source: libcfa/src/concurrency/ready_queue.cfa@ becb85b9

//
// Cforall Version 1.0.0 Copyright (C) 2019 University of Waterloo
//
// The contents of this file are covered under the licence agreement in the
// file "LICENCE" distributed with Cforall.
//
// ready_queue.cfa --
//
// Author           : Thierry Delisle
// Created On       : Mon Nov dd 16:29:18 2019
// Last Modified By :
// Last Modified On :
// Update Count     :
//

#define __cforall_thread__
#define _GNU_SOURCE

// #define __CFA_DEBUG_PRINT_READY_QUEUE__


// #define USE_RELAXED_FIFO
// #define USE_WORK_STEALING
// #define USE_CPU_WORK_STEALING
#define USE_AWARE_STEALING

#include "bits/defs.hfa"
#include "device/cpu.hfa"
#include "kernel_private.hfa"

#include "stdlib.hfa"
#include "limits.hfa"
#include "math.hfa"

#include <errno.h>
#include <unistd.h>

extern "C" {
        #include <sys/syscall.h>  // __NR_xxx
}

#include "ready_subqueue.hfa"

static const size_t cache_line_size = 64;

#if !defined(__CFA_NO_STATISTICS__)
        #define __STATS(...) __VA_ARGS__
#else
        #define __STATS(...)
#endif

// No overriden function, no environment variable, no define
// fall back to a magic number
#ifndef __CFA_MAX_PROCESSORS__
        #define __CFA_MAX_PROCESSORS__ 1024
#endif

#if   defined(USE_AWARE_STEALING)
        #define READYQ_SHARD_FACTOR 2
        #define SEQUENTIAL_SHARD 2
#elif defined(USE_CPU_WORK_STEALING)
        #define READYQ_SHARD_FACTOR 2
#elif defined(USE_RELAXED_FIFO)
        #define BIAS 4
        #define READYQ_SHARD_FACTOR 4
        #define SEQUENTIAL_SHARD 1
#elif defined(USE_WORK_STEALING)
        #define READYQ_SHARD_FACTOR 2
        #define SEQUENTIAL_SHARD 2
#else
        #error no scheduling strategy selected
#endif

static inline struct thread$ * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats));
static inline struct thread$ * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats));
static inline struct thread$ * search(struct cluster * cltr);
static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred);


// returns the maximum number of processors the RWLock support
__attribute__((weak)) unsigned __max_processors() {
        const char * max_cores_s = getenv("CFA_MAX_PROCESSORS");
        if(!max_cores_s) {
                __cfadbg_print_nolock(ready_queue, "No CFA_MAX_PROCESSORS in ENV\n");
                return __CFA_MAX_PROCESSORS__;
        }

        char * endptr = 0p;
        long int max_cores_l = strtol(max_cores_s, &endptr, 10);
        if(max_cores_l < 1 || max_cores_l > 65535) {
                __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS out of range : %ld\n", max_cores_l);
                return __CFA_MAX_PROCESSORS__;
        }
        if('\0' != *endptr) {
                __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS not a decimal number : %s\n", max_cores_s);
                return __CFA_MAX_PROCESSORS__;
        }

        return max_cores_l;
}

#if   defined(CFA_HAVE_LINUX_LIBRSEQ)
        // No forward declaration needed
        #define __kernel_rseq_register rseq_register_current_thread
        #define __kernel_rseq_unregister rseq_unregister_current_thread
#elif defined(CFA_HAVE_LINUX_RSEQ_H)
        static void __kernel_raw_rseq_register  (void);
        static void __kernel_raw_rseq_unregister(void);

        #define __kernel_rseq_register __kernel_raw_rseq_register
        #define __kernel_rseq_unregister __kernel_raw_rseq_unregister
#else
        // No forward declaration needed
        // No initialization needed
        static inline void noop(void) {}

        #define __kernel_rseq_register noop
        #define __kernel_rseq_unregister noop
#endif

//=======================================================================
// Cluster wide reader-writer lock
//=======================================================================
void  ?{}(__scheduler_RWLock_t & this) {
        this.max   = __max_processors();
        this.alloc = 0;
        this.ready = 0;
        this.data  = alloc(this.max);
        this.write_lock  = false;

        /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.alloc), &this.alloc));
        /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.ready), &this.ready));

}
void ^?{}(__scheduler_RWLock_t & this) {
        free(this.data);
}


//=======================================================================
// Lock-Free registering/unregistering of threads
unsigned register_proc_id( void ) with(*__scheduler_lock) {
        __kernel_rseq_register();

        bool * handle = (bool *)&kernelTLS().sched_lock;

        // Step - 1 : check if there is already space in the data
        uint_fast32_t s = ready;

        // Check among all the ready
        for(uint_fast32_t i = 0; i < s; i++) {
                bool * volatile * cell = (bool * volatile *)&data[i]; // Cforall is bugged and the double volatiles causes problems
                /* paranoid */ verify( handle != *cell );

                bool * null = 0p; // Re-write every loop since compare thrashes it
                if( __atomic_load_n(cell, (int)__ATOMIC_RELAXED) == null
                        && __atomic_compare_exchange_n( cell, &null, handle, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
                        /* paranoid */ verify(i < ready);
                        /* paranoid */ verify( (kernelTLS().sched_id = i, true) );
                        return i;
                }
        }

        if(max <= alloc) abort("Trying to create more than %ud processors", __scheduler_lock->max);

