source: libcfa/src/concurrency/ready_queue.cfa @ e0d6748

arm-ehjacob/cs343-translationnew-ast-unique-expr
Last change on this file since e0d6748 was e0d6748, checked in by Thierry Delisle <tdelisle@…>, 8 months ago

Removed unused functions

  • Property mode set to 100644
File size: 18.2 KB
Line 
1//
2// Cforall Version 1.0.0 Copyright (C) 2019 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// ready_queue.cfa --
8//
9// Author           : Thierry Delisle
10// Created On       : Mon Nov dd 16:29:18 2019
11// Last Modified By :
12// Last Modified On :
13// Update Count     :
14//
15
16#define __cforall_thread__
17// #define __CFA_DEBUG_PRINT_READY_QUEUE__
18
19// #define USE_MPSC
20
21#include "bits/defs.hfa"
22#include "kernel_private.hfa"
23
24#define _GNU_SOURCE
25#include "stdlib.hfa"
26#include "math.hfa"
27
28#include <unistd.h>
29
30#include "ready_subqueue.hfa"
31
32static const size_t cache_line_size = 64;
33
34// No overriden function, no environment variable, no define
35// fall back to a magic number
36#ifndef __CFA_MAX_PROCESSORS__
37        #define __CFA_MAX_PROCESSORS__ 1024
38#endif
39
40#define BIAS 4
41
42// returns the maximum number of processors the RWLock support
43__attribute__((weak)) unsigned __max_processors() {
44        const char * max_cores_s = getenv("CFA_MAX_PROCESSORS");
45        if(!max_cores_s) {
46                __cfadbg_print_nolock(ready_queue, "No CFA_MAX_PROCESSORS in ENV\n");
47                return __CFA_MAX_PROCESSORS__;
48        }
49
50        char * endptr = 0p;
51        long int max_cores_l = strtol(max_cores_s, &endptr, 10);
52        if(max_cores_l < 1 || max_cores_l > 65535) {
53                __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS out of range : %ld\n", max_cores_l);
54                return __CFA_MAX_PROCESSORS__;
55        }
56        if('\0' != *endptr) {
57                __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS not a decimal number : %s\n", max_cores_s);
58                return __CFA_MAX_PROCESSORS__;
59        }
60
61        return max_cores_l;
62}
63
64//=======================================================================
65// Cluster wide reader-writer lock
66//=======================================================================
67void  ?{}(__scheduler_RWLock_t & this) {
68        this.max   = __max_processors();
69        this.alloc = 0;
70        this.ready = 0;
71        this.lock  = false;
72        this.data  = alloc(this.max);
73
74        /*paranoid*/ verify( 0 == (((uintptr_t)(this.data    )) % 64) );
75        /*paranoid*/ verify( 0 == (((uintptr_t)(this.data + 1)) % 64) );
76        /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.alloc), &this.alloc));
77        /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.ready), &this.ready));
78
79}
80void ^?{}(__scheduler_RWLock_t & this) {
81        free(this.data);
82}
83
84void ?{}( __scheduler_lock_id_t & this, __processor_id_t * proc ) {
85        this.handle = proc;
86        this.lock   = false;
87        #ifdef __CFA_WITH_VERIFY__
88                this.owned  = false;
89        #endif
90}
91
92//=======================================================================
93// Lock-Free registering/unregistering of threads
94void register_proc_id( struct __processor_id_t * proc ) with(*__scheduler_lock) {
95        __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p for RW-Lock\n", proc);
96
97        // Step - 1 : check if there is already space in the data
98        uint_fast32_t s = ready;
99
100        // Check among all the ready
101        for(uint_fast32_t i = 0; i < s; i++) {
102                __processor_id_t * null = 0p; // Re-write every loop since compare thrashes it
103                if( __atomic_load_n(&data[i].handle, (int)__ATOMIC_RELAXED) == null
104                        && __atomic_compare_exchange_n( &data[i].handle, &null, proc, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
105                        /*paranoid*/ verify(i < ready);
106                        /*paranoid*/ verify(0 == (__alignof__(data[i]) % cache_line_size));
107                        /*paranoid*/ verify((((uintptr_t)&data[i]) % cache_line_size) == 0);
108                        proc->id = i;
109                }
110        }
111
112        if(max <= alloc) abort("Trying to create more than %ud processors", __scheduler_lock->max);
113
114        // Step - 2 : F&A to get a new spot in the array.
