Context Navigation

ready_queue.cfa@ 1f3d212

Visit:

ADT ast-experimental enum forall-pointer-decay jacob/cs343-translation new-ast-unique-expr pthread-emulation qualifiedEnum

Last change on this file since 1f3d212 was 07b4970, checked in by Thierry Delisle <tdelisle@…>, 4 years ago
Fix incorrect change that disabled all lists
Property mode set to `100644`
File size: 27.4 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 _GNU_SOURCE
18
19	// #define __CFA_DEBUG_PRINT_READY_QUEUE__
20
21
22	#define USE_RELAXED_FIFO
23	// #define USE_WORK_STEALING
24
25	#include "bits/defs.hfa"
26	#include "device/cpu.hfa"
27	#include "kernel_private.hfa"
28
29	#include "stdlib.hfa"
30	#include "math.hfa"
31
32	#include <unistd.h>
33
34	#include "ready_subqueue.hfa"
35
36	static const size_t cache_line_size = 64;
37
38	#if !defined(__CFA_NO_STATISTICS__)
39	#define __STATS(...) __VA_ARGS__
40	#else
41	#define __STATS(...)
42	#endif
43
44	// No overriden function, no environment variable, no define
45	// fall back to a magic number
46	#ifndef __CFA_MAX_PROCESSORS__
47	#define __CFA_MAX_PROCESSORS__ 1024
48	#endif
49
50	#if defined(USE_CPU_WORK_STEALING)
51	#define READYQ_SHARD_FACTOR 2
52	#elif defined(USE_RELAXED_FIFO)
53	#define BIAS 4
54	#define READYQ_SHARD_FACTOR 4
55	#define SEQUENTIAL_SHARD 1
56	#elif defined(USE_WORK_STEALING)
57	#define READYQ_SHARD_FACTOR 2
58	#define SEQUENTIAL_SHARD 2
59	#else
60	#error no scheduling strategy selected
61	#endif
62
63	static inline struct $thread * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats));
64	static inline struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats));
65	static inline struct $thread * search(struct cluster * cltr);
66	static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred);
67
68
69	// returns the maximum number of processors the RWLock support
70	__attribute__((weak)) unsigned __max_processors() {
71	const char * max_cores_s = getenv("CFA_MAX_PROCESSORS");
72	if(!max_cores_s) {
73	__cfadbg_print_nolock(ready_queue, "No CFA_MAX_PROCESSORS in ENV\n");
74	return __CFA_MAX_PROCESSORS__;
75	}
76
77	char * endptr = 0p;
78	long int max_cores_l = strtol(max_cores_s, &endptr, 10);
79	if(max_cores_l < 1 \|\| max_cores_l > 65535) {
80	__cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS out of range : %ld\n", max_cores_l);
81	return __CFA_MAX_PROCESSORS__;
82	}
83	if('\0' != *endptr) {
84	__cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS not a decimal number : %s\n", max_cores_s);
85	return __CFA_MAX_PROCESSORS__;
86	}
87
88	return max_cores_l;
89	}
90
91	//=======================================================================
92	// Cluster wide reader-writer lock
93	//=======================================================================
94	void ?{}(__scheduler_RWLock_t & this) {
95	this.max = __max_processors();
96	this.alloc = 0;
97	this.ready = 0;
98	this.data = alloc(this.max);
99	this.write_lock = false;
100
101	/paranoid/ verify(__atomic_is_lock_free(sizeof(this.alloc), &this.alloc));
102	/paranoid/ verify(__atomic_is_lock_free(sizeof(this.ready), &this.ready));
103
104	}
105	void ^?{}(__scheduler_RWLock_t & this) {
106	free(this.data);
107	}
108
109
110	//=======================================================================
111	// Lock-Free registering/unregistering of threads
112	unsigned register_proc_id( void ) with(*__scheduler_lock) {
113	__cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p for RW-Lock\n", proc);
114	bool * handle = (bool *)&kernelTLS().sched_lock;
115
116	// Step - 1 : check if there is already space in the data
117	uint_fast32_t s = ready;
118
119	// Check among all the ready
120	for(uint_fast32_t i = 0; i < s; i++) {
121	bool * volatile * cell = (bool * volatile *)&data[i]; // Cforall is bugged and the double volatiles causes problems
122	/* paranoid / verify( handle != cell );
123
124	bool * null = 0p; // Re-write every loop since compare thrashes it
125	if( __atomic_load_n(cell, (int)__ATOMIC_RELAXED) == null
126	&& __atomic_compare_exchange_n( cell, &null, handle, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
127	/* paranoid */ verify(i < ready);
128	/* paranoid */ verify( (kernelTLS().sched_id = i, true) );
129	return i;
130	}
131	}
132
133	if(max <= alloc) abort("Trying to create more than %ud processors", __scheduler_lock->max);
134
135	// Step - 2 : F&A to get a new spot in the array.
