint32_t i = 0; i < L2_SIZE * 4; i += LINE_SIZE) \
            buffer[i]++;
#else
// On x86 call the wbinvld instruction (it's in a kernel module due to it being ring-0)
#define FLUSH_CACHES \
    FILE *fp = fopen("/proc/wbinvd", "r");\
    if (fp == NULL) {\
        perror("Cache flush module interface cannot be opened");\
        exit(1);\
    }\
    char dummy;\
    if (fread(&dummy, 1, 1, fp) == 0) {\
        perror("Unable to access cache flush module interface");\
        exit(1);\
    }\
    fclose(fp);
#endif

#if MMDC_PROF
#define START_TIMER \
    /* This disables profiling, resets the counters, clears the overflow bit, and enables profiling */ \
    start_mmdc_profiling(mmdc); \
    /*nanosleep(&(struct timespec){0, ms2ns(999)}, NULL);*/ \
    clock_gettime(CLOCK_MONOTONIC, &_start);
#else
#define START_TIMER clock_gettime(CLOCK_MONOTONIC, &_start);
#endif

#if MMDC_PROF
#define STOP_TIMER \
    clock_gettime(CLOCK_MONOTONIC, &_end); \
    /* This freezes the profiling and makes results available */ \
    pause_mmdc_profiling(mmdc); \
    get_mmdc_profiling_results(mmdc, &mmdc_res);
#else
#define STOP_TIMER clock_gettime(CLOCK_MONOTONIC, &_end);
#endif

#define SLEEP nanosleep((const struct timespec[]){{0, 1000000}}, NULL);

#if LITMUS
#define START_LOOP \
    if (sleep_next_period() != 0) { \
        perror("Unable to sleep for next period"); \
    } \
    FLUSH_CACHES START_TIMER
#else
#define START_LOOP sched_yield(); FLUSH_CACHES START_TIMER
#endif

#define STOP_LOOP STOP_TIMER jobs_complete++; SAVE_RESULTS


/*
Intended structure

main
SET_UP
notice that STOP LOOP negates the ++ if outout=0
for (jobsComplete=0; jobsComplete<maxJobs; jobsComplete++){
    START_LOOP
    tacleInit();
    tacleMain();
    STOP_LOOP
}
WRITE_TO_FILE
tacleReturn 
*/