        // Step - 2 : F&A to get a new spot in the array.
        uint_fast32_t n = __atomic_fetch_add(&alloc, 1, __ATOMIC_SEQ_CST);
        if(max <= n) abort("Trying to create more than %ud processors", __scheduler_lock->max);

        // Step - 3 : Mark space as used and then publish it.
        data[n] = handle;
        while() {
                unsigned copy = n;
                if( __atomic_load_n(&ready, __ATOMIC_RELAXED) == n
                        && __atomic_compare_exchange_n(&ready, &copy, n + 1, true, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))
                        break;
                Pause();
        }

        // Return new spot.
        /* paranoid */ verify(n < ready);
        /* paranoid */ verify( (kernelTLS().sched_id = n, true) );
        return n;
}

void unregister_proc_id( unsigned id ) with(*__scheduler_lock) {
        /* paranoid */ verify(id < ready);
        /* paranoid */ verify(id == kernelTLS().sched_id);
        /* paranoid */ verify(data[id] == &kernelTLS().sched_lock);

        bool * volatile * cell = (bool * volatile *)&data[id]; // Cforall is bugged and the double volatiles causes problems

        __atomic_store_n(cell, 0p, __ATOMIC_RELEASE);

        __kernel_rseq_unregister();
}

//-----------------------------------------------------------------------
// Writer side : acquire when changing the ready queue, e.g. adding more
//  queues or removing them.
uint_fast32_t ready_mutate_lock( void ) with(*__scheduler_lock) {
        /* paranoid */ verify( ! __preemption_enabled() );

        // Step 1 : lock global lock
        // It is needed to avoid processors that register mid Critical-Section
        //   to simply lock their own lock and enter.
        __atomic_acquire( &write_lock );

        // Make sure we won't deadlock ourself
        // Checking before acquiring the writer lock isn't safe
        // because someone else could have locked us.
        /* paranoid */ verify( ! kernelTLS().sched_lock );

        // Step 2 : lock per-proc lock
        // Processors that are currently being registered aren't counted
        //   but can't be in read_lock or in the critical section.
        // All other processors are counted
        uint_fast32_t s = ready;
        for(uint_fast32_t i = 0; i < s; i++) {
                volatile bool * llock = data[i];
                if(llock) __atomic_acquire( llock );
        }

        /* paranoid */ verify( ! __preemption_enabled() );
        return s;
}

void ready_mutate_unlock( uint_fast32_t last_s ) with(*__scheduler_lock) {
        /* paranoid */ verify( ! __preemption_enabled() );

        // Step 1 : release local locks
        // This must be done while the global lock is held to avoid
        //   threads that where created mid critical section
        //   to race to lock their local locks and have the writer
        //   immidiately unlock them
        // Alternative solution : return s in write_lock and pass it to write_unlock
        for(uint_fast32_t i = 0; i < last_s; i++) {
                volatile bool * llock = data[i];
                if(llock) __atomic_store_n(llock, (bool)false, __ATOMIC_RELEASE);
        }

        // Step 2 : release global lock
        /*paranoid*/ assert(true == write_lock);
        __atomic_store_n(&write_lock, (bool)false, __ATOMIC_RELEASE);

        /* paranoid */ verify( ! __preemption_enabled() );
}

//=======================================================================
// caches handling

struct __attribute__((aligned(128))) __ready_queue_caches_t {
        // Count States:
        // - 0  : No one is looking after this cache
        // - 1  : No one is looking after this cache, BUT it's not empty
        // - 2+ : At least one processor is looking after this cache
        volatile unsigned count;
};

void  ?{}(__ready_queue_caches_t & this) { this.count = 0; }
void ^?{}(__ready_queue_caches_t & this) {}

static inline void depart(__ready_queue_caches_t & cache) {
        /* paranoid */ verify( cache.count > 1);
        __atomic_fetch_add(&cache.count, -1, __ATOMIC_SEQ_CST);
        /* paranoid */ verify( cache.count != 0);
        /* paranoid */ verify( cache.count < 65536 ); // This verify assumes no cluster will have more than 65000 kernel threads mapped to a single cache, which could be correct but is super weird.
}

static inline void arrive(__ready_queue_caches_t & cache) {
        // for() {
        //      unsigned expected = cache.count;
        //      unsigned desired  = 0 == expected ? 2 : expected + 1;
        // }
}

//=======================================================================
// Cforall Ready Queue used for scheduling
//=======================================================================
unsigned long long moving_average(unsigned long long currtsc, unsigned long long instsc, unsigned long long old_avg) {
        /* paranoid */ verifyf( currtsc < 45000000000000000, "Suspiciously large current time: %'llu (%llx)\n", currtsc, currtsc );
        /* paranoid */ verifyf( instsc  < 45000000000000000, "Suspiciously large insert time: %'llu (%llx)\n", instsc, instsc );
        /* paranoid */ verifyf( old_avg < 15000000000000, "Suspiciously large previous average: %'llu (%llx)\n", old_avg, old_avg );

        const unsigned long long new_val = currtsc > instsc ? currtsc - instsc : 0;
        const unsigned long long total_weight = 16;
        const unsigned long long new_weight   = 4;
        const unsigned long long old_weight = total_weight - new_weight;
        const unsigned long long ret = ((new_weight * new_val) + (old_weight * old_avg)) / total_weight;
        return ret;
}

void ?{}(__ready_queue_t & this) with (this) {
        #if defined(USE_CPU_WORK_STEALING)
                lanes.count = cpu_info.hthrd_count * READYQ_SHARD_FACTOR;
                lanes.data = alloc( lanes.count );
                lanes.tscs = alloc( lanes.count );
                lanes.help = alloc( cpu_info.hthrd_count );