136	uint_fast32_t n = __atomic_fetch_add(&alloc, 1, __ATOMIC_SEQ_CST);
137	if(max <= n) abort("Trying to create more than %ud processors", __scheduler_lock->max);
138
139	// Step - 3 : Mark space as used and then publish it.
140	data[n] = handle;
141	while() {
142	unsigned copy = n;
143	if( __atomic_load_n(&ready, __ATOMIC_RELAXED) == n
144	&& __atomic_compare_exchange_n(&ready, &copy, n + 1, true, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))
145	break;
146	Pause();
147	}
148
149	__cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p done, id %lu\n", proc, n);
150
151	// Return new spot.
152	/* paranoid */ verify(n < ready);
153	/* paranoid */ verify( (kernelTLS().sched_id = n, true) );
154	return n;
155	}
156
157	void unregister_proc_id( unsigned id ) with(*__scheduler_lock) {
158	/* paranoid */ verify(id < ready);
159	/* paranoid */ verify(id == kernelTLS().sched_id);
160	/* paranoid */ verify(data[id] == &kernelTLS().sched_lock);
161
162	bool * volatile * cell = (bool * volatile *)&data[id]; // Cforall is bugged and the double volatiles causes problems
163
164	__atomic_store_n(cell, 0p, __ATOMIC_RELEASE);
165
166	__cfadbg_print_safe(ready_queue, "Kernel : Unregister proc %p\n", proc);
167	}
168
169	//-----------------------------------------------------------------------
170	// Writer side : acquire when changing the ready queue, e.g. adding more
171	// queues or removing them.
172	uint_fast32_t ready_mutate_lock( void ) with(*__scheduler_lock) {
173	/* paranoid */ verify( ! __preemption_enabled() );
174	/* paranoid */ verify( ! kernelTLS().sched_lock );
175
176	// Step 1 : lock global lock
177	// It is needed to avoid processors that register mid Critical-Section
178	// to simply lock their own lock and enter.
179	__atomic_acquire( &write_lock );
180
181	// Step 2 : lock per-proc lock
182	// Processors that are currently being registered aren't counted
183	// but can't be in read_lock or in the critical section.
184	// All other processors are counted
185	uint_fast32_t s = ready;
186	for(uint_fast32_t i = 0; i < s; i++) {
187	volatile bool * llock = data[i];
188	if(llock) __atomic_acquire( llock );
189	}
190
191	/* paranoid */ verify( ! __preemption_enabled() );
192	return s;
193	}
194
195	void ready_mutate_unlock( uint_fast32_t last_s ) with(*__scheduler_lock) {
196	/* paranoid */ verify( ! __preemption_enabled() );
197
198	// Step 1 : release local locks
199	// This must be done while the global lock is held to avoid
200	// threads that where created mid critical section
201	// to race to lock their local locks and have the writer
202	// immidiately unlock them
203	// Alternative solution : return s in write_lock and pass it to write_unlock
204	for(uint_fast32_t i = 0; i < last_s; i++) {
205	volatile bool * llock = data[i];
206	if(llock) __atomic_store_n(llock, (bool)false, __ATOMIC_RELEASE);
207	}
208
209	// Step 2 : release global lock
210	/paranoid/ assert(true == write_lock);
211	__atomic_store_n(&write_lock, (bool)false, __ATOMIC_RELEASE);
212
213	/* paranoid */ verify( ! __preemption_enabled() );
214	}
215
216	//=======================================================================
217	// Cforall Ready Queue used for scheduling
218	//=======================================================================
219	void ?{}(__ready_queue_t & this) with (this) {
220	#if defined(USE_CPU_WORK_STEALING)
221	lanes.count = cpu_info.hthrd_count * READYQ_SHARD_FACTOR;
222	lanes.data = alloc( lanes.count );
223	lanes.tscs = alloc( lanes.count );
224
225	for( idx; (size_t)lanes.count ) {
226	(lanes.data[idx]){};
227	lanes.tscs[idx].tv = rdtscl();
228	}
229	#else
230	lanes.data = 0p;
231	lanes.tscs = 0p;
232	lanes.count = 0;
233	#endif
234	}
235
236	void ^?{}(__ready_queue_t & this) with (this) {
237	#if !defined(USE_CPU_WORK_STEALING)
238	verify( SEQUENTIAL_SHARD == lanes.count );
239	#endif
240
241	free(lanes.data);
242	free(lanes.tscs);
243	}
244
245	//-----------------------------------------------------------------------
