Context Navigation

ready_queue.cfa@ ef94ae7

Visit:

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

Last change on this file since ef94ae7 was ef94ae7, checked in by Thierry Delisle <tdelisle@…>, 4 years ago
Changed ready-queue to use -1 for empty ts.
Property mode set to `100644`
File size: 30.4 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 <errno.h>
33	#include <unistd.h>
34
35	extern "C" {
36	#include <sys/syscall.h> // __NR_xxx
37	}
38
39	#include "ready_subqueue.hfa"
40
41	static const size_t cache_line_size = 64;
42
43	#if !defined(__CFA_NO_STATISTICS__)
44	#define __STATS(...) __VA_ARGS__
45	#else
46	#define __STATS(...)
47	#endif
48
49	// No overriden function, no environment variable, no define
50	// fall back to a magic number
51	#ifndef __CFA_MAX_PROCESSORS__
52	#define __CFA_MAX_PROCESSORS__ 1024
53	#endif
54
55	#if defined(USE_CPU_WORK_STEALING)
56	#define READYQ_SHARD_FACTOR 2
57	#elif defined(USE_RELAXED_FIFO)
58	#define BIAS 4
59	#define READYQ_SHARD_FACTOR 4
60	#define SEQUENTIAL_SHARD 1
61	#elif defined(USE_WORK_STEALING)
62	#define READYQ_SHARD_FACTOR 2
63	#define SEQUENTIAL_SHARD 2
64	#else
65	#error no scheduling strategy selected
66	#endif
67
68	static inline struct $thread * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats));
69	static inline struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats));
70	static inline struct $thread * search(struct cluster * cltr);
71	static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred);
72
73
74	// returns the maximum number of processors the RWLock support
75	__attribute__((weak)) unsigned __max_processors() {
76	const char * max_cores_s = getenv("CFA_MAX_PROCESSORS");
77	if(!max_cores_s) {
78	__cfadbg_print_nolock(ready_queue, "No CFA_MAX_PROCESSORS in ENV\n");
79	return __CFA_MAX_PROCESSORS__;
80	}
81
82	char * endptr = 0p;
83	long int max_cores_l = strtol(max_cores_s, &endptr, 10);
84	if(max_cores_l < 1 \|\| max_cores_l > 65535) {
85	__cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS out of range : %ld\n", max_cores_l);
86	return __CFA_MAX_PROCESSORS__;
87	}
88	if('\0' != *endptr) {
89	__cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS not a decimal number : %s\n", max_cores_s);
90	return __CFA_MAX_PROCESSORS__;
91	}
92
93	return max_cores_l;
94	}
95
96	#if defined(CFA_HAVE_LINUX_LIBRSEQ)
97	// No forward declaration needed
98	#define __kernel_rseq_register rseq_register_current_thread
99	#define __kernel_rseq_unregister rseq_unregister_current_thread
100	#elif defined(CFA_HAVE_LINUX_RSEQ_H)
101	void __kernel_raw_rseq_register (void);
102	void __kernel_raw_rseq_unregister(void);
103
104	#define __kernel_rseq_register __kernel_raw_rseq_register
105	#define __kernel_rseq_unregister __kernel_raw_rseq_unregister
106	#else
107	// No forward declaration needed
108	// No initialization needed
109	static inline void noop(void) {}
110
111	#define __kernel_rseq_register noop
112	#define __kernel_rseq_unregister noop
113	#endif
114
115	//=======================================================================
116	// Cluster wide reader-writer lock
117	//=======================================================================
118	void ?{}(__scheduler_RWLock_t & this) {
119	this.max = __max_processors();
120	this.alloc = 0;
121	this.ready = 0;
122	this.data = alloc(this.max);
123	this.write_lock = false;
124
125	/paranoid/ verify(__atomic_is_lock_free(sizeof(this.alloc), &this.alloc));
