Context Navigation

ready_queue.cfa @ 44856ed

ADTast-experimentalenumforall-pointer-decayjacob/cs343-translationnew-ast-unique-exprpthread-emulationqualifiedEnum

Last change on this file since 44856ed was 6ba6846, checked in by Thierry Delisle <tdelisle@…>, 3 years ago
Minor ready-queue fixes
Property mode set to `100644`
File size: 31.3 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	// #define USE_CPU_WORK_STEALING
25
26	#include "bits/defs.hfa"
27	#include "device/cpu.hfa"
28	#include "kernel_private.hfa"
29
30	#include "stdlib.hfa"
31	#include "math.hfa"
32
33	#include <errno.h>
34	#include <unistd.h>
35
36	extern "C" {
37	#include <sys/syscall.h> // __NR_xxx
38	}
39
40	#include "ready_subqueue.hfa"
41
42	static const size_t cache_line_size = 64;
43
44	#if !defined(__CFA_NO_STATISTICS__)
45	#define __STATS(...) __VA_ARGS__
46	#else
47	#define __STATS(...)
48	#endif
49
50	// No overriden function, no environment variable, no define
51	// fall back to a magic number
52	#ifndef __CFA_MAX_PROCESSORS__
53	#define __CFA_MAX_PROCESSORS__ 1024
54	#endif
55
56	#if defined(USE_CPU_WORK_STEALING)
57	#define READYQ_SHARD_FACTOR 2
58	#elif defined(USE_RELAXED_FIFO)
59	#define BIAS 4
60	#define READYQ_SHARD_FACTOR 4
61	#define SEQUENTIAL_SHARD 1
62	#elif defined(USE_WORK_STEALING)
63	#define READYQ_SHARD_FACTOR 2
64	#define SEQUENTIAL_SHARD 2
65	#else
66	#error no scheduling strategy selected
67	#endif
68
69	static inline struct $thread * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats));
70	static inline struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats));
71	static inline struct $thread * search(struct cluster * cltr);
72	static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred);
73
74
75	// returns the maximum number of processors the RWLock support
76	__attribute__((weak)) unsigned __max_processors() {
77	const char * max_cores_s = getenv("CFA_MAX_PROCESSORS");
78	if(!max_cores_s) {
79	__cfadbg_print_nolock(ready_queue, "No CFA_MAX_PROCESSORS in ENV\n");
80	return __CFA_MAX_PROCESSORS__;
81	}
82
83	char * endptr = 0p;
84	long int max_cores_l = strtol(max_cores_s, &endptr, 10);
85	if(max_cores_l < 1 \|\| max_cores_l > 65535) {
86	__cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS out of range : %ld\n", max_cores_l);
87	return __CFA_MAX_PROCESSORS__;
88	}
89	if('\0' != *endptr) {
90	__cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS not a decimal number : %s\n", max_cores_s);
91	return __CFA_MAX_PROCESSORS__;
92	}
93
94	return max_cores_l;
95	}
96
97	#if defined(CFA_HAVE_LINUX_LIBRSEQ)
98	// No forward declaration needed
99	#define __kernel_rseq_register rseq_register_current_thread
100	#define __kernel_rseq_unregister rseq_unregister_current_thread
101	#elif defined(CFA_HAVE_LINUX_RSEQ_H)
102	void __kernel_raw_rseq_register (void);
103	void __kernel_raw_rseq_unregister(void);
104
105	#define __kernel_rseq_register __kernel_raw_rseq_register
106	#define __kernel_rseq_unregister __kernel_raw_rseq_unregister
107	#else
108	// No forward declaration needed
109	// No initialization needed
110	static inline void noop(void) {}
111
112	#define __kernel_rseq_register noop
113	#define __kernel_rseq_unregister noop
114	#endif
115
116	//=======================================================================
117	// Cluster wide reader-writer lock
118	//=======================================================================
119	void ?{}(__scheduler_RWLock_t & this) {
120	this.max = __max_processors();
121	this.alloc = 0;
122	this.ready = 0;
123	this.data = alloc(this.max);
124	this.write_lock = false;
125
126	/paranoid/ verify(__atomic_is_lock_free(sizeof(this.alloc), &this.alloc));
127	/paranoid/ verify(__atomic_is_lock_free(sizeof(this.ready), &this.ready));
128
129	}
130	void ^?{}(__scheduler_RWLock_t & this) {
131	free(this.data);
132	}
133
134
135	//=======================================================================
136	// Lock-Free registering/unregistering of threads
137	unsigned register_proc_id( void ) with(*__scheduler_lock) {
138	__kernel_rseq_register();
139
140	__cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p for RW-Lock\n", proc);
141	bool * handle = (bool *)&kernelTLS().sched_lock;
142
143	// Step - 1 : check if there is already space in the data
144	uint_fast32_t s = ready;
145
146	// Check among all the ready
147	for(uint_fast32_t i = 0; i < s; i++) {
148	bool * volatile * cell = (bool * volatile *)&data[i]; // Cforall is bugged and the double volatiles causes problems
149	/* paranoid / verify( handle != cell );
150
151	bool * null = 0p; // Re-write every loop since compare thrashes it
152	if( __atomic_load_n(cell, (int)__ATOMIC_RELAXED) == null
153	&& __atomic_compare_exchange_n( cell, &null, handle, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
154	/* paranoid */ verify(i < ready);
155	/* paranoid */ verify( (kernelTLS().sched_id = i, true) );
156	return i;
157	}
158	}
159
160	if(max <= alloc) abort("Trying to create more than %ud processors", __scheduler_lock->max);
161
162	// Step - 2 : F&A to get a new spot in the array.