115        uint_fast32_t n = __atomic_fetch_add(&alloc, 1, __ATOMIC_SEQ_CST);
116        if(max <= n) abort("Trying to create more than %ud processors", __scheduler_lock->max);
117
118        // Step - 3 : Mark space as used and then publish it.
119        __scheduler_lock_id_t * storage = (__scheduler_lock_id_t *)&data[n];
120        (*storage){ proc };
121        while() {
122                unsigned copy = n;
123                if( __atomic_load_n(&ready, __ATOMIC_RELAXED) == n
124                        && __atomic_compare_exchange_n(&ready, &copy, n + 1, true, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))
125                        break;
126                Pause();
127        }
128
129        __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p done, id %lu\n", proc, n);
130
131        // Return new spot.
132        /*paranoid*/ verify(n < ready);
133        /*paranoid*/ verify(__alignof__(data[n]) == (2 * cache_line_size));
134        /*paranoid*/ verify((((uintptr_t)&data[n]) % cache_line_size) == 0);
135        proc->id = n;
136}
137
138void unregister_proc_id( struct __processor_id_t * proc ) with(*__scheduler_lock) {
139        unsigned id = proc->id;
140        /*paranoid*/ verify(id < ready);
141        /*paranoid*/ verify(proc == __atomic_load_n(&data[id].handle, __ATOMIC_RELAXED));
142        __atomic_store_n(&data[id].handle, 0p, __ATOMIC_RELEASE);
143
144        __cfadbg_print_safe(ready_queue, "Kernel : Unregister proc %p\n", proc);
145}
146
147//-----------------------------------------------------------------------
148// Writer side : acquire when changing the ready queue, e.g. adding more
149//  queues or removing them.
150uint_fast32_t ready_mutate_lock( void ) with(*__scheduler_lock) {
151        /* paranoid */ verify( ! __preemption_enabled() );
152
153        // Step 1 : lock global lock
154        // It is needed to avoid processors that register mid Critical-Section
155        //   to simply lock their own lock and enter.
156        __atomic_acquire( &lock );
157
158        // Step 2 : lock per-proc lock
159        // Processors that are currently being registered aren't counted
160        //   but can't be in read_lock or in the critical section.
161        // All other processors are counted
162        uint_fast32_t s = ready;
163        for(uint_fast32_t i = 0; i < s; i++) {
164                __atomic_acquire( &data[i].lock );
165        }
166
167        /* paranoid */ verify( ! __preemption_enabled() );
168        return s;
169}
170
171void ready_mutate_unlock( uint_fast32_t last_s ) with(*__scheduler_lock) {
172        /* paranoid */ verify( ! __preemption_enabled() );
173
174        // Step 1 : release local locks
175        // This must be done while the global lock is held to avoid
176        //   threads that where created mid critical section
177        //   to race to lock their local locks and have the writer
178        //   immidiately unlock them
179        // Alternative solution : return s in write_lock and pass it to write_unlock
180        for(uint_fast32_t i = 0; i < last_s; i++) {
181                verify(data[i].lock);
182                __atomic_store_n(&data[i].lock, (bool)false, __ATOMIC_RELEASE);
183        }
184
185        // Step 2 : release global lock
186        /*paranoid*/ assert(true == lock);
187        __atomic_store_n(&lock, (bool)false, __ATOMIC_RELEASE);
188
189        /* paranoid */ verify( ! __preemption_enabled() );
190}
191
192//=======================================================================
193// Cforall Reqdy Queue used for scheduling
194//=======================================================================
195void ?{}(__ready_queue_t & this) with (this) {
196        lanes.data  = 0p;
197        lanes.count = 0;
198}
199
200void ^?{}(__ready_queue_t & this) with (this) {
201        verify( 1 == lanes.count );
202        free(lanes.data);
203}
204
205//-----------------------------------------------------------------------
206static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred) {
207        unsigned i;
208        bool local;
209        #if defined(BIAS)
210                unsigned rlow  = r % BIAS;
211                unsigned rhigh = r / BIAS;
212                if((0 != rlow) && preferred >= 0) {
213                        // (BIAS - 1) out of BIAS chances
214                        // Use perferred queues
215                        i = preferred + (rhigh % 4);
216                        local = true;
217                }
218                else {
219                        // 1 out of BIAS chances
220                        // Use all queues
221                        i = rhigh;
222                        local = false;
223                }
224        #else
225                i = r;
226                local = false;
227        #endif
228        return [i, local];
229}
230
231//-----------------------------------------------------------------------
232__attribute__((hot)) bool push(struct cluster * cltr, struct $thread * thrd) with (cltr->ready_queue) {
233        __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
234
235        const bool external = (!kernelTLS().this_processor) || (cltr != kernelTLS().this_processor->cltr);
236
237        // write timestamp
238        thrd->link.ts = rdtscl();
