source: libcfa/src/concurrency/ready_queue.cfa @ 43784ac

jacob/cs343-translationnew-ast-unique-expr
Last change on this file since 43784ac was 43784ac, checked in by Thierry Delisle <tdelisle@…>, 4 months ago

Changed libcfathread to consistently define _GNU_SOURCE

  • Property mode set to 100644
File size: 22.3 KB
RevLine 
[7768b8d]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__
[43784ac]17#define _GNU_SOURCE
18
[1b143de]19// #define __CFA_DEBUG_PRINT_READY_QUEUE__
[7768b8d]20
[1eb239e4]21
[9cc3a18]22#define USE_RELAXED_FIFO
23// #define USE_WORK_STEALING
24
[7768b8d]25#include "bits/defs.hfa"
26#include "kernel_private.hfa"
27
28#include "stdlib.hfa"
[61d7bec]29#include "math.hfa"
[7768b8d]30
[04b5cef]31#include <unistd.h>
32
[13c5e19]33#include "ready_subqueue.hfa"
34
[7768b8d]35static const size_t cache_line_size = 64;
36
[d2fadeb]37#if !defined(__CFA_NO_STATISTICS__)
38        #define __STATS(...) __VA_ARGS__
39#else
40        #define __STATS(...)
41#endif
42
[dca5802]43// No overriden function, no environment variable, no define
44// fall back to a magic number
45#ifndef __CFA_MAX_PROCESSORS__
[b388ee8]46        #define __CFA_MAX_PROCESSORS__ 1024
[dca5802]47#endif
[7768b8d]48
[9cc3a18]49#if   defined(USE_RELAXED_FIFO)
50        #define BIAS 4
51        #define READYQ_SHARD_FACTOR 4
[5f6a172]52        #define SEQUENTIAL_SHARD 1
[9cc3a18]53#elif defined(USE_WORK_STEALING)
54        #define READYQ_SHARD_FACTOR 2
[5f6a172]55        #define SEQUENTIAL_SHARD 2
[9cc3a18]56#else
57        #error no scheduling strategy selected
58#endif
59
[d2fadeb]60static inline struct $thread * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats));
61static inline struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats));
[431cd4f]62static inline struct $thread * search(struct cluster * cltr);
[d2fadeb]63static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred);
[9cc3a18]64
[04b5cef]65
[dca5802]66// returns the maximum number of processors the RWLock support
[7768b8d]67__attribute__((weak)) unsigned __max_processors() {
68        const char * max_cores_s = getenv("CFA_MAX_PROCESSORS");
69        if(!max_cores_s) {
[504a7dc]70                __cfadbg_print_nolock(ready_queue, "No CFA_MAX_PROCESSORS in ENV\n");
[dca5802]71                return __CFA_MAX_PROCESSORS__;
[7768b8d]72        }
73
74        char * endptr = 0p;
75        long int max_cores_l = strtol(max_cores_s, &endptr, 10);
76        if(max_cores_l < 1 || max_cores_l > 65535) {
[504a7dc]77                __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS out of range : %ld\n", max_cores_l);
[dca5802]78                return __CFA_MAX_PROCESSORS__;
[7768b8d]79        }
80        if('\0' != *endptr) {
[504a7dc]81                __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS not a decimal number : %s\n", max_cores_s);
[dca5802]82                return __CFA_MAX_PROCESSORS__;
[7768b8d]83        }
84
85        return max_cores_l;
86}
87
88//=======================================================================
89// Cluster wide reader-writer lock
90//=======================================================================
[b388ee8]91void  ?{}(__scheduler_RWLock_t & this) {
[7768b8d]92        this.max   = __max_processors();
93        this.alloc = 0;
94        this.ready = 0;
95        this.data  = alloc(this.max);
[c993b15]96        this.write_lock  = false;
[7768b8d]97
98        /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.alloc), &this.alloc));
99        /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.ready), &this.ready));
100
101}
[b388ee8]102void ^?{}(__scheduler_RWLock_t & this) {
[7768b8d]103        free(this.data);
104}
105
106
107//=======================================================================
108// Lock-Free registering/unregistering of threads
[c993b15]109unsigned register_proc_id( void ) with(*__scheduler_lock) {
[b388ee8]110        __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p for RW-Lock\n", proc);
[c993b15]111        bool * handle = (bool *)&kernelTLS().sched_lock;
[504a7dc]112
[7768b8d]113        // Step - 1 : check if there is already space in the data
114        uint_fast32_t s = ready;
115
116        // Check among all the ready
117        for(uint_fast32_t i = 0; i < s; i++) {
[c993b15]118                bool * volatile * cell = (bool * volatile *)&data[i]; // Cforall is bugged and the double volatiles causes problems
119                /* paranoid */ verify( handle != *cell );
120
121                bool * null = 0p; // Re-write every loop since compare thrashes it
122                if( __atomic_load_n(cell, (int)__ATOMIC_RELAXED) == null
123                        && __atomic_compare_exchange_n( cell, &null, handle, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
124                        /* paranoid */ verify(i < ready);
125                        /* paranoid */ verify( (kernelTLS().sched_id = i, true) );
126                        return i;
[7768b8d]127                }
128        }
129
[b388ee8]130        if(max <= alloc) abort("Trying to create more than %ud processors", __scheduler_lock->max);
[7768b8d]131
132        // Step - 2 : F&A to get a new spot in the array.