                for( idx; (size_t)lanes.count ) {
                        (lanes.data[idx]){};
                        lanes.tscs[idx].tv = rdtscl();
                        lanes.tscs[idx].ma = rdtscl();
                }
                for( idx; (size_t)cpu_info.hthrd_count ) {
                        lanes.help[idx].src = 0;
                        lanes.help[idx].dst = 0;
                        lanes.help[idx].tri = 0;
                }
        #else
                lanes.data   = 0p;
                lanes.tscs   = 0p;
                lanes.caches = 0p;
                lanes.help   = 0p;
                lanes.count  = 0;
        #endif
}

void ^?{}(__ready_queue_t & this) with (this) {
        #if !defined(USE_CPU_WORK_STEALING)
                verify( SEQUENTIAL_SHARD == lanes.count );
        #endif

        free(lanes.data);
        free(lanes.tscs);
        free(lanes.caches);
        free(lanes.help);
}

//-----------------------------------------------------------------------
#if defined(USE_AWARE_STEALING)
        __attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint) with (cltr->ready_queue) {
                processor * const proc = kernelTLS().this_processor;
                const bool external = (!proc) || (cltr != proc->cltr);
                const bool remote   = hint == UNPARK_REMOTE;

                unsigned i;
                if( external || remote ) {
                        // Figure out where thread was last time and make sure it's valid
                        /* paranoid */ verify(thrd->preferred >= 0);
                        if(thrd->preferred * READYQ_SHARD_FACTOR < lanes.count) {
                                /* paranoid */ verify(thrd->preferred * READYQ_SHARD_FACTOR < lanes.count);
                                unsigned start = thrd->preferred * READYQ_SHARD_FACTOR;
                                do {
                                        unsigned r = __tls_rand();
                                        i = start + (r % READYQ_SHARD_FACTOR);
                                        /* paranoid */ verify( i < lanes.count );
                                        // If we can't lock it retry
                                } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
                        } else {
                                do {
                                        i = __tls_rand() % lanes.count;
                                } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
                        }
                } else {
                        do {
                                unsigned r = proc->rdq.its++;
                                i = proc->rdq.id + (r % READYQ_SHARD_FACTOR);
                                /* paranoid */ verify( i < lanes.count );
                                // If we can't lock it retry
                        } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
                }

                // Actually push it
                push(lanes.data[i], thrd);

                // Unlock and return
                __atomic_unlock( &lanes.data[i].lock );

                #if !defined(__CFA_NO_STATISTICS__)
                        if(unlikely(external || remote)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
                        else __tls_stats()->ready.push.local.success++;
                #endif
        }

        static inline unsigned long long calc_cutoff(const unsigned long long ctsc, const processor * proc, __ready_queue_t & rdq) {
                unsigned start = proc->rdq.id;
                unsigned long long max = 0;
                for(i; READYQ_SHARD_FACTOR) {
                        unsigned long long ptsc = ts(rdq.lanes.data[start + i]);
                        if(ptsc != -1ull) {
                                /* paranoid */ verify( start + i < rdq.lanes.count );
                                unsigned long long tsc = moving_average(ctsc, ptsc, rdq.lanes.tscs[start + i].ma);
                                if(tsc > max) max = tsc;
                        }
                }
                return (max + 2 * max) / 2;
        }

        __attribute__((hot)) struct thread$ * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
                /* paranoid */ verify( lanes.count > 0 );
                /* paranoid */ verify( kernelTLS().this_processor );
                /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );

                processor * const proc = kernelTLS().this_processor;
                unsigned this = proc->rdq.id;
                /* paranoid */ verify( this < lanes.count );
                __cfadbg_print_safe(ready_queue, "Kernel : pop from %u\n", this);

                // Figure out the current cpu and make sure it is valid
                const int cpu = __kernel_getcpu();
                /* paranoid */ verify(cpu >= 0);
                /* paranoid */ verify(cpu < cpu_info.hthrd_count);
                unsigned this_cache = cpu_info.llc_map[cpu].cache;

                // Super important: don't write the same value over and over again
                // We want to maximise our chances that his particular values stays in cache
                if(lanes.caches[this / READYQ_SHARD_FACTOR].id != this_cache)
                        __atomic_store_n(&lanes.caches[this / READYQ_SHARD_FACTOR].id, this_cache, __ATOMIC_RELAXED);

                const unsigned long long ctsc = rdtscl();

                if(proc->rdq.target == MAX) {
                        uint64_t chaos = __tls_rand();
                        unsigned ext = chaos & 0xff;
                        unsigned other  = (chaos >> 8) % (lanes.count);

                        if(ext < 3 || __atomic_load_n(&lanes.caches[other / READYQ_SHARD_FACTOR].id, __ATOMIC_RELAXED) == this_cache) {
                                proc->rdq.target = other;
                        }
                }
                else {
                        const unsigned target = proc->rdq.target;
                        __cfadbg_print_safe(ready_queue, "Kernel : %u considering helping %u, tcsc %llu\n", this, target, lanes.tscs[target].tv);
                        /* paranoid */ verify( lanes.tscs[target].tv != MAX );
                        if(target < lanes.count) {
                                const unsigned long long cutoff = calc_cutoff(ctsc, proc, cltr->ready_queue);
                                const unsigned long long age = moving_average(ctsc, lanes.tscs[target].tv, lanes.tscs[target].ma);
                                __cfadbg_print_safe(ready_queue, "Kernel : Help attempt on %u from %u, age %'llu vs cutoff %'llu, %s\n", target, this, age, cutoff, age > cutoff ? "yes" : "no");
                                if(age > cutoff) {
                                        thread$ * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
                                        if(t) return t;
                                }
                        }
                        proc->rdq.target = MAX;
                }

                for(READYQ_SHARD_FACTOR) {
                        unsigned i = this + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
                        if(thread$ * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
                }

