source: libcfa/src/concurrency/ready_queue.cfa @ 0ee224b

ADTast-experimentalenumforall-pointer-decayjacob/cs343-translationnew-ast-unique-exprpthread-emulationqualifiedEnum
Last change on this file since 0ee224b was 0ee224b, checked in by Thierry Delisle <tdelisle@…>, 3 years ago

Fixed rseq so it is initilizaed even for non-processor threads.

  • Property mode set to 100644
File size: 30.3 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
[07b4970]22#define USE_RELAXED_FIFO
[9cc3a18]23// #define USE_WORK_STEALING
24
[7768b8d]25#include "bits/defs.hfa"
[12daa43]26#include "device/cpu.hfa"
[7768b8d]27#include "kernel_private.hfa"
28
29#include "stdlib.hfa"
[61d7bec]30#include "math.hfa"
[7768b8d]31
[0ee224b]32#include <errno.h>
[04b5cef]33#include <unistd.h>
34
[0ee224b]35extern "C" {
36        #include <sys/syscall.h>  // __NR_xxx
37}
38
[13c5e19]39#include "ready_subqueue.hfa"
40
[7768b8d]41static const size_t cache_line_size = 64;
42
[d2fadeb]43#if !defined(__CFA_NO_STATISTICS__)
44        #define __STATS(...) __VA_ARGS__
45#else
46        #define __STATS(...)
47#endif
48
[dca5802]49// No overriden function, no environment variable, no define
50// fall back to a magic number
51#ifndef __CFA_MAX_PROCESSORS__
[b388ee81]52        #define __CFA_MAX_PROCESSORS__ 1024
[dca5802]53#endif
[7768b8d]54
[12daa43]55#if   defined(USE_CPU_WORK_STEALING)
56        #define READYQ_SHARD_FACTOR 2
57#elif defined(USE_RELAXED_FIFO)
[9cc3a18]58        #define BIAS 4
59        #define READYQ_SHARD_FACTOR 4
[5f6a172]60        #define SEQUENTIAL_SHARD 1
[9cc3a18]61#elif defined(USE_WORK_STEALING)
62        #define READYQ_SHARD_FACTOR 2
[5f6a172]63        #define SEQUENTIAL_SHARD 2
[9cc3a18]64#else
65        #error no scheduling strategy selected
66#endif
67
[d2fadeb]68static inline struct $thread * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats));
69static inline struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats));
[431cd4f]70static inline struct $thread * search(struct cluster * cltr);
[d2fadeb]71static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred);
[9cc3a18]72
[04b5cef]73
[dca5802]74// returns the maximum number of processors the RWLock support
[7768b8d]75__attribute__((weak)) unsigned __max_processors() {
76        const char * max_cores_s = getenv("CFA_MAX_PROCESSORS");
77        if(!max_cores_s) {
[504a7dc]78                __cfadbg_print_nolock(ready_queue, "No CFA_MAX_PROCESSORS in ENV\n");
[dca5802]79                return __CFA_MAX_PROCESSORS__;
[7768b8d]80        }
81
82        char * endptr = 0p;
83        long int max_cores_l = strtol(max_cores_s, &endptr, 10);
84        if(max_cores_l < 1 || max_cores_l > 65535) {
[504a7dc]85                __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS out of range : %ld\n", max_cores_l);
[dca5802]86                return __CFA_MAX_PROCESSORS__;
[7768b8d]87        }
88        if('\0' != *endptr) {
[504a7dc]89                __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS not a decimal number : %s\n", max_cores_s);
[dca5802]90                return __CFA_MAX_PROCESSORS__;
[7768b8d]91        }
92
93        return max_cores_l;
94}
95
[0ee224b]96#if   defined(CFA_HAVE_LINUX_LIBRSEQ)
97        // No forward declaration needed
98        #define __kernel_rseq_register rseq_register_current_thread
99        #define __kernel_rseq_unregister rseq_unregister_current_thread
100#elif defined(CFA_HAVE_LINUX_RSEQ_H)
101        void __kernel_raw_rseq_register  (void);
102        void __kernel_raw_rseq_unregister(void);
103
104        #define __kernel_rseq_register __kernel_raw_rseq_register
105        #define __kernel_rseq_unregister __kernel_raw_rseq_unregister
106#else
107        // No forward declaration needed
108        // No initialization needed
109        static inline void noop(void) {}
110
111        #define __kernel_rseq_register noop
112        #define __kernel_rseq_unregister noop
113#endif
114
[7768b8d]115//=======================================================================
116// Cluster wide reader-writer lock
117//=======================================================================
[b388ee81]118void  ?{}(__scheduler_RWLock_t & this) {
[7768b8d]119        this.max   = __max_processors();
120        this.alloc = 0;
121        this.ready = 0;
122        this.data  = alloc(this.max);
[c993b15]123        this.write_lock  = false;