133        uint_fast32_t n = __atomic_fetch_add(&alloc, 1, __ATOMIC_SEQ_CST);
[b388ee8]134        if(max <= n) abort("Trying to create more than %ud processors", __scheduler_lock->max);
[7768b8d]135
136        // Step - 3 : Mark space as used and then publish it.
[c993b15]137        data[n] = handle;
[fd9b524]138        while() {
[7768b8d]139                unsigned copy = n;
140                if( __atomic_load_n(&ready, __ATOMIC_RELAXED) == n
141                        && __atomic_compare_exchange_n(&ready, &copy, n + 1, true, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))
142                        break;
[fd9b524]143                Pause();
[7768b8d]144        }
145
[1b143de]146        __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p done, id %lu\n", proc, n);
[504a7dc]147
[7768b8d]148        // Return new spot.
[c993b15]149        /* paranoid */ verify(n < ready);
150        /* paranoid */ verify( (kernelTLS().sched_id = n, true) );
151        return n;
[7768b8d]152}
153
[c993b15]154void unregister_proc_id( unsigned id ) with(*__scheduler_lock) {
155        /* paranoid */ verify(id < ready);
156        /* paranoid */ verify(id == kernelTLS().sched_id);
157        /* paranoid */ verify(data[id] == &kernelTLS().sched_lock);
158
159        bool * volatile * cell = (bool * volatile *)&data[id]; // Cforall is bugged and the double volatiles causes problems
160
161        __atomic_store_n(cell, 0p, __ATOMIC_RELEASE);
[504a7dc]162
163        __cfadbg_print_safe(ready_queue, "Kernel : Unregister proc %p\n", proc);
[7768b8d]164}
165
166//-----------------------------------------------------------------------
167// Writer side : acquire when changing the ready queue, e.g. adding more
168//  queues or removing them.
[b388ee8]169uint_fast32_t ready_mutate_lock( void ) with(*__scheduler_lock) {
[8fc652e0]170        /* paranoid */ verify( ! __preemption_enabled() );
[c993b15]171        /* paranoid */ verify( ! kernelTLS().sched_lock );
[62502cc4]172
[7768b8d]173        // Step 1 : lock global lock
174        // It is needed to avoid processors that register mid Critical-Section
175        //   to simply lock their own lock and enter.
[c993b15]176        __atomic_acquire( &write_lock );
[7768b8d]177
178        // Step 2 : lock per-proc lock
179        // Processors that are currently being registered aren't counted
180        //   but can't be in read_lock or in the critical section.
