source: libcfa/src/concurrency/preemption.cfa @ ecfd758

ADTarm-ehast-experimentalenumforall-pointer-decayjacob/cs343-translationnew-ast-unique-exprpthread-emulationqualifiedEnum
Last change on this file since ecfd758 was b0904bf, checked in by Thierry Delisle <tdelisle@…>, 3 years ago

Removed alarm stats, they are now tallied directly into the cluster
  • Property mode set to 100644
File size: 26.7 KB
Line 
1//
2// Cforall Version 1.0.0 Copyright (C) 2016 University of Waterloo
3//
4// The contents of this file are covered under the licence agreement in the
5// file "LICENCE" distributed with Cforall.
6//
7// signal.c --
8//
9// Author           : Thierry Delisle
10// Created On       : Mon Jun 5 14:20:42 2017
11// Last Modified By : Peter A. Buhr
12// Last Modified On : Fri Nov  6 07:42:13 2020
13// Update Count     : 54
14//
15
16#define __cforall_thread__
17// #define __CFA_DEBUG_PRINT_PREEMPTION__
18
19#include "preemption.hfa"
20#include <assert.h>
21
22#include <errno.h>
23#include <stdio.h>
24#include <string.h>
25#include <unistd.h>
26#include <limits.h>                                                                             // PTHREAD_STACK_MIN
27
28#include "bits/signal.hfa"
29#include "kernel_private.hfa"
30
31
32#if !defined(__CFA_DEFAULT_PREEMPTION__)
33#define __CFA_DEFAULT_PREEMPTION__ 10`ms
34#endif
35
36__attribute__((weak)) Duration default_preemption() {
37        const char * preempt_rate_s = getenv("CFA_DEFAULT_PREEMPTION");
38        if(!preempt_rate_s) {
39                __cfadbg_print_safe(preemption, "No CFA_DEFAULT_PREEMPTION in ENV\n");
40                return __CFA_DEFAULT_PREEMPTION__;
41        }
42
43        char * endptr = 0p;
44        long int preempt_rate_l = strtol(preempt_rate_s, &endptr, 10);
45        if(preempt_rate_l < 0 || preempt_rate_l > 65535) {
46                __cfadbg_print_safe(preemption, "CFA_DEFAULT_PREEMPTION out of range : %ld\n", preempt_rate_l);
47                return __CFA_DEFAULT_PREEMPTION__;
48        }
49        if('\0' != *endptr) {
50                __cfadbg_print_safe(preemption, "CFA_DEFAULT_PREEMPTION not a decimal number : %s\n", preempt_rate_s);
51                return __CFA_DEFAULT_PREEMPTION__;
52        }
53
54        return preempt_rate_l`ms;
55}
56
57// FwdDeclarations : timeout handlers
58static void preempt( processor   * this );
59static void timeout( $thread * this );
60
61// FwdDeclarations : Signal handlers
62static void sigHandler_ctxSwitch( __CFA_SIGPARMS__ );
63static void sigHandler_alarm    ( __CFA_SIGPARMS__ );
64static void sigHandler_segv     ( __CFA_SIGPARMS__ );
65static void sigHandler_ill      ( __CFA_SIGPARMS__ );
66static void sigHandler_fpe      ( __CFA_SIGPARMS__ );
67static void sigHandler_abort    ( __CFA_SIGPARMS__ );
68
69// FwdDeclarations : alarm thread main
70static void * alarm_loop( __attribute__((unused)) void * args );
71
72// Machine specific register name
73#if   defined( __i386 )
74#define CFA_REG_IP gregs[REG_EIP]
75#elif defined( __x86_64 )
76#define CFA_REG_IP gregs[REG_RIP]
77#elif defined( __arm__ )
78#define CFA_REG_IP arm_pc
79#elif defined( __aarch64__ )
80#define CFA_REG_IP pc
81#else
82#error unsupported hardware architecture
83#endif
84
85KERNEL_STORAGE(event_kernel_t, event_kernel);         // private storage for event kernel
86event_kernel_t * event_kernel;                        // kernel public handle to even kernel
87static pthread_t alarm_thread;                        // pthread handle to alarm thread
88static void * alarm_stack;                                                        // pthread stack for alarm thread
89
90static void ?{}(event_kernel_t & this) with( this ) {
91        alarms{};
92        lock{};
93}
94
95enum {
96        PREEMPT_NORMAL    = 0,
97        PREEMPT_TERMINATE = 1,
98};
99
100//=============================================================================================
101// Kernel Preemption logic
102//=============================================================================================
103
104// Get next expired node
105static inline alarm_node_t * get_expired( alarm_list_t * alarms, Time currtime ) {
106        if( ! & (*alarms)`first ) return 0p;                                            // If no alarms return null