126	/paranoid/ verify(__atomic_is_lock_free(sizeof(this.ready), &this.ready));
127
128	}
129	void ^?{}(__scheduler_RWLock_t & this) {
130	free(this.data);
131	}
132
133
134	//=======================================================================
135	// Lock-Free registering/unregistering of threads
136	unsigned register_proc_id( void ) with(*__scheduler_lock) {
137	__kernel_rseq_register();
138
139	__cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p for RW-Lock\n", proc);
140	bool * handle = (bool *)&kernelTLS().sched_lock;
141
142	// Step - 1 : check if there is already space in the data
143	uint_fast32_t s = ready;
144
145	// Check among all the ready
146	for(uint_fast32_t i = 0; i < s; i++) {
147	bool * volatile * cell = (bool * volatile *)&data[i]; // Cforall is bugged and the double volatiles causes problems
148	/* paranoid / verify( handle != cell );
149
150	bool * null = 0p; // Re-write every loop since compare thrashes it
151	if( __atomic_load_n(cell, (int)__ATOMIC_RELAXED) == null
152	&& __atomic_compare_exchange_n( cell, &null, handle, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
153	/* paranoid */ verify(i < ready);
154	/* paranoid */ verify( (kernelTLS().sched_id = i, true) );
155	return i;
156	}
157	}
158
159	if(max <= alloc) abort("Trying to create more than %ud processors", __scheduler_lock->max);
160
161	// Step - 2 : F&A to get a new spot in the array.
162	uint_fast32_t n = __atomic_fetch_add(&alloc, 1, __ATOMIC_SEQ_CST);
163	if(max <= n) abort("Trying to create more than %ud processors", __scheduler_lock->max);
164
165	// Step - 3 : Mark space as used and then publish it.
166	data[n] = handle;
167	while() {
168	unsigned copy = n;
169	if( __atomic_load_n(&ready, __ATOMIC_RELAXED) == n
170	&& __atomic_compare_exchange_n(&ready, &copy, n + 1, true, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))
171	break;
172	Pause();
173	}
174
175	__cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p done, id %lu\n", proc, n);
176
177	// Return new spot.
178	/* paranoid */ verify(n < ready);
179	/* paranoid */ verify( (kernelTLS().sched_id = n, true) );
180	return n;
181	}
182
183	void unregister_proc_id( unsigned id ) with(*__scheduler_lock) {
184	/* paranoid */ verify(id < ready);
185	/* paranoid */ verify(id == kernelTLS().sched_id);
186	/* paranoid */ verify(data[id] == &kernelTLS().sched_lock);
187
188	bool * volatile * cell = (bool * volatile *)&data[id]; // Cforall is bugged and the double volatiles causes problems
189
190	__atomic_store_n(cell, 0p, __ATOMIC_RELEASE);
191
192	__cfadbg_print_safe(ready_queue, "Kernel : Unregister proc %p\n", proc);
193
194	__kernel_rseq_unregister();
195	}
196
197	//-----------------------------------------------------------------------
198	// Writer side : acquire when changing the ready queue, e.g. adding more
199	// queues or removing them.
200	uint_fast32_t ready_mutate_lock( void ) with(*__scheduler_lock) {
201	/* paranoid */ verify( ! __preemption_enabled() );
202	/* paranoid */ verify( ! kernelTLS().sched_lock );
203
204	// Step 1 : lock global lock
205	// It is needed to avoid processors that register mid Critical-Section
206	// to simply lock their own lock and enter.
207	__atomic_acquire( &write_lock );
208
209	// Step 2 : lock per-proc lock
210	// Processors that are currently being registered aren't counted
211	// but can't be in read_lock or in the critical section.