163	uint_fast32_t n = __atomic_fetch_add(&alloc, 1, __ATOMIC_SEQ_CST);
164	if(max <= n) abort("Trying to create more than %ud processors", __scheduler_lock->max);
165
166	// Step - 3 : Mark space as used and then publish it.
167	data[n] = handle;
168	while() {
169	unsigned copy = n;
170	if( __atomic_load_n(&ready, __ATOMIC_RELAXED) == n
171	&& __atomic_compare_exchange_n(&ready, &copy, n + 1, true, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))
172	break;
173	Pause();
174	}
175
176	__cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p done, id %lu\n", proc, n);
177
178	// Return new spot.
179	/* paranoid */ verify(n < ready);
180	/* paranoid */ verify( (kernelTLS().sched_id = n, true) );
181	return n;
182	}
183
184	void unregister_proc_id( unsigned id ) with(*__scheduler_lock) {
185	/* paranoid */ verify(id < ready);
186	/* paranoid */ verify(id == kernelTLS().sched_id);
187	/* paranoid */ verify(data[id] == &kernelTLS().sched_lock);
188
189	bool * volatile * cell = (bool * volatile *)&data[id]; // Cforall is bugged and the double volatiles causes problems
190
191	__atomic_store_n(cell, 0p, __ATOMIC_RELEASE);
192
193	__cfadbg_print_safe(ready_queue, "Kernel : Unregister proc %p\n", proc);
194
195	__kernel_rseq_unregister();
196	}
197
198	//-----------------------------------------------------------------------
199	// Writer side : acquire when changing the ready queue, e.g. adding more
200	// queues or removing them.
201	uint_fast32_t ready_mutate_lock( void ) with(*__scheduler_lock) {
202	/* paranoid */ verify( ! __preemption_enabled() );
203	/* paranoid */ verify( ! kernelTLS().sched_lock );
204
205	// Step 1 : lock global lock
206	// It is needed to avoid processors that register mid Critical-Section
207	// to simply lock their own lock and enter.
208	__atomic_acquire( &write_lock );
209
210	// Step 2 : lock per-proc lock
211	// Processors that are currently being registered aren't counted
212	// but can't be in read_lock or in the critical section.