246	#if defined(USE_CPU_WORK_STEALING)
247	__attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd, bool push_local) with (cltr->ready_queue) {
248	__cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
249
250	processor * const proc = kernelTLS().this_processor;
251	const bool external = !push_local \|\| (!proc) \|\| (cltr != proc->cltr);
252
253	const int cpu = __kernel_getcpu();
254	/* paranoid */ verify(cpu >= 0);
255	/* paranoid */ verify(cpu < cpu_info.hthrd_count);
256	/* paranoid / verify(cpu READYQ_SHARD_FACTOR < lanes.count);
257
258	const cpu_map_entry_t & map = cpu_info.llc_map[cpu];
259	/* paranoid / verify(map.start READYQ_SHARD_FACTOR < lanes.count);
260	/* paranoid / verify(map.self READYQ_SHARD_FACTOR < lanes.count);
261	/* 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);
262
263	const int start = map.self * READYQ_SHARD_FACTOR;
264	unsigned i;
265	do {
266	unsigned r;
267	if(unlikely(external)) { r = __tls_rand(); }
268	else { r = proc->rdq.its++; }
269	i = start + (r % READYQ_SHARD_FACTOR);
270	// If we can't lock it retry
271	} while( !__atomic_try_acquire( &lanes.data[i].lock ) );
272
273	// Actually push it
274	push(lanes.data[i], thrd);
275
276	// Unlock and return
277	__atomic_unlock( &lanes.data[i].lock );
278
279	#if !defined(__CFA_NO_STATISTICS__)
280	if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
281	else __tls_stats()->ready.push.local.success++;
282	#endif
283
284	__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);
285
286	}
287
288	// Pop from the ready queue from a given cluster
289	__attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
290	/* paranoid */ verify( lanes.count > 0 );
291	/* paranoid */ verify( kernelTLS().this_processor );
292
293	const int cpu = __kernel_getcpu();
294	/* paranoid */ verify(cpu >= 0);
295	/* paranoid */ verify(cpu < cpu_info.hthrd_count);
296	/* paranoid / verify(cpu READYQ_SHARD_FACTOR < lanes.count);
297
298	const cpu_map_entry_t & map = cpu_info.llc_map[cpu];
299	/* paranoid / verify(map.start READYQ_SHARD_FACTOR < lanes.count);
300	/* paranoid / verify(map.self READYQ_SHARD_FACTOR < lanes.count);
301	/* 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);
302
303	processor * const proc = kernelTLS().this_processor;
304	const int start = map.self * READYQ_SHARD_FACTOR;
305
306	// Did we already have a help target
307	if(proc->rdq.target == -1u) {
308	// if We don't have a
309	unsigned long long min = ts(lanes.data[start]);
310	for(i; READYQ_SHARD_FACTOR) {
311	unsigned long long tsc = ts(lanes.data[start + i]);
312	if(tsc < min) min = tsc;
313	}
314	proc->rdq.cutoff = min;
315	proc->rdq.target = (map.start * READYQ_SHARD_FACTOR) + (__tls_rand() % (map.count* READYQ_SHARD_FACTOR));
316	}
317	else {
318	const unsigned long long bias = 0; //2_500_000_000;
319	const unsigned long long cutoff = proc->rdq.cutoff > bias ? proc->rdq.cutoff - bias : proc->rdq.cutoff;
320	{
321	unsigned target = proc->rdq.target;
322	proc->rdq.target = -1u;
323	if(lanes.tscs[target].tv < cutoff && ts(lanes.data[target]) < cutoff) {
324	$thread * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
325	proc->rdq.last = target;
326	if(t) return t;
327	}
328	}
329
330	unsigned last = proc->rdq.last;
331	if(last != -1u && lanes.tscs[last].tv < cutoff && ts(lanes.data[last]) < cutoff) {
332	$thread * t = try_pop(cltr, last __STATS(, __tls_stats()->ready.pop.help));
333	if(t) return t;
334	}
335	else {
336	proc->rdq.last = -1u;
337	}
338	}
339
340	for(READYQ_SHARD_FACTOR) {
341	unsigned i = start + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
342	if($thread * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
343	}
344
345	// All lanes where empty return 0p
346	return 0p;
347	}
348
349	__attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
350	processor * const proc = kernelTLS().this_processor;
351	unsigned last = proc->rdq.last;
352
353	unsigned i = __tls_rand() % lanes.count;
354	return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
355	}