107        if( (*alarms)`first.alarm >= currtime ) return 0p;      // If alarms head not expired return null
108        return pop(alarms);                                                                     // Otherwise just pop head
109}
110
111// Tick one frame of the Discrete Event Simulation for alarms
112static void tick_preemption(void) {
113        alarm_node_t * node = 0p;                                                       // Used in the while loop but cannot be declared in the while condition
114        alarm_list_t * alarms = &event_kernel->alarms;          // Local copy for ease of reading
115        Time currtime = __kernel_get_time();                            // Check current time once so everything "happens at once"
116
117        //Loop throught every thing expired
118        while( node = get_expired( alarms, currtime ) ) {
119                __cfadbg_print_buffer_decl( preemption, " KERNEL: preemption tick %lu\n", currtime.tn);
120                Duration period = node->period;
121                if( period == 0) {
122                        node->set = false;                  // Node is one-shot, just mark it as not pending
123                }
124
125                __cfadbg_print_buffer_local( preemption, " KERNEL: alarm ticking node %p.\n", node );
126
127
128                // Check if this is a kernel
129                if( node->type == Kernel ) {
130                        preempt( node->proc );
131                }
132                else if( node->type == User ) {
133                        __cfadbg_print_buffer_local( preemption, " KERNEL: alarm unparking %p.\n", node->thrd );
134                        timeout( node->thrd );
135                }
136                else {
137                        node->callback(*node);
138                }
139
140                // Check if this is a periodic alarm
141                if( period > 0 ) {
142                        __cfadbg_print_buffer_local( preemption, " KERNEL: alarm period is %lu.\n", period`ns );
143                        node->alarm = currtime + period;    // Alarm is periodic, add currtime to it (used cached current time)
144                        insert( alarms, node );             // Reinsert the node for the next time it triggers
145                }
146        }
147
148        // If there are still alarms pending, reset the timer
149        if( & (*alarms)`first ) {
150                Duration delta = (*alarms)`first.alarm - currtime;
151                Duration capped = max(delta, 50`us);
152                __kernel_set_timer( capped );
153        }
154}
155
156// Update the preemption of a processor and notify interested parties
157void update_preemption( processor * this, Duration duration ) {
158        alarm_node_t * alarm = this->preemption_alarm;
159
160        // Alarms need to be enabled
161        if ( duration > 0 && ! alarm->set ) {
162                alarm->alarm = __kernel_get_time() + duration;
163                alarm->period = duration;
164                register_self( alarm );
165        }
166        // Zero duration but alarm is set
167        else if ( duration == 0 && alarm->set ) {
168                unregister_self( alarm );
169                alarm->alarm = 0;
170                alarm->period = 0;
171        }
172        // If alarm is different from previous, change it
173        else if ( duration > 0 && alarm->period != duration ) {
174                unregister_self( alarm );
175                alarm->alarm = __kernel_get_time() + duration;
176                alarm->period = duration;
177                register_self( alarm );
178        }
179}
180
181//=============================================================================================
182// Kernel Signal Tools
183//=============================================================================================
184// In a user-level threading system, there are handful of thread-local variables where this problem occurs on the ARM.
185//
186// For each kernel thread running user-level threads, there is a flag variable to indicate if interrupts are
187// enabled/disabled for that kernel thread. Therefore, this variable is made thread local.
188//
189// For example, this code fragment sets the state of the "interrupt" variable in thread-local memory.