213	// All other processors are counted
214	uint_fast32_t s = ready;
215	for(uint_fast32_t i = 0; i < s; i++) {
216	volatile bool * llock = data[i];
217	if(llock) __atomic_acquire( llock );
218	}
219
220	/* paranoid */ verify( ! __preemption_enabled() );
221	return s;
222	}
223
224	void ready_mutate_unlock( uint_fast32_t last_s ) with(*__scheduler_lock) {
225	/* paranoid */ verify( ! __preemption_enabled() );
226
227	// Step 1 : release local locks
228	// This must be done while the global lock is held to avoid
229	// threads that where created mid critical section
230	// to race to lock their local locks and have the writer
231	// immidiately unlock them
232	// Alternative solution : return s in write_lock and pass it to write_unlock
233	for(uint_fast32_t i = 0; i < last_s; i++) {
234	volatile bool * llock = data[i];
235	if(llock) __atomic_store_n(llock, (bool)false, __ATOMIC_RELEASE);
236	}
237
238	// Step 2 : release global lock
239	/paranoid/ assert(true == write_lock);
240	__atomic_store_n(&write_lock, (bool)false, __ATOMIC_RELEASE);
241
242	/* paranoid */ verify( ! __preemption_enabled() );
243	}
244
245	//=======================================================================
246	// Cforall Ready Queue used for scheduling
247	//=======================================================================
248	void ?{}(__ready_queue_t & this) with (this) {
249	#if defined(USE_CPU_WORK_STEALING)
250	lanes.count = cpu_info.hthrd_count * READYQ_SHARD_FACTOR;
251	lanes.data = alloc( lanes.count );
252	lanes.tscs = alloc( lanes.count );
253
254	for( idx; (size_t)lanes.count ) {
255	(lanes.data[idx]){};
256	lanes.tscs[idx].tv = rdtscl();
257	}
258	#else
259	lanes.data = 0p;
260	lanes.tscs = 0p;
261	lanes.count = 0;
262	#endif
263	}
264
265	void ^?{}(__ready_queue_t & this) with (this) {
266	#if !defined(USE_CPU_WORK_STEALING)
267	verify( SEQUENTIAL_SHARD == lanes.count );
268	#endif
269
270	free(lanes.data);
271	free(lanes.tscs);
272	}
273
274	//-----------------------------------------------------------------------
275	#if defined(USE_CPU_WORK_STEALING)
276	__attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd, bool push_local) with (cltr->ready_queue) {
277	__cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
278
279	processor * const proc = kernelTLS().this_processor;
280	const bool external = !push_local \|\| (!proc) \|\| (cltr != proc->cltr);
281
282	const int cpu = __kernel_getcpu();
283	/* paranoid */ verify(cpu >= 0);
284	/* paranoid */ verify(cpu < cpu_info.hthrd_count);
285	/* paranoid / verify(cpu READYQ_SHARD_FACTOR < lanes.count);
286
287	const cpu_map_entry_t & map = cpu_info.llc_map[cpu];
288	/* paranoid / verify(map.start READYQ_SHARD_FACTOR < lanes.count);
289	/* paranoid / verify(map.self READYQ_SHARD_FACTOR < lanes.count);
290	/* paranoid / verifyf((map.start + map.count) READYQ_SHARD_FACTOR <= lanes.count, "have %zu lanes but map can go up to %u", lanes.count, (map.start + map.count) * READYQ_SHARD_FACTOR);
291
292	const int start = map.self * READYQ_SHARD_FACTOR;
293	unsigned i;
294	do {
295	unsigned r;
296	if(unlikely(external)) { r = __tls_rand(); }
297	else { r = proc->rdq.its++; }
298	i = start + (r % READYQ_SHARD_FACTOR);
299	// If we can't lock it retry
300	} while( !__atomic_try_acquire( &lanes.data[i].lock ) );
301
302	// Actually push it
303	push(lanes.data[i], thrd);
304
305	// Unlock and return
306	__atomic_unlock( &lanes.data[i].lock );
307
308	#if !defined(__CFA_NO_STATISTICS__)
309	if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
310	else __tls_stats()->ready.push.local.success++;
311	#endif
312
313	__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);
314
315	}
316
317	// Pop from the ready queue from a given cluster
318	__attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
319	/* paranoid */ verify( lanes.count > 0 );
320	/* paranoid */ verify( kernelTLS().this_processor );
321
322	const int cpu = __kernel_getcpu();
323	/* paranoid */ verify(cpu >= 0);
324	/* paranoid */ verify(cpu < cpu_info.hthrd_count);
325	/* paranoid / verify(cpu READYQ_SHARD_FACTOR < lanes.count);
326
327	const cpu_map_entry_t & map = cpu_info.llc_map[cpu];
328	/* paranoid / verify(map.start READYQ_SHARD_FACTOR < lanes.count);
329	/* paranoid / verify(map.self READYQ_SHARD_FACTOR < lanes.count);
330	/* paranoid / verifyf((map.start + map.count) READYQ_SHARD_FACTOR <= lanes.count, "have %zu lanes but map can go up to %u", lanes.count, (map.start + map.count) * READYQ_SHARD_FACTOR);
331
332	processor * const proc = kernelTLS().this_processor;
333	const int start = map.self * READYQ_SHARD_FACTOR;
334
335	// Did we already have a help target
336	if(proc->rdq.target == -1u) {
337	// if We don't have a
338	unsigned long long min = ts(lanes.data[start]);
339	for(i; READYQ_SHARD_FACTOR) {
340	unsigned long long tsc = ts(lanes.data[start + i]);
341	if(tsc < min) min = tsc;
342	}
343	proc->rdq.cutoff = min;
344
345	/* paranoid */ verify(lanes.count < 65536); // The following code assumes max 65536 cores.
