source: libcfa/src/concurrency/preemption.cfa@ 231b18f

//
// 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.
//
// signal.c --
//
// Author           : Thierry Delisle
// Created On       : Mon Jun 5 14:20:42 2017
// Last Modified By : Peter A. Buhr
// Last Modified On : Fri Nov  6 07:42:13 2020
// Update Count     : 54
//

#define __cforall_thread__

#include "preemption.hfa"
#include <assert.h>

#include <errno.h>
#include <stdio.h>
#include <string.h>
#include <unistd.h>
#include <limits.h>                                                                             // PTHREAD_STACK_MIN

#include "bits/signal.hfa"
#include "kernel_private.hfa"

#if !defined(__CFA_DEFAULT_PREEMPTION__)
#define __CFA_DEFAULT_PREEMPTION__ 10`ms
#endif

Duration default_preemption() __attribute__((weak)) {
        return __CFA_DEFAULT_PREEMPTION__;
}

// FwdDeclarations : timeout handlers
static void preempt( processor   * this );
static void timeout( $thread * this );

// FwdDeclarations : Signal handlers
static void sigHandler_ctxSwitch( __CFA_SIGPARMS__ );
static void sigHandler_alarm    ( __CFA_SIGPARMS__ );
static void sigHandler_segv     ( __CFA_SIGPARMS__ );
static void sigHandler_ill      ( __CFA_SIGPARMS__ );
static void sigHandler_fpe      ( __CFA_SIGPARMS__ );
static void sigHandler_abort    ( __CFA_SIGPARMS__ );

// FwdDeclarations : alarm thread main
static void * alarm_loop( __attribute__((unused)) void * args );

// Machine specific register name
#if   defined( __i386 )
#define CFA_REG_IP gregs[REG_EIP]
#elif defined( __x86_64 )
#define CFA_REG_IP gregs[REG_RIP]
#elif defined( __arm__ )
#define CFA_REG_IP arm_pc
#elif defined( __aarch64__ )
#define CFA_REG_IP pc
#else
#error unsupported hardware architecture
#endif

KERNEL_STORAGE(event_kernel_t, event_kernel);         // private storage for event kernel
event_kernel_t * event_kernel;                        // kernel public handle to even kernel
static pthread_t alarm_thread;                        // pthread handle to alarm thread
static void * alarm_stack;                                                        // pthread stack for alarm thread

static void ?{}(event_kernel_t & this) with( this ) {
        alarms{};
        lock{};
}

enum {
        PREEMPT_NORMAL    = 0,
        PREEMPT_TERMINATE = 1,
};

//=============================================================================================
// Kernel Preemption logic
//=============================================================================================

// Get next expired node
static inline alarm_node_t * get_expired( alarm_list_t * alarms, Time currtime ) {
        if( ! & (*alarms)`first ) return 0p;                                            // If no alarms return null
        if( (*alarms)`first.alarm >= currtime ) return 0p;      // If alarms head not expired return null
        return pop(alarms);                                                                     // Otherwise just pop head
}

// Tick one frame of the Discrete Event Simulation for alarms
static void tick_preemption(void) {
        alarm_node_t * node = 0p;                                                       // Used in the while loop but cannot be declared in the while condition
        alarm_list_t * alarms = &event_kernel->alarms;          // Local copy for ease of reading
        Time currtime = __kernel_get_time();                            // Check current time once so everything "happens at once"

        //Loop throught every thing expired
        while( node = get_expired( alarms, currtime ) ) {
                // __cfaabi_dbg_print_buffer_decl( " KERNEL: preemption tick.\n" );
                Duration period = node->period;
                if( period == 0) {
                        node->set = false;                  // Node is one-shot, just mark it as not pending
                }

                // Check if this is a kernel
                if( node->type == Kernel ) {
                        preempt( node->proc );
                }
                else if( node->type == User ) {
                        timeout( node->thrd );
                }
                else {
                        node->callback(*node);
                }

                // Check if this is a periodic alarm
                if( period > 0 ) {
                        // __cfaabi_dbg_print_buffer_local( " KERNEL: alarm period is %lu.\n", period.tv );
                        node->alarm = currtime + period;    // Alarm is periodic, add currtime to it (used cached current time)
                        insert( alarms, node );             // Reinsert the node for the next time it triggers
                }
        }