356	__attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) {
357	return search(cltr);
358	}
359	#endif
360	#if defined(USE_RELAXED_FIFO)
361	//-----------------------------------------------------------------------
362	// get index from random number with or without bias towards queues
363	static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred) {
364	unsigned i;
365	bool local;
366	unsigned rlow = r % BIAS;
367	unsigned rhigh = r / BIAS;
368	if((0 != rlow) && preferred >= 0) {
369	// (BIAS - 1) out of BIAS chances
370	// Use perferred queues
371	i = preferred + (rhigh % READYQ_SHARD_FACTOR);
372	local = true;
373	}
374	else {
375	// 1 out of BIAS chances
376	// Use all queues
377	i = rhigh;
378	local = false;
379	}
380	return [i, local];
381	}
382
383	__attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd, bool push_local) with (cltr->ready_queue) {
384	__cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
385
386	const bool external = !push_local \|\| (!kernelTLS().this_processor) \|\| (cltr != kernelTLS().this_processor->cltr);
387	/* paranoid */ verify(external \|\| kernelTLS().this_processor->rdq.id < lanes.count );
388
389	bool local;
390	int preferred = external ? -1 : kernelTLS().this_processor->rdq.id;
391
392	// Try to pick a lane and lock it
393	unsigned i;
394	do {
395	// Pick the index of a lane
396	unsigned r = __tls_rand_fwd();
397	[i, local] = idx_from_r(r, preferred);
398
399	i %= __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
400
401	#if !defined(__CFA_NO_STATISTICS__)
402	if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
403	else if(local) __tls_stats()->ready.push.local.attempt++;
404	else __tls_stats()->ready.push.share.attempt++;
405	#endif
406
407	// If we can't lock it retry
408	} while( !__atomic_try_acquire( &lanes.data[i].lock ) );
409
410	// Actually push it
411	push(lanes.data[i], thrd);
412
413	// Unlock and return
414	__atomic_unlock( &lanes.data[i].lock );
415
416	// Mark the current index in the tls rng instance as having an item
417	__tls_rand_advance_bck();
418
419	__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);
420
421	// Update statistics
422	#if !defined(__CFA_NO_STATISTICS__)
423	if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
424	else if(local) __tls_stats()->ready.push.local.success++;
425	else __tls_stats()->ready.push.share.success++;
426	#endif
427	}
428
429	// Pop from the ready queue from a given cluster
430	__attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
431	/* paranoid */ verify( lanes.count > 0 );
432	/* paranoid */ verify( kernelTLS().this_processor );
433	/* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
434
435	unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
436	int preferred = kernelTLS().this_processor->rdq.id;
437
438
439	// As long as the list is not empty, try finding a lane that isn't empty and pop from it
440	for(25) {
441	// Pick two lists at random
442	unsigned ri = __tls_rand_bck();
443	unsigned rj = __tls_rand_bck();
444
445	unsigned i, j;
446	__attribute__((unused)) bool locali, localj;
447	[i, locali] = idx_from_r(ri, preferred);
448	[j, localj] = idx_from_r(rj, preferred);
449
450	i %= count;
451	j %= count;
452
453	// try popping from the 2 picked lists
454	struct $thread * thrd = try_pop(cltr, i, j __STATS(, *(locali \|\| localj ? &__tls_stats()->ready.pop.local : &__tls_stats()->ready.pop.help)));
455	if(thrd) {
456	return thrd;
457	}
458	}
459
460	// All lanes where empty return 0p
461	return 0p;
462	}
463
464	__attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) { return pop_fast(cltr); }
465	__attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) {
466	return search(cltr);
467	}
468	#endif
469	#if defined(USE_WORK_STEALING)
470	__attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd, bool push_local) with (cltr->ready_queue) {
471	__cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
472
473	// #define USE_PREFERRED
474	#if !defined(USE_PREFERRED)