                // All lanes where empty return 0p
                return 0p;

        }
        __attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
                unsigned i = __tls_rand() % lanes.count;
                return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
        }
        __attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr) {
                return search(cltr);
        }
#endif
#if defined(USE_CPU_WORK_STEALING)
        __attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint) with (cltr->ready_queue) {
                __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);

                processor * const proc = kernelTLS().this_processor;
                const bool external = (!proc) || (cltr != proc->cltr);

                // Figure out the current cpu and make sure it is valid
                const int cpu = __kernel_getcpu();
                /* paranoid */ verify(cpu >= 0);
                /* paranoid */ verify(cpu < cpu_info.hthrd_count);
                /* paranoid */ verify(cpu * READYQ_SHARD_FACTOR < lanes.count);

                // Figure out where thread was last time and make sure it's
                /* paranoid */ verify(thrd->preferred >= 0);
                /* paranoid */ verify(thrd->preferred < cpu_info.hthrd_count);
                /* paranoid */ verify(thrd->preferred * READYQ_SHARD_FACTOR < lanes.count);
                const int prf = thrd->preferred * READYQ_SHARD_FACTOR;

                const cpu_map_entry_t & map;
                choose(hint) {
                        case UNPARK_LOCAL : &map = &cpu_info.llc_map[cpu];
                        case UNPARK_REMOTE: &map = &cpu_info.llc_map[prf];
                }
                /* paranoid */ verify(map.start * READYQ_SHARD_FACTOR < lanes.count);
                /* paranoid */ verify(map.self * READYQ_SHARD_FACTOR < lanes.count);
                /* paranoid */ verifyf((map.start + map.count) * READYQ_SHARD_FACTOR <= lanes.count, "have %zu lanes but map can go up to %u", lanes.count, (map.start + map.count) * READYQ_SHARD_FACTOR);

                const int start = map.self * READYQ_SHARD_FACTOR;
                unsigned i;
                do {
                        unsigned r;
                        if(unlikely(external)) { r = __tls_rand(); }
                        else { r = proc->rdq.its++; }
                        choose(hint) {
                                case UNPARK_LOCAL : i = start + (r % READYQ_SHARD_FACTOR);
                                case UNPARK_REMOTE: i = prf   + (r % READYQ_SHARD_FACTOR);
                        }
                        // If we can't lock it retry
                } while( !__atomic_try_acquire( &lanes.data[i].lock ) );

                // Actually push it
                push(lanes.data[i], thrd);

                // Unlock and return
                __atomic_unlock( &lanes.data[i].lock );

                #if !defined(__CFA_NO_STATISTICS__)
                        if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
                        else __tls_stats()->ready.push.local.success++;
                #endif

                __cfadbg_print_safe(ready_queue, "Kernel : Pushed %p on cluster %p (idx: %u, mask %llu, first %d)\n", thrd, cltr, i, used.mask[0], lane_first);

        }

        // Pop from the ready queue from a given cluster
        __attribute__((hot)) thread$ * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
                /* paranoid */ verify( lanes.count > 0 );
                /* paranoid */ verify( kernelTLS().this_processor );

                processor * const proc = kernelTLS().this_processor;
                const int cpu = __kernel_getcpu();
                /* paranoid */ verify(cpu >= 0);
                /* paranoid */ verify(cpu < cpu_info.hthrd_count);
                /* paranoid */ verify(cpu * READYQ_SHARD_FACTOR < lanes.count);

                const cpu_map_entry_t & map = cpu_info.llc_map[cpu];
                /* paranoid */ verify(map.start * READYQ_SHARD_FACTOR < lanes.count);
                /* paranoid */ verify(map.self * READYQ_SHARD_FACTOR < lanes.count);
                /* paranoid */ verifyf((map.start + map.count) * READYQ_SHARD_FACTOR <= lanes.count, "have %zu lanes but map can go up to %u", lanes.count, (map.start + map.count) * READYQ_SHARD_FACTOR);

                const int start = map.self * READYQ_SHARD_FACTOR;
                const unsigned long long ctsc = rdtscl();

                // Did we already have a help target
                if(proc->rdq.target == MAX) {
                        unsigned long long max = 0;
                        for(i; READYQ_SHARD_FACTOR) {
                                unsigned long long tsc = moving_average(ctsc, ts(lanes.data[start + i]), lanes.tscs[start + i].ma);
                                if(tsc > max) max = tsc;
                        }
                        //  proc->rdq.cutoff = (max + 2 * max) / 2;
                        /* paranoid */ verify(lanes.count < 65536); // The following code assumes max 65536 cores.
                        /* paranoid */ verify(map.count < 65536); // The following code assumes max 65536 cores.