190//
191// _Thread_local volatile int interrupts;
192// int main() {
193//     interrupts = 0; // disable interrupts }
194//
195// which generates the following code on the ARM
196//
197// (gdb) disassemble main
198// Dump of assembler code for function main:
199//    0x0000000000000610 <+0>:  mrs     x1, tpidr_el0
200//    0x0000000000000614 <+4>:  mov     w0, #0x0                        // #0
201//    0x0000000000000618 <+8>:  add     x1, x1, #0x0, lsl #12
202//    0x000000000000061c <+12>: add     x1, x1, #0x10
203//    0x0000000000000620 <+16>: str     wzr, [x1]
204//    0x0000000000000624 <+20>: ret
205//
206// The mrs moves a pointer from coprocessor register tpidr_el0 into register x1.  Register w0 is set to 0. The two adds
207// increase the TLS pointer with the displacement (offset) 0x10, which is the location in the TSL of variable
208// "interrupts".  Finally, 0 is stored into "interrupts" through the pointer in register x1 that points into the
209// TSL. Now once x1 has the pointer to the location of the TSL for kernel thread N, it can be be preempted at a
210// user-level and the user thread is put on the user-level ready-queue. When the preempted thread gets to the front of
211// the user-level ready-queue it is run on kernel thread M. It now stores 0 into "interrupts" back on kernel thread N,
212// turning off interrupt on the wrong kernel thread.
213//
214// On the x86, the following code is generated for the same code fragment.
215//
216// (gdb) disassemble main
217// Dump of assembler code for function main:
218//    0x0000000000400420 <+0>:  movl   $0x0,%fs:0xfffffffffffffffc
219//    0x000000000040042c <+12>: xor    %eax,%eax
220//    0x000000000040042e <+14>: retq
221//
222// and there is base-displacement addressing used to atomically reset variable "interrupts" off of the TSL pointer in
223// register "fs".
224//
225// Hence, the ARM has base-displacement address for the general purpose registers, BUT not to the coprocessor
226// registers. As a result, generating the address for the write into variable "interrupts" is no longer atomic.
227//
228// Note this problem does NOT occur when just using multiple kernel threads because the preemption ALWAYS restarts the
229// thread on the same kernel thread.
230//
231// The obvious question is why does ARM use a coprocessor register to store the TSL pointer given that coprocessor
232// registers are second-class registers with respect to the instruction set. One possible answer is that they did not
233// want to dedicate one of the general registers to hold the TLS pointer and there was a free coprocessor register
234// available.
235
236//-----------------------------------------------------------------------------
237// Some assembly required
238#define __cfaasm_label(label, when) when: asm volatile goto(".global __cfaasm_" #label "_" #when "\n" "__cfaasm_" #label "_" #when ":":::"memory":when)
239
240//----------
241// special case for preemption since used often
242bool __preemption_enabled() {
243        // create a assembler label before
244        // marked as clobber all to avoid movement
245        __cfaasm_label(check, before);
246
247        // access tls as normal
248        bool enabled = __cfaabi_tls.preemption_state.enabled;
249
250        // create a assembler label after
251        // marked as clobber all to avoid movement
252        __cfaasm_label(check, after);
253        return enabled;
254}
255
256struct asm_region {
257        void * before;
258        void * after;
259};
260
261static inline bool __cfaasm_in( void * ip, struct asm_region & region ) {
262        return ip >= region.before && ip <= region.after;
263}
264
265
266//----------
267// Get data from the TLS block
268// struct asm_region __cfaasm_get;
269uintptr_t __cfatls_get( unsigned long int offset ) __attribute__((__noinline__)); //no inline to avoid problems
270uintptr_t __cfatls_get( unsigned long int offset ) {
271        // create a assembler label before
272        // marked as clobber all to avoid movement
273        __cfaasm_label(get, before);
274
275        // access tls as normal (except for pointer arithmetic)
276        uintptr_t val = *(uintptr_t*)((uintptr_t)&__cfaabi_tls + offset);
277
278        // create a assembler label after
279        // marked as clobber all to avoid movement
280        __cfaasm_label(get, after);
281        return val;
282}
283
284extern "C" {
285        // Disable interrupts by incrementing the counter
286        void disable_interrupts() {
287                // create a assembler label before
288                // marked as clobber all to avoid movement
289                __cfaasm_label(dsable, before);
290
291                with( __cfaabi_tls.preemption_state ) {
292                        #if GCC_VERSION > 50000
293                        static_assert(__atomic_always_lock_free(sizeof(enabled), &enabled), "Must be lock-free");
294                        #endif
295
296                        // Set enabled flag to false
297                        // should be atomic to avoid preemption in the middle of the operation.