475	const bool external = !push_local \|\| (!kernelTLS().this_processor) \|\| (cltr != kernelTLS().this_processor->cltr);
476	/* paranoid */ verify(external \|\| kernelTLS().this_processor->rdq.id < lanes.count );
477	#else
478	unsigned preferred = thrd->preferred;
479	const bool external = push_local \|\| (!kernelTLS().this_processor) \|\| preferred == -1u \|\| thrd->curr_cluster != cltr;
480	/* paranoid */ verifyf(external \|\| preferred < lanes.count, "Invalid preferred queue %u for %u lanes", preferred, lanes.count );
481
482	unsigned r = preferred % READYQ_SHARD_FACTOR;
483	const unsigned start = preferred - r;
484	#endif
485
486	// Try to pick a lane and lock it
487	unsigned i;
488	do {
489	#if !defined(__CFA_NO_STATISTICS__)
490	if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
491	else __tls_stats()->ready.push.local.attempt++;
492	#endif
493
494	if(unlikely(external)) {
495	i = __tls_rand() % lanes.count;
496	}
497	else {
498	#if !defined(USE_PREFERRED)
499	processor * proc = kernelTLS().this_processor;
500	unsigned r = proc->rdq.its++;
501	i = proc->rdq.id + (r % READYQ_SHARD_FACTOR);
502	#else
503	i = start + (r++ % READYQ_SHARD_FACTOR);
504	#endif
505	}
506	// If we can't lock it retry
507	} while( !__atomic_try_acquire( &lanes.data[i].lock ) );
508
509	// Actually push it
510	push(lanes.data[i], thrd);
511
512	// Unlock and return
513	__atomic_unlock( &lanes.data[i].lock );
514
515	#if !defined(__CFA_NO_STATISTICS__)
516	if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
517	else __tls_stats()->ready.push.local.success++;
518	#endif
519
520	__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);
521	}
522
523	// Pop from the ready queue from a given cluster
524	__attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
525	/* paranoid */ verify( lanes.count > 0 );
526	/* paranoid */ verify( kernelTLS().this_processor );
527	/* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
528
529	processor * proc = kernelTLS().this_processor;
530
531	if(proc->rdq.target == -1u) {
532	unsigned long long min = ts(lanes.data[proc->rdq.id]);
533	for(int i = 0; i < READYQ_SHARD_FACTOR; i++) {
534	unsigned long long tsc = ts(lanes.data[proc->rdq.id + i]);
535	if(tsc < min) min = tsc;
536	}
537	proc->rdq.cutoff = min;
538	proc->rdq.target = __tls_rand() % lanes.count;
539	}
540	else {
541	unsigned target = proc->rdq.target;
542	proc->rdq.target = -1u;
543	const unsigned long long bias = 0; //2_500_000_000;
544	const unsigned long long cutoff = proc->rdq.cutoff > bias ? proc->rdq.cutoff - bias : proc->rdq.cutoff;
545	if(lanes.tscs[target].tv < cutoff && ts(lanes.data[target]) < cutoff) {
546	$thread * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
547	if(t) return t;
548	}
549	}
550
551	for(READYQ_SHARD_FACTOR) {
552	unsigned i = proc->rdq.id + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
553	if($thread * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
554	}
555	return 0p;
556	}
557
558	__attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
559	unsigned i = __tls_rand() % lanes.count;
560	return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
561	}
562
563	__attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) with (cltr->ready_queue) {
564	return search(cltr);
565	}
566	#endif
567
568	//=======================================================================
569	// Various Ready Queue utilities
570	//=======================================================================
571	// these function work the same or almost the same
572	// whether they are using work-stealing or relaxed fifo scheduling
573
574	//-----------------------------------------------------------------------
575	// try to pop from a lane given by index w
576	static inline struct $thread * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
577	__STATS( stats.attempt++; )
578
579	// Get relevant elements locally
580	__intrusive_lane_t & lane = lanes.data[w];
581
582	// If list looks empty retry
583	if( is_empty(lane) ) {
584	return 0p;
585	}
586
587	// If we can't get the lock retry
