source: libcfa/src/concurrency/kernel.hfa @ f90c092

//
// Cforall Version 1.0.0 Copyright (C) 2016 University of Waterloo
//
// The contents of this file are covered under the licence agreement in the
// file "LICENCE" distributed with Cforall.
//
// kernel -- Header containing the core of the kernel API
//
// Author           : Thierry Delisle
// Created On       : Tue Jan 17 12:27:26 2017
// Last Modified By : Peter A. Buhr
// Last Modified On : Wed Aug 30 21:28:46 2023
// Update Count     : 23
//

#pragma once

#include "invoke.h"
#include "time_t.hfa"
#include "coroutine.hfa"

#include "collections/list.hfa"

extern "C" {
        #include <bits/pthreadtypes.h>
        #include <pthread.h>
        #include <linux/types.h>
}

#ifdef __CFA_WITH_VERIFY__
        extern bool __cfaabi_dbg_in_kernel();
#endif

//-----------------------------------------------------------------------------
// I/O
struct cluster;
struct io_context$;
struct io_arbiter$;

struct io_context_params {
        int num_entries;
};

void  ?{}(io_context_params & this);

//-----------------------------------------------------------------------------
// Processor
extern struct cluster * mainCluster;

// Coroutine used py processors for the 2-step context switch

struct processorCtx_t {
        struct coroutine$ self;
        struct processor * proc;
};

struct io_future_t;

// Information needed for idle sleep
struct __fd_waitctx {
        // semaphore/future like object
        // values can be 0, 1 or some file descriptor.
        // 0 - is the default state
        // 1 - means the proc should wake-up immediately
        // FD - means the proc is going asleep and should be woken by writing to the FD.
        //      The FD value should always be the evfd field just below.
        volatile int sem;

        // The event FD that corresponds to this processor
        int evfd;

        // Used for debugging, should be removed eventually.
        volatile unsigned long long wake__time;
        volatile unsigned long long sleep_time;
        volatile unsigned long long drain_time;
};

// Wrapper around kernel threads
struct __attribute__((aligned(64))) processor {
        // Cluster from which to get threads
        struct cluster * cltr;

        // Ready Queue state per processor
        struct {
                unsigned short its;
                unsigned short itr;
                unsigned id;
                unsigned target;
                unsigned last;
                signed   cpu;
        } rdq;

        // Set to true to notify the processor should terminate
        volatile bool do_terminate;

        // Coroutine ctx who does keeps the state of the processor
        struct processorCtx_t runner;

        // Name of the processor
        const char * name;

        // Handle to pthreads
        pthread_t kernel_thread;

        // Unique id for the processor (not per cluster)
        unsigned unique_id;

        struct {
                io_context$ * ctx;
                unsigned target;
                volatile bool pending;
                volatile bool dirty;
        } io;

        // Preemption data
        // Node which is added in the discrete event simulaiton
        struct alarm_node_t * preemption_alarm;

        // If true, a preemption was triggered in an unsafe region, the processor must preempt as soon as possible
        bool pending_preemption;

        // context for idle sleep
        struct __fd_waitctx idle_wctx;

        // Termination synchronisation (user semaphore)
        oneshot terminated;

        // pthread Stack
        void * stack;

        // Link lists fields
        dlink(processor) link;

        // special init fields
        // This is needed for memcached integration
        // once memcached experiments are done this should probably be removed
        // it is not a particularly safe scheme as it can make processors less homogeneous
        struct {
                thread$ * thrd;
        } init;

        struct KernelThreadData * local_data;

        #if !defined(__CFA_NO_STATISTICS__)
                int print_stats;
                bool print_halts;
        #endif

#ifdef __CFA_DEBUG__
        // Last function to enable preemption on this processor
        const char * last_enable;
#endif
};
// P9_EMBEDDED( processor, dlink(processor) )
static inline tytagref( dlink(processor), dlink(processor) ) ?`inner( processor & this ) {
        dlink(processor) & b = this.link;
        tytagref( dlink(processor), dlink(processor) ) result = { b };
        return result;
}

void  ?{}(processor & this, const char name[], struct cluster & cltr);
void ^?{}(processor & this);

static inline void  ?{}(processor & this)                        { this{ "Anonymous Processor", *mainCluster}; }
static inline void  ?{}(processor & this, struct cluster & cltr) { this{ "Anonymous Processor", cltr}; }
static inline void  ?{}(processor & this, const char name[])     { this{name, *mainCluster}; }

//-----------------------------------------------------------------------------
// Cluster Tools

// Intrusives lanes which are used by the ready queue
union __attribute__((aligned(64))) __intrusive_lane_t;
void  ?{}(__intrusive_lane_t & this);
void ^?{}(__intrusive_lane_t & this);

// Aligned timestamps which are used by the ready queue and io subsystem
union __attribute__((aligned(64))) __timestamp_t;

void  ?{}(__timestamp_t & this);
void ^?{}(__timestamp_t &);


struct __attribute__((aligned(16))) __cache_id_t {
        volatile unsigned id;
};

// Idle Sleep
struct __cluster_proc_list {
        // Spin lock protecting the queue
        __spinlock_t lock;

