Context Navigation

ready_queue.cfa @ db614d0

ADTast-experimentalenumforall-pointer-decaypthread-emulationqualifiedEnum

Last change on this file since db614d0 was c86ee4c, checked in by Thierry Delisle <tdelisle@…>, 3 years ago
Merge branch 'master' of plg.uwaterloo.ca:software/cfa/cfa-cc
Property mode set to `100644`
File size: 31.1 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
348	if(0 == (__tls_rand() % 10_000)) {
349	proc->rdq.target = __tls_rand() % lanes.count;
350	} else {
351	unsigned cpu_chaos = map.start + (__tls_rand() % map.count);
352	proc->rdq.target = (cpu_chaos * READYQ_SHARD_FACTOR) + (__tls_rand() % READYQ_SHARD_FACTOR);
353	/* paranoid / verify(proc->rdq.target >= (map.start READYQ_SHARD_FACTOR));
354	/* paranoid / verify(proc->rdq.target < ((map.start + map.count) READYQ_SHARD_FACTOR));
355	}
356
357	/* paranoid */ verify(proc->rdq.target != -1u);
358	}
359	else {
360	const unsigned long long bias = 0; //2_500_000_000;
361	const unsigned long long cutoff = proc->rdq.cutoff > bias ? proc->rdq.cutoff - bias : proc->rdq.cutoff;
362	{
363	unsigned target = proc->rdq.target;
364	proc->rdq.target = -1u;
365	if(lanes.tscs[target].tv < cutoff && ts(lanes.data[target]) < cutoff) {
366	thread$ * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
367	proc->rdq.last = target;
368	if(t) return t;
369	}
370	}
371
372	unsigned last = proc->rdq.last;
373	if(last != -1u && lanes.tscs[last].tv < cutoff && ts(lanes.data[last]) < cutoff) {
374	thread$ * t = try_pop(cltr, last __STATS(, __tls_stats()->ready.pop.help));
375	if(t) return t;
376	}
377	else {
378	proc->rdq.last = -1u;
379	}
380	}
381
382	for(READYQ_SHARD_FACTOR) {
383	unsigned i = start + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
384	if(thread$ * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
385	}
386
387	// All lanes where empty return 0p
388	return 0p;
389	}
390
391	__attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
392	processor * const proc = kernelTLS().this_processor;
393	unsigned last = proc->rdq.last;
394	if(last != -1u) {
395	struct thread$ * t = try_pop(cltr, last __STATS(, __tls_stats()->ready.pop.steal));
396	if(t) return t;
397	proc->rdq.last = -1u;
398	}
399
400	unsigned i = __tls_rand() % lanes.count;
401	return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
402	}
403	__attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr) {
404	return search(cltr);
405	}
406	#endif
407	#if defined(USE_RELAXED_FIFO)
408	//-----------------------------------------------------------------------
409	// get index from random number with or without bias towards queues
410	static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred) {
411	unsigned i;
412	bool local;
413	unsigned rlow = r % BIAS;
414	unsigned rhigh = r / BIAS;
415	if((0 != rlow) && preferred >= 0) {
416	// (BIAS - 1) out of BIAS chances
417	// Use perferred queues
418	i = preferred + (rhigh % READYQ_SHARD_FACTOR);
419	local = true;
420	}
421	else {
422	// 1 out of BIAS chances
423	// Use all queues
424	i = rhigh;
425	local = false;
426	}
427	return [i, local];
428	}
429
430	__attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, bool push_local) with (cltr->ready_queue) {
431	__cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
432
433	const bool external = !push_local \|\| (!kernelTLS().this_processor) \|\| (cltr != kernelTLS().this_processor->cltr);
434	/* paranoid */ verify(external \|\| kernelTLS().this_processor->rdq.id < lanes.count );
435
436	bool local;
437	int preferred = external ? -1 : kernelTLS().this_processor->rdq.id;
438
439	// Try to pick a lane and lock it
440	unsigned i;
441	do {
442	// Pick the index of a lane
443	unsigned r = __tls_rand_fwd();
444	[i, local] = idx_from_r(r, preferred);
445
446	i %= __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
447
448	#if !defined(__CFA_NO_STATISTICS__)
449	if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
450	else if(local) __tls_stats()->ready.push.local.attempt++;
451	else __tls_stats()->ready.push.share.attempt++;
452	#endif
453
454	// If we can't lock it retry
455	} while( !__atomic_try_acquire( &lanes.data[i].lock ) );
456
457	// Actually push it