346	/* paranoid */ verify(map.count < 65536); // The following code assumes max 65536 cores.
347	uint64_t chaos = __tls_rand();
348	uint64_t high_chaos = (chaos >> 32);
349	uint64_t mid_chaos = (chaos >> 16) & 0xffff;
350	uint64_t low_chaos = chaos & 0xffff;
351
352	unsigned me = map.self;
353	unsigned cpu_chaos = map.start + (mid_chaos % map.count);
354	bool global = cpu_chaos == me;
355
356	if(global) {
357	proc->rdq.target = high_chaos % lanes.count;
358	} else {
359	proc->rdq.target = (cpu_chaos * READYQ_SHARD_FACTOR) + (low_chaos % READYQ_SHARD_FACTOR);
360	/* paranoid / verify(proc->rdq.target >= (map.start READYQ_SHARD_FACTOR));
361	/* paranoid / verify(proc->rdq.target < ((map.start + map.count) READYQ_SHARD_FACTOR));
362	}
363
364	/* paranoid */ verify(proc->rdq.target != -1u);
365	}
366	else {
367	const unsigned long long bias = 0; //2_500_000_000;
368	const unsigned long long cutoff = proc->rdq.cutoff > bias ? proc->rdq.cutoff - bias : proc->rdq.cutoff;
369	{
370	unsigned target = proc->rdq.target;
371	proc->rdq.target = -1u;
372	if(lanes.tscs[target].tv < cutoff && ts(lanes.data[target]) < cutoff) {
373	$thread * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
374	proc->rdq.last = target;
375	if(t) return t;
376	}
377	}
378
379	unsigned last = proc->rdq.last;
380	if(last != -1u && lanes.tscs[last].tv < cutoff && ts(lanes.data[last]) < cutoff) {
381	$thread * t = try_pop(cltr, last __STATS(, __tls_stats()->ready.pop.help));
382	if(t) return t;
383	}
384	else {
385	proc->rdq.last = -1u;
386	}
387	}
388
389	for(READYQ_SHARD_FACTOR) {
390	unsigned i = start + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
391	if($thread * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
392	}
393
394	// All lanes where empty return 0p
395	return 0p;
396	}
397
398	__attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
399	processor * const proc = kernelTLS().this_processor;
400	unsigned last = proc->rdq.last;
401	if(last != -1u) {
402	struct $thread * t = try_pop(cltr, last __STATS(, __tls_stats()->ready.pop.steal));
403	if(t) return t;
404	proc->rdq.last = -1u;
405	}
406
407	unsigned i = __tls_rand() % lanes.count;
408	return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
409	}
410	__attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) {
411	return search(cltr);
412	}
413	#endif
414	#if defined(USE_RELAXED_FIFO)
415	//-----------------------------------------------------------------------
416	// get index from random number with or without bias towards queues
417	static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred) {
418	unsigned i;
419	bool local;
420	unsigned rlow = r % BIAS;
421	unsigned rhigh = r / BIAS;
422	if((0 != rlow) && preferred >= 0) {
423	// (BIAS - 1) out of BIAS chances
424	// Use perferred queues
425	i = preferred + (rhigh % READYQ_SHARD_FACTOR);
426	local = true;
427	}
428	else {
429	// 1 out of BIAS chances
430	// Use all queues
431	i = rhigh;
432	local = false;
433	}
434	return [i, local];
435	}
436
437	__attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd, bool push_local) with (cltr->ready_queue) {
438	__cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
439
440	const bool external = !push_local \|\| (!kernelTLS().this_processor) \|\| (cltr != kernelTLS().this_processor->cltr);
441	/* paranoid */ verify(external \|\| kernelTLS().this_processor->rdq.id < lanes.count );
442
443	bool local;
444	int preferred = external ? -1 : kernelTLS().this_processor->rdq.id;
445
446	// Try to pick a lane and lock it
447	unsigned i;
448	do {
449	// Pick the index of a lane
450	unsigned r = __tls_rand_fwd();
451	[i, local] = idx_from_r(r, preferred);
452
453	i %= __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
454
455	#if !defined(__CFA_NO_STATISTICS__)
456	if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
457	else if(local) __tls_stats()->ready.push.local.attempt++;
458	else __tls_stats()->ready.push.share.attempt++;
459	#endif
460
461	// If we can't lock it retry
462	} while( !__atomic_try_acquire( &lanes.data[i].lock ) );
463
464	// Actually push it
465	push(lanes.data[i], thrd);
466
467	// Unlock and return