                        if(0 == (__tls_rand() % 100)) {
                                proc->rdq.target = __tls_rand() % lanes.count;
                        } else {
                                unsigned cpu_chaos = map.start + (__tls_rand() % map.count);
                                proc->rdq.target = (cpu_chaos * READYQ_SHARD_FACTOR) + (__tls_rand() % READYQ_SHARD_FACTOR);
                                /* paranoid */ verify(proc->rdq.target >= (map.start * READYQ_SHARD_FACTOR));
                                /* paranoid */ verify(proc->rdq.target <  ((map.start + map.count) * READYQ_SHARD_FACTOR));
                        }

                        /* paranoid */ verify(proc->rdq.target != MAX);
                }
                else {
                        unsigned long long max = 0;
                        for(i; READYQ_SHARD_FACTOR) {
                                unsigned long long tsc = moving_average(ctsc, ts(lanes.data[start + i]), lanes.tscs[start + i].ma);
                                if(tsc > max) max = tsc;
                        }
                        const unsigned long long cutoff = (max + 2 * max) / 2;
                        {
                                unsigned target = proc->rdq.target;
                                proc->rdq.target = MAX;
                                lanes.help[target / READYQ_SHARD_FACTOR].tri++;
                                if(moving_average(ctsc, lanes.tscs[target].tv, lanes.tscs[target].ma) > cutoff) {
                                        thread$ * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
                                        proc->rdq.last = target;
                                        if(t) return t;
                                }
                                proc->rdq.target = MAX;
                        }

                        unsigned last = proc->rdq.last;
                        if(last != MAX && moving_average(ctsc, lanes.tscs[last].tv, lanes.tscs[last].ma) > cutoff) {
                                thread$ * t = try_pop(cltr, last __STATS(, __tls_stats()->ready.pop.help));
                                if(t) return t;
                        }
                        else {
                                proc->rdq.last = MAX;
                        }
                }

                for(READYQ_SHARD_FACTOR) {
                        unsigned i = start + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
                        if(thread$ * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
                }

                // All lanes where empty return 0p
                return 0p;
        }

        __attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
                processor * const proc = kernelTLS().this_processor;
                unsigned last = proc->rdq.last;
                if(last != MAX) {
                        struct thread$ * t = try_pop(cltr, last __STATS(, __tls_stats()->ready.pop.steal));
                        if(t) return t;
                        proc->rdq.last = MAX;
                }

                unsigned i = __tls_rand() % lanes.count;
                return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
        }
        __attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr) {
                return search(cltr);
        }
#endif
#if defined(USE_RELAXED_FIFO)
        //-----------------------------------------------------------------------
        // get index from random number with or without bias towards queues
        static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred) {
                unsigned i;
                bool local;
                unsigned rlow  = r % BIAS;
                unsigned rhigh = r / BIAS;
                if((0 != rlow) && preferred >= 0) {
                        // (BIAS - 1) out of BIAS chances
                        // Use perferred queues
                        i = preferred + (rhigh % READYQ_SHARD_FACTOR);
                        local = true;
                }
                else {
                        // 1 out of BIAS chances
                        // Use all queues
                        i = rhigh;
                        local = false;
                }
                return [i, local];
        }

        __attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint) with (cltr->ready_queue) {
                __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);

                const bool external = (hint != UNPARK_LOCAL) || (!kernelTLS().this_processor) || (cltr != kernelTLS().this_processor->cltr);
                /* paranoid */ verify(external || kernelTLS().this_processor->rdq.id < lanes.count );

                bool local;
                int preferred = external ? -1 : kernelTLS().this_processor->rdq.id;

                // Try to pick a lane and lock it
                unsigned i;
                do {
                        // Pick the index of a lane
                        unsigned r = __tls_rand_fwd();
                        [i, local] = idx_from_r(r, preferred);

                        i %= __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );

                        #if !defined(__CFA_NO_STATISTICS__)
                                if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
                                else if(local) __tls_stats()->ready.push.local.attempt++;
                                else __tls_stats()->ready.push.share.attempt++;
                        #endif

                        // If we can't lock it retry
                } while( !__atomic_try_acquire( &lanes.data[i].lock ) );

                // Actually push it
                push(lanes.data[i], thrd);

                // Unlock and return
                __atomic_unlock( &lanes.data[i].lock );

                // Mark the current index in the tls rng instance as having an item
                __tls_rand_advance_bck();

                __cfadbg_print_safe(ready_queue, "Kernel : Pushed %p on cluster %p (idx: %u, mask %llu, first %d)\n", thrd, cltr, i, used.mask[0], lane_first);

                // Update statistics
                #if !defined(__CFA_NO_STATISTICS__)
                        if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
                        else if(local) __tls_stats()->ready.push.local.success++;
                        else __tls_stats()->ready.push.share.success++;
                #endif
        }

        // Pop from the ready queue from a given cluster
        __attribute__((hot)) thread$ * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
                /* paranoid */ verify( lanes.count > 0 );
                /* paranoid */ verify( kernelTLS().this_processor );
                /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );

                unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
                int preferred = kernelTLS().this_processor->rdq.id;


                // As long as the list is not empty, try finding a lane that isn't empty and pop from it
                for(25) {
                        // Pick two lists at random
                        unsigned ri = __tls_rand_bck();
                        unsigned rj = __tls_rand_bck();

                        unsigned i, j;
                        __attribute__((unused)) bool locali, localj;
                        [i, locali] = idx_from_r(ri, preferred);
                        [j, localj] = idx_from_r(rj, preferred);

                        i %= count;
                        j %= count;

                        // try popping from the 2 picked lists
                        struct thread$ * thrd = try_pop(cltr, i, j __STATS(, *(locali || localj ? &__tls_stats()->ready.pop.local : &__tls_stats()->ready.pop.help)));
                        if(thrd) {
                                return thrd;
                        }
                }

                // All lanes where empty return 0p
                return 0p;
        }

        __attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr) { return pop_fast(cltr); }
        __attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr) {
                return search(cltr);
        }
#endif
#if defined(USE_WORK_STEALING)
        __attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint) with (cltr->ready_queue) {
                __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);