        // If there are still alarms pending, reset the timer
        if( & (*alarms)`first ) {
                __cfadbg_print_buffer_decl(preemption, " KERNEL: @%ju(%ju) resetting alarm to %ju.\n", currtime.tv, __kernel_get_time().tv, (alarms->head->alarm - currtime).tv);
                Duration delta = (*alarms)`first.alarm - currtime;
                Duration capped = max(delta, 50`us);
                // itimerval tim  = { caped };
                // __cfaabi_dbg_print_buffer_local( "    Values are %lu, %lu, %lu %lu.\n", delta.tv, caped.tv, tim.it_value.tv_sec, tim.it_value.tv_usec);

                __kernel_set_timer( capped );
        }
}

// Update the preemption of a processor and notify interested parties
void update_preemption( processor * this, Duration duration ) {
        alarm_node_t * alarm = this->preemption_alarm;

        // Alarms need to be enabled
        if ( duration > 0 && ! alarm->set ) {
                alarm->alarm = __kernel_get_time() + duration;
                alarm->period = duration;
                register_self( alarm );
        }
        // Zero duration but alarm is set
        else if ( duration == 0 && alarm->set ) {
                unregister_self( alarm );
                alarm->alarm = 0;
                alarm->period = 0;
        }
        // If alarm is different from previous, change it
        else if ( duration > 0 && alarm->period != duration ) {
                unregister_self( alarm );
                alarm->alarm = __kernel_get_time() + duration;
                alarm->period = duration;
                register_self( alarm );
        }
}

//=============================================================================================
// Kernel Signal Tools
//=============================================================================================

// In a user-level threading system, there are handful of thread-local variables where this problem occurs on the ARM.
// 
// For each kernel thread running user-level threads, there is a flag variable to indicate if interrupts are
// enabled/disabled for that kernel thread. Therefore, this variable is made thread local.
// 
// For example, this code fragment sets the state of the "interrupt" variable in thread-local memory.
// 
// _Thread_local volatile int interrupts;
// int main() {
//     interrupts = 0; // disable interrupts }
// 
// which generates the following code on the ARM
// 
// (gdb) disassemble main
// Dump of assembler code for function main:
//    0x0000000000000610 <+0>:  mrs     x1, tpidr_el0
//    0x0000000000000614 <+4>:  mov     w0, #0x0                        // #0
//    0x0000000000000618 <+8>:  add     x1, x1, #0x0, lsl #12
//    0x000000000000061c <+12>: add     x1, x1, #0x10
//    0x0000000000000620 <+16>: str     wzr, [x1]
//    0x0000000000000624 <+20>: ret
// 
// The mrs moves a pointer from coprocessor register tpidr_el0 into register x1.  Register w0 is set to 0. The two adds
// increase the TLS pointer with the displacement (offset) 0x10, which is the location in the TSL of variable
// "interrupts".  Finally, 0 is stored into "interrupts" through the pointer in register x1 that points into the
// TSL. Now once x1 has the pointer to the location of the TSL for kernel thread N, it can be be preempted at a
// user-level and the user thread is put on the user-level ready-queue. When the preempted thread gets to the front of
// the user-level ready-queue it is run on kernel thread M. It now stores 0 into "interrupts" back on kernel thread N,
// turning off interrupt on the wrong kernel thread.
// 
// On the x86, the following code is generated for the same code fragment.
// 
// (gdb) disassemble main
// Dump of assembler code for function main:
//    0x0000000000400420 <+0>:  movl   $0x0,%fs:0xfffffffffffffffc
//    0x000000000040042c <+12>: xor    %eax,%eax
//    0x000000000040042e <+14>: retq   
// 
// and there is base-displacement addressing used to atomically reset variable "interrupts" off of the TSL pointer in
// register "fs".
// 
// Hence, the ARM has base-displacement address for the general purpose registers, BUT not to the coprocessor
// registers. As a result, generating the address for the write into variable "interrupts" is no longer atomic.
// 
// Note this problem does NOT occur when just using multiple kernel threads because the preemption ALWAYS restarts the
// thread on the same kernel thread.
// 
// The obvious question is why does ARM use a coprocessor register to store the TSL pointer given that coprocessor
// registers are second-class registers with respect to the instruction set. One possible answer is that they did not
// want to dedicate one of the general registers to hold the TLS pointer and there was a free coprocessor register
// available.