468	__atomic_unlock( &lanes.data[i].lock );
469
470	// Mark the current index in the tls rng instance as having an item
471	__tls_rand_advance_bck();
472
473	__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);
474
475	// Update statistics
476	#if !defined(__CFA_NO_STATISTICS__)
477	if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
478	else if(local) __tls_stats()->ready.push.local.success++;
479	else __tls_stats()->ready.push.share.success++;
480	#endif
481	}
482
483	// Pop from the ready queue from a given cluster
484	__attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
485	/* paranoid */ verify( lanes.count > 0 );
486	/* paranoid */ verify( kernelTLS().this_processor );
487	/* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
488
489	unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
490	int preferred = kernelTLS().this_processor->rdq.id;
491
492
493	// As long as the list is not empty, try finding a lane that isn't empty and pop from it
494	for(25) {
495	// Pick two lists at random
496	unsigned ri = __tls_rand_bck();
497	unsigned rj = __tls_rand_bck();
498
499	unsigned i, j;
500	__attribute__((unused)) bool locali, localj;
501	[i, locali] = idx_from_r(ri, preferred);
502	[j, localj] = idx_from_r(rj, preferred);
503
504	i %= count;
505	j %= count;
506
507	// try popping from the 2 picked lists
508	struct $thread * thrd = try_pop(cltr, i, j __STATS(, *(locali \|\| localj ? &__tls_stats()->ready.pop.local : &__tls_stats()->ready.pop.help)));
509	if(thrd) {
510	return thrd;
511	}
512	}
513
514	// All lanes where empty return 0p
515	return 0p;
516	}
517
518	__attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) { return pop_fast(cltr); }
519	__attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) {
520	return search(cltr);
521	}
522	#endif
523	#if defined(USE_WORK_STEALING)
524	__attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd, bool push_local) with (cltr->ready_queue) {
525	__cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
526
527	// #define USE_PREFERRED
528	#if !defined(USE_PREFERRED)
529	const bool external = !push_local \|\| (!kernelTLS().this_processor) \|\| (cltr != kernelTLS().this_processor->cltr);
530	/* paranoid */ verify(external \|\| kernelTLS().this_processor->rdq.id < lanes.count );
531	#else
532	unsigned preferred = thrd->preferred;
533	const bool external = push_local \|\| (!kernelTLS().this_processor) \|\| preferred == -1u \|\| thrd->curr_cluster != cltr;
534	/* paranoid */ verifyf(external \|\| preferred < lanes.count, "Invalid preferred queue %u for %u lanes", preferred, lanes.count );
535
536	unsigned r = preferred % READYQ_SHARD_FACTOR;
537	const unsigned start = preferred - r;
538	#endif
539
540	// Try to pick a lane and lock it
541	unsigned i;
542	do {
543	#if !defined(__CFA_NO_STATISTICS__)
544	if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
545	else __tls_stats()->ready.push.local.attempt++;
546	#endif
547
548	if(unlikely(external)) {
549	i = __tls_rand() % lanes.count;
550	}
551	else {
552	#if !defined(USE_PREFERRED)
553	processor * proc = kernelTLS().this_processor;
554	unsigned r = proc->rdq.its++;
555	i = proc->rdq.id + (r % READYQ_SHARD_FACTOR);
556	#else
557	i = start + (r++ % READYQ_SHARD_FACTOR);
558	#endif
559	}
560	// If we can't lock it retry
561	} while( !__atomic_try_acquire( &lanes.data[i].lock ) );
562
563	// Actually push it
564	push(lanes.data[i], thrd);
565
566	// Unlock and return
567	__atomic_unlock( &lanes.data[i].lock );
568
569	#if !defined(__CFA_NO_STATISTICS__)
570	if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
571	else __tls_stats()->ready.push.local.success++;
572	#endif
573
574	__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);
575	}
576
577	// Pop from the ready queue from a given cluster
578	__attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
579	/* paranoid */ verify( lanes.count > 0 );
580	/* paranoid */ verify( kernelTLS().this_processor );
581	/* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
582
583	processor * proc = kernelTLS().this_processor;
584
585	if(proc->rdq.target == -1u) {
586	unsigned long long min = ts(lanes.data[proc->rdq.id]);
