source: libcfa/src/concurrency/kernel.hfa @ d3ab183

//
// 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 --
//
// Author           : Thierry Delisle
// Created On       : Tue Jan 17 12:27:26 2017
// Last Modified By : Peter A. Buhr
// Last Modified On : Tue Feb  4 12:29:26 2020
// Update Count     : 22
//

#pragma once

#include <stdbool.h>
#include <stdint.h>

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

extern "C" {
#include <pthread.h>
#include <semaphore.h>
}

//-----------------------------------------------------------------------------
// Locks
struct semaphore {
        __spinlock_t lock;
        int count;
        __queue_t($thread) waiting;
};

void  ?{}(semaphore & this, int count = 1);
void ^?{}(semaphore & this);
void   P (semaphore & this);
bool   V (semaphore & this);
bool   V (semaphore & this, unsigned count);


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

// Processor
coroutine processorCtx_t {
        struct processor * proc;
};

// Wrapper around kernel threads
struct processor {
        // Main state
        // Coroutine ctx who does keeps the state of the processor
        struct processorCtx_t runner;

        // Cluster from which to get threads
        struct cluster * cltr;

        // Name of the processor
        const char * name;

        // Handle to pthreads
        pthread_t kernel_thread;

        // RunThread data
        // Action to do after a thread is ran
        $thread * destroyer;

        // 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;

        // Idle lock (kernel semaphore)
        __bin_sem_t idle;

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

        // Termination synchronisation (user semaphore)
        semaphore terminated;

        // pthread Stack
        void * stack;

        // Link lists fields
        struct __dbg_node_proc {
                struct processor * next;
                struct processor * prev;
        } node;

#ifdef __CFA_DEBUG__
        // Last function to enable preemption on this processor
        const char * last_enable;
#endif
};

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 }; }

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

//-----------------------------------------------------------------------------
// I/O
#if defined(HAVE_LINUX_IO_URING_H)
struct io_uring_sq {
        // Head and tail of the ring (associated with array)
        volatile uint32_t * head;
        volatile uint32_t * tail;

        // The actual kernel ring which uses head/tail
        // indexes into the sqes arrays
        uint32_t * array;

        // number of entries and mask to go with it
        const uint32_t * num;
        const uint32_t * mask;

        // Submission flags (Not sure what for)
        uint32_t * flags;

        // number of sqes not submitted (whatever that means)
        uint32_t * dropped;

        // Like head/tail but not seen by the kernel
        volatile uint32_t alloc;
        volatile uint32_t ready;

        __spinlock_t lock;

        // A buffer of sqes (not the actual ring)
        struct io_uring_sqe * sqes;

        // The location and size of the mmaped area
        void * ring_ptr;
        size_t ring_sz;

        // Statistics
        #if !defined(__CFA_NO_STATISTICS__)
                struct {
                        struct {
                                unsigned long long int val;
                                unsigned long long int cnt;
                        } submit_avg;
                } stats;
        #endif
};

struct io_uring_cq {
        // Head and tail of the ring
        volatile uint32_t * head;
        volatile uint32_t * tail;

        // number of entries and mask to go with it
        const uint32_t * mask;
        const uint32_t * num;

        // number of cqes not submitted (whatever that means)
        uint32_t * overflow;

        // the kernel ring
        struct io_uring_cqe * cqes;

        // The location and size of the mmaped area
        void * ring_ptr;
        size_t ring_sz;

        // Statistics
        #if !defined(__CFA_NO_STATISTICS__)
                struct {
                        struct {
                                unsigned long long int val;
                                unsigned long long int cnt;
                        } completed_avg;
                } stats;
        #endif
};

#if defined(__CFA_IO_POLLING_USER__)
        struct __io_poller_fast {
                struct io_ring * ring;
                $thread thrd;
        };
#endif

struct io_ring {
        struct io_uring_sq submit_q;
        struct io_uring_cq completion_q;
        uint32_t flags;
        int fd;
        semaphore submit;
        volatile bool done;
        struct {
                struct {
                        void * stack;
                        pthread_t kthrd;
                } slow;
                #if defined(__CFA_IO_POLLING_USER__)
                        __io_poller_fast fast;
                        __bin_sem_t sem;
                #endif
        } poller;
};
#endif

//-----------------------------------------------------------------------------
// Cluster
struct cluster {
        // Ready queue locks
        __spinlock_t ready_queue_lock;

        // Ready queue for threads
        __queue_t($thread) ready_queue;

        // Name of the cluster
        const char * name;

        // Preemption rate on this cluster
        Duration preemption_rate;

        // List of processors
        __spinlock_t idle_lock;
        __dllist_t(struct processor) procs;
        __dllist_t(struct processor) idles;
        unsigned int nprocessors;

        // List of threads
        __spinlock_t thread_list_lock;
        __dllist_t(struct $thread) threads;
        unsigned int nthreads;

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

        #if defined(HAVE_LINUX_IO_URING_H)
                struct io_ring io;
        #endif

        #if !defined(__CFA_NO_STATISTICS__)
                bool print_stats;
        #endif
};
extern Duration default_preemption();

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

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

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

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

#if !defined(__CFA_NO_STATISTICS__)
        static inline void print_stats_at_exit( cluster & this ) {
                this.print_stats = true;
        }
#endif

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