298                        // use memory order RELAXED since there is no inter-thread on this variable requirements
299                        __atomic_store_n(&enabled, false, __ATOMIC_RELAXED);
300
301                        // Signal the compiler that a fence is needed but only for signal handlers
302                        __atomic_signal_fence(__ATOMIC_ACQUIRE);
303
304                        __attribute__((unused)) unsigned short new_val = disable_count + 1;
305                        disable_count = new_val;
306                        verify( new_val < 65_000u );              // If this triggers someone is disabling interrupts without enabling them
307                }
308
309                // create a assembler label after
310                // marked as clobber all to avoid movement
311                __cfaasm_label(dsable, after);
312
313        }
314
315        // Enable interrupts by decrementing the counter
316        // If counter reaches 0, execute any pending __cfactx_switch
317        void enable_interrupts( __cfaabi_dbg_ctx_param ) {
318                // Cache the processor now since interrupts can start happening after the atomic store
319                processor   * proc = __cfaabi_tls.this_processor;
320                /* paranoid */ verify( proc );
321
322                with( __cfaabi_tls.preemption_state ){
323                        unsigned short prev = disable_count;
324                        disable_count -= 1;
325
326                        // If this triggers someone is enabled already enabled interruptsverify( prev != 0u );
327                        /* paranoid */ verify( prev != 0u );
328
329                        // Check if we need to prempt the thread because an interrupt was missed
330                        if( prev == 1 ) {
331                                #if GCC_VERSION > 50000
332                                        static_assert(__atomic_always_lock_free(sizeof(enabled), &enabled), "Must be lock-free");
333                                #endif
334
335                                // Set enabled flag to true
336                                // should be atomic to avoid preemption in the middle of the operation.
337                                // use memory order RELAXED since there is no inter-thread on this variable requirements
338                                __atomic_store_n(&enabled, true, __ATOMIC_RELAXED);
339
340                                // Signal the compiler that a fence is needed but only for signal handlers
341                                __atomic_signal_fence(__ATOMIC_RELEASE);
342                                if( proc->pending_preemption ) {
343                                        proc->pending_preemption = false;
344                                        force_yield( __POLL_PREEMPTION );
345                                }
346                        }
347                }
348
349                // For debugging purposes : keep track of the last person to enable the interrupts
350                __cfaabi_dbg_debug_do( proc->last_enable = caller; )
351        }
352
353        // Disable interrupts by incrementint the counter
354        // Don't execute any pending __cfactx_switch even if counter reaches 0
355        void enable_interrupts_noPoll() {
356                unsigned short prev = __cfaabi_tls.preemption_state.disable_count;
357                __cfaabi_tls.preemption_state.disable_count -= 1;
358                // If this triggers someone is enabled already enabled interrupts
359                /* paranoid */ verifyf( prev != 0u, "Incremented from %u\n", prev );
360                if( prev == 1 ) {
361                        #if GCC_VERSION > 50000
362                                static_assert(__atomic_always_lock_free(sizeof(__cfaabi_tls.preemption_state.enabled), &__cfaabi_tls.preemption_state.enabled), "Must be lock-free");
363                        #endif
364                        // Set enabled flag to true
365                        // should be atomic to avoid preemption in the middle of the operation.