212	// All other processors are counted
213	uint_fast32_t s = ready;
214	for(uint_fast32_t i = 0; i < s; i++) {
215	volatile bool * llock = data[i];
216	if(llock) __atomic_acquire( llock );
217	}
218
219	/* paranoid */ verify( ! __preemption_enabled() );
220	return s;
221	}
222
223	void ready_mutate_unlock( uint_fast32_t last_s ) with(*__scheduler_lock) {
224	/* paranoid */ verify( ! __preemption_enabled() );
225
226	// Step 1 : release local locks
227	// This must be done while the global lock is held to avoid
228	// threads that where created mid critical section
229	// to race to lock their local locks and have the writer
230	// immidiately unlock them
231	// Alternative solution : return s in write_lock and pass it to write_unlock
232	for(uint_fast32_t i = 0; i < last_s; i++) {
233	volatile bool * llock = data[i];
234	if(llock) __atomic_store_n(llock, (bool)false, __ATOMIC_RELEASE);
235	}
236
237	// Step 2 : release global lock
238	/paranoid/ assert(true == write_lock);
239	__atomic_store_n(&write_lock, (bool)false, __ATOMIC_RELEASE);
240
241	/* paranoid */ verify( ! __preemption_enabled() );
242	}
243
244	//=======================================================================
245	// Cforall Ready Queue used for scheduling
246	//=======================================================================
247	void ?{}(__ready_queue_t & this) with (this) {
248	#if defined(USE_CPU_WORK_STEALING)
249	lanes.count = cpu_info.hthrd_count * READYQ_SHARD_FACTOR;
250	lanes.data = alloc( lanes.count );
251	lanes.tscs = alloc( lanes.count );
252
253	for( idx; (size_t)lanes.count ) {
254	(lanes.data[idx]){};
255	lanes.tscs[idx].tv = rdtscl();
256	}
257	#else
258	lanes.data = 0p;
259	lanes.tscs = 0p;
260	lanes.count = 0;
261	#endif
262	}
263
264	void ^?{}(__ready_queue_t & this) with (this) {
265	#if !defined(USE_CPU_WORK_STEALING)
266	verify( SEQUENTIAL_SHARD == lanes.count );
267	#endif
268
269	free(lanes.data);
270	free(lanes.tscs);
271	}
272
273	//-----------------------------------------------------------------------
274	#if defined(USE_CPU_WORK_STEALING)
275	__attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd, bool push_local) with (cltr->ready_queue) {
276	__cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
277
278	processor * const proc = kernelTLS().this_processor;
279	const bool external = !push_local \|\| (!proc) \|\| (cltr != proc->cltr);
280
281	const int cpu = __kernel_getcpu();
282	/* paranoid */ verify(cpu >= 0);
283	/* paranoid */ verify(cpu < cpu_info.hthrd_count);
284	/* paranoid / verify(cpu READYQ_SHARD_FACTOR < lanes.count);
285
286	const cpu_map_entry_t & map = cpu_info.llc_map[cpu];
287	/* paranoid / verify(map.start READYQ_SHARD_FACTOR < lanes.count);
288	/* paranoid / verify(map.self READYQ_SHARD_FACTOR < lanes.count);
289	/* paranoid / verifyf((map.start + map.count) READYQ_SHARD_FACTOR <= lanes.count, "have %u lanes but map can go up to %u", lanes.count, (map.start + map.count) * READYQ_SHARD_FACTOR);
290
291	const int start = map.self * READYQ_SHARD_FACTOR;
292	unsigned i;
293	do {
294	unsigned r;
295	if(unlikely(external)) { r = __tls_rand(); }
296	else { r = proc->rdq.its++; }
297	i = start + (r % READYQ_SHARD_FACTOR);
298	// If we can't lock it retry
299	} while( !__atomic_try_acquire( &lanes.data[i].lock ) );
300
301	// Actually push it
302	push(lanes.data[i], thrd);
303
304	// Unlock and return
305	__atomic_unlock( &lanes.data[i].lock );
306
307	#if !defined(__CFA_NO_STATISTICS__)
308	if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