__cfaabi_dbg_debug_do( static thread_local void * last_interrupt = 0; )

extern "C" {
        // Disable interrupts by incrementing the counter
        void disable_interrupts() {
                with( kernelTLS.preemption_state ) {
                        #if GCC_VERSION > 50000
                        static_assert(__atomic_always_lock_free(sizeof(enabled), &enabled), "Must be lock-free");
                        #endif

                        // Set enabled flag to false
                        // should be atomic to avoid preemption in the middle of the operation.
                        // use memory order RELAXED since there is no inter-thread on this variable requirements
                        __atomic_store_n(&enabled, false, __ATOMIC_RELAXED);

                        // Signal the compiler that a fence is needed but only for signal handlers
                        __atomic_signal_fence(__ATOMIC_ACQUIRE);

                        __attribute__((unused)) unsigned short new_val = disable_count + 1;
                        disable_count = new_val;
                        verify( new_val < 65_000u );              // If this triggers someone is disabling interrupts without enabling them
                }
        }

        // Enable interrupts by decrementing the counter
        // If counter reaches 0, execute any pending __cfactx_switch
        void enable_interrupts( __cfaabi_dbg_ctx_param ) {
                processor   * proc = kernelTLS.this_processor; // Cache the processor now since interrupts can start happening after the atomic store
                /* paranoid */ verify( proc );

                with( kernelTLS.preemption_state ){
                        unsigned short prev = disable_count;
                        disable_count -= 1;
                        verify( prev != 0u );                     // If this triggers someone is enabled already enabled interruptsverify( prev != 0u );

                        // Check if we need to prempt the thread because an interrupt was missed
                        if( prev == 1 ) {
                                #if GCC_VERSION > 50000
                                static_assert(__atomic_always_lock_free(sizeof(enabled), &enabled), "Must be lock-free");
                                #endif

                                // Set enabled flag to true
                                // should be atomic to avoid preemption in the middle of the operation.
                                // use memory order RELAXED since there is no inter-thread on this variable requirements
                                __atomic_store_n(&enabled, true, __ATOMIC_RELAXED);

                                // Signal the compiler that a fence is needed but only for signal handlers
                                __atomic_signal_fence(__ATOMIC_RELEASE);
                                if( proc->pending_preemption ) {
                                        proc->pending_preemption = false;
                                        force_yield( __POLL_PREEMPTION );
                                }
                        }
                }

                // For debugging purposes : keep track of the last person to enable the interrupts
                __cfaabi_dbg_debug_do( proc->last_enable = caller; )
        }

        // Disable interrupts by incrementint the counter
        // Don't execute any pending __cfactx_switch even if counter reaches 0
        void enable_interrupts_noPoll() {
                unsigned short prev = kernelTLS.preemption_state.disable_count;
                kernelTLS.preemption_state.disable_count -= 1;
                verifyf( prev != 0u, "Incremented from %u\n", prev );                     // If this triggers someone is enabled already enabled interrupts
                if( prev == 1 ) {
                        #if GCC_VERSION > 50000
                        static_assert(__atomic_always_lock_free(sizeof(kernelTLS.preemption_state.enabled), &kernelTLS.preemption_state.enabled), "Must be lock-free");
                        #endif
                        // Set enabled flag to true
                        // should be atomic to avoid preemption in the middle of the operation.
                        // use memory order RELAXED since there is no inter-thread on this variable requirements
                        __atomic_store_n(&kernelTLS.preemption_state.enabled, true, __ATOMIC_RELAXED);

                        // Signal the compiler that a fence is needed but only for signal handlers
                        __atomic_signal_fence(__ATOMIC_RELEASE);
                }
        }
}

// sigprocmask wrapper : unblock a single signal
static inline void signal_unblock( int sig ) {
        sigset_t mask;
        sigemptyset( &mask );
        sigaddset( &mask, sig );

        if ( pthread_sigmask( SIG_UNBLOCK, &mask, 0p ) == -1 ) {
            abort( "internal error, pthread_sigmask" );
        }
}

// sigprocmask wrapper : block a single signal
static inline void signal_block( int sig ) {
        sigset_t mask;
        sigemptyset( &mask );
        sigaddset( &mask, sig );

        if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
                abort( "internal error, pthread_sigmask" );
        }
}

// kill wrapper : signal a processor
static void preempt( processor * this ) {
        sigval_t value = { PREEMPT_NORMAL };
        pthread_sigqueue( this->kernel_thread, SIGUSR1, value );
}

// reserved for future use
static void timeout( $thread * this ) {
        #if !defined( __CFA_NO_STATISTICS__ )
                kernelTLS.this_stats = this->curr_cluster->stats;
        #endif
        unpark( this );
}