[7768b8d]124
125        /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.alloc), &this.alloc));
126        /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.ready), &this.ready));
127
128}
[b388ee81]129void ^?{}(__scheduler_RWLock_t & this) {
[7768b8d]130        free(this.data);
131}
132
133
134//=======================================================================
135// Lock-Free registering/unregistering of threads
[c993b15]136unsigned register_proc_id( void ) with(*__scheduler_lock) {
[0ee224b]137        __kernel_rseq_register();
138
[b388ee81]139        __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p for RW-Lock\n", proc);
[c993b15]140        bool * handle = (bool *)&kernelTLS().sched_lock;
[504a7dc]141
[7768b8d]142        // Step - 1 : check if there is already space in the data
143        uint_fast32_t s = ready;
144
145        // Check among all the ready
146        for(uint_fast32_t i = 0; i < s; i++) {
[c993b15]147                bool * volatile * cell = (bool * volatile *)&data[i]; // Cforall is bugged and the double volatiles causes problems
148                /* paranoid */ verify( handle != *cell );
149
150                bool * null = 0p; // Re-write every loop since compare thrashes it
151                if( __atomic_load_n(cell, (int)__ATOMIC_RELAXED) == null
152                        && __atomic_compare_exchange_n( cell, &null, handle, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
153                        /* paranoid */ verify(i < ready);
154                        /* paranoid */ verify( (kernelTLS().sched_id = i, true) );
155                        return i;
[7768b8d]156                }
157        }
158
[b388ee81]159        if(max <= alloc) abort("Trying to create more than %ud processors", __scheduler_lock->max);
[7768b8d]160
161        // Step - 2 : F&A to get a new spot in the array.
162        uint_fast32_t n = __atomic_fetch_add(&alloc, 1, __ATOMIC_SEQ_CST);
[b388ee81]163        if(max <= n) abort("Trying to create more than %ud processors", __scheduler_lock->max);
[7768b8d]164
165        // Step - 3 : Mark space as used and then publish it.
[c993b15]166        data[n] = handle;
[fd9b524]167        while() {
[7768b8d]168                unsigned copy = n;
169                if( __atomic_load_n(&ready, __ATOMIC_RELAXED) == n
170                        && __atomic_compare_exchange_n(&ready, &copy, n + 1, true, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))
171                        break;
[fd9b524]172                Pause();
[7768b8d]173        }
174
[1b143de]175        __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p done, id %lu\n", proc, n);
[504a7dc]176
[7768b8d]177        // Return new spot.
[c993b15]178        /* paranoid */ verify(n < ready);
179        /* paranoid */ verify( (kernelTLS().sched_id = n, true) );
180        return n;
[7768b8d]181}
182
[c993b15]183void unregister_proc_id( unsigned id ) with(*__scheduler_lock) {
184        /* paranoid */ verify(id < ready);
185        /* paranoid */ verify(id == kernelTLS().sched_id);
186        /* paranoid */ verify(data[id] == &kernelTLS().sched_lock);
187
188        bool * volatile * cell = (bool * volatile *)&data[id]; // Cforall is bugged and the double volatiles causes problems
189
190        __atomic_store_n(cell, 0p, __ATOMIC_RELEASE);
[504a7dc]191
192        __cfadbg_print_safe(ready_queue, "Kernel : Unregister proc %p\n", proc);
[0ee224b]193
194        __kernel_rseq_unregister();
[7768b8d]195}
196
197//-----------------------------------------------------------------------
198// Writer side : acquire when changing the ready queue, e.g. adding more
199//  queues or removing them.
[b388ee81]200uint_fast32_t ready_mutate_lock( void ) with(*__scheduler_lock) {
[8fc652e0]201        /* paranoid */ verify( ! __preemption_enabled() );
[c993b15]202        /* paranoid */ verify( ! kernelTLS().sched_lock );
[62502cc4]203
[7768b8d]204        // Step 1 : lock global lock
205        // It is needed to avoid processors that register mid Critical-Section
206        //   to simply lock their own lock and enter.
[c993b15]207        __atomic_acquire( &write_lock );
[7768b8d]208
209        // Step 2 : lock per-proc lock
210        // Processors that are currently being registered aren't counted
211        //   but can't be in read_lock or in the critical section.