                // #define USE_PREFERRED
                #if !defined(USE_PREFERRED)
                const bool external = (hint != UNPARK_LOCAL) || (!kernelTLS().this_processor) || (cltr != kernelTLS().this_processor->cltr);
                /* paranoid */ verify(external || kernelTLS().this_processor->rdq.id < lanes.count );
                #else
                        unsigned preferred = thrd->preferred;
                        const bool external = (hint != UNPARK_LOCAL) || (!kernelTLS().this_processor) || preferred == MAX || thrd->curr_cluster != cltr;
                        /* paranoid */ verifyf(external || preferred < lanes.count, "Invalid preferred queue %u for %u lanes", preferred, lanes.count );

                        unsigned r = preferred % READYQ_SHARD_FACTOR;
                        const unsigned start = preferred - r;
                #endif

                // Try to pick a lane and lock it
                unsigned i;
                do {
                        #if !defined(__CFA_NO_STATISTICS__)
                                if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
                                else __tls_stats()->ready.push.local.attempt++;
                        #endif

                        if(unlikely(external)) {
                                i = __tls_rand() % lanes.count;
                        }
                        else {
                                #if !defined(USE_PREFERRED)
                                        processor * proc = kernelTLS().this_processor;
                                        unsigned r = proc->rdq.its++;
                                        i =  proc->rdq.id + (r % READYQ_SHARD_FACTOR);
                                #else
                                        i = start + (r++ % READYQ_SHARD_FACTOR);
                                #endif
                        }
                        // If we can't lock it retry
                } while( !__atomic_try_acquire( &lanes.data[i].lock ) );

                // Actually push it
                push(lanes.data[i], thrd);

                // Unlock and return
                __atomic_unlock( &lanes.data[i].lock );

                #if !defined(__CFA_NO_STATISTICS__)
                        if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
                        else __tls_stats()->ready.push.local.success++;
                #endif

                __cfadbg_print_safe(ready_queue, "Kernel : Pushed %p on cluster %p (idx: %u, mask %llu, first %d)\n", thrd, cltr, i, used.mask[0], lane_first);
        }

        // Pop from the ready queue from a given cluster
        __attribute__((hot)) thread$ * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
                /* paranoid */ verify( lanes.count > 0 );
                /* paranoid */ verify( kernelTLS().this_processor );
                /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );

                processor * proc = kernelTLS().this_processor;

                if(proc->rdq.target == MAX) {
                        unsigned long long min = ts(lanes.data[proc->rdq.id]);
                        for(int i = 0; i < READYQ_SHARD_FACTOR; i++) {
                                unsigned long long tsc = ts(lanes.data[proc->rdq.id + i]);
                                if(tsc < min) min = tsc;
                        }
                        proc->rdq.cutoff = min;
                        proc->rdq.target = __tls_rand() % lanes.count;
                }
                else {
                        unsigned target = proc->rdq.target;
                        proc->rdq.target = MAX;
                        const unsigned long long bias = 0; //2_500_000_000;
                        const unsigned long long cutoff = proc->rdq.cutoff > bias ? proc->rdq.cutoff - bias : proc->rdq.cutoff;
                        if(lanes.tscs[target].tv < cutoff && ts(lanes.data[target]) < cutoff) {
                                thread$ * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
                                if(t) return t;
                        }
                }

                for(READYQ_SHARD_FACTOR) {
                        unsigned i = proc->rdq.id + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
                        if(thread$ * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
                }
                return 0p;
        }

        __attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
                unsigned i = __tls_rand() % lanes.count;
                return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
        }

        __attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr) with (cltr->ready_queue) {
                return search(cltr);
        }
#endif

//=======================================================================
// Various Ready Queue utilities
//=======================================================================
// these function work the same or almost the same
// whether they are using work-stealing or relaxed fifo scheduling

//-----------------------------------------------------------------------
// try to pop from a lane given by index w
static inline struct thread$ * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
        /* paranoid */ verify( w < lanes.count );
        __STATS( stats.attempt++; )

        // Get relevant elements locally
        __intrusive_lane_t & lane = lanes.data[w];

        // If list looks empty retry
        if( is_empty(lane) ) {
                return 0p;
        }

        // If we can't get the lock retry
        if( !__atomic_try_acquire(&lane.lock) ) {
                return 0p;
        }

        // If list is empty, unlock and retry
        if( is_empty(lane) ) {
                __atomic_unlock(&lane.lock);
                return 0p;
        }

        // Actually pop the list
        struct thread$ * thrd;
        #if defined(USE_AWARE_STEALING) || defined(USE_WORK_STEALING) || defined(USE_CPU_WORK_STEALING)
                unsigned long long tsc_before = ts(lane);
        #endif
        unsigned long long tsv;
        [thrd, tsv] = pop(lane);

        /* paranoid */ verify(thrd);
        /* paranoid */ verify(tsv);
        /* paranoid */ verify(lane.lock);

        // Unlock and return
        __atomic_unlock(&lane.lock);

        // Update statistics
        __STATS( stats.success++; )

        #if defined(USE_AWARE_STEALING) || defined(USE_WORK_STEALING) || defined(USE_CPU_WORK_STEALING)
                if (tsv != MAX) {
                        unsigned long long now = rdtscl();
                        unsigned long long pma = __atomic_load_n(&lanes.tscs[w].ma, __ATOMIC_RELAXED);
                        __atomic_store_n(&lanes.tscs[w].tv, tsv, __ATOMIC_RELAXED);
                        __atomic_store_n(&lanes.tscs[w].ma, moving_average(now, tsc_before, pma), __ATOMIC_RELAXED);
                }
        #endif