587	for(int i = 0; i < READYQ_SHARD_FACTOR; i++) {
588	unsigned long long tsc = ts(lanes.data[proc->rdq.id + i]);
589	if(tsc < min) min = tsc;
590	}
591	proc->rdq.cutoff = min;
592	proc->rdq.target = __tls_rand() % lanes.count;
593	}
594	else {
595	unsigned target = proc->rdq.target;
596	proc->rdq.target = -1u;
597	const unsigned long long bias = 0; //2_500_000_000;
598	const unsigned long long cutoff = proc->rdq.cutoff > bias ? proc->rdq.cutoff - bias : proc->rdq.cutoff;
599	if(lanes.tscs[target].tv < cutoff && ts(lanes.data[target]) < cutoff) {
600	$thread * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
601	if(t) return t;
602	}
603	}
604
605	for(READYQ_SHARD_FACTOR) {
606	unsigned i = proc->rdq.id + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
607	if($thread * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
608	}
609	return 0p;
610	}
611
612	__attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
613	unsigned i = __tls_rand() % lanes.count;
614	return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
615	}
616
617	__attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) with (cltr->ready_queue) {
618	return search(cltr);
619	}
620	#endif
621
622	//=======================================================================
623	// Various Ready Queue utilities
624	//=======================================================================
625	// these function work the same or almost the same
626	// whether they are using work-stealing or relaxed fifo scheduling
627
628	//-----------------------------------------------------------------------
629	// try to pop from a lane given by index w
630	static inline struct $thread * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
631	__STATS( stats.attempt++; )
632
633	// Get relevant elements locally
634	__intrusive_lane_t & lane = lanes.data[w];
635
636	// If list looks empty retry
637	if( is_empty(lane) ) {
638	return 0p;
639	}
640
641	// If we can't get the lock retry
642	if( !__atomic_try_acquire(&lane.lock) ) {
643	return 0p;
644	}
645
646	// If list is empty, unlock and retry
647	if( is_empty(lane) ) {
648	__atomic_unlock(&lane.lock);
649	return 0p;
650	}
651
652	// Actually pop the list
653	struct $thread * thrd;
654	unsigned long long tsv;
655	[thrd, tsv] = pop(lane);
656
657	/* paranoid */ verify(thrd);
658	/* paranoid */ verify(tsv);
659	/* paranoid */ verify(lane.lock);
660
661	// Unlock and return
662	__atomic_unlock(&lane.lock);
663
664	// Update statistics
665	__STATS( stats.success++; )
666
667	#if defined(USE_WORK_STEALING)
668	lanes.tscs[w].tv = tsv;
669	#endif
670
671	thrd->preferred = w;
672
673	// return the popped thread
674	return thrd;
675	}
676
677	//-----------------------------------------------------------------------
678	// try to pop from any lanes making sure you don't miss any threads push
679	// before the start of the function
680	static inline struct $thread * search(struct cluster * cltr) with (cltr->ready_queue) {
681	/* paranoid */ verify( lanes.count > 0 );
682	unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
683	unsigned offset = __tls_rand();
684	for(i; count) {
685	unsigned idx = (offset + i) % count;
686	struct $thread * thrd = try_pop(cltr, idx __STATS(, __tls_stats()->ready.pop.search));
687	if(thrd) {
688	return thrd;
689	}
690	}
691
692	// All lanes where empty return 0p
693	return 0p;
694	}
695
696	//-----------------------------------------------------------------------
697	// Check that all the intrusive queues in the data structure are still consistent
698	static void check( __ready_queue_t & q ) with (q) {
699	#if defined(__CFA_WITH_VERIFY__)
700	{
701	for( idx ; lanes.count ) {
702	__intrusive_lane_t & sl = lanes.data[idx];
703	assert(!lanes.data[idx].lock);
704
705	if(is_empty(sl)) {
706	assert( sl.anchor.next == 0p );
707	assert( sl.anchor.ts == -1llu );
708	assert( mock_head(sl) == sl.prev );
709	} else {
710	assert( sl.anchor.next != 0p );
711	assert( sl.anchor.ts != -1llu );
712	assert( mock_head(sl) != sl.prev );
713	}
714	}
715	}
716	#endif
717	}
718
719	//-----------------------------------------------------------------------