212        // All other processors are counted
213        uint_fast32_t s = ready;
214        for(uint_fast32_t i = 0; i < s; i++) {
[c993b15]215                volatile bool * llock = data[i];
216                if(llock) __atomic_acquire( llock );
[7768b8d]217        }
218
[8fc652e0]219        /* paranoid */ verify( ! __preemption_enabled() );
[7768b8d]220        return s;
221}
222
[b388ee81]223void ready_mutate_unlock( uint_fast32_t last_s ) with(*__scheduler_lock) {
[8fc652e0]224        /* paranoid */ verify( ! __preemption_enabled() );
[62502cc4]225
[7768b8d]226        // Step 1 : release local locks
227        // This must be done while the global lock is held to avoid
228        //   threads that where created mid critical section
229        //   to race to lock their local locks and have the writer
230        //   immidiately unlock them
231        // Alternative solution : return s in write_lock and pass it to write_unlock
232        for(uint_fast32_t i = 0; i < last_s; i++) {
[c993b15]233                volatile bool * llock = data[i];
234                if(llock) __atomic_store_n(llock, (bool)false, __ATOMIC_RELEASE);
[7768b8d]235        }
236
237        // Step 2 : release global lock
[c993b15]238        /*paranoid*/ assert(true == write_lock);
239        __atomic_store_n(&write_lock, (bool)false, __ATOMIC_RELEASE);
[62502cc4]240
[8fc652e0]241        /* paranoid */ verify( ! __preemption_enabled() );
[7768b8d]242}
243
244//=======================================================================
[9cc3a18]245// Cforall Ready Queue used for scheduling
[b798713]246//=======================================================================
247void ?{}(__ready_queue_t & this) with (this) {
[12daa43]248        #if defined(USE_CPU_WORK_STEALING)
249                lanes.count = cpu_info.hthrd_count * READYQ_SHARD_FACTOR;
250                lanes.data = alloc( lanes.count );
251                lanes.tscs = alloc( lanes.count );
252
253                for( idx; (size_t)lanes.count ) {
254                        (lanes.data[idx]){};
255                        lanes.tscs[idx].tv = rdtscl();
256                }
257        #else
258                lanes.data  = 0p;
259                lanes.tscs  = 0p;
260                lanes.count = 0;
261        #endif
[b798713]262}
263
264void ^?{}(__ready_queue_t & this) with (this) {
[12daa43]265        #if !defined(USE_CPU_WORK_STEALING)
266                verify( SEQUENTIAL_SHARD == lanes.count );
267        #endif
268
[dca5802]269        free(lanes.data);
[9cc3a18]270        free(lanes.tscs);
[dca5802]271}
272
[64a7146]273//-----------------------------------------------------------------------
[12daa43]274#if defined(USE_CPU_WORK_STEALING)
275        __attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd, bool push_local) with (cltr->ready_queue) {
276                __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
277
278                processor * const proc = kernelTLS().this_processor;
279                const bool external = !push_local || (!proc) || (cltr != proc->cltr);
280
281                const int cpu = __kernel_getcpu();
282                /* paranoid */ verify(cpu >= 0);
283                /* paranoid */ verify(cpu < cpu_info.hthrd_count);
284                /* paranoid */ verify(cpu * READYQ_SHARD_FACTOR < lanes.count);
285
[df7597e0]286                const cpu_map_entry_t & map = cpu_info.llc_map[cpu];
287                /* paranoid */ verify(map.start * READYQ_SHARD_FACTOR < lanes.count);
288                /* paranoid */ verify(map.self * READYQ_SHARD_FACTOR < lanes.count);
289                /* paranoid */ verifyf((map.start + map.count) * READYQ_SHARD_FACTOR <= lanes.count, "have %u lanes but map can go up to %u", lanes.count, (map.start + map.count) * READYQ_SHARD_FACTOR);
290
291                const int start = map.self * READYQ_SHARD_FACTOR;
[12daa43]292                unsigned i;
293                do {
294                        unsigned r;
295                        if(unlikely(external)) { r = __tls_rand(); }
296                        else { r = proc->rdq.its++; }
297                        i = start + (r % READYQ_SHARD_FACTOR);
298                        // If we can't lock it retry
299                } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
300
301                // Actually push it
302                push(lanes.data[i], thrd);
303
304                // Unlock and return
305                __atomic_unlock( &lanes.data[i].lock );
306
307                #if !defined(__CFA_NO_STATISTICS__)
308                        if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
309                        else __tls_stats()->ready.push.local.success++;
310                #endif
311
312                __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);
313
314        }
315
316        // Pop from the ready queue from a given cluster
317        __attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
318                /* paranoid */ verify( lanes.count > 0 );
319                /* paranoid */ verify( kernelTLS().this_processor );
320
321                const int cpu = __kernel_getcpu();
322                /* paranoid */ verify(cpu >= 0);
323                /* paranoid */ verify(cpu < cpu_info.hthrd_count);
[df7597e0]324                /* paranoid */ verify(cpu * READYQ_SHARD_FACTOR < lanes.count);
325
326                const cpu_map_entry_t & map = cpu_info.llc_map[cpu];
327                /* paranoid */ verify(map.start * READYQ_SHARD_FACTOR < lanes.count);
328                /* paranoid */ verify(map.self * READYQ_SHARD_FACTOR < lanes.count);
329                /* paranoid */ verifyf((map.start + map.count) * READYQ_SHARD_FACTOR <= lanes.count, "have %u lanes but map can go up to %u", lanes.count, (map.start + map.count) * READYQ_SHARD_FACTOR);
[12daa43]330
331                processor * const proc = kernelTLS().this_processor;
[df7597e0]332                const int start = map.self * READYQ_SHARD_FACTOR;
[12daa43]333
334                // Did we already have a help target