588	if( !__atomic_try_acquire(&lane.lock) ) {
589	return 0p;
590	}
591
592	// If list is empty, unlock and retry
593	if( is_empty(lane) ) {
594	__atomic_unlock(&lane.lock);
595	return 0p;
596	}
597
598	// Actually pop the list
599	struct $thread * thrd;
600	unsigned long long tsv;
601	[thrd, tsv] = pop(lane);
602
603	/* paranoid */ verify(thrd);
604	/* paranoid */ verify(tsv);
605	/* paranoid */ verify(lane.lock);
606
607	// Unlock and return
608	__atomic_unlock(&lane.lock);
609
610	// Update statistics
611	__STATS( stats.success++; )
612
613	#if defined(USE_WORK_STEALING)
614	lanes.tscs[w].tv = tsv;
615	#endif
616
617	thrd->preferred = w;
618
619	// return the popped thread
620	return thrd;
621	}
622
623	//-----------------------------------------------------------------------
624	// try to pop from any lanes making sure you don't miss any threads push
625	// before the start of the function
626	static inline struct $thread * search(struct cluster * cltr) with (cltr->ready_queue) {
627	/* paranoid */ verify( lanes.count > 0 );
628	unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
629	unsigned offset = __tls_rand();
630	for(i; count) {
631	unsigned idx = (offset + i) % count;
632	struct $thread * thrd = try_pop(cltr, idx __STATS(, __tls_stats()->ready.pop.search));
633	if(thrd) {
634	return thrd;
635	}
636	}
637
638	// All lanes where empty return 0p
639	return 0p;
640	}
641
642	//-----------------------------------------------------------------------
643	// Check that all the intrusive queues in the data structure are still consistent
644	static void check( __ready_queue_t & q ) with (q) {
645	#if defined(__CFA_WITH_VERIFY__)
646	{
647	for( idx ; lanes.count ) {
648	__intrusive_lane_t & sl = lanes.data[idx];
649	assert(!lanes.data[idx].lock);
650
651	if(is_empty(sl)) {
652	assert( sl.anchor.next == 0p );
653	assert( sl.anchor.ts == 0 );
654	assert( mock_head(sl) == sl.prev );
655	} else {
656	assert( sl.anchor.next != 0p );
657	assert( sl.anchor.ts != 0 );
658	assert( mock_head(sl) != sl.prev );
659	}
660	}
661	}
662	#endif
663	}
664
665	//-----------------------------------------------------------------------
666	// Given 2 indexes, pick the list with the oldest push an try to pop from it
667	static inline struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
668	// Pick the bet list
669	int w = i;
670	if( __builtin_expect(!is_empty(lanes.data[j]), true) ) {
671	w = (ts(lanes.data[i]) < ts(lanes.data[j])) ? i : j;
672	}
673
674	return try_pop(cltr, w __STATS(, stats));
675	}
676
677	// Call this function of the intrusive list was moved using memcpy
678	// fixes the list so that the pointers back to anchors aren't left dangling
679	static inline void fix(__intrusive_lane_t & ll) {
680	if(is_empty(ll)) {
681	verify(ll.anchor.next == 0p);
682	ll.prev = mock_head(ll);
683	}
684	}
685
686	static void assign_list(unsigned & value, dlist(processor) & list, unsigned count) {
687	processor * it = &list`first;
688	for(unsigned i = 0; i < count; i++) {
689	/* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
690	it->rdq.id = value;
691	it->rdq.target = -1u;
692	value += READYQ_SHARD_FACTOR;
693	it = &(*it)`next;
694	}
695	}
696
697	static void reassign_cltr_id(struct cluster * cltr) {
698	unsigned preferred = 0;
699	assign_list(preferred, cltr->procs.actives, cltr->procs.total - cltr->procs.idle);
700	assign_list(preferred, cltr->procs.idles , cltr->procs.idle );
701	}
702
703	static void fix_times( struct cluster * cltr ) with( cltr->ready_queue ) {
704	#if defined(USE_WORK_STEALING)
705	lanes.tscs = alloc(lanes.count, lanes.tscs`realloc);
706	for(i; lanes.count) {
707	unsigned long long tsc = ts(lanes.data[i]);
708	lanes.tscs[i].tv = tsc != 0 ? tsc : rdtscl();
709	}
710	#endif
711	}
712
713	#if defined(USE_CPU_WORK_STEALING)
714	// ready_queue size is fixed in this case
715	void ready_queue_grow(struct cluster * cltr) {}
716	void ready_queue_shrink(struct cluster * cltr) {}
717	#else
718	// Grow the ready queue
719	void ready_queue_grow(struct cluster * cltr) {