239
240        bool first = false;
241        __attribute__((unused)) bool local;
242        __attribute__((unused)) int preferred;
243        #if defined(BIAS)
244                /* paranoid */ verify(external || kernelTLS().this_processor->cltr_id < lanes.count );
245                preferred =
246                        //*
247                        external ? -1 : kernelTLS().this_processor->cltr_id;
248                        /*/
249                        thrd->link.preferred * 4;
250                        //*/
251        #endif
252
253        // Try to pick a lane and lock it
254        unsigned i;
255        do {
256                // Pick the index of a lane
257                // unsigned r = __tls_rand();
258                unsigned r = __tls_rand_fwd();
259                [i, local] = idx_from_r(r, preferred);
260
261                i %= __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
262
263                #if !defined(__CFA_NO_STATISTICS__)
264                        if(external) {
265                                if(local) __atomic_fetch_add(&cltr->stats->ready.pick.ext.local, 1, __ATOMIC_RELAXED);
266                                __atomic_fetch_add(&cltr->stats->ready.pick.ext.attempt, 1, __ATOMIC_RELAXED);
267                        }
268                        else {
269                                if(local) __tls_stats()->ready.pick.push.local++;
270                                __tls_stats()->ready.pick.push.attempt++;
271                        }
272                #endif
273
274        #if defined(USE_MPSC)
275                // mpsc always succeeds
276        } while( false );
277        #else
278                // If we can't lock it retry
279        } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
280        #endif
281
282        // Actually push it
283        push(lanes.data[i], thrd);
284
285        #if !defined(USE_MPSC)
286                // Unlock and return
287                __atomic_unlock( &lanes.data[i].lock );
288        #endif
289
290        // Mark the current index in the tls rng instance as having an item
291        __tls_rand_advance_bck();
292
293        __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);
294
295        // Update statistics
296        #if !defined(__CFA_NO_STATISTICS__)
297                if(external) {
298                        if(local) __atomic_fetch_add(&cltr->stats->ready.pick.ext.lsuccess, 1, __ATOMIC_RELAXED);
299                        __atomic_fetch_add(&cltr->stats->ready.pick.ext.success, 1, __ATOMIC_RELAXED);
300                }
301                else {
302                        if(local) __tls_stats()->ready.pick.push.lsuccess++;
303                        __tls_stats()->ready.pick.push.success++;
304                }
305        #endif
306
307        // return whether or not the list was empty before this push
308        return first;
309}
310
311static struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j);
312static struct $thread * try_pop(struct cluster * cltr, unsigned i);
313
314// Pop from the ready queue from a given cluster
315__attribute__((hot)) $thread * pop(struct cluster * cltr) with (cltr->ready_queue) {
316        /* paranoid */ verify( lanes.count > 0 );
317        unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
318        int preferred;
319        #if defined(BIAS)
320                /* paranoid */ verify(kernelTLS().this_processor->cltr_id < lanes.count );
321                preferred = kernelTLS().this_processor->cltr_id;
322        #endif
323
324
325        // As long as the list is not empty, try finding a lane that isn't empty and pop from it
326        for(25) {
327                // Pick two lists at random
328                // unsigned ri = __tls_rand();
329                // unsigned rj = __tls_rand();
330                unsigned ri = __tls_rand_bck();
331                unsigned rj = __tls_rand_bck();
332
333                unsigned i, j;
334                __attribute__((unused)) bool locali, localj;
335                [i, locali] = idx_from_r(ri, preferred);
336                [j, localj] = idx_from_r(rj, preferred);
337
338                #if !defined(__CFA_NO_STATISTICS__)
339                        if(locali && localj) {
340                                __tls_stats()->ready.pick.pop.local++;
341                        }
342                #endif
343
344                i %= count;
345                j %= count;
346
347                // try popping from the 2 picked lists
348                struct $thread * thrd = try_pop(cltr, i, j);
349                if(thrd) {
350                        #if defined(BIAS) && !defined(__CFA_NO_STATISTICS__)
351                                if( locali || localj ) __tls_stats()->ready.pick.pop.lsuccess++;
352                        #endif
353                        return thrd;
354                }
355        }
356
357        // All lanes where empty return 0p
358        return 0p;
359}
360
361__attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
362        /* paranoid */ verify( lanes.count > 0 );
363        unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
364        unsigned offset = __tls_rand();
365        for(i; count) {
366                unsigned idx = (offset + i) % count;