335                if(proc->rdq.target == -1u) {
336                        // if We don't have a
337                        unsigned long long min = ts(lanes.data[start]);
338                        for(i; READYQ_SHARD_FACTOR) {
339                                unsigned long long tsc = ts(lanes.data[start + i]);
340                                if(tsc < min) min = tsc;
341                        }
342                        proc->rdq.cutoff = min;
[df7597e0]343                        proc->rdq.target = (map.start * READYQ_SHARD_FACTOR) + (__tls_rand() % (map.count* READYQ_SHARD_FACTOR));
[12daa43]344                }
345                else {
346                        const unsigned long long bias = 0; //2_500_000_000;
347                        const unsigned long long cutoff = proc->rdq.cutoff > bias ? proc->rdq.cutoff - bias : proc->rdq.cutoff;
348                        {
349                                unsigned target = proc->rdq.target;
350                                proc->rdq.target = -1u;
351                                if(lanes.tscs[target].tv < cutoff && ts(lanes.data[target]) < cutoff) {
352                                        $thread * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
353                                        proc->rdq.last = target;
354                                        if(t) return t;
355                                }
356                        }
357
358                        unsigned last = proc->rdq.last;
359                        if(last != -1u && lanes.tscs[last].tv < cutoff && ts(lanes.data[last]) < cutoff) {
360                                $thread * t = try_pop(cltr, last __STATS(, __tls_stats()->ready.pop.help));
361                                if(t) return t;
362                        }
363                        else {
364                                proc->rdq.last = -1u;
365                        }
366                }
367
368                for(READYQ_SHARD_FACTOR) {
369                        unsigned i = start + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
370                        if($thread * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
371                }
372
373                // All lanes where empty return 0p
374                return 0p;
375        }
376
377        __attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
378                processor * const proc = kernelTLS().this_processor;
379                unsigned last = proc->rdq.last;
380
381                unsigned i = __tls_rand() % lanes.count;
382                return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
383        }
384        __attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) {
385                return search(cltr);
386        }
387#endif
[431cd4f]388#if defined(USE_RELAXED_FIFO)
389        //-----------------------------------------------------------------------
390        // get index from random number with or without bias towards queues
391        static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred) {
392                unsigned i;
393                bool local;
394                unsigned rlow  = r % BIAS;
395                unsigned rhigh = r / BIAS;
396                if((0 != rlow) && preferred >= 0) {
397                        // (BIAS - 1) out of BIAS chances
398                        // Use perferred queues
399                        i = preferred + (rhigh % READYQ_SHARD_FACTOR);
400                        local = true;
401                }
402                else {
403                        // 1 out of BIAS chances
404                        // Use all queues
405                        i = rhigh;
406                        local = false;
407                }
408                return [i, local];
409        }
410
[b808625]411        __attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd, bool push_local) with (cltr->ready_queue) {
[431cd4f]412                __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
[1b143de]413
[b808625]414                const bool external = !push_local || (!kernelTLS().this_processor) || (cltr != kernelTLS().this_processor->cltr);
[431cd4f]415                /* paranoid */ verify(external || kernelTLS().this_processor->rdq.id < lanes.count );
[fd1f65e]416
[431cd4f]417                bool local;
418                int preferred = external ? -1 : kernelTLS().this_processor->rdq.id;
[52769ba]419
[431cd4f]420                // Try to pick a lane and lock it
421                unsigned i;
422                do {
423                        // Pick the index of a lane
424                        unsigned r = __tls_rand_fwd();
425                        [i, local] = idx_from_r(r, preferred);
[772411a]426
[431cd4f]427                        i %= __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
428
429                        #if !defined(__CFA_NO_STATISTICS__)
[d2fadeb]430                                if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
431                                else if(local) __tls_stats()->ready.push.local.attempt++;
432                                else __tls_stats()->ready.push.share.attempt++;
[431cd4f]433                        #endif
[b798713]434
[431cd4f]435                        // If we can't lock it retry
436                } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
437
438                // Actually push it
439                push(lanes.data[i], thrd);
440
[b808625]441                // Unlock and return
442                __atomic_unlock( &lanes.data[i].lock );
[431cd4f]443
444                // Mark the current index in the tls rng instance as having an item
445                __tls_rand_advance_bck();
446
447                __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);
448
449                // Update statistics
[b798713]450                #if !defined(__CFA_NO_STATISTICS__)
[d2fadeb]451                        if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