720	// Given 2 indexes, pick the list with the oldest push an try to pop from it
721	static inline struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
722	// Pick the bet list
723	int w = i;
724	if( __builtin_expect(!is_empty(lanes.data[j]), true) ) {
725	w = (ts(lanes.data[i]) < ts(lanes.data[j])) ? i : j;
726	}
727
728	return try_pop(cltr, w __STATS(, stats));
729	}
730
731	// Call this function of the intrusive list was moved using memcpy
732	// fixes the list so that the pointers back to anchors aren't left dangling
733	static inline void fix(__intrusive_lane_t & ll) {
734	if(is_empty(ll)) {
735	verify(ll.anchor.next == 0p);
736	ll.prev = mock_head(ll);
737	}
738	}
739
740	static void assign_list(unsigned & value, dlist(processor) & list, unsigned count) {
741	processor * it = &list`first;
742	for(unsigned i = 0; i < count; i++) {
743	/* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
744	it->rdq.id = value;
745	it->rdq.target = -1u;
746	value += READYQ_SHARD_FACTOR;
747	it = &(*it)`next;
748	}
749	}
750
751	static void reassign_cltr_id(struct cluster * cltr) {
752	unsigned preferred = 0;
753	assign_list(preferred, cltr->procs.actives, cltr->procs.total - cltr->procs.idle);
754	assign_list(preferred, cltr->procs.idles , cltr->procs.idle );
755	}
756
757	static void fix_times( struct cluster * cltr ) with( cltr->ready_queue ) {
758	#if defined(USE_WORK_STEALING)
759	lanes.tscs = alloc(lanes.count, lanes.tscs`realloc);
760	for(i; lanes.count) {
761	unsigned long long tsc1 = ts(lanes.data[i]);
762	unsigned long long tsc2 = rdtscl();
763	lanes.tscs[i].tv = min(tsc1, tsc2);
764	}
765	#endif
766	}
767
768	#if defined(USE_CPU_WORK_STEALING)
769	// ready_queue size is fixed in this case
770	void ready_queue_grow(struct cluster * cltr) {}
771	void ready_queue_shrink(struct cluster * cltr) {}
772	#else
773	// Grow the ready queue
774	void ready_queue_grow(struct cluster * cltr) {
775	size_t ncount;
776	int target = cltr->procs.total;
777
778	/* paranoid */ verify( ready_mutate_islocked() );
779	__cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue\n");
780
781	// Make sure that everything is consistent
782	/* paranoid */ check( cltr->ready_queue );
783
784	// grow the ready queue
785	with( cltr->ready_queue ) {
786	// Find new count
787	// Make sure we always have atleast 1 list
788	if(target >= 2) {
789	ncount = target * READYQ_SHARD_FACTOR;
790	} else {
791	ncount = SEQUENTIAL_SHARD;
792	}
793
794	// Allocate new array (uses realloc and memcpies the data)
795	lanes.data = alloc( ncount, lanes.data`realloc );
796
797	// Fix the moved data
798	for( idx; (size_t)lanes.count ) {
799	fix(lanes.data[idx]);
800	}
801
802	// Construct new data
803	for( idx; (size_t)lanes.count ~ ncount) {
804	(lanes.data[idx]){};
805	}
806
807	// Update original
808	lanes.count = ncount;
809	}
810
811	fix_times(cltr);
812
813	reassign_cltr_id(cltr);
814
815	// Make sure that everything is consistent
816	/* paranoid */ check( cltr->ready_queue );
817
818	__cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue done\n");
819
820	/* paranoid */ verify( ready_mutate_islocked() );
821	}
822
823	// Shrink the ready queue
824	void ready_queue_shrink(struct cluster * cltr) {
825	/* paranoid */ verify( ready_mutate_islocked() );
826	__cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue\n");
827
828	// Make sure that everything is consistent
829	/* paranoid */ check( cltr->ready_queue );
830
831	int target = cltr->procs.total;
832
833	with( cltr->ready_queue ) {
834	// Remember old count
835	size_t ocount = lanes.count;
836
837	// Find new count
838	// Make sure we always have atleast 1 list
839	lanes.count = target >= 2 ? target * READYQ_SHARD_FACTOR: SEQUENTIAL_SHARD;
840	/* paranoid */ verify( ocount >= lanes.count );
841	/* paranoid / verify( lanes.count == target READYQ_SHARD_FACTOR \|\| target < 2 );
842
843	// for printing count the number of displaced threads
844	#if defined(__CFA_DEBUG_PRINT__) \|\| defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
845	__attribute__((unused)) size_t displaced = 0;
846	#endif
847
848	// redistribute old data