// KERNEL ONLY
// Check if a __cfactx_switch signal handler shoud defer
// If true  : preemption is safe
// If false : preemption is unsafe and marked as pending
static inline bool preemption_ready() {
        // Check if preemption is safe
        bool ready = kernelTLS.preemption_state.enabled && ! kernelTLS.preemption_state.in_progress;

        // Adjust the pending flag accordingly
        kernelTLS.this_processor->pending_preemption = !ready;
        return ready;
}

//=============================================================================================
// Kernel Signal Startup/Shutdown logic
//=============================================================================================

// Startup routine to activate preemption
// Called from kernel_startup
void __kernel_alarm_startup() {
        __cfaabi_dbg_print_safe( "Kernel : Starting preemption\n" );

        // Start with preemption disabled until ready
        kernelTLS.preemption_state.enabled = false;
        kernelTLS.preemption_state.disable_count = 1;

        // Initialize the event kernel
        event_kernel = (event_kernel_t *)&storage_event_kernel;
        (*event_kernel){};

        // Setup proper signal handlers
        __cfaabi_sigaction( SIGUSR1, sigHandler_ctxSwitch, SA_SIGINFO | SA_RESTART ); // __cfactx_switch handler
        __cfaabi_sigaction( SIGALRM, sigHandler_alarm    , SA_SIGINFO | SA_RESTART ); // debug handler

        signal_block( SIGALRM );

        alarm_stack = __create_pthread( &alarm_thread, alarm_loop, 0p );
}

// Shutdown routine to deactivate preemption
// Called from kernel_shutdown
void __kernel_alarm_shutdown() {
        __cfaabi_dbg_print_safe( "Kernel : Preemption stopping\n" );

        // Block all signals since we are already shutting down
        sigset_t mask;
        sigfillset( &mask );
        sigprocmask( SIG_BLOCK, &mask, 0p );

        // Notify the alarm thread of the shutdown
        sigval val = { 1 };
        pthread_sigqueue( alarm_thread, SIGALRM, val );

        // Wait for the preemption thread to finish

        pthread_join( alarm_thread, 0p );
        free( alarm_stack );

        // Preemption is now fully stopped

        __cfaabi_dbg_print_safe( "Kernel : Preemption stopped\n" );
}

// Raii ctor/dtor for the preemption_scope
// Used by thread to control when they want to receive preemption signals
void ?{}( preemption_scope & this, processor * proc ) {
        (this.alarm){ proc, (Time){ 0 }, 0`s };
        this.proc = proc;
        this.proc->preemption_alarm = &this.alarm;

        update_preemption( this.proc, this.proc->cltr->preemption_rate );
}

void ^?{}( preemption_scope & this ) {
        disable_interrupts();

        update_preemption( this.proc, 0`s );
}

//=============================================================================================
// Kernel Signal Handlers
//=============================================================================================

// Context switch signal handler
// Receives SIGUSR1 signal and causes the current thread to yield
static void sigHandler_ctxSwitch( __CFA_SIGPARMS__ ) {
        __cfaabi_dbg_debug_do( last_interrupt = (void *)(cxt->uc_mcontext.CFA_REG_IP); )

        // SKULLDUGGERY: if a thread creates a processor and the immediately deletes it,
        // the interrupt that is supposed to force the kernel thread to preempt might arrive
        // before the kernel thread has even started running. When that happens, an interrupt
        // with a null 'this_processor' will be caught, just ignore it.
        if(! kernelTLS.this_processor ) return;

        choose(sfp->si_value.sival_int) {
                case PREEMPT_NORMAL   : ;// Normal case, nothing to do here
                case PREEMPT_TERMINATE: verify( __atomic_load_n( &kernelTLS.this_processor->do_terminate, __ATOMIC_SEQ_CST ) );
                default:
                        abort( "internal error, signal value is %d", sfp->si_value.sival_int );
        }

        // Check if it is safe to preempt here
        if( !preemption_ready() ) { return; }

        __cfaabi_dbg_print_buffer_decl( " KERNEL: preempting core %p (%p @ %p).\n", kernelTLS.this_processor, kernelTLS.this_thread, (void *)(cxt->uc_mcontext.CFA_REG_IP) );

        // Sync flag : prevent recursive calls to the signal handler
        kernelTLS.preemption_state.in_progress = true;