181        // All other processors are counted
182        uint_fast32_t s = ready;
183        for(uint_fast32_t i = 0; i < s; i++) {
[c993b15]184                volatile bool * llock = data[i];
185                if(llock) __atomic_acquire( llock );
[7768b8d]186        }
187
[8fc652e0]188        /* paranoid */ verify( ! __preemption_enabled() );
[7768b8d]189        return s;
190}
191
[b388ee8]192void ready_mutate_unlock( uint_fast32_t last_s ) with(*__scheduler_lock) {
[8fc652e0]193        /* paranoid */ verify( ! __preemption_enabled() );
[62502cc4]194
[7768b8d]195        // Step 1 : release local locks
196        // This must be done while the global lock is held to avoid
197        //   threads that where created mid critical section
198        //   to race to lock their local locks and have the writer
199        //   immidiately unlock them
200        // Alternative solution : return s in write_lock and pass it to write_unlock
201        for(uint_fast32_t i = 0; i < last_s; i++) {
[c993b15]202                volatile bool * llock = data[i];
203                if(llock) __atomic_store_n(llock, (bool)false, __ATOMIC_RELEASE);
[7768b8d]204        }
205
206        // Step 2 : release global lock
[c993b15]207        /*paranoid*/ assert(true == write_lock);
208        __atomic_store_n(&write_lock, (bool)false, __ATOMIC_RELEASE);
[62502cc4]209
[8fc652e0]210        /* paranoid */ verify( ! __preemption_enabled() );
[7768b8d]211}
212
213//=======================================================================
[9cc3a18]214// Cforall Ready Queue used for scheduling
[b798713]215//=======================================================================
216void ?{}(__ready_queue_t & this) with (this) {
[28d73c1]217        lanes.data  = 0p;
[9cc3a18]218        lanes.tscs  = 0p;
[28d73c1]219        lanes.count = 0;
[b798713]220}
221
222void ^?{}(__ready_queue_t & this) with (this) {
[5f6a172]223        verify( SEQUENTIAL_SHARD == lanes.count );
[dca5802]224        free(lanes.data);
[9cc3a18]225        free(lanes.tscs);
[dca5802]226}
227
[64a7146]228//-----------------------------------------------------------------------
[431cd4f]229#if defined(USE_RELAXED_FIFO)
230        //-----------------------------------------------------------------------
231        // get index from random number with or without bias towards queues
232        static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred) {
233                unsigned i;
234                bool local;
235                unsigned rlow  = r % BIAS;
236                unsigned rhigh = r / BIAS;
237                if((0 != rlow) && preferred >= 0) {
238                        // (BIAS - 1) out of BIAS chances
239                        // Use perferred queues
240                        i = preferred + (rhigh % READYQ_SHARD_FACTOR);
241                        local = true;
242                }
243                else {
244                        // 1 out of BIAS chances
245                        // Use all queues
246                        i = rhigh;
247                        local = false;
248                }
249                return [i, local];
250        }
251
[b808625]252        __attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd, bool push_local) with (cltr->ready_queue) {
[431cd4f]253                __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
[1b143de]254
[b808625]255                const bool external = !push_local || (!kernelTLS().this_processor) || (cltr != kernelTLS().this_processor->cltr);
[431cd4f]256                /* paranoid */ verify(external || kernelTLS().this_processor->rdq.id < lanes.count );
[fd1f65e]257
[431cd4f]258                bool local;
259                int preferred = external ? -1 : kernelTLS().this_processor->rdq.id;
[52769ba]260
[431cd4f]261                // Try to pick a lane and lock it
262                unsigned i;
263                do {
264                        // Pick the index of a lane
265                        unsigned r = __tls_rand_fwd();
266                        [i, local] = idx_from_r(r, preferred);
[772411a]267
[431cd4f]268                        i %= __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
269
270                        #if !defined(__CFA_NO_STATISTICS__)
[d2fadeb]271                                if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
272                                else if(local) __tls_stats()->ready.push.local.attempt++;
273                                else __tls_stats()->ready.push.share.attempt++;
[431cd4f]274                        #endif
[b798713]275
[431cd4f]276                        // If we can't lock it retry
277                } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
278
279                // Actually push it
280                push(lanes.data[i], thrd);
281
[b808625]282                // Unlock and return
283                __atomic_unlock( &lanes.data[i].lock );
[431cd4f]284
285                // Mark the current index in the tls rng instance as having an item
286                __tls_rand_advance_bck();
287
288                __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);