367                struct $thread * thrd = try_pop(cltr, idx);
368                if(thrd) {
369                        return thrd;
370                }
371        }
372
373        // All lanes where empty return 0p
374        return 0p;
375}
376
377
378//-----------------------------------------------------------------------
379// Given 2 indexes, pick the list with the oldest push an try to pop from it
380static inline struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j) with (cltr->ready_queue) {
381        #if !defined(__CFA_NO_STATISTICS__)
382                __tls_stats()->ready.pick.pop.attempt++;
383        #endif
384
385        // Pick the bet list
386        int w = i;
387        if( __builtin_expect(!is_empty(lanes.data[j]), true) ) {
388                w = (ts(lanes.data[i]) < ts(lanes.data[j])) ? i : j;
389        }
390
391        return try_pop(cltr, w);
392}
393
394static inline struct $thread * try_pop(struct cluster * cltr, unsigned w) with (cltr->ready_queue) {
395        // Get relevant elements locally
396        __intrusive_lane_t & lane = lanes.data[w];
397
398        // If list looks empty retry
399        if( is_empty(lane) ) return 0p;
400
401        // If we can't get the lock retry
402        if( !__atomic_try_acquire(&lane.lock) ) return 0p;
403
404
405        // If list is empty, unlock and retry
406        if( is_empty(lane) ) {
407                __atomic_unlock(&lane.lock);
408                return 0p;
409        }
410
411        // Actually pop the list
412        struct $thread * thrd;
413        thrd = pop(lane);
414
415        /* paranoid */ verify(thrd);
416        /* paranoid */ verify(lane.lock);
417
418        // Unlock and return
419        __atomic_unlock(&lane.lock);
420
421        // Update statistics
422        #if !defined(__CFA_NO_STATISTICS__)
423                __tls_stats()->ready.pick.pop.success++;
424        #endif
425
426        // Update the thread bias
427        thrd->link.preferred = w / 4;
428
429        // return the popped thread
430        return thrd;
431}
432//-----------------------------------------------------------------------
433
434bool remove_head(struct cluster * cltr, struct $thread * thrd) with (cltr->ready_queue) {
435        for(i; lanes.count) {
436                __intrusive_lane_t & lane = lanes.data[i];
437
438                bool removed = false;
439
440                __atomic_acquire(&lane.lock);
441                        if(head(lane)->link.next == thrd) {
442                                $thread * pthrd;
443                                pthrd = pop(lane);
444
445                                /* paranoid */ verify( pthrd == thrd );
446
447                                removed = true;
448                        }
449                __atomic_unlock(&lane.lock);
450
451                if( removed ) return true;
452        }
453        return false;
454}
455
456//-----------------------------------------------------------------------
457
458static void check( __ready_queue_t & q ) with (q) {
459        #if defined(__CFA_WITH_VERIFY__) && !defined(USE_MPSC)
460                {
461                        for( idx ; lanes.count ) {
462                                __intrusive_lane_t & sl = lanes.data[idx];
463                                assert(!lanes.data[idx].lock);
464
465                                assert(head(sl)->link.prev == 0p );
466                                assert(head(sl)->link.next->link.prev == head(sl) );
467                                assert(tail(sl)->link.next == 0p );
468                                assert(tail(sl)->link.prev->link.next == tail(sl) );
469
470                                if(is_empty(sl)) {
471                                        assert(tail(sl)->link.prev == head(sl));
472                                        assert(head(sl)->link.next == tail(sl));
473                                } else {
474                                        assert(tail(sl)->link.prev != head(sl));
475                                        assert(head(sl)->link.next != tail(sl));
476                                }
477                        }
478                }
479        #endif
480}
481
482// Call this function of the intrusive list was moved using memcpy
483// fixes the list so that the pointers back to anchors aren't left dangling
484static inline void fix(__intrusive_lane_t & ll) {
485        #if !defined(USE_MPSC)
486                // if the list is not empty then follow he pointer and fix its reverse
487                if(!is_empty(ll)) {
488                        head(ll)->link.next->link.prev = head(ll);
489                        tail(ll)->link.prev->link.next = tail(ll);
490                }
491                // Otherwise just reset the list
492                else {
493                        verify(tail(ll)->link.next == 0p);
494                        tail(ll)->link.prev = head(ll);
495                        head(ll)->link.next = tail(ll);
496                        verify(head(ll)->link.prev == 0p);
497                }
498        #endif
499}
500
501static void assign_list(unsigned & value, const int inc, dlist(processor, processor) & list, unsigned count) {
502        processor * it = &list`first;
503        for(unsigned i = 0; i < count; i++) {