        #if defined(USE_AWARE_STEALING) || defined(USE_CPU_WORK_STEALING)
                thrd->preferred = w / READYQ_SHARD_FACTOR;
        #else
                thrd->preferred = w;
        #endif

        // return the popped thread
        return thrd;
}

//-----------------------------------------------------------------------
// try to pop from any lanes making sure you don't miss any threads push
// before the start of the function
static inline struct thread$ * search(struct cluster * cltr) with (cltr->ready_queue) {
        /* paranoid */ verify( lanes.count > 0 );
        unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
        unsigned offset = __tls_rand();
        for(i; count) {
                unsigned idx = (offset + i) % count;
                struct thread$ * thrd = try_pop(cltr, idx __STATS(, __tls_stats()->ready.pop.search));
                if(thrd) {
                        return thrd;
                }
        }

        // All lanes where empty return 0p
        return 0p;
}

//-----------------------------------------------------------------------
// get preferred ready for new thread
unsigned ready_queue_new_preferred() {
        unsigned pref = 0;
        if(struct thread$ * thrd = publicTLS_get( this_thread )) {
                pref = thrd->preferred;
        }
        else {
                #if defined(USE_CPU_WORK_STEALING)
                        pref = __kernel_getcpu();
                #endif
        }

        #if defined(USE_CPU_WORK_STEALING)
                /* paranoid */ verify(pref >= 0);
                /* paranoid */ verify(pref < cpu_info.hthrd_count);
        #endif

        return pref;
}

//-----------------------------------------------------------------------
// Check that all the intrusive queues in the data structure are still consistent
static void check( __ready_queue_t & q ) with (q) {
        #if defined(__CFA_WITH_VERIFY__)
                {
                        for( idx ; lanes.count ) {
                                __intrusive_lane_t & sl = lanes.data[idx];
                                assert(!lanes.data[idx].lock);

                                        if(is_empty(sl)) {
                                                assert( sl.anchor.next == 0p );
                                                assert( sl.anchor.ts   == -1llu );
                                                assert( mock_head(sl)  == sl.prev );
                                        } else {
                                                assert( sl.anchor.next != 0p );
                                                assert( sl.anchor.ts   != -1llu );
                                                assert( mock_head(sl)  != sl.prev );
                                        }
                        }
                }
        #endif
}

//-----------------------------------------------------------------------
// Given 2 indexes, pick the list with the oldest push an try to pop from it
static inline struct thread$ * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
        // Pick the bet list
        int w = i;
        if( __builtin_expect(!is_empty(lanes.data[j]), true) ) {
                w = (ts(lanes.data[i]) < ts(lanes.data[j])) ? i : j;
        }

        return try_pop(cltr, w __STATS(, stats));
}

// Call this function of the intrusive list was moved using memcpy
// fixes the list so that the pointers back to anchors aren't left dangling
static inline void fix(__intrusive_lane_t & ll) {
                        if(is_empty(ll)) {
                                verify(ll.anchor.next == 0p);
                                ll.prev = mock_head(ll);
                        }
}

static void assign_list(unsigned & value, dlist(processor) & list, unsigned count) {
        processor * it = &list`first;
        for(unsigned i = 0; i < count; i++) {
                /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
                it->rdq.id = value;
                it->rdq.target = MAX;
                value += READYQ_SHARD_FACTOR;
                it = &(*it)`next;
        }
}

static void reassign_cltr_id(struct cluster * cltr) {
        unsigned preferred = 0;
        assign_list(preferred, cltr->procs.actives, cltr->procs.total - cltr->procs.idle);
        assign_list(preferred, cltr->procs.idles  , cltr->procs.idle );
}

static void fix_times( struct cluster * cltr ) with( cltr->ready_queue ) {
        #if defined(USE_AWARE_STEALING) || defined(USE_WORK_STEALING)
                lanes.tscs = alloc(lanes.count, lanes.tscs`realloc);
                for(i; lanes.count) {
                        lanes.tscs[i].tv = rdtscl();
                        lanes.tscs[i].ma = 0;
                }
        #endif
}

#if defined(USE_CPU_WORK_STEALING)
        // ready_queue size is fixed in this case
        void ready_queue_grow(struct cluster * cltr) {}
        void ready_queue_shrink(struct cluster * cltr) {}
#else
        // Grow the ready queue
        void ready_queue_grow(struct cluster * cltr) {
                size_t ncount;
                int target = cltr->procs.total;

                /* paranoid */ verify( ready_mutate_islocked() );
                __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue\n");

                // Make sure that everything is consistent
                /* paranoid */ check( cltr->ready_queue );

                // grow the ready queue
                with( cltr->ready_queue ) {
                        // Find new count
                        // Make sure we always have atleast 1 list
                        if(target >= 2) {
                                ncount = target * READYQ_SHARD_FACTOR;
                        } else {
                                ncount = SEQUENTIAL_SHARD;
                        }

                        // Allocate new array (uses realloc and memcpies the data)
                        lanes.data = alloc( ncount, lanes.data`realloc );

                        // Fix the moved data
                        for( idx; (size_t)lanes.count ) {
                                fix(lanes.data[idx]);
                        }

                        // Construct new data
                        for( idx; (size_t)lanes.count ~ ncount) {
                                (lanes.data[idx]){};
                        }

                        // Update original
                        lanes.count = ncount;

                        lanes.caches = alloc( target, lanes.caches`realloc );
                }

                fix_times(cltr);

                reassign_cltr_id(cltr);