        // FD to use to wake a processor
        struct __fd_waitctx * volatile fdw;

        // Total number of processors
        unsigned total;

    // Number of processors constructed
    //  incremented at construction time, total is incremented once the processor asyncronously registers
        unsigned constructed;

        // Total number of idle processors
        unsigned idle;

        // List of idle processors
        dlist(processor) idles;

        // List of active processors
        dlist(processor) actives;
};

//-----------------------------------------------------------------------------
// Cluster
struct __attribute__((aligned(64))) cluster {
        struct {
                struct {
                        // Arary of subqueues
                        __intrusive_lane_t * data;

                        // Time since subqueues were processed
                        __timestamp_t * tscs;

                        // Number of subqueue / timestamps
                        size_t count;
                } readyQ;

                struct {
                        // Array of $io_
                        io_context$ ** data;

                        // Time since subqueues were processed
                        __timestamp_t * tscs;

                        // Number of I/O subqueues
                        size_t count;
                } io;

                // Cache each kernel thread belongs to
                __cache_id_t * caches;
        } sched;

        // // Ready queue for threads
        // __ready_queue_t ready_queue;

        // Name of the cluster
        const char * name;

        // Preemption rate on this cluster
        Duration preemption_rate;

        // List of idle processors
        __cluster_proc_list procs;

        // List of threads
        __spinlock_t thread_list_lock;
        dlist(struct thread$, struct __thread_user_link) threads;
        unsigned int nthreads;

        // Link lists fields
        struct __dbg_node_cltr {
                cluster * next;
                cluster * prev;
        } node;

        struct {
                io_arbiter$ * arbiter;
                io_context_params params;
        } io;

        struct {
                struct processor ** procs;
                unsigned cnt;
        } managed;

        #if !defined(__CFA_NO_STATISTICS__)
                struct __stats_t * stats;
                int print_stats;
        #endif
};
extern Duration default_preemption();

void ?{} (cluster & this, const char name[], Duration preemption_rate, unsigned num_io, const io_context_params & io_params);
void ^?{}(cluster & this);

static inline void ?{} (cluster & this)                                            { io_context_params default_params;    this{"Anonymous Cluster", default_preemption(), 1, default_params}; }
static inline void ?{} (cluster & this, Duration preemption_rate)                  { io_context_params default_params;    this{"Anonymous Cluster", preemption_rate, 1, default_params}; }
static inline void ?{} (cluster & this, const char name[])                         { io_context_params default_params;    this{name, default_preemption(), 1, default_params}; }
static inline void ?{} (cluster & this, unsigned num_io)                           { io_context_params default_params;    this{"Anonymous Cluster", default_preemption(), num_io, default_params}; }
static inline void ?{} (cluster & this, Duration preemption_rate, unsigned num_io) { io_context_params default_params;    this{"Anonymous Cluster", preemption_rate, num_io, default_params}; }
static inline void ?{} (cluster & this, const char name[], unsigned num_io)        { io_context_params default_params;    this{name, default_preemption(), num_io, default_params}; }
static inline void ?{} (cluster & this, const io_context_params & io_params)                                            { this{"Anonymous Cluster", default_preemption(), 1, io_params}; }
static inline void ?{} (cluster & this, Duration preemption_rate, const io_context_params & io_params)                  { this{"Anonymous Cluster", preemption_rate, 1, io_params}; }
static inline void ?{} (cluster & this, const char name[], const io_context_params & io_params)                         { this{name, default_preemption(), 1, io_params}; }
static inline void ?{} (cluster & this, unsigned num_io, const io_context_params & io_params)                           { this{"Anonymous Cluster", default_preemption(), num_io, io_params}; }
static inline void ?{} (cluster & this, Duration preemption_rate, unsigned num_io, const io_context_params & io_params) { this{"Anonymous Cluster", preemption_rate, num_io, io_params}; }
static inline void ?{} (cluster & this, const char name[], unsigned num_io, const io_context_params & io_params)        { this{name, default_preemption(), num_io, io_params}; }

static inline [cluster *&, cluster *& ] __get( cluster & this ) __attribute__((const)) { return this.node.[next, prev]; }

static inline struct processor * active_processor() { return publicTLS_get( this_processor ); } // UNSAFE
static inline struct cluster   * active_cluster  () { return publicTLS_get( this_processor )->cltr; }

// gets the number of constructed processors on the cluster
static inline unsigned get_proc_count( cluster & this ) { return this.procs.constructed; }
static inline unsigned get_proc_count() { return publicTLS_get( this_processor )->cltr->procs.constructed; }

// set the number of internal processors
// these processors are in addition to any explicitly declared processors
unsigned set_concurrency( cluster & this, unsigned new_count );

#if !defined(__CFA_NO_STATISTICS__)
        void print_stats_now( cluster & this, int flags );

        static inline void print_stats_at_exit( cluster & this, int flags ) {
                this.print_stats |= flags;
        }

        static inline void print_stats_at_exit( processor & this, int flags ) {
                this.print_stats |= flags;
        }

        void print_halts( processor & this );
#endif

// Local Variables: //
// mode: c //
// tab-width: 4 //
// End: //