366                        // use memory order RELAXED since there is no inter-thread on this variable requirements
367                        __atomic_store_n(&__cfaabi_tls.preemption_state.enabled, true, __ATOMIC_RELAXED);
368
369                        // Signal the compiler that a fence is needed but only for signal handlers
370                        __atomic_signal_fence(__ATOMIC_RELEASE);
371                }
372        }
373}
374
375//-----------------------------------------------------------------------------
376// Kernel Signal Debug
377void __cfaabi_check_preemption() {
378        bool ready = __preemption_enabled();
379        if(!ready) { abort("Preemption should be ready"); }
380
381        __cfaasm_label(debug, before);
382
383                sigset_t oldset;
384                int ret;
385                ret = pthread_sigmask(0, ( const sigset_t * ) 0p, &oldset);  // workaround trac#208: cast should be unnecessary
386                if(ret != 0) { abort("ERROR sigprocmask returned %d", ret); }
387
388                ret = sigismember(&oldset, SIGUSR1);
389                if(ret <  0) { abort("ERROR sigismember returned %d", ret); }
390                if(ret == 1) { abort("ERROR SIGUSR1 is disabled"); }
391
392                ret = sigismember(&oldset, SIGALRM);
393                if(ret <  0) { abort("ERROR sigismember returned %d", ret); }
394                if(ret == 0) { abort("ERROR SIGALRM is enabled"); }
395
396                ret = sigismember(&oldset, SIGTERM);
397                if(ret <  0) { abort("ERROR sigismember returned %d", ret); }
398                if(ret == 1) { abort("ERROR SIGTERM is disabled"); }
399
400        __cfaasm_label(debug, after);
401}
402
403#ifdef __CFA_WITH_VERIFY__
404bool __cfaabi_dbg_in_kernel() {
405        return !__preemption_enabled();
406}
407#endif
408
409#undef __cfaasm_label
410
411//-----------------------------------------------------------------------------
412// Signal handling
413
414// sigprocmask wrapper : unblock a single signal
415static inline void signal_unblock( int sig ) {
416        sigset_t mask;
417        sigemptyset( &mask );
418        sigaddset( &mask, sig );
419
420        if ( pthread_sigmask( SIG_UNBLOCK, &mask, 0p ) == -1 ) {
421            abort( "internal error, pthread_sigmask" );
422        }
423}
424
425// sigprocmask wrapper : block a single signal
426static inline void signal_block( int sig ) {
427        sigset_t mask;
428        sigemptyset( &mask );
429        sigaddset( &mask, sig );
430
431        if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
432                abort( "internal error, pthread_sigmask" );
433        }
434}
435
436// kill wrapper : signal a processor
437static void preempt( processor * this ) {
438        sigval_t value = { PREEMPT_NORMAL };
439        pthread_sigqueue( this->kernel_thread, SIGUSR1, value );
440}
441
442// reserved for future use
443static void timeout( $thread * this ) {
444        unpark( this );
445}
446
447void __disable_interrupts_hard() {
448        sigset_t oldset;
449        int ret;
450        ret = pthread_sigmask(0, ( const sigset_t * ) 0p, &oldset);  // workaround trac#208: cast should be unnecessary
451        if(ret != 0) { abort("ERROR sigprocmask returned %d", ret); }
452
453        ret = sigismember(&oldset, SIGUSR1);
454        if(ret <  0) { abort("ERROR sigismember returned %d", ret); }
455        if(ret == 1) { abort("ERROR SIGUSR1 is disabled"); }
456
457        ret = sigismember(&oldset, SIGALRM);
458        if(ret <  0) { abort("ERROR sigismember returned %d", ret); }
459        if(ret == 0) { abort("ERROR SIGALRM is enabled"); }
460
461        signal_block( SIGUSR1 );
462}
463
464void __enable_interrupts_hard() {
465        signal_unblock( SIGUSR1 );
466
467        sigset_t oldset;
468        int ret;
469        ret = pthread_sigmask(0, ( const sigset_t * ) 0p, &oldset);  // workaround trac#208: cast should be unnecessary
470        if(ret != 0) { abort("ERROR sigprocmask returned %d", ret); }
471
472        ret = sigismember(&oldset, SIGUSR1);
473        if(ret <  0) { abort("ERROR sigismember returned %d", ret); }
474        if(ret == 1) { abort("ERROR SIGUSR1 is disabled"); }
475
476        ret = sigismember(&oldset, SIGALRM);
477        if(ret <  0) { abort("ERROR sigismember returned %d", ret); }
478        if(ret == 0) { abort("ERROR SIGALRM is enabled"); }
479}
480