458	push(lanes.data[i], thrd);
459
460	// Unlock and return
461	__atomic_unlock( &lanes.data[i].lock );
462
463	// Mark the current index in the tls rng instance as having an item
464	__tls_rand_advance_bck();
465
466	__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);
467
468	// Update statistics
469	#if !defined(__CFA_NO_STATISTICS__)
470	if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
471	else if(local) __tls_stats()->ready.push.local.success++;
472	else __tls_stats()->ready.push.share.success++;
473	#endif
474	}
475
476	// Pop from the ready queue from a given cluster
477	__attribute__((hot)) thread$ * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
478	/* paranoid */ verify( lanes.count > 0 );
479	/* paranoid */ verify( kernelTLS().this_processor );
480	/* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
481
482	unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
483	int preferred = kernelTLS().this_processor->rdq.id;
484
485
486	// As long as the list is not empty, try finding a lane that isn't empty and pop from it
487	for(25) {
488	// Pick two lists at random
489	unsigned ri = __tls_rand_bck();
490	unsigned rj = __tls_rand_bck();
491
492	unsigned i, j;
493	__attribute__((unused)) bool locali, localj;
494	[i, locali] = idx_from_r(ri, preferred);
495	[j, localj] = idx_from_r(rj, preferred);
496
497	i %= count;
498	j %= count;
499
500	// try popping from the 2 picked lists
501	struct thread$ * thrd = try_pop(cltr, i, j __STATS(, *(locali \|\| localj ? &__tls_stats()->ready.pop.local : &__tls_stats()->ready.pop.help)));
502	if(thrd) {
503	return thrd;
504	}
505	}
506
507	// All lanes where empty return 0p
508	return 0p;
509	}
510
511	__attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr) { return pop_fast(cltr); }
512	__attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr) {
513	return search(cltr);
514	}
515	#endif
516	#if defined(USE_WORK_STEALING)
517	__attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, bool push_local) with (cltr->ready_queue) {
518	__cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
519
520	// #define USE_PREFERRED
521	#if !defined(USE_PREFERRED)
522	const bool external = !push_local \|\| (!kernelTLS().this_processor) \|\| (cltr != kernelTLS().this_processor->cltr);
523	/* paranoid */ verify(external \|\| kernelTLS().this_processor->rdq.id < lanes.count );
524	#else
525	unsigned preferred = thrd->preferred;
526	const bool external = push_local \|\| (!kernelTLS().this_processor) \|\| preferred == -1u \|\| thrd->curr_cluster != cltr;
527	/* paranoid */ verifyf(external \|\| preferred < lanes.count, "Invalid preferred queue %u for %u lanes", preferred, lanes.count );
528
529	unsigned r = preferred % READYQ_SHARD_FACTOR;
530	const unsigned start = preferred - r;
531	#endif
532
533	// Try to pick a lane and lock it
534	unsigned i;
535	do {
536	#if !defined(__CFA_NO_STATISTICS__)
537	if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
538	else __tls_stats()->ready.push.local.attempt++;
539	#endif
540
541	if(unlikely(external)) {
542	i = __tls_rand() % lanes.count;
543	}
544	else {
545	#if !defined(USE_PREFERRED)
546	processor * proc = kernelTLS().this_processor;
547	unsigned r = proc->rdq.its++;
548	i = proc->rdq.id + (r % READYQ_SHARD_FACTOR);
549	#else
550	i = start + (r++ % READYQ_SHARD_FACTOR);
551	#endif
552	}
553	// If we can't lock it retry
554	} while( !__atomic_try_acquire( &lanes.data[i].lock ) );
555
556	// Actually push it
557	push(lanes.data[i], thrd);
558
559	// Unlock and return
560	__atomic_unlock( &lanes.data[i].lock );
561
562	#if !defined(__CFA_NO_STATISTICS__)
563	if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
564	else __tls_stats()->ready.push.local.success++;
565	#endif
566
567	__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);
568	}
569
570	// Pop from the ready queue from a given cluster
571	__attribute__((hot)) thread$ * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
572	/* paranoid */ verify( lanes.count > 0 );
573	/* paranoid */ verify( kernelTLS().this_processor );
574	/* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
575
576	processor * proc = kernelTLS().this_processor;