289
290                // Update statistics
[b798713]291                #if !defined(__CFA_NO_STATISTICS__)
[d2fadeb]292                        if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
293                        else if(local) __tls_stats()->ready.push.local.success++;
294                        else __tls_stats()->ready.push.share.success++;
[b798713]295                #endif
[431cd4f]296        }
[b798713]297
[431cd4f]298        // Pop from the ready queue from a given cluster
299        __attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
300                /* paranoid */ verify( lanes.count > 0 );
301                /* paranoid */ verify( kernelTLS().this_processor );
302                /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
[b798713]303
[431cd4f]304                unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
305                int preferred = kernelTLS().this_processor->rdq.id;
[dca5802]306
307
[431cd4f]308                // As long as the list is not empty, try finding a lane that isn't empty and pop from it
309                for(25) {
310                        // Pick two lists at random
311                        unsigned ri = __tls_rand_bck();
312                        unsigned rj = __tls_rand_bck();
[c426b03]313
[431cd4f]314                        unsigned i, j;
315                        __attribute__((unused)) bool locali, localj;
316                        [i, locali] = idx_from_r(ri, preferred);
317                        [j, localj] = idx_from_r(rj, preferred);
[1b143de]318
[431cd4f]319                        i %= count;
320                        j %= count;
[9cc3a18]321
[431cd4f]322                        // try popping from the 2 picked lists
[d2fadeb]323                        struct $thread * thrd = try_pop(cltr, i, j __STATS(, *(locali || localj ? &__tls_stats()->ready.pop.local : &__tls_stats()->ready.pop.help)));
[431cd4f]324                        if(thrd) {
325                                return thrd;
326                        }
327                }
[13c5e19]328
[431cd4f]329                // All lanes where empty return 0p
330                return 0p;
331        }
[772411a]332
[fc59df78]333        __attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) { return pop_fast(cltr); }
334        __attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) {
[431cd4f]335                return search(cltr);
336        }
337#endif
338#if defined(USE_WORK_STEALING)
[b808625]339        __attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd, bool push_local) with (cltr->ready_queue) {
[431cd4f]340                __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
[772411a]341
[d3ba775]342                // #define USE_PREFERRED
343                #if !defined(USE_PREFERRED)
[b808625]344                const bool external = !push_local || (!kernelTLS().this_processor) || (cltr != kernelTLS().this_processor->cltr);
[431cd4f]345                /* paranoid */ verify(external || kernelTLS().this_processor->rdq.id < lanes.count );
[d3ba775]346                #else
347                        unsigned preferred = thrd->preferred;
[b808625]348                        const bool external = push_local || (!kernelTLS().this_processor) || preferred == -1u || thrd->curr_cluster != cltr;
[d3ba775]349                        /* paranoid */ verifyf(external || preferred < lanes.count, "Invalid preferred queue %u for %u lanes", preferred, lanes.count );
[772411a]350
[d3ba775]351                        unsigned r = preferred % READYQ_SHARD_FACTOR;
352                        const unsigned start = preferred - r;
[2b96031]353                #endif
[431cd4f]354
355                // Try to pick a lane and lock it
356                unsigned i;
357                do {
[d2fadeb]358                        #if !defined(__CFA_NO_STATISTICS__)
359                                if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
360                                else __tls_stats()->ready.push.local.attempt++;
361                        #endif
362
[431cd4f]363                        if(unlikely(external)) {
364                                i = __tls_rand() % lanes.count;
365                        }
366                        else {
[d3ba775]367                                #if !defined(USE_PREFERRED)
[b808625]368                                        processor * proc = kernelTLS().this_processor;
369                                        unsigned r = proc->rdq.its++;
370                                        i =  proc->rdq.id + (r % READYQ_SHARD_FACTOR);
371                                #else
[d3ba775]372                                        i = start + (r++ % READYQ_SHARD_FACTOR);
373                                #endif
374                        }
[431cd4f]375                        // If we can't lock it retry
376                } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
[13c5e19]377
[431cd4f]378                // Actually push it
379                push(lanes.data[i], thrd);
[13c5e19]380
[b808625]381                // Unlock and return