481//-----------------------------------------------------------------------------
482// Some assembly required
483#if defined( __i386 )
484        #ifdef __PIC__
485                #define RELOC_PRELUDE( label ) \
486                        "calll   .Lcfaasm_prelude_" #label "$pb\n\t" \
487                        ".Lcfaasm_prelude_" #label "$pb:\n\t" \
488                        "popl    %%eax\n\t" \
489                        ".Lcfaasm_prelude_" #label "_end:\n\t" \
490                        "addl    $_GLOBAL_OFFSET_TABLE_+(.Lcfaasm_prelude_" #label "_end-.Lcfaasm_prelude_" #label "$pb), %%eax\n\t"
491                #define RELOC_PREFIX ""
492                #define RELOC_SUFFIX "@GOT(%%eax)"
493        #else
494                #define RELOC_PREFIX "$"
495                #define RELOC_SUFFIX ""
496        #endif
497        #define __cfaasm_label( label ) struct asm_region label = \
498                ({ \
499                        struct asm_region region; \
500                        asm( \
501                                RELOC_PRELUDE( label ) \
502                                "movl " RELOC_PREFIX "__cfaasm_" #label "_before" RELOC_SUFFIX ", %[vb]\n\t" \
503                                "movl " RELOC_PREFIX "__cfaasm_" #label "_after"  RELOC_SUFFIX ", %[va]\n\t" \
504                                 : [vb]"=r"(region.before), [va]"=r"(region.after) \
505                        ); \
506                        region; \
507                });
508#elif defined( __x86_64 )
509        #ifdef __PIC__
510                #define RELOC_PREFIX ""
511                #define RELOC_SUFFIX "@GOTPCREL(%%rip)"
512        #else
513                #define RELOC_PREFIX "$"
514                #define RELOC_SUFFIX ""
515        #endif
516        #define __cfaasm_label( label ) struct asm_region label = \
517                ({ \
518                        struct asm_region region; \
519                        asm( \
520                                "movq " RELOC_PREFIX "__cfaasm_" #label "_before" RELOC_SUFFIX ", %[vb]\n\t" \
521                                "movq " RELOC_PREFIX "__cfaasm_" #label "_after"  RELOC_SUFFIX ", %[va]\n\t" \
522                                 : [vb]"=r"(region.before), [va]"=r"(region.after) \
523                        ); \
524                        region; \
525                });
526#elif defined( __aarch64__ )
527        #ifdef __PIC__
528                // Note that this works only for gcc
529                #define __cfaasm_label( label ) struct asm_region label = \
530                ({ \
531                        struct asm_region region; \
532                        asm( \
533                                "adrp %[vb], _GLOBAL_OFFSET_TABLE_"                              "\n\t" \
534                                "ldr  %[vb], [%[vb], #:gotpage_lo15:__cfaasm_" #label "_before]" "\n\t" \
535                                "adrp %[va], _GLOBAL_OFFSET_TABLE_"                              "\n\t" \
536                                "ldr  %[va], [%[va], #:gotpage_lo15:__cfaasm_" #label "_after]"  "\n\t" \
537                                 : [vb]"=r"(region.before), [va]"=r"(region.after) \
538                        ); \
539                        region; \
540                });
541        #else
542                #error this is not the right thing to do
543                /*
544                #define __cfaasm_label( label ) struct asm_region label = \
545                ({ \
546                        struct asm_region region; \
547                        asm( \
548                                "adrp %[vb], __cfaasm_" #label "_before"              "\n\t" \
549                                "add  %[vb], %[vb], :lo12:__cfaasm_" #label "_before" "\n\t" \
550                                "adrp %[va], :got:__cfaasm_" #label "_after"          "\n\t" \
551                                "add  %[va], %[va], :lo12:__cfaasm_" #label "_after"  "\n\t" \
552                                 : [vb]"=r"(region.before), [va]"=r"(region.after) \
553                        ); \
554                        region; \
555                });
556                */
557        #endif
558#else
559        #error unknown hardware architecture
560#endif
561
562// KERNEL ONLY
563// Check if a __cfactx_switch signal handler shoud defer
564// If true  : preemption is safe
565// If false : preemption is unsafe and marked as pending
566static inline bool preemption_ready( void * ip ) {