577
578	if(proc->rdq.target == -1u) {
579	unsigned long long min = ts(lanes.data[proc->rdq.id]);
580	for(int i = 0; i < READYQ_SHARD_FACTOR; i++) {
581	unsigned long long tsc = ts(lanes.data[proc->rdq.id + i]);
582	if(tsc < min) min = tsc;
583	}
584	proc->rdq.cutoff = min;
585	proc->rdq.target = __tls_rand() % lanes.count;
586	}
587	else {
588	unsigned target = proc->rdq.target;
589	proc->rdq.target = -1u;
590	const unsigned long long bias = 0; //2_500_000_000;
591	const unsigned long long cutoff = proc->rdq.cutoff > bias ? proc->rdq.cutoff - bias : proc->rdq.cutoff;
592	if(lanes.tscs[target].tv < cutoff && ts(lanes.data[target]) < cutoff) {
593	thread$ * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
594	if(t) return t;
595	}
596	}
597
598	for(READYQ_SHARD_FACTOR) {
599	unsigned i = proc->rdq.id + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
600	if(thread$ * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
601	}
602	return 0p;
603	}
604
605	__attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
606	unsigned i = __tls_rand() % lanes.count;
607	return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
608	}
609
610	__attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr) with (cltr->ready_queue) {
611	return search(cltr);
612	}
613	#endif
614
615	//=======================================================================
616	// Various Ready Queue utilities
617	//=======================================================================
618	// these function work the same or almost the same
619	// whether they are using work-stealing or relaxed fifo scheduling
620
621	//-----------------------------------------------------------------------
622	// try to pop from a lane given by index w
623	static inline struct thread$ * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
624	__STATS( stats.attempt++; )
625
626	// Get relevant elements locally
627	__intrusive_lane_t & lane = lanes.data[w];
628
629	// If list looks empty retry
630	if( is_empty(lane) ) {
631	return 0p;
632	}
633
634	// If we can't get the lock retry
635	if( !__atomic_try_acquire(&lane.lock) ) {
636	return 0p;
637	}
638
639	// If list is empty, unlock and retry
640	if( is_empty(lane) ) {
641	__atomic_unlock(&lane.lock);
642	return 0p;
643	}
644
645	// Actually pop the list
646	struct thread$ * thrd;
647	unsigned long long tsv;
648	[thrd, tsv] = pop(lane);
649
650	/* paranoid */ verify(thrd);
651	/* paranoid */ verify(tsv);
652	/* paranoid */ verify(lane.lock);
653
654	// Unlock and return
655	__atomic_unlock(&lane.lock);
656
657	// Update statistics
658	__STATS( stats.success++; )
659
660	#if defined(USE_WORK_STEALING)
661	lanes.tscs[w].tv = tsv;
662	#endif
663
664	thrd->preferred = w;
665
666	// return the popped thread
667	return thrd;
668	}
669
670	//-----------------------------------------------------------------------
671	// try to pop from any lanes making sure you don't miss any threads push
672	// before the start of the function
673	static inline struct thread$ * search(struct cluster * cltr) with (cltr->ready_queue) {
674	/* paranoid */ verify( lanes.count > 0 );
675	unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
676	unsigned offset = __tls_rand();
677	for(i; count) {
678	unsigned idx = (offset + i) % count;
679	struct thread$ * thrd = try_pop(cltr, idx __STATS(, __tls_stats()->ready.pop.search));
680	if(thrd) {
681	return thrd;
682	}
683	}
684
685	// All lanes where empty return 0p
686	return 0p;
687	}
688
689	//-----------------------------------------------------------------------
690	// Check that all the intrusive queues in the data structure are still consistent
691	static void check( __ready_queue_t & q ) with (q) {
692	#if defined(__CFA_WITH_VERIFY__)
693	{
694	for( idx ; lanes.count ) {
695	__intrusive_lane_t & sl = lanes.data[idx];
696	assert(!lanes.data[idx].lock);
697
698	if(is_empty(sl)) {
699	assert( sl.anchor.next == 0p );
700	assert( sl.anchor.ts == -1llu );
701	assert( mock_head(sl) == sl.prev );
702	} else {
703	assert( sl.anchor.next != 0p );
704	assert( sl.anchor.ts != -1llu );
705	assert( mock_head(sl) != sl.prev );
706	}
707	}
708	}
709	#endif
710	}
711