382                __atomic_unlock( &lanes.data[i].lock );
[431cd4f]383
[d2fadeb]384                #if !defined(__CFA_NO_STATISTICS__)
385                        if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
386                        else __tls_stats()->ready.push.local.success++;
387                #endif
388
[431cd4f]389                __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);
[13c5e19]390        }
391
[431cd4f]392        // Pop from the ready queue from a given cluster
393        __attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
394                /* paranoid */ verify( lanes.count > 0 );
395                /* paranoid */ verify( kernelTLS().this_processor );
396                /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
397
398                processor * proc = kernelTLS().this_processor;
399
400                if(proc->rdq.target == -1u) {
[1680072]401                        unsigned long long min = ts(lanes.data[proc->rdq.id]);
402                        for(int i = 0; i < READYQ_SHARD_FACTOR; i++) {
403                                unsigned long long tsc = ts(lanes.data[proc->rdq.id + i]);
404                                if(tsc < min) min = tsc;
405                        }
406                        proc->rdq.cutoff = min;
[f55d54d]407                        proc->rdq.target = __tls_rand() % lanes.count;
[431cd4f]408                }
[341aa39]409                else {
410                        unsigned target = proc->rdq.target;
[431cd4f]411                        proc->rdq.target = -1u;
[9cac0da]412                        const unsigned long long bias = 0; //2_500_000_000;
413                        const unsigned long long cutoff = proc->rdq.cutoff > bias ? proc->rdq.cutoff - bias : proc->rdq.cutoff;
414                        if(lanes.tscs[target].tv < cutoff && ts(lanes.data[target]) < cutoff) {
[341aa39]415                                $thread * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
416                                if(t) return t;
417                        }
[431cd4f]418                }
[13c5e19]419
[431cd4f]420                for(READYQ_SHARD_FACTOR) {
[f55d54d]421                        unsigned i = proc->rdq.id + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
[d2fadeb]422                        if($thread * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
[431cd4f]423                }
424                return 0p;
[1eb239e4]425        }
426
[431cd4f]427        __attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
[fc59df78]428                unsigned i = __tls_rand() % lanes.count;
429                return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
430        }
[431cd4f]431
[fc59df78]432        __attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) with (cltr->ready_queue) {
[431cd4f]433                return search(cltr);
434        }
435#endif
[1eb239e4]436
[9cc3a18]437//=======================================================================
438// Various Ready Queue utilities
439//=======================================================================
440// these function work the same or almost the same
441// whether they are using work-stealing or relaxed fifo scheduling
[1eb239e4]442
[9cc3a18]443//-----------------------------------------------------------------------
444// try to pop from a lane given by index w
[d2fadeb]445static inline struct $thread * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
446        __STATS( stats.attempt++; )
447
[dca5802]448        // Get relevant elements locally
449        __intrusive_lane_t & lane = lanes.data[w];
450
[b798713]451        // If list looks empty retry
[d2fadeb]452        if( is_empty(lane) ) {
453                return 0p;
454        }
[b798713]455
456        // If we can't get the lock retry
[d2fadeb]457        if( !__atomic_try_acquire(&lane.lock) ) {
458                return 0p;
459        }
[b798713]460
461        // If list is empty, unlock and retry
[dca5802]462        if( is_empty(lane) ) {
463                __atomic_unlock(&lane.lock);
[b798713]464                return 0p;
465        }
466
467        // Actually pop the list
[504a7dc]468        struct $thread * thrd;
[f302d80]469        unsigned long long tsv;
470        [thrd, tsv] = pop(lane);
[b798713]471
[dca5802]472        /* paranoid */ verify(thrd);
[78ea291]473        /* paranoid */ verify(tsv);
[dca5802]474        /* paranoid */ verify(lane.lock);
[b798713]475
476        // Unlock and return
[dca5802]477        __atomic_unlock(&lane.lock);
[b798713]478
[dca5802]479        // Update statistics
[d2fadeb]480        __STATS( stats.success++; )
[b798713]481
[431cd4f]482        #if defined(USE_WORK_STEALING)
[f302d80]483                lanes.tscs[w].tv = tsv;
[9cc3a18]484        #endif
[d72c074]485
[d3ba775]486        thrd->preferred = w;
487
[dca5802]488        // return the popped thread
[b798713]489        return thrd;
490}
[04b5cef]491
[9cc3a18]492//-----------------------------------------------------------------------
493// try to pop from any lanes making sure you don't miss any threads push
494// before the start of the function
