#define _GNU_SOURCE 1
#include "VoxelForgeCore/Linux77Profiler.h"

#if defined(VF_L77_PROFILE)

#include <algorithm>
#include <atomic>
#include <cerrno>
#include <cinttypes>
#include <cstdio>
#include <cstdlib>
#include <mutex>
#include <signal.h>
#include <csignal>
#include <sys/resource.h>
#include <sys/syscall.h>
#include <time.h>
#include <ucontext.h>
#include <unistd.h>
#include <pthread.h>

namespace VoxelForgeCore::Linux77Profiler
{
    namespace
    {
        constexpr std::uint32_t Capacity=1u<<20;
        constexpr std::size_t MaxRegisteredThreads=32;
        struct FSample
        {
            std::uintptr_t pc=0;
            std::uint8_t lod=0xff,route=0,stage=0,phase=0,halo=0,weight=1;
        };
        static_assert(std::atomic<std::uint32_t>::is_always_lock_free);
        static_assert(std::atomic<std::uint64_t>::is_always_lock_free);
        std::array<FSample,Capacity> Samples{};
        std::atomic<std::uint32_t> NextSample{0};
        std::atomic<std::uint64_t> LostSamples{0},HandlerSamples{0},HandlerNanoseconds{0};
        std::atomic<std::uint64_t> ExpandedTicks{0};
        std::atomic<bool> Enabled{false};
        volatile sig_atomic_t TimerActive=0;
        std::atomic<std::uint32_t> ActiveHandlers{0};
        std::uint32_t PeriodMicroseconds=1000;
        std::mutex TimerRegistryMutex;
        std::array<timer_t,MaxRegisteredThreads> ThreadTimers{};
        std::array<pid_t,MaxRegisteredThreads> ThreadIds{};
        std::array<clockid_t,MaxRegisteredThreads> ThreadCpuClockIds{};
        std::array<std::uint64_t,MaxRegisteredThreads> ThreadCpuStartNanoseconds{};
        std::array<std::uint64_t,MaxRegisteredThreads> ThreadProfileCpuNanoseconds{};
        std::array<std::atomic<std::uint64_t>,MaxRegisteredThreads> ThreadSampleCounts{};
        std::size_t ThreadTimerCount=0;
        inline thread_local bool ThreadTimerRegistered=false;
        std::atomic<std::uint64_t> TimerOverruns{0};

        std::uint64_t ThreadCpuNanoseconds() noexcept
        {
            timespec time{};
            if(clock_gettime(CLOCK_THREAD_CPUTIME_ID,&time)!=0)return 0;
            return static_cast<std::uint64_t>(time.tv_sec)*1000000000ull
                +static_cast<std::uint64_t>(time.tv_nsec);
        }
        std::uint64_t UserCpuNanoseconds(const timeval& value) noexcept
        {
            return static_cast<std::uint64_t>(value.tv_sec)*1000000000ull
                +static_cast<std::uint64_t>(value.tv_usec)*1000ull;
        }
        void SigprofHandler(int, siginfo_t* information, void* context) noexcept
        {
            if(!TimerActive)return;
            ActiveHandlers.fetch_add(1,std::memory_order_relaxed);
            const std::uint64_t start=ThreadCpuNanoseconds();
            HandlerSamples.fetch_add(1,std::memory_order_relaxed);
            const int timerIndex=information?information->si_value.sival_int:-1;
            if(timerIndex>=0&&static_cast<std::size_t>(timerIndex)<MaxRegisteredThreads)
            {
                ThreadSampleCounts[static_cast<std::size_t>(timerIndex)].fetch_add(
                    1,std::memory_order_relaxed);
                if(information&&information->si_overrun>0)
                    TimerOverruns.fetch_add(static_cast<std::uint64_t>(information->si_overrun),
                        std::memory_order_relaxed);
            }
            const std::uint32_t index=NextSample.fetch_add(1,std::memory_order_relaxed);
            const std::uint64_t overrun=information&&information->si_overrun>0
                ?static_cast<std::uint64_t>(information->si_overrun):0;
            const std::uint8_t weight=static_cast<std::uint8_t>(
                std::min<std::uint64_t>(255,overrun+1));
            ExpandedTicks.fetch_add(weight,std::memory_order_relaxed);
            if(index<Capacity)
            {
                std::uintptr_t pc=0;
#if defined(__x86_64__) && defined(REG_RIP)
                if(context)
                    pc=static_cast<std::uintptr_t>(
                        static_cast<ucontext_t*>(context)->uc_mcontext.gregs[REG_RIP]);
#endif
                const FThreadContext& tag=CurrentContext;
                Samples[index]={pc,tag.lod,static_cast<std::uint8_t>(tag.route),
                    static_cast<std::uint8_t>(tag.stage),
                    static_cast<std::uint8_t>(tag.phase),static_cast<std::uint8_t>(tag.halo),weight};
            }
            else LostSamples.fetch_add(1,std::memory_order_relaxed);
            const std::uint64_t end=ThreadCpuNanoseconds();
            if(end>=start)HandlerNanoseconds.fetch_add(end-start,std::memory_order_relaxed);
            ActiveHandlers.fetch_sub(1,std::memory_order_relaxed);
        }
    }