452                        else if(local) __tls_stats()->ready.push.local.success++;
453                        else __tls_stats()->ready.push.share.success++;
[b798713]454                #endif
[431cd4f]455        }
[b798713]456
[431cd4f]457        // Pop from the ready queue from a given cluster
458        __attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
459                /* paranoid */ verify( lanes.count > 0 );
460                /* paranoid */ verify( kernelTLS().this_processor );
461                /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
[b798713]462
[431cd4f]463                unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
464                int preferred = kernelTLS().this_processor->rdq.id;
[dca5802]465
466
[431cd4f]467                // As long as the list is not empty, try finding a lane that isn't empty and pop from it
468                for(25) {
469                        // Pick two lists at random
470                        unsigned ri = __tls_rand_bck();
471                        unsigned rj = __tls_rand_bck();
[c426b03]472
[431cd4f]473                        unsigned i, j;
474                        __attribute__((unused)) bool locali, localj;
475                        [i, locali] = idx_from_r(ri, preferred);
476                        [j, localj] = idx_from_r(rj, preferred);
[1b143de]477
[431cd4f]478                        i %= count;
479                        j %= count;
[9cc3a18]480
[431cd4f]481                        // try popping from the 2 picked lists
[d2fadeb]482                        struct $thread * thrd = try_pop(cltr, i, j __STATS(, *(locali || localj ? &__tls_stats()->ready.pop.local : &__tls_stats()->ready.pop.help)));
[431cd4f]483                        if(thrd) {
484                                return thrd;
485                        }
486                }
[13c5e19]487
[431cd4f]488                // All lanes where empty return 0p
489                return 0p;
490        }
[772411a]491
[fc59df78]492        __attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) { return pop_fast(cltr); }
493        __attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) {
[431cd4f]494                return search(cltr);
495        }
496#endif
497#if defined(USE_WORK_STEALING)
[b808625]498        __attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd, bool push_local) with (cltr->ready_queue) {
[431cd4f]499                __cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
[772411a]500
[d3ba775]501                // #define USE_PREFERRED
502                #if !defined(USE_PREFERRED)
[b808625]503                const bool external = !push_local || (!kernelTLS().this_processor) || (cltr != kernelTLS().this_processor->cltr);
[431cd4f]504                /* paranoid */ verify(external || kernelTLS().this_processor->rdq.id < lanes.count );
[d3ba775]505                #else
506                        unsigned preferred = thrd->preferred;
[b808625]507                        const bool external = push_local || (!kernelTLS().this_processor) || preferred == -1u || thrd->curr_cluster != cltr;
[d3ba775]508                        /* paranoid */ verifyf(external || preferred < lanes.count, "Invalid preferred queue %u for %u lanes", preferred, lanes.count );
[772411a]509
[d3ba775]510                        unsigned r = preferred % READYQ_SHARD_FACTOR;
511                        const unsigned start = preferred - r;
[2b96031]512                #endif
[431cd4f]513
514                // Try to pick a lane and lock it
515                unsigned i;
516                do {
[d2fadeb]517                        #if !defined(__CFA_NO_STATISTICS__)
518                                if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
519                                else __tls_stats()->ready.push.local.attempt++;
520                        #endif
521
[431cd4f]522                        if(unlikely(external)) {
523                                i = __tls_rand() % lanes.count;
524                        }
525                        else {
[d3ba775]526                                #if !defined(USE_PREFERRED)
[b808625]527                                        processor * proc = kernelTLS().this_processor;
528                                        unsigned r = proc->rdq.its++;
529                                        i =  proc->rdq.id + (r % READYQ_SHARD_FACTOR);
530                                #else
[d3ba775]531                                        i = start + (r++ % READYQ_SHARD_FACTOR);
532                                #endif
533                        }
[431cd4f]534                        // If we can't lock it retry
535                } while( !__atomic_try_acquire( &lanes.data[i].lock ) );
[13c5e19]536
[431cd4f]537                // Actually push it
538                push(lanes.data[i], thrd);
[13c5e19]539
[b808625]540                // Unlock and return
541                __atomic_unlock( &lanes.data[i].lock );
[431cd4f]542
[d2fadeb]543                #if !defined(__CFA_NO_STATISTICS__)
544                        if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
545                        else __tls_stats()->ready.push.local.success++;
546                #endif
547
[431cd4f]548                __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]549        }
550
[431cd4f]551        // Pop from the ready queue from a given cluster
552        __attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
553                /* paranoid */ verify( lanes.count > 0 );
554                /* paranoid */ verify( kernelTLS().this_processor );
555                /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
556
557                processor * proc = kernelTLS().this_processor;
558
559                if(proc->rdq.target == -1u) {
[1680072]560                        unsigned long long min = ts(lanes.data[proc->rdq.id]);
561                        for(int i = 0; i < READYQ_SHARD_FACTOR; i++) {
562                                unsigned long long tsc = ts(lanes.data[proc->rdq.id + i]);