849	for( idx; (size_t)lanes.count ~ ocount) {
850	// Lock is not strictly needed but makes checking invariants much easier
851	__attribute__((unused)) bool locked = __atomic_try_acquire(&lanes.data[idx].lock);
852	verify(locked);
853
854	// As long as we can pop from this lane to push the threads somewhere else in the queue
855	while(!is_empty(lanes.data[idx])) {
856	struct $thread * thrd;
857	unsigned long long _;
858	[thrd, _] = pop(lanes.data[idx]);
859
860	push(cltr, thrd, true);
861
862	// for printing count the number of displaced threads
863	#if defined(__CFA_DEBUG_PRINT__) \|\| defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
864	displaced++;
865	#endif
866	}
867
868	// Unlock the lane
869	__atomic_unlock(&lanes.data[idx].lock);
870
871	// TODO print the queue statistics here
872
873	^(lanes.data[idx]){};
874	}
875
876	__cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue displaced %zu threads\n", displaced);
877
878	// Allocate new array (uses realloc and memcpies the data)
879	lanes.data = alloc( lanes.count, lanes.data`realloc );
880
881	// Fix the moved data
882	for( idx; (size_t)lanes.count ) {
883	fix(lanes.data[idx]);
884	}
885	}
886
887	fix_times(cltr);
888
889	reassign_cltr_id(cltr);
890
891	// Make sure that everything is consistent
892	/* paranoid */ check( cltr->ready_queue );
893
894	__cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue done\n");
895	/* paranoid */ verify( ready_mutate_islocked() );
896	}
897	#endif
898
899	#if !defined(__CFA_NO_STATISTICS__)
900	unsigned cnt(const __ready_queue_t & this, unsigned idx) {
901	/* paranoid */ verify(this.lanes.count > idx);
902	return this.lanes.data[idx].cnt;
903	}
904	#endif
905
906
907	#if defined(CFA_HAVE_LINUX_LIBRSEQ)
908	// No definition needed
909	#elif defined(CFA_HAVE_LINUX_RSEQ_H)
910
911	#if defined( __x86_64 ) \|\| defined( __i386 )
912	#define RSEQ_SIG 0x53053053
913	#elif defined( __ARM_ARCH )
914	#ifdef __ARMEB__
915	#define RSEQ_SIG 0xf3def5e7 /* udf #24035 ; 0x5de3 (ARMv6+) */
916	#else
917	#define RSEQ_SIG 0xe7f5def3 /* udf #24035 ; 0x5de3 */
918	#endif
919	#endif
920
921	extern void __disable_interrupts_hard();
922	extern void __enable_interrupts_hard();
923
924	void __kernel_raw_rseq_register (void) {
925	/* paranoid */ verify( __cfaabi_rseq.cpu_id == RSEQ_CPU_ID_UNINITIALIZED );
926
927	// int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), 0, (sigset_t *)0p, _NSIG / 8);
928	int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), 0, RSEQ_SIG);
929	if(ret != 0) {
930	int e = errno;
931	switch(e) {
932	case EINVAL: abort("KERNEL ERROR: rseq register invalid argument");
933	case ENOSYS: abort("KERNEL ERROR: rseq register no supported");
934	case EFAULT: abort("KERNEL ERROR: rseq register with invalid argument");
935	case EBUSY : abort("KERNEL ERROR: rseq register already registered");
936	case EPERM : abort("KERNEL ERROR: rseq register sig argument on unregistration does not match the signature received on registration");
937	default: abort("KERNEL ERROR: rseq register unexpected return %d", e);
938	}
939	}
940	}
941
942	void __kernel_raw_rseq_unregister(void) {
943	/* paranoid */ verify( __cfaabi_rseq.cpu_id >= 0 );
944
945	// int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), RSEQ_FLAG_UNREGISTER, (sigset_t *)0p, _NSIG / 8);
946	int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), RSEQ_FLAG_UNREGISTER, RSEQ_SIG);
947	if(ret != 0) {
948	int e = errno;
949	switch(e) {
950	case EINVAL: abort("KERNEL ERROR: rseq unregister invalid argument");
951	case ENOSYS: abort("KERNEL ERROR: rseq unregister no supported");
952	case EFAULT: abort("KERNEL ERROR: rseq unregister with invalid argument");
953	case EBUSY : abort("KERNEL ERROR: rseq unregister already registered");
954	case EPERM : abort("KERNEL ERROR: rseq unregister sig argument on unregistration does not match the signature received on registration");
955	default: abort("KERNEL ERROR: rseq unregisteunexpected return %d", e);
956	}
957	}
958	}
959	#else
960	// No definition needed
961	#endif

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

Download in other formats:

Original Format