504                /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
505                it->cltr_id = value;
506                value += inc;
507                it = &(*it)`next;
508        }
509}
510
511static void reassign_cltr_id(struct cluster * cltr, const int inc) {
512        unsigned preferred = 0;
513        assign_list(preferred, inc, cltr->procs.actives, cltr->procs.total - cltr->procs.idle);
514        assign_list(preferred, inc, cltr->procs.idles  , cltr->procs.idle );
515}
516
517// Grow the ready queue
518void ready_queue_grow(struct cluster * cltr) {
519        size_t ncount;
520        int target = cltr->procs.total;
521
522        /* paranoid */ verify( ready_mutate_islocked() );
523        __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue\n");
524
525        // Make sure that everything is consistent
526        /* paranoid */ check( cltr->ready_queue );
527
528        // grow the ready queue
529        with( cltr->ready_queue ) {
530                // Find new count
531                // Make sure we always have atleast 1 list
532                if(target >= 2) {
533                        ncount = target * 4;
534                } else {
535                        ncount = 1;
536                }
537
538                // Allocate new array (uses realloc and memcpies the data)
539                lanes.data = alloc( ncount, lanes.data`realloc );
540
541                // Fix the moved data
542                for( idx; (size_t)lanes.count ) {
543                        fix(lanes.data[idx]);
544                }
545
546                // Construct new data
547                for( idx; (size_t)lanes.count ~ ncount) {
548                        (lanes.data[idx]){};
549                }
550
551                // Update original
552                lanes.count = ncount;
553        }
554
555        reassign_cltr_id(cltr, 4);
556
557        // Make sure that everything is consistent
558        /* paranoid */ check( cltr->ready_queue );
559
560        __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue done\n");
561
562        /* paranoid */ verify( ready_mutate_islocked() );
563}
564
565// Shrink the ready queue
566void ready_queue_shrink(struct cluster * cltr) {
567        /* paranoid */ verify( ready_mutate_islocked() );
568        __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue\n");
569
570        // Make sure that everything is consistent
571        /* paranoid */ check( cltr->ready_queue );
572
573        int target = cltr->procs.total;
574
575        with( cltr->ready_queue ) {
576                // Remember old count
577                size_t ocount = lanes.count;
578
579                // Find new count
580                // Make sure we always have atleast 1 list
581                lanes.count = target >= 2 ? target * 4: 1;
582                /* paranoid */ verify( ocount >= lanes.count );
583                /* paranoid */ verify( lanes.count == target * 4 || target < 2 );
584
585                // for printing count the number of displaced threads
586                #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
587                        __attribute__((unused)) size_t displaced = 0;
588                #endif
589
590                // redistribute old data
591                for( idx; (size_t)lanes.count ~ ocount) {
592                        // Lock is not strictly needed but makes checking invariants much easier
593                        __attribute__((unused)) bool locked = __atomic_try_acquire(&lanes.data[idx].lock);
594                        verify(locked);
595
596                        // As long as we can pop from this lane to push the threads somewhere else in the queue
597                        while(!is_empty(lanes.data[idx])) {
598                                struct $thread * thrd;
599                                thrd = pop(lanes.data[idx]);
600
601                                push(cltr, thrd);
602
603                                // for printing count the number of displaced threads
604                                #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
605                                        displaced++;
606                                #endif
607                        }
608
609                        // Unlock the lane
610                        __atomic_unlock(&lanes.data[idx].lock);
611
612                        // TODO print the queue statistics here
613
614                        ^(lanes.data[idx]){};
615                }
616
617                __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue displaced %zu threads\n", displaced);
618
619                // Allocate new array (uses realloc and memcpies the data)
620                lanes.data = alloc( lanes.count, lanes.data`realloc );
621
622                // Fix the moved data
623                for( idx; (size_t)lanes.count ) {
624                        fix(lanes.data[idx]);
625                }
626        }
627
628        reassign_cltr_id(cltr, 4);
629
630        // Make sure that everything is consistent
631        /* paranoid */ check( cltr->ready_queue );
632
633        __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue done\n");
634        /* paranoid */ verify( ready_mutate_islocked() );
635}
Note: See TracBrowser for help on using the repository browser.