567        // Get all the region for which it is not safe to preempt
568        __cfaasm_label( get    );
569        __cfaasm_label( check  );
570        __cfaasm_label( dsable );
571        __cfaasm_label( debug  );
572
573        // Check if preemption is safe
574        bool ready = true;
575        if( __cfaasm_in( ip, get    ) ) { ready = false; goto EXIT; };
576        if( __cfaasm_in( ip, check  ) ) { ready = false; goto EXIT; };
577        if( __cfaasm_in( ip, dsable ) ) { ready = false; goto EXIT; };
578        if( __cfaasm_in( ip, debug  ) ) { ready = false; goto EXIT; };
579        if( !__cfaabi_tls.preemption_state.enabled) { ready = false; goto EXIT; };
580        if( __cfaabi_tls.preemption_state.in_progress ) { ready = false; goto EXIT; };
581
582EXIT:
583        // Adjust the pending flag accordingly
584        __cfaabi_tls.this_processor->pending_preemption = !ready;
585        return ready;
586}
587
588//=============================================================================================
589// Kernel Signal Startup/Shutdown logic
590//=============================================================================================
591
592// Startup routine to activate preemption
593// Called from kernel_startup
594void __kernel_alarm_startup() {
595        __cfaabi_dbg_print_safe( "Kernel : Starting preemption\n" );
596
597        // Start with preemption disabled until ready
598        __cfaabi_tls.preemption_state.enabled = false;
599        __cfaabi_tls.preemption_state.disable_count = 1;
600
601        // Initialize the event kernel
602        event_kernel = (event_kernel_t *)&storage_event_kernel;
603        (*event_kernel){};
604
605        // Setup proper signal handlers
606        __cfaabi_sigaction( SIGUSR1, sigHandler_ctxSwitch, SA_SIGINFO ); // __cfactx_switch handler
607        __cfaabi_sigaction( SIGALRM, sigHandler_alarm    , SA_SIGINFO ); // debug handler
608
609        signal_block( SIGALRM );
610
611        alarm_stack = __create_pthread( &alarm_thread, alarm_loop, 0p );
612}
613
614// Shutdown routine to deactivate preemption
615// Called from kernel_shutdown
616void __kernel_alarm_shutdown() {
617        __cfaabi_dbg_print_safe( "Kernel : Preemption stopping\n" );
618
619        // Block all signals since we are already shutting down
620        sigset_t mask;
621        sigfillset( &mask );
622        sigprocmask( SIG_BLOCK, &mask, 0p );
623
624        // Notify the alarm thread of the shutdown
625        sigval val = { 1 };
626        pthread_sigqueue( alarm_thread, SIGALRM, val );
627
628        // Wait for the preemption thread to finish
629
630        __destroy_pthread( alarm_thread, alarm_stack, 0p );
631
632        // Preemption is now fully stopped
633
634        __cfaabi_dbg_print_safe( "Kernel : Preemption stopped\n" );
635}
636
637// Prevent preemption since we are about to start terminating things
638void __kernel_abort_lock(void) {
639        signal_block( SIGUSR1 );
640}
641
642// Raii ctor/dtor for the preemption_scope
643// Used by thread to control when they want to receive preemption signals
644void ?{}( preemption_scope & this, processor * proc ) {
645        (this.alarm){ proc, (Time){ 0 }, 0`s };
646        this.proc = proc;
647        this.proc->preemption_alarm = &this.alarm;
648
649        update_preemption( this.proc, this.proc->cltr->preemption_rate );
650}
651
652void ^?{}( preemption_scope & this ) {
653        disable_interrupts();
654
655        update_preemption( this.proc, 0`s );
656}
657
658//=============================================================================================
659// Kernel Signal Handlers
660//=============================================================================================
661__cfaabi_dbg_debug_do( static thread_local void * last_interrupt = 0; )
662
663// Context switch signal handler
664// Receives SIGUSR1 signal and causes the current thread to yield
665static void sigHandler_ctxSwitch( __CFA_SIGPARMS__ ) {
666        void * ip = (void *)(cxt->uc_mcontext.CFA_REG_IP);
667        __cfaabi_dbg_debug_do( last_interrupt = ip; )
668
669        // SKULLDUGGERY: if a thread creates a processor and the immediately deletes it,
670        // the interrupt that is supposed to force the kernel thread to preempt might arrive
671        // before the kernel thread has even started running. When that happens, an interrupt
672        // with a null 'this_processor' will be caught, just ignore it.