[431cd4f]495static inline struct $thread * search(struct cluster * cltr) with (cltr->ready_queue) {
[9cc3a18]496        /* paranoid */ verify( lanes.count > 0 );
497        unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
498        unsigned offset = __tls_rand();
499        for(i; count) {
500                unsigned idx = (offset + i) % count;
[d2fadeb]501                struct $thread * thrd = try_pop(cltr, idx __STATS(, __tls_stats()->ready.pop.search));
[9cc3a18]502                if(thrd) {
503                        return thrd;
504                }
[13c5e19]505        }
[9cc3a18]506
507        // All lanes where empty return 0p
508        return 0p;
[b798713]509}
510
511//-----------------------------------------------------------------------
[9cc3a18]512// Check that all the intrusive queues in the data structure are still consistent
[b798713]513static void check( __ready_queue_t & q ) with (q) {
[d3ba775]514        #if defined(__CFA_WITH_VERIFY__)
[b798713]515                {
[dca5802]516                        for( idx ; lanes.count ) {
517                                __intrusive_lane_t & sl = lanes.data[idx];
518                                assert(!lanes.data[idx].lock);
[b798713]519
[2b96031]520                                        if(is_empty(sl)) {
521                                                assert( sl.anchor.next == 0p );
522                                                assert( sl.anchor.ts   == 0  );
523                                                assert( mock_head(sl)  == sl.prev );
524                                        } else {
525                                                assert( sl.anchor.next != 0p );
526                                                assert( sl.anchor.ts   != 0  );
527                                                assert( mock_head(sl)  != sl.prev );
528                                        }
[b798713]529                        }
530                }
531        #endif
532}
533
[9cc3a18]534//-----------------------------------------------------------------------
535// Given 2 indexes, pick the list with the oldest push an try to pop from it
[d2fadeb]536static inline struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
[9cc3a18]537        // Pick the bet list
538        int w = i;
539        if( __builtin_expect(!is_empty(lanes.data[j]), true) ) {
540                w = (ts(lanes.data[i]) < ts(lanes.data[j])) ? i : j;
541        }
542
[d2fadeb]543        return try_pop(cltr, w __STATS(, stats));
[9cc3a18]544}
545
[b798713]546// Call this function of the intrusive list was moved using memcpy
[dca5802]547// fixes the list so that the pointers back to anchors aren't left dangling
548static inline void fix(__intrusive_lane_t & ll) {
[2b96031]549                        if(is_empty(ll)) {
550                                verify(ll.anchor.next == 0p);
551                                ll.prev = mock_head(ll);
552                        }
[b798713]553}
554
[69914cbc]555static void assign_list(unsigned & value, dlist(processor) & list, unsigned count) {
[a017ee7]556        processor * it = &list`first;
557        for(unsigned i = 0; i < count; i++) {
558                /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
[431cd4f]559                it->rdq.id = value;
560                it->rdq.target = -1u;
[9cc3a18]561                value += READYQ_SHARD_FACTOR;
[a017ee7]562                it = &(*it)`next;
563        }
564}
565
[9cc3a18]566static void reassign_cltr_id(struct cluster * cltr) {
[a017ee7]567        unsigned preferred = 0;
[9cc3a18]568        assign_list(preferred, cltr->procs.actives, cltr->procs.total - cltr->procs.idle);
569        assign_list(preferred, cltr->procs.idles  , cltr->procs.idle );
[a017ee7]570}
571
[431cd4f]572static void fix_times( struct cluster * cltr ) with( cltr->ready_queue ) {
573        #if defined(USE_WORK_STEALING)
574                lanes.tscs = alloc(lanes.count, lanes.tscs`realloc);
575                for(i; lanes.count) {
[9cac0da]576                        unsigned long long tsc = ts(lanes.data[i]);
577                        lanes.tscs[i].tv = tsc != 0 ? tsc : rdtscl();
[431cd4f]578                }
579        #endif
580}
581
[dca5802]582// Grow the ready queue
[a017ee7]583void ready_queue_grow(struct cluster * cltr) {
[bd0bdd37]584        size_t ncount;
[a017ee7]585        int target = cltr->procs.total;
[bd0bdd37]586
[64a7146]587        /* paranoid */ verify( ready_mutate_islocked() );
[504a7dc]588        __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue\n");
[b798713]589
[dca5802]590        // Make sure that everything is consistent
591        /* paranoid */ check( cltr->ready_queue );
592
593        // grow the ready queue
[b798713]594        with( cltr->ready_queue ) {
[39fc03e]595                // Find new count
596                // Make sure we always have atleast 1 list
[bd0bdd37]597                if(target >= 2) {
[9cc3a18]598                        ncount = target * READYQ_SHARD_FACTOR;
[bd0bdd37]599                } else {
[5f6a172]600                        ncount = SEQUENTIAL_SHARD;