563                                if(tsc < min) min = tsc;
564                        }
565                        proc->rdq.cutoff = min;
[f55d54d]566                        proc->rdq.target = __tls_rand() % lanes.count;
[431cd4f]567                }
[341aa39]568                else {
569                        unsigned target = proc->rdq.target;
[431cd4f]570                        proc->rdq.target = -1u;
[9cac0da]571                        const unsigned long long bias = 0; //2_500_000_000;
572                        const unsigned long long cutoff = proc->rdq.cutoff > bias ? proc->rdq.cutoff - bias : proc->rdq.cutoff;
573                        if(lanes.tscs[target].tv < cutoff && ts(lanes.data[target]) < cutoff) {
[341aa39]574                                $thread * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
575                                if(t) return t;
576                        }
[431cd4f]577                }
[13c5e19]578
[431cd4f]579                for(READYQ_SHARD_FACTOR) {
[f55d54d]580                        unsigned i = proc->rdq.id + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
[d2fadeb]581                        if($thread * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
[431cd4f]582                }
583                return 0p;
[1eb239e4]584        }
585
[431cd4f]586        __attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
[fc59df78]587                unsigned i = __tls_rand() % lanes.count;
588                return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
589        }
[431cd4f]590
[fc59df78]591        __attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) with (cltr->ready_queue) {
[431cd4f]592                return search(cltr);
593        }
594#endif
[1eb239e4]595
[9cc3a18]596//=======================================================================
597// Various Ready Queue utilities
598//=======================================================================
599// these function work the same or almost the same
600// whether they are using work-stealing or relaxed fifo scheduling
[1eb239e4]601
[9cc3a18]602//-----------------------------------------------------------------------
603// try to pop from a lane given by index w
[d2fadeb]604static inline struct $thread * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
605        __STATS( stats.attempt++; )
606
[dca5802]607        // Get relevant elements locally
608        __intrusive_lane_t & lane = lanes.data[w];
609
[b798713]610        // If list looks empty retry
[d2fadeb]611        if( is_empty(lane) ) {
612                return 0p;
613        }
[b798713]614
615        // If we can't get the lock retry
[d2fadeb]616        if( !__atomic_try_acquire(&lane.lock) ) {
617                return 0p;
618        }
[b798713]619
620        // If list is empty, unlock and retry
[dca5802]621        if( is_empty(lane) ) {
622                __atomic_unlock(&lane.lock);
[b798713]623                return 0p;
624        }
625
626        // Actually pop the list
[504a7dc]627        struct $thread * thrd;
[f302d80]628        unsigned long long tsv;
629        [thrd, tsv] = pop(lane);
[b798713]630
[dca5802]631        /* paranoid */ verify(thrd);
[78ea291]632        /* paranoid */ verify(tsv);
[dca5802]633        /* paranoid */ verify(lane.lock);
[b798713]634
635        // Unlock and return
[dca5802]636        __atomic_unlock(&lane.lock);
[b798713]637
[dca5802]638        // Update statistics
[d2fadeb]639        __STATS( stats.success++; )
[b798713]640
[431cd4f]641        #if defined(USE_WORK_STEALING)
[f302d80]642                lanes.tscs[w].tv = tsv;
[9cc3a18]643        #endif
[d72c074]644
[d3ba775]645        thrd->preferred = w;
646
[dca5802]647        // return the popped thread
[b798713]648        return thrd;
649}
[04b5cef]650
[9cc3a18]651//-----------------------------------------------------------------------
652// try to pop from any lanes making sure you don't miss any threads push
653// before the start of the function
[431cd4f]654static inline struct $thread * search(struct cluster * cltr) with (cltr->ready_queue) {
[9cc3a18]655        /* paranoid */ verify( lanes.count > 0 );
656        unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
657        unsigned offset = __tls_rand();
658        for(i; count) {
659                unsigned idx = (offset + i) % count;
[d2fadeb]660                struct $thread * thrd = try_pop(cltr, idx __STATS(, __tls_stats()->ready.pop.search));
[9cc3a18]661                if(thrd) {
662                        return thrd;
663                }
[13c5e19]664        }
[9cc3a18]665
666        // All lanes where empty return 0p
667        return 0p;
[b798713]668}
669
670//-----------------------------------------------------------------------
[9cc3a18]671// Check that all the intrusive queues in the data structure are still consistent
[b798713]672static void check( __ready_queue_t & q ) with (q) {
[d3ba775]673        #if defined(__CFA_WITH_VERIFY__)
[b798713]674                {
[dca5802]675                        for( idx ; lanes.count ) {
676                                __intrusive_lane_t & sl = lanes.data[idx];
677                                assert(!lanes.data[idx].lock);
[b798713]678
[2b96031]679                                        if(is_empty(sl)) {
680                                                assert( sl.anchor.next == 0p );
681                                                assert( sl.anchor.ts   == 0  );
682                                                assert( mock_head(sl)  == sl.prev );
683                                        } else {
684                                                assert( sl.anchor.next != 0p );
685                                                assert( sl.anchor.ts   != 0  );
686                                                assert( mock_head(sl)  != sl.prev );
687                                        }
[b798713]688                        }
689                }
690        #endif
691}
692
[9cc3a18]693//-----------------------------------------------------------------------