720	size_t ncount;
721	int target = cltr->procs.total;
722
723	/* paranoid */ verify( ready_mutate_islocked() );
724	__cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue\n");
725
726	// Make sure that everything is consistent
727	/* paranoid */ check( cltr->ready_queue );
728
729	// grow the ready queue
730	with( cltr->ready_queue ) {
731	// Find new count
732	// Make sure we always have atleast 1 list
733	if(target >= 2) {
734	ncount = target * READYQ_SHARD_FACTOR;
735	} else {
736	ncount = SEQUENTIAL_SHARD;
737	}
738
739	// Allocate new array (uses realloc and memcpies the data)
740	lanes.data = alloc( ncount, lanes.data`realloc );
741
742	// Fix the moved data
743	for( idx; (size_t)lanes.count ) {
744	fix(lanes.data[idx]);
745	}
746
747	// Construct new data
748	for( idx; (size_t)lanes.count ~ ncount) {
749	(lanes.data[idx]){};
750	}
751
752	// Update original
753	lanes.count = ncount;
754	}
755
756	fix_times(cltr);
757
758	reassign_cltr_id(cltr);
759
760	// Make sure that everything is consistent
761	/* paranoid */ check( cltr->ready_queue );
762
763	__cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue done\n");
764
765	/* paranoid */ verify( ready_mutate_islocked() );
766	}
767
768	// Shrink the ready queue
769	void ready_queue_shrink(struct cluster * cltr) {
770	/* paranoid */ verify( ready_mutate_islocked() );
771	__cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue\n");
772
773	// Make sure that everything is consistent
774	/* paranoid */ check( cltr->ready_queue );
775
776	int target = cltr->procs.total;
777
778	with( cltr->ready_queue ) {
779	// Remember old count
780	size_t ocount = lanes.count;
781
782	// Find new count
783	// Make sure we always have atleast 1 list
784	lanes.count = target >= 2 ? target * READYQ_SHARD_FACTOR: SEQUENTIAL_SHARD;
785	/* paranoid */ verify( ocount >= lanes.count );
786	/* paranoid / verify( lanes.count == target READYQ_SHARD_FACTOR \|\| target < 2 );
787
788	// for printing count the number of displaced threads
789	#if defined(__CFA_DEBUG_PRINT__) \|\| defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
790	__attribute__((unused)) size_t displaced = 0;
791	#endif
792
793	// redistribute old data
794	for( idx; (size_t)lanes.count ~ ocount) {
795	// Lock is not strictly needed but makes checking invariants much easier
796	__attribute__((unused)) bool locked = __atomic_try_acquire(&lanes.data[idx].lock);
797	verify(locked);
798
799	// As long as we can pop from this lane to push the threads somewhere else in the queue
800	while(!is_empty(lanes.data[idx])) {
801	struct $thread * thrd;
802	unsigned long long _;
803	[thrd, _] = pop(lanes.data[idx]);
804
805	push(cltr, thrd, true);
806
807	// for printing count the number of displaced threads
808	#if defined(__CFA_DEBUG_PRINT__) \|\| defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
809	displaced++;
810	#endif
811	}
812
813	// Unlock the lane
814	__atomic_unlock(&lanes.data[idx].lock);
815
816	// TODO print the queue statistics here
817
818	^(lanes.data[idx]){};
819	}
820
821	__cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue displaced %zu threads\n", displaced);
822
823	// Allocate new array (uses realloc and memcpies the data)
824	lanes.data = alloc( lanes.count, lanes.data`realloc );
825
826	// Fix the moved data
827	for( idx; (size_t)lanes.count ) {
828	fix(lanes.data[idx]);
829	}
830	}
831
832	fix_times(cltr);
833
834	reassign_cltr_id(cltr);
835
836	// Make sure that everything is consistent
837	/* paranoid */ check( cltr->ready_queue );
838
839	__cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue done\n");
840	/* paranoid */ verify( ready_mutate_islocked() );
841	}
842	#endif
843
844	#if !defined(__CFA_NO_STATISTICS__)
845	unsigned cnt(const __ready_queue_t & this, unsigned idx) {
846	/* paranoid */ verify(this.lanes.count > idx);
847	return this.lanes.data[idx].cnt;
848	}
849	#endif

Note: See TracBrowser for help on using the repository browser.

Download in other formats:

Original Format