712	//-----------------------------------------------------------------------
713	// Given 2 indexes, pick the list with the oldest push an try to pop from it
714	static inline struct thread$ * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
715	// Pick the bet list
716	int w = i;
717	if( __builtin_expect(!is_empty(lanes.data[j]), true) ) {
718	w = (ts(lanes.data[i]) < ts(lanes.data[j])) ? i : j;
719	}
720
721	return try_pop(cltr, w __STATS(, stats));
722	}
723
724	// Call this function of the intrusive list was moved using memcpy
725	// fixes the list so that the pointers back to anchors aren't left dangling
726	static inline void fix(__intrusive_lane_t & ll) {
727	if(is_empty(ll)) {
728	verify(ll.anchor.next == 0p);
729	ll.prev = mock_head(ll);
730	}
731	}
732
733	static void assign_list(unsigned & value, dlist(processor) & list, unsigned count) {
734	processor * it = &list`first;
735	for(unsigned i = 0; i < count; i++) {
736	/* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
737	it->rdq.id = value;
738	it->rdq.target = -1u;
739	value += READYQ_SHARD_FACTOR;
740	it = &(*it)`next;
741	}
742	}
743
744	static void reassign_cltr_id(struct cluster * cltr) {
745	unsigned preferred = 0;
746	assign_list(preferred, cltr->procs.actives, cltr->procs.total - cltr->procs.idle);
747	assign_list(preferred, cltr->procs.idles , cltr->procs.idle );
748	}
749
750	static void fix_times( struct cluster * cltr ) with( cltr->ready_queue ) {
751	#if defined(USE_WORK_STEALING)
752	lanes.tscs = alloc(lanes.count, lanes.tscs`realloc);
753	for(i; lanes.count) {
754	unsigned long long tsc1 = ts(lanes.data[i]);
755	unsigned long long tsc2 = rdtscl();
756	lanes.tscs[i].tv = min(tsc1, tsc2);
757	}
758	#endif
759	}
760
761	#if defined(USE_CPU_WORK_STEALING)
762	// ready_queue size is fixed in this case
763	void ready_queue_grow(struct cluster * cltr) {}
764	void ready_queue_shrink(struct cluster * cltr) {}
765	#else
766	// Grow the ready queue
767	void ready_queue_grow(struct cluster * cltr) {
768	size_t ncount;
769	int target = cltr->procs.total;
770
771	/* paranoid */ verify( ready_mutate_islocked() );
772	__cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue\n");
773
774	// Make sure that everything is consistent
775	/* paranoid */ check( cltr->ready_queue );
776
777	// grow the ready queue
778	with( cltr->ready_queue ) {
779	// Find new count
780	// Make sure we always have atleast 1 list
781	if(target >= 2) {
782	ncount = target * READYQ_SHARD_FACTOR;
783	} else {
784	ncount = SEQUENTIAL_SHARD;
785	}
786
787	// Allocate new array (uses realloc and memcpies the data)
788	lanes.data = alloc( ncount, lanes.data`realloc );
789
790	// Fix the moved data
791	for( idx; (size_t)lanes.count ) {
792	fix(lanes.data[idx]);
793	}
794
795	// Construct new data
796	for( idx; (size_t)lanes.count ~ ncount) {
797	(lanes.data[idx]){};
798	}
799
800	// Update original
801	lanes.count = ncount;
802	}
803
804	fix_times(cltr);
805
806	reassign_cltr_id(cltr);
807
808	// Make sure that everything is consistent
809	/* paranoid */ check( cltr->ready_queue );
810
811	__cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue done\n");
812
813	/* paranoid */ verify( ready_mutate_islocked() );
814	}
815
816	// Shrink the ready queue
817	void ready_queue_shrink(struct cluster * cltr) {
818	/* paranoid */ verify( ready_mutate_islocked() );
819	__cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue\n");
820
821	// Make sure that everything is consistent
822	/* paranoid */ check( cltr->ready_queue );
823
824	int target = cltr->procs.total;
825
826	with( cltr->ready_queue ) {
827	// Remember old count
828	size_t ocount = lanes.count;
829
830	// Find new count
831	// Make sure we always have atleast 1 list
832	lanes.count = target >= 2 ? target * READYQ_SHARD_FACTOR: SEQUENTIAL_SHARD;
833	/* paranoid */ verify( ocount >= lanes.count );
834	/* paranoid / verify( lanes.count == target READYQ_SHARD_FACTOR \|\| target < 2 );
835
836	// for printing count the number of displaced threads
837	#if defined(__CFA_DEBUG_PRINT__) \|\| defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
838	__attribute__((unused)) size_t displaced = 0;
839	#endif