694// Given 2 indexes, pick the list with the oldest push an try to pop from it
[d2fadeb]695static inline struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
[9cc3a18]696        // Pick the bet list
697        int w = i;
698        if( __builtin_expect(!is_empty(lanes.data[j]), true) ) {
699                w = (ts(lanes.data[i]) < ts(lanes.data[j])) ? i : j;
700        }
701
[d2fadeb]702        return try_pop(cltr, w __STATS(, stats));
[9cc3a18]703}
704
[b798713]705// Call this function of the intrusive list was moved using memcpy
[dca5802]706// fixes the list so that the pointers back to anchors aren't left dangling
707static inline void fix(__intrusive_lane_t & ll) {
[2b96031]708                        if(is_empty(ll)) {
709                                verify(ll.anchor.next == 0p);
710                                ll.prev = mock_head(ll);
711                        }
[b798713]712}
713
[69914cbc]714static void assign_list(unsigned & value, dlist(processor) & list, unsigned count) {
[a017ee7]715        processor * it = &list`first;
716        for(unsigned i = 0; i < count; i++) {
717                /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
[431cd4f]718                it->rdq.id = value;
719                it->rdq.target = -1u;
[9cc3a18]720                value += READYQ_SHARD_FACTOR;
[a017ee7]721                it = &(*it)`next;
722        }
723}
724
[9cc3a18]725static void reassign_cltr_id(struct cluster * cltr) {
[a017ee7]726        unsigned preferred = 0;
[9cc3a18]727        assign_list(preferred, cltr->procs.actives, cltr->procs.total - cltr->procs.idle);
728        assign_list(preferred, cltr->procs.idles  , cltr->procs.idle );
[a017ee7]729}
730
[431cd4f]731static void fix_times( struct cluster * cltr ) with( cltr->ready_queue ) {
732        #if defined(USE_WORK_STEALING)
733                lanes.tscs = alloc(lanes.count, lanes.tscs`realloc);
734                for(i; lanes.count) {
[9cac0da]735                        unsigned long long tsc = ts(lanes.data[i]);
736                        lanes.tscs[i].tv = tsc != 0 ? tsc : rdtscl();
[431cd4f]737                }
738        #endif
739}
740
[12daa43]741#if defined(USE_CPU_WORK_STEALING)
742        // ready_queue size is fixed in this case
743        void ready_queue_grow(struct cluster * cltr) {}
744        void ready_queue_shrink(struct cluster * cltr) {}
745#else
746        // Grow the ready queue
747        void ready_queue_grow(struct cluster * cltr) {
748                size_t ncount;
749                int target = cltr->procs.total;
750
751                /* paranoid */ verify( ready_mutate_islocked() );
752                __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue\n");
753
754                // Make sure that everything is consistent
755                /* paranoid */ check( cltr->ready_queue );
756
757                // grow the ready queue
758                with( cltr->ready_queue ) {
759                        // Find new count
760                        // Make sure we always have atleast 1 list
761                        if(target >= 2) {
762                                ncount = target * READYQ_SHARD_FACTOR;
763                        } else {
764                                ncount = SEQUENTIAL_SHARD;
765                        }
[b798713]766
[12daa43]767                        // Allocate new array (uses realloc and memcpies the data)
768                        lanes.data = alloc( ncount, lanes.data`realloc );
[b798713]769
[12daa43]770                        // Fix the moved data
771                        for( idx; (size_t)lanes.count ) {
772                                fix(lanes.data[idx]);
773                        }
[b798713]774
[12daa43]775                        // Construct new data
776                        for( idx; (size_t)lanes.count ~ ncount) {
777                                (lanes.data[idx]){};
778                        }
[b798713]779
[12daa43]780                        // Update original
781                        lanes.count = ncount;
782                }
[b798713]783
[12daa43]784                fix_times(cltr);
[9cc3a18]785
[12daa43]786                reassign_cltr_id(cltr);
[a017ee7]787
[12daa43]788                // Make sure that everything is consistent
789                /* paranoid */ check( cltr->ready_queue );
[dca5802]790
[12daa43]791                __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue done\n");
[dca5802]792
[12daa43]793                /* paranoid */ verify( ready_mutate_islocked() );
794        }
[b798713]795
[12daa43]796        // Shrink the ready queue
797        void ready_queue_shrink(struct cluster * cltr) {
798                /* paranoid */ verify( ready_mutate_islocked() );
799                __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue\n");
[dca5802]800
[12daa43]801                // Make sure that everything is consistent
802                /* paranoid */ check( cltr->ready_queue );
[dca5802]803
[12daa43]804                int target = cltr->procs.total;
[a017ee7]805
[12daa43]806                with( cltr->ready_queue ) {
807                        // Remember old count
808                        size_t ocount = lanes.count;
[b798713]809
[12daa43]810                        // Find new count
811                        // Make sure we always have atleast 1 list
812                        lanes.count = target >= 2 ? target * READYQ_SHARD_FACTOR: SEQUENTIAL_SHARD;
813                        /* paranoid */ verify( ocount >= lanes.count );
814                        /* paranoid */ verify( lanes.count == target * READYQ_SHARD_FACTOR || target < 2 );
[dca5802]815
[12daa43]816                        // for printing count the number of displaced threads
817                        #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
818                                __attribute__((unused)) size_t displaced = 0;
819                        #endif
[b798713]820
[12daa43]821                        // redistribute old data