[bd0bdd37]601                }
[b798713]602
[dca5802]603                // Allocate new array (uses realloc and memcpies the data)
[ceb7db8]604                lanes.data = alloc( ncount, lanes.data`realloc );
[b798713]605
606                // Fix the moved data
[dca5802]607                for( idx; (size_t)lanes.count ) {
608                        fix(lanes.data[idx]);
[b798713]609                }
610
611                // Construct new data
[dca5802]612                for( idx; (size_t)lanes.count ~ ncount) {
613                        (lanes.data[idx]){};
[b798713]614                }
615
616                // Update original
[dca5802]617                lanes.count = ncount;
[b798713]618        }
619
[9cc3a18]620        fix_times(cltr);
621
622        reassign_cltr_id(cltr);
[a017ee7]623
[b798713]624        // Make sure that everything is consistent
[dca5802]625        /* paranoid */ check( cltr->ready_queue );
626
[504a7dc]627        __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue done\n");
[dca5802]628
[64a7146]629        /* paranoid */ verify( ready_mutate_islocked() );
[b798713]630}
631
[dca5802]632// Shrink the ready queue
[a017ee7]633void ready_queue_shrink(struct cluster * cltr) {
[64a7146]634        /* paranoid */ verify( ready_mutate_islocked() );
[504a7dc]635        __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue\n");
[dca5802]636
637        // Make sure that everything is consistent
638        /* paranoid */ check( cltr->ready_queue );
639
[a017ee7]640        int target = cltr->procs.total;
641
[b798713]642        with( cltr->ready_queue ) {
[39fc03e]643                // Remember old count
[dca5802]644                size_t ocount = lanes.count;
[b798713]645
[39fc03e]646                // Find new count
647                // Make sure we always have atleast 1 list
[5f6a172]648                lanes.count = target >= 2 ? target * READYQ_SHARD_FACTOR: SEQUENTIAL_SHARD;
[39fc03e]649                /* paranoid */ verify( ocount >= lanes.count );
[9cc3a18]650                /* paranoid */ verify( lanes.count == target * READYQ_SHARD_FACTOR || target < 2 );
[dca5802]651
652                // for printing count the number of displaced threads
[504a7dc]653                #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
[dca5802]654                        __attribute__((unused)) size_t displaced = 0;
655                #endif
[b798713]656
657                // redistribute old data
[dca5802]658                for( idx; (size_t)lanes.count ~ ocount) {
659                        // Lock is not strictly needed but makes checking invariants much easier
[1b143de]660                        __attribute__((unused)) bool locked = __atomic_try_acquire(&lanes.data[idx].lock);
[b798713]661                        verify(locked);
[dca5802]662
663                        // As long as we can pop from this lane to push the threads somewhere else in the queue
664                        while(!is_empty(lanes.data[idx])) {
[504a7dc]665                                struct $thread * thrd;
[f302d80]666                                unsigned long long _;
667                                [thrd, _] = pop(lanes.data[idx]);
[dca5802]668
[b808625]669                                push(cltr, thrd, true);
[dca5802]670
671                                // for printing count the number of displaced threads
[504a7dc]672                                #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
[dca5802]673                                        displaced++;
674                                #endif
[b798713]675                        }
676
[dca5802]677                        // Unlock the lane
678                        __atomic_unlock(&lanes.data[idx].lock);
[b798713]679
680                        // TODO print the queue statistics here
681
[dca5802]682                        ^(lanes.data[idx]){};
[b798713]683                }
684
[504a7dc]685                __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue displaced %zu threads\n", displaced);
[c84b4be]686
[dca5802]687                // Allocate new array (uses realloc and memcpies the data)
[ceb7db8]688                lanes.data = alloc( lanes.count, lanes.data`realloc );
[b798713]689
690                // Fix the moved data
[dca5802]691                for( idx; (size_t)lanes.count ) {
692                        fix(lanes.data[idx]);
[b798713]693                }
694        }
695
[9cc3a18]696        fix_times(cltr);
697
698        reassign_cltr_id(cltr);
[a017ee7]699
[b798713]700        // Make sure that everything is consistent
[dca5802]701        /* paranoid */ check( cltr->ready_queue );
702
[504a7dc]703        __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue done\n");
[64a7146]704        /* paranoid */ verify( ready_mutate_islocked() );
[fd9b524]705}
[8cd5434]706
707#if !defined(__CFA_NO_STATISTICS__)
708        unsigned cnt(const __ready_queue_t & this, unsigned idx) {
709                /* paranoid */ verify(this.lanes.count > idx);
710                return this.lanes.data[idx].cnt;
711        }
712#endif
Note: See TracBrowser for help on using the repository browser.