    bool IsEnabled() noexcept{return Enabled.load(std::memory_order_relaxed);}
    void SetEnabled(bool enabled) noexcept{Enabled.store(enabled,std::memory_order_relaxed);}

    void RegisterThread()
    {
        if(ThreadTimerRegistered||!IsEnabled())return;
        std::lock_guard<std::mutex> lock(TimerRegistryMutex);
        if(ThreadTimerRegistered||!IsEnabled()||ThreadTimerCount>=MaxRegisteredThreads)return;
        struct sigevent event{};
        event.sigev_notify=SIGEV_THREAD_ID;
        event.sigev_signo=SIGPROF;
        event.sigev_value.sival_int=static_cast<int>(ThreadTimerCount);
        event._sigev_un._tid=static_cast<int>(syscall(SYS_gettid));
        timer_t timer{};
        if(timer_create(CLOCK_THREAD_CPUTIME_ID,&event,&timer)!=0)
        {
            std::fprintf(stderr,"[L77SamplerError] action=thread_timer_create errno=%d\n",errno);
            return;
        }
        itimerspec interval{};
        interval.it_value.tv_nsec=static_cast<long>(PeriodMicroseconds%1000000u)*1000l;
        interval.it_value.tv_sec=static_cast<time_t>(PeriodMicroseconds/1000000u);
        interval.it_interval=interval.it_value;
        const std::size_t timerIndex=ThreadTimerCount;
        clockid_t threadCpuClock{};
        if(pthread_getcpuclockid(pthread_self(),&threadCpuClock)!=0)
        {
            timer_delete(timer);
            std::fprintf(stderr,"[L77SamplerError] action=thread_clock errno=%d\n",errno);
            return;
        }
        timespec threadCpuStart{};
        if(clock_gettime(threadCpuClock,&threadCpuStart)!=0)
        {
            timer_delete(timer);
            std::fprintf(stderr,"[L77SamplerError] action=thread_clock_read errno=%d\n",errno);
            return;
        }
        ThreadCpuClockIds[timerIndex]=threadCpuClock;
        ThreadCpuStartNanoseconds[timerIndex]=
            static_cast<std::uint64_t>(threadCpuStart.tv_sec)*1000000000ull
                +static_cast<std::uint64_t>(threadCpuStart.tv_nsec);
        ThreadIds[timerIndex]=static_cast<pid_t>(event._sigev_un._tid);
        if(timer_settime(timer,0,&interval,nullptr)!=0)
        {
            timer_delete(timer);
            std::fprintf(stderr,"[L77SamplerError] action=thread_timer_start errno=%d\n",errno);
            return;
        }
        ThreadTimers[ThreadTimerCount++]=timer;
        ThreadTimerRegistered=true;
    }

    void Start(std::uint32_t periodMicroseconds)
    {
        if(TimerActive)return;
        PeriodMicroseconds=periodMicroseconds?periodMicroseconds:1000;
        NextSample.store(0,std::memory_order_relaxed);
        LostSamples.store(0,std::memory_order_relaxed);
        HandlerSamples.store(0,std::memory_order_relaxed);
        HandlerNanoseconds.store(0,std::memory_order_relaxed);
        ExpandedTicks.store(0,std::memory_order_relaxed);
        TimerOverruns.store(0,std::memory_order_relaxed);
        struct sigaction action{};
        action.sa_sigaction=&SigprofHandler;
        action.sa_flags=SA_SIGINFO|SA_RESTART;
        sigemptyset(&action.sa_mask);
        if(sigaction(SIGPROF,&action,nullptr)!=0)
        {
            std::fprintf(stderr,"[L77SamplerError] action=install errno=%d\n",errno);
            return;
        }
        SetEnabled(true);
        TimerActive=1;
        RegisterThread();
        std::printf("[L77Sampler] action=start method=thread_cpu_timer period_us=%u capacity=%u\n",
            PeriodMicroseconds,Capacity);
    }