673        if(! __cfaabi_tls.this_processor ) return;
674
675        choose(sfp->si_value.sival_int) {
676                case PREEMPT_NORMAL   : ;// Normal case, nothing to do here
677                case PREEMPT_TERMINATE: verify( __atomic_load_n( &__cfaabi_tls.this_processor->do_terminate, __ATOMIC_SEQ_CST ) );
678                default:
679                        abort( "internal error, signal value is %d", sfp->si_value.sival_int );
680        }
681
682        // Check if it is safe to preempt here
683        if( !preemption_ready( ip ) ) { return; }
684
685        __cfaabi_dbg_print_buffer_decl( " KERNEL: preempting core %p (%p @ %p).\n", __cfaabi_tls.this_processor, __cfaabi_tls.this_thread, (void *)(cxt->uc_mcontext.CFA_REG_IP) );
686
687        // Sync flag : prevent recursive calls to the signal handler
688        __cfaabi_tls.preemption_state.in_progress = true;
689
690        // Clear sighandler mask before context switching.
691        #if GCC_VERSION > 50000
692        static_assert( sizeof( sigset_t ) == sizeof( cxt->uc_sigmask ), "Expected cxt->uc_sigmask to be of sigset_t" );
693        #endif
694        if ( pthread_sigmask( SIG_SETMASK, (sigset_t *)&(cxt->uc_sigmask), 0p ) == -1 ) {
695                abort( "internal error, sigprocmask" );
696        }
697
698        // Clear the in progress flag
699        __cfaabi_tls.preemption_state.in_progress = false;
700
701        // Preemption can occur here
702
703        force_yield( __ALARM_PREEMPTION ); // Do the actual __cfactx_switch
704}
705
706static void sigHandler_alarm( __CFA_SIGPARMS__ ) {
707        abort("SIGALRM should never reach the signal handler");
708}
709
710// Main of the alarm thread
711// Waits on SIGALRM and send SIGUSR1 to whom ever needs it
712static void * alarm_loop( __attribute__((unused)) void * args ) {
713        __processor_id_t id;
714        id.id = doregister(&id);
715        __cfaabi_tls.this_proc_id = &id;
716
717
718        // Block sigalrms to control when they arrive
719        sigset_t mask;
720        sigfillset(&mask);
721        if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
722            abort( "internal error, pthread_sigmask" );
723        }
724
725        sigemptyset( &mask );
726        sigaddset( &mask, SIGALRM );
727
728        // Main loop
729        while( true ) {
730                // Wait for a sigalrm
731                siginfo_t info;
732                int sig = sigwaitinfo( &mask, &info );
733
734                if( sig < 0 ) {
735                        //Error!
736                        int err = errno;
737                        switch( err ) {
738                                case EAGAIN :
739                                case EINTR :
740                                        {__cfaabi_dbg_print_buffer_decl( " KERNEL: Spurious wakeup %d.\n", err );}
741                                        continue;
742                                case EINVAL :
743                                        abort( "Timeout was invalid." );
744                                default:
745                                        abort( "Unhandled error %d", err);
746                        }
747                }
748
749                // If another signal arrived something went wrong
750                assertf(sig == SIGALRM, "Kernel Internal Error, sigwait: Unexpected signal %d (%d : %d)\n", sig, info.si_code, info.si_value.sival_int);
751
752                // __cfaabi_dbg_print_safe( "Kernel : Caught alarm from %d with %d\n", info.si_code, info.si_value.sival_int );
753                // Switch on the code (a.k.a. the sender) to
754                switch( info.si_code )
755                {
756                // Timers can apparently be marked as sent for the kernel
757                // In either case, tick preemption
758                case SI_TIMER:
759                case SI_KERNEL:
760                        // __cfaabi_dbg_print_safe( "Kernel : Preemption thread tick\n" );
761                        lock( event_kernel->lock __cfaabi_dbg_ctx2 );
762                        tick_preemption();
763                        unlock( event_kernel->lock );
764                        break;
765                // Signal was not sent by the kernel but by an other thread
766                case SI_QUEUE:
767                        // For now, other thread only signal the alarm thread to shut it down
768                        // If this needs to change use info.si_value and handle the case here
769                        goto EXIT;
770                }
771        }
772
773EXIT:
774        __cfaabi_dbg_print_safe( "Kernel : Preemption thread stopping\n" );
775        unregister(&id);
776
777        return 0p;
778}
779
780// Local Variables: //
781// mode: c //
782// tab-width: 4 //
783// End: //
Note: See TracBrowser for help on using the repository browser.