                // Make sure that everything is consistent
                /* paranoid */ check( cltr->ready_queue );

                __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue done\n");

                /* paranoid */ verify( ready_mutate_islocked() );
        }

        // Shrink the ready queue
        void ready_queue_shrink(struct cluster * cltr) {
                /* paranoid */ verify( ready_mutate_islocked() );
                __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue\n");

                // Make sure that everything is consistent
                /* paranoid */ check( cltr->ready_queue );

                int target = cltr->procs.total;

                with( cltr->ready_queue ) {
                        // Remember old count
                        size_t ocount = lanes.count;

                        // Find new count
                        // Make sure we always have atleast 1 list
                        lanes.count = target >= 2 ? target * READYQ_SHARD_FACTOR: SEQUENTIAL_SHARD;
                        /* paranoid */ verify( ocount >= lanes.count );
                        /* paranoid */ verify( lanes.count == target * READYQ_SHARD_FACTOR || target < 2 );

                        // for printing count the number of displaced threads
                        #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
                                __attribute__((unused)) size_t displaced = 0;
                        #endif

                        // redistribute old data
                        for( idx; (size_t)lanes.count ~ ocount) {
                                // Lock is not strictly needed but makes checking invariants much easier
                                __attribute__((unused)) bool locked = __atomic_try_acquire(&lanes.data[idx].lock);
                                verify(locked);

                                // As long as we can pop from this lane to push the threads somewhere else in the queue
                                while(!is_empty(lanes.data[idx])) {
                                        struct thread$ * thrd;
                                        unsigned long long _;
                                        [thrd, _] = pop(lanes.data[idx]);

                                        push(cltr, thrd, true);

                                        // for printing count the number of displaced threads
                                        #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
                                                displaced++;
                                        #endif
                                }

                                // Unlock the lane
                                __atomic_unlock(&lanes.data[idx].lock);

                                // TODO print the queue statistics here

                                ^(lanes.data[idx]){};
                        }

                        __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue displaced %zu threads\n", displaced);

                        // Allocate new array (uses realloc and memcpies the data)
                        lanes.data = alloc( lanes.count, lanes.data`realloc );

                        // Fix the moved data
                        for( idx; (size_t)lanes.count ) {
                                fix(lanes.data[idx]);
                        }

                        lanes.caches = alloc( target, lanes.caches`realloc );
                }

                fix_times(cltr);


                reassign_cltr_id(cltr);

                // Make sure that everything is consistent
                /* paranoid */ check( cltr->ready_queue );

                __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue done\n");
                /* paranoid */ verify( ready_mutate_islocked() );
        }
#endif

#if !defined(__CFA_NO_STATISTICS__)
        unsigned cnt(const __ready_queue_t & this, unsigned idx) {
                /* paranoid */ verify(this.lanes.count > idx);
                return this.lanes.data[idx].cnt;
        }
#endif


#if   defined(CFA_HAVE_LINUX_LIBRSEQ)
        // No definition needed
#elif defined(CFA_HAVE_LINUX_RSEQ_H)

        #if defined( __x86_64 ) || defined( __i386 )
                #define RSEQ_SIG        0x53053053
        #elif defined( __ARM_ARCH )
                #ifdef __ARMEB__
                #define RSEQ_SIG    0xf3def5e7      /* udf    #24035    ; 0x5de3 (ARMv6+) */
                #else
                #define RSEQ_SIG    0xe7f5def3      /* udf    #24035    ; 0x5de3 */
                #endif
        #endif

        extern void __disable_interrupts_hard();
        extern void __enable_interrupts_hard();

        static void __kernel_raw_rseq_register  (void) {
                /* paranoid */ verify( __cfaabi_rseq.cpu_id == RSEQ_CPU_ID_UNINITIALIZED );

                // int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), 0, (sigset_t *)0p, _NSIG / 8);
                int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), 0, RSEQ_SIG);
                if(ret != 0) {
                        int e = errno;
                        switch(e) {
                        case EINVAL: abort("KERNEL ERROR: rseq register invalid argument");
                        case ENOSYS: abort("KERNEL ERROR: rseq register no supported");
                        case EFAULT: abort("KERNEL ERROR: rseq register with invalid argument");
                        case EBUSY : abort("KERNEL ERROR: rseq register already registered");
                        case EPERM : abort("KERNEL ERROR: rseq register sig  argument  on unregistration does not match the signature received on registration");
                        default: abort("KERNEL ERROR: rseq register unexpected return %d", e);
                        }
                }
        }

        static void __kernel_raw_rseq_unregister(void) {
                /* paranoid */ verify( __cfaabi_rseq.cpu_id >= 0 );

                // int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), RSEQ_FLAG_UNREGISTER, (sigset_t *)0p, _NSIG / 8);
                int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), RSEQ_FLAG_UNREGISTER, RSEQ_SIG);
                if(ret != 0) {
                        int e = errno;
                        switch(e) {
                        case EINVAL: abort("KERNEL ERROR: rseq unregister invalid argument");
                        case ENOSYS: abort("KERNEL ERROR: rseq unregister no supported");
                        case EFAULT: abort("KERNEL ERROR: rseq unregister with invalid argument");
                        case EBUSY : abort("KERNEL ERROR: rseq unregister already registered");
                        case EPERM : abort("KERNEL ERROR: rseq unregister sig  argument  on unregistration does not match the signature received on registration");
                        default: abort("KERNEL ERROR: rseq unregisteunexpected return %d", e);
                        }
                }
        }
#else
        // No definition needed
#endif