    void StopAndPrint()
    {
        if(!TimerActive)return;
        SetEnabled(false);
        std::size_t registeredThreadCount=0;
        std::uint64_t profiledThreadCpuNanoseconds=0;
        {
            std::lock_guard<std::mutex> lock(TimerRegistryMutex);
            registeredThreadCount=ThreadTimerCount;
            for(std::size_t index=0;index<ThreadTimerCount;++index)
            {
                itimerspec stopped{};
                timer_settime(ThreadTimers[index],0,&stopped,nullptr);
                timespec now{};
                if(clock_gettime(ThreadCpuClockIds[index],&now)==0)
                {
                    const std::uint64_t elapsed=static_cast<std::uint64_t>(now.tv_sec)
                        *1000000000ull+static_cast<std::uint64_t>(now.tv_nsec)
                        -ThreadCpuStartNanoseconds[index];
                    ThreadProfileCpuNanoseconds[index]=elapsed;
                    profiledThreadCpuNanoseconds+=elapsed;
                }
                timer_delete(ThreadTimers[index]);
            }
            ThreadTimerCount=0;
        }
        TimerActive=0;
        while(ActiveHandlers.load(std::memory_order_relaxed)!=0){}
        rusage usage{};
        getrusage(RUSAGE_SELF,&usage);
        const std::uint64_t userNs=UserCpuNanoseconds(usage.ru_utime);
        const std::uint64_t systemNs=UserCpuNanoseconds(usage.ru_stime);
        const std::uint32_t stored=std::min(NextSample.load(std::memory_order_relaxed),Capacity);
        const std::uint64_t handlerCount=HandlerSamples.load(std::memory_order_relaxed);
        const std::uint64_t handlerNs=HandlerNanoseconds.load(std::memory_order_relaxed);
        const std::uint64_t expandedTicks=ExpandedTicks.load(std::memory_order_relaxed);
        std::printf("[L77Sampler] action=stop samples=%u handlers=%llu expanded_ticks=%llu lost=%llu timer_overruns=%llu timer_threads=%zu requested_period_us=%u profiled_thread_cpu_s=%.6f whole_process_cpu_s=%.6f handler_cpu_ms=%.3f handler_share_pct=%.3f observed_us_per_tick=%.3f status=%s\n",
            stored,static_cast<unsigned long long>(handlerCount),
            static_cast<unsigned long long>(expandedTicks),
            static_cast<unsigned long long>(LostSamples.load(std::memory_order_relaxed)),
            static_cast<unsigned long long>(TimerOverruns.load(std::memory_order_relaxed)),
            registeredThreadCount,PeriodMicroseconds,
            static_cast<double>(profiledThreadCpuNanoseconds)/1.0e9,
            static_cast<double>(userNs+systemNs)/1.0e9,
            static_cast<double>(handlerNs)/1.0e6,
            profiledThreadCpuNanoseconds?100.0*static_cast<double>(handlerNs)
                /static_cast<double>(profiledThreadCpuNanoseconds):0.0,
            expandedTicks?static_cast<double>(profiledThreadCpuNanoseconds)
                /static_cast<double>(expandedTicks)/1000.0:0.0,
            stored>0&&handlerCount>0?"PASS":"NO_SAMPLES");
        for(std::size_t index=0;index<registeredThreadCount;++index)
            std::printf("[L77ThreadTimer] index=%zu tid=%d cpu_ms=%.3f samples=%llu\n",index,
                static_cast<int>(ThreadIds[index]),
                static_cast<double>(ThreadProfileCpuNanoseconds[index])/1.0e6,
                static_cast<unsigned long long>(ThreadSampleCounts[index].load(
                    std::memory_order_relaxed)));
        for(std::uint32_t index=0;index<stored;++index)
        {
            const FSample& sample=Samples[index];
            const auto route=static_cast<ERoute>(sample.route);
            const auto stage=static_cast<EStage>(sample.stage);
            const auto phase=static_cast<EPhase>(sample.phase);
            std::printf("[L77Sample] pc=0x%" PRIxPTR " lod=%u stage=%s route=%s phase=%s halo=%u weight=%u\n",
                sample.pc,static_cast<unsigned>(sample.lod),StageName(stage),
                RouteName(route),PhaseName(phase),static_cast<unsigned>(sample.halo),
                static_cast<unsigned>(sample.weight));
        }
        std::fflush(stdout);
    }
}

#endif