822                        for( idx; (size_t)lanes.count ~ ocount) {
823                                // Lock is not strictly needed but makes checking invariants much easier
824                                __attribute__((unused)) bool locked = __atomic_try_acquire(&lanes.data[idx].lock);
825                                verify(locked);
[dca5802]826
[12daa43]827                                // As long as we can pop from this lane to push the threads somewhere else in the queue
828                                while(!is_empty(lanes.data[idx])) {
829                                        struct $thread * thrd;
830                                        unsigned long long _;
831                                        [thrd, _] = pop(lanes.data[idx]);
[dca5802]832
[12daa43]833                                        push(cltr, thrd, true);
[dca5802]834
[12daa43]835                                        // for printing count the number of displaced threads
836                                        #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
837                                                displaced++;
838                                        #endif
839                                }
[b798713]840
[12daa43]841                                // Unlock the lane
842                                __atomic_unlock(&lanes.data[idx].lock);
[b798713]843
[12daa43]844                                // TODO print the queue statistics here
[b798713]845
[12daa43]846                                ^(lanes.data[idx]){};
847                        }
[b798713]848
[12daa43]849                        __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue displaced %zu threads\n", displaced);
[c84b4be]850
[12daa43]851                        // Allocate new array (uses realloc and memcpies the data)
852                        lanes.data = alloc( lanes.count, lanes.data`realloc );
[b798713]853
[12daa43]854                        // Fix the moved data
855                        for( idx; (size_t)lanes.count ) {
856                                fix(lanes.data[idx]);
857                        }
[b798713]858                }
859
[12daa43]860                fix_times(cltr);
[9cc3a18]861
[12daa43]862                reassign_cltr_id(cltr);
[a017ee7]863
[12daa43]864                // Make sure that everything is consistent
865                /* paranoid */ check( cltr->ready_queue );
[dca5802]866
[12daa43]867                __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue done\n");
868                /* paranoid */ verify( ready_mutate_islocked() );
869        }
870#endif
[8cd5434]871
872#if !defined(__CFA_NO_STATISTICS__)
873        unsigned cnt(const __ready_queue_t & this, unsigned idx) {
874                /* paranoid */ verify(this.lanes.count > idx);
875                return this.lanes.data[idx].cnt;
876        }
877#endif
[0ee224b]878
879
880#if   defined(CFA_HAVE_LINUX_LIBRSEQ)
881        // No definition needed
882#elif defined(CFA_HAVE_LINUX_RSEQ_H)
883
884        #if defined( __x86_64 ) || defined( __i386 )
885                #define RSEQ_SIG        0x53053053
886        #elif defined( __ARM_ARCH )
887                #ifdef __ARMEB__
888                #define RSEQ_SIG    0xf3def5e7      /* udf    #24035    ; 0x5de3 (ARMv6+) */
889                #else
890                #define RSEQ_SIG    0xe7f5def3      /* udf    #24035    ; 0x5de3 */
891                #endif
892        #endif
893
894        extern void __disable_interrupts_hard();
895        extern void __enable_interrupts_hard();
896
897        void __kernel_raw_rseq_register  (void) {
898                /* paranoid */ verify( __cfaabi_rseq.cpu_id == RSEQ_CPU_ID_UNINITIALIZED );
899
900                // int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), 0, (sigset_t *)0p, _NSIG / 8);
901                int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), 0, RSEQ_SIG);
902                if(ret != 0) {
903                        int e = errno;
904                        switch(e) {
905                        case EINVAL: abort("KERNEL ERROR: rseq register invalid argument");
906                        case ENOSYS: abort("KERNEL ERROR: rseq register no supported");
907                        case EFAULT: abort("KERNEL ERROR: rseq register with invalid argument");
908                        case EBUSY : abort("KERNEL ERROR: rseq register already registered");
909                        case EPERM : abort("KERNEL ERROR: rseq register sig  argument  on unregistration does not match the signature received on registration");
910                        default: abort("KERNEL ERROR: rseq register unexpected return %d", e);
911                        }
912                }
913        }
914
915        void __kernel_raw_rseq_unregister(void) {
916                /* paranoid */ verify( __cfaabi_rseq.cpu_id >= 0 );
917
918                // int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), RSEQ_FLAG_UNREGISTER, (sigset_t *)0p, _NSIG / 8);
919                int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), RSEQ_FLAG_UNREGISTER, RSEQ_SIG);
920                if(ret != 0) {
921                        int e = errno;
922                        switch(e) {
923                        case EINVAL: abort("KERNEL ERROR: rseq unregister invalid argument");
924                        case ENOSYS: abort("KERNEL ERROR: rseq unregister no supported");
925                        case EFAULT: abort("KERNEL ERROR: rseq unregister with invalid argument");
926                        case EBUSY : abort("KERNEL ERROR: rseq unregister already registered");
927                        case EPERM : abort("KERNEL ERROR: rseq unregister sig  argument  on unregistration does not match the signature received on registration");
928                        default: abort("KERNEL ERROR: rseq unregisteunexpected return %d", e);
929                        }
930                }
931        }
932#else
933        // No definition needed
934#endif
Note: See TracBrowser for help on using the repository browser.