        // Clear sighandler mask before context switching.
        #if GCC_VERSION > 50000
        static_assert( sizeof( sigset_t ) == sizeof( cxt->uc_sigmask ), "Expected cxt->uc_sigmask to be of sigset_t" );
        #endif
        if ( pthread_sigmask( SIG_SETMASK, (sigset_t *)&(cxt->uc_sigmask), 0p ) == -1 ) {
                abort( "internal error, sigprocmask" );
        }

        // TODO: this should go in finish action
        // Clear the in progress flag
        kernelTLS.preemption_state.in_progress = false;

        // Preemption can occur here

        force_yield( __ALARM_PREEMPTION ); // Do the actual __cfactx_switch
}

static void sigHandler_alarm( __CFA_SIGPARMS__ ) {
        abort("SIGALRM should never reach the signal handler");
}

// Main of the alarm thread
// Waits on SIGALRM and send SIGUSR1 to whom ever needs it
static void * alarm_loop( __attribute__((unused)) void * args ) {
        __processor_id_t id;
        id.full_proc = false;
        id.id = doregister(&id);
        kernelTLS.this_proc_id = &id;

        // Block sigalrms to control when they arrive
        sigset_t mask;
        sigfillset(&mask);
        if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
            abort( "internal error, pthread_sigmask" );
        }

        sigemptyset( &mask );
        sigaddset( &mask, SIGALRM );

        // Main loop
        while( true ) {
                // Wait for a sigalrm
                siginfo_t info;
                int sig = sigwaitinfo( &mask, &info );

                if( sig < 0 ) {
                        //Error!
                        int err = errno;
                        switch( err ) {
                                case EAGAIN :
                                case EINTR :
                                        {__cfaabi_dbg_print_buffer_decl( " KERNEL: Spurious wakeup %d.\n", err );}
                                        continue;
                                case EINVAL :
                                        abort( "Timeout was invalid." );
                                default:
                                        abort( "Unhandled error %d", err);
                        }
                }

                // If another signal arrived something went wrong
                assertf(sig == SIGALRM, "Kernel Internal Error, sigwait: Unexpected signal %d (%d : %d)\n", sig, info.si_code, info.si_value.sival_int);

                // __cfaabi_dbg_print_safe( "Kernel : Caught alarm from %d with %d\n", info.si_code, info.si_value.sival_int );
                // Switch on the code (a.k.a. the sender) to
                switch( info.si_code )
                {
                // Timers can apparently be marked as sent for the kernel
                // In either case, tick preemption
                case SI_TIMER:
                case SI_KERNEL:
                        // __cfaabi_dbg_print_safe( "Kernel : Preemption thread tick\n" );
                        lock( event_kernel->lock __cfaabi_dbg_ctx2 );
                        tick_preemption();
                        unlock( event_kernel->lock );
                        break;
                // Signal was not sent by the kernel but by an other thread
                case SI_QUEUE:
                        // For now, other thread only signal the alarm thread to shut it down
                        // If this needs to change use info.si_value and handle the case here
                        goto EXIT;
                }
        }

EXIT:
        __cfaabi_dbg_print_safe( "Kernel : Preemption thread stopping\n" );
        unregister(&id);
        return 0p;
}

//=============================================================================================
// Kernel Signal Debug
//=============================================================================================

void __cfaabi_check_preemption() {
        bool ready = kernelTLS.preemption_state.enabled;
        if(!ready) { abort("Preemption should be ready"); }

        sigset_t oldset;
        int ret;
        ret = pthread_sigmask(0, ( const sigset_t * ) 0p, &oldset);  // workaround trac#208: cast should be unnecessary
        if(ret != 0) { abort("ERROR sigprocmask returned %d", ret); }

        ret = sigismember(&oldset, SIGUSR1);
        if(ret <  0) { abort("ERROR sigismember returned %d", ret); }
        if(ret == 1) { abort("ERROR SIGUSR1 is disabled"); }

        ret = sigismember(&oldset, SIGALRM);
        if(ret <  0) { abort("ERROR sigismember returned %d", ret); }
        if(ret == 0) { abort("ERROR SIGALRM is enabled"); }

        ret = sigismember(&oldset, SIGTERM);
        if(ret <  0) { abort("ERROR sigismember returned %d", ret); }
        if(ret == 1) { abort("ERROR SIGTERM is disabled"); }
}

#ifdef __CFA_WITH_VERIFY__
bool __cfaabi_dbg_in_kernel() {
        return !kernelTLS.preemption_state.enabled;
}
#endif

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