840
841	// redistribute old data
842	for( idx; (size_t)lanes.count ~ ocount) {
843	// Lock is not strictly needed but makes checking invariants much easier
844	__attribute__((unused)) bool locked = __atomic_try_acquire(&lanes.data[idx].lock);
845	verify(locked);
846
847	// As long as we can pop from this lane to push the threads somewhere else in the queue
848	while(!is_empty(lanes.data[idx])) {
849	struct thread$ * thrd;
850	unsigned long long _;
851	[thrd, _] = pop(lanes.data[idx]);
852
853	push(cltr, thrd, true);
854
855	// for printing count the number of displaced threads
856	#if defined(__CFA_DEBUG_PRINT__) \|\| defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
857	displaced++;
858	#endif
859	}
860
861	// Unlock the lane
862	__atomic_unlock(&lanes.data[idx].lock);
863
864	// TODO print the queue statistics here
865
866	^(lanes.data[idx]){};
867	}
868
869	__cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue displaced %zu threads\n", displaced);
870
871	// Allocate new array (uses realloc and memcpies the data)
872	lanes.data = alloc( lanes.count, lanes.data`realloc );
873
874	// Fix the moved data
875	for( idx; (size_t)lanes.count ) {
876	fix(lanes.data[idx]);
877	}
878	}
879
880	fix_times(cltr);
881
882	reassign_cltr_id(cltr);
883
884	// Make sure that everything is consistent
885	/* paranoid */ check( cltr->ready_queue );
886
887	__cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue done\n");
888	/* paranoid */ verify( ready_mutate_islocked() );
889	}
890	#endif
891
892	#if !defined(__CFA_NO_STATISTICS__)
893	unsigned cnt(const __ready_queue_t & this, unsigned idx) {
894	/* paranoid */ verify(this.lanes.count > idx);
895	return this.lanes.data[idx].cnt;
896	}
897	#endif
898
899
900	#if defined(CFA_HAVE_LINUX_LIBRSEQ)
901	// No definition needed
902	#elif defined(CFA_HAVE_LINUX_RSEQ_H)
903
904	#if defined( __x86_64 ) \|\| defined( __i386 )
905	#define RSEQ_SIG 0x53053053
906	#elif defined( __ARM_ARCH )
907	#ifdef __ARMEB__
908	#define RSEQ_SIG 0xf3def5e7 /* udf #24035 ; 0x5de3 (ARMv6+) */
909	#else
910	#define RSEQ_SIG 0xe7f5def3 /* udf #24035 ; 0x5de3 */
911	#endif
912	#endif
913
914	extern void __disable_interrupts_hard();
915	extern void __enable_interrupts_hard();
916
917	void __kernel_raw_rseq_register (void) {
918	/* paranoid */ verify( __cfaabi_rseq.cpu_id == RSEQ_CPU_ID_UNINITIALIZED );
919
920	// int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), 0, (sigset_t *)0p, _NSIG / 8);
921	int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), 0, RSEQ_SIG);
922	if(ret != 0) {
923	int e = errno;
924	switch(e) {
925	case EINVAL: abort("KERNEL ERROR: rseq register invalid argument");
926	case ENOSYS: abort("KERNEL ERROR: rseq register no supported");
927	case EFAULT: abort("KERNEL ERROR: rseq register with invalid argument");
928	case EBUSY : abort("KERNEL ERROR: rseq register already registered");
929	case EPERM : abort("KERNEL ERROR: rseq register sig argument on unregistration does not match the signature received on registration");
930	default: abort("KERNEL ERROR: rseq register unexpected return %d", e);
931	}
932	}
933	}
934
935	void __kernel_raw_rseq_unregister(void) {
936	/* paranoid */ verify( __cfaabi_rseq.cpu_id >= 0 );
937
938	// int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), RSEQ_FLAG_UNREGISTER, (sigset_t *)0p, _NSIG / 8);
939	int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), RSEQ_FLAG_UNREGISTER, RSEQ_SIG);
940	if(ret != 0) {
941	int e = errno;
942	switch(e) {
943	case EINVAL: abort("KERNEL ERROR: rseq unregister invalid argument");
944	case ENOSYS: abort("KERNEL ERROR: rseq unregister no supported");
945	case EFAULT: abort("KERNEL ERROR: rseq unregister with invalid argument");
946	case EBUSY : abort("KERNEL ERROR: rseq unregister already registered");
947	case EPERM : abort("KERNEL ERROR: rseq unregister sig argument on unregistration does not match the signature received on registration");
948	default: abort("KERNEL ERROR: rseq unregisteunexpected return %d", e);
949	}
950	}
951	}
952	#else
953	// No definition needed
954	#endif

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

Download in other formats:

Original Format