309	else __tls_stats()->ready.push.local.success++;
310	#endif
311
312	__cfadbg_print_safe(ready_queue, "Kernel : Pushed %p on cluster %p (idx: %u, mask %llu, first %d)\n", thrd, cltr, i, used.mask[0], lane_first);
313
314	}
315
316	// Pop from the ready queue from a given cluster
317	__attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
318	/* paranoid */ verify( lanes.count > 0 );
319	/* paranoid */ verify( kernelTLS().this_processor );
320
321	const int cpu = __kernel_getcpu();
322	/* paranoid */ verify(cpu >= 0);
323	/* paranoid */ verify(cpu < cpu_info.hthrd_count);
324	/* paranoid / verify(cpu READYQ_SHARD_FACTOR < lanes.count);
325
326	const cpu_map_entry_t & map = cpu_info.llc_map[cpu];
327	/* paranoid / verify(map.start READYQ_SHARD_FACTOR < lanes.count);
328	/* paranoid / verify(map.self READYQ_SHARD_FACTOR < lanes.count);
329	/* paranoid / verifyf((map.start + map.count) READYQ_SHARD_FACTOR <= lanes.count, "have %u lanes but map can go up to %u", lanes.count, (map.start + map.count) * READYQ_SHARD_FACTOR);
330
331	processor * const proc = kernelTLS().this_processor;
332	const int start = map.self * READYQ_SHARD_FACTOR;
333
334	// Did we already have a help target
335	if(proc->rdq.target == -1u) {
336	// if We don't have a
337	unsigned long long min = ts(lanes.data[start]);
338	for(i; READYQ_SHARD_FACTOR) {
339	unsigned long long tsc = ts(lanes.data[start + i]);
340	if(tsc < min) min = tsc;
341	}
342	proc->rdq.cutoff = min;
343	proc->rdq.target = (map.start * READYQ_SHARD_FACTOR) + (__tls_rand() % (map.count* READYQ_SHARD_FACTOR));
344	}
345	else {
346	const unsigned long long bias = 0; //2_500_000_000;
347	const unsigned long long cutoff = proc->rdq.cutoff > bias ? proc->rdq.cutoff - bias : proc->rdq.cutoff;
348	{
349	unsigned target = proc->rdq.target;
350	proc->rdq.target = -1u;
351	if(lanes.tscs[target].tv < cutoff && ts(lanes.data[target]) < cutoff) {
352	$thread * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
353	proc->rdq.last = target;
354	if(t) return t;
355	}
356	}
357
358	unsigned last = proc->rdq.last;
359	if(last != -1u && lanes.tscs[last].tv < cutoff && ts(lanes.data[last]) < cutoff) {
360	$thread * t = try_pop(cltr, last __STATS(, __tls_stats()->ready.pop.help));
361	if(t) return t;
362	}
363	else {
364	proc->rdq.last = -1u;
365	}
366	}
367
368	for(READYQ_SHARD_FACTOR) {
369	unsigned i = start + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
370	if($thread * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
371	}
372
373	// All lanes where empty return 0p
374	return 0p;
375	}
376
377	__attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
378	processor * const proc = kernelTLS().this_processor;
379	unsigned last = proc->rdq.last;
380
381	unsigned i = __tls_rand() % lanes.count;
382	return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
383	}
384	__attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) {
385	return search(cltr);
386	}
387	#endif
388	#if defined(USE_RELAXED_FIFO)
389	//-----------------------------------------------------------------------
390	// get index from random number with or without bias towards queues
391	static inline [unsigned, bool] idx_from_r(unsigned r, unsigned preferred) {
392	unsigned i;
393	bool local;
394	unsigned rlow = r % BIAS;
395	unsigned rhigh = r / BIAS;
396	if((0 != rlow) && preferred >= 0) {
397	// (BIAS - 1) out of BIAS chances
398	// Use perferred queues
399	i = preferred + (rhigh % READYQ_SHARD_FACTOR);
400	local = true;
401	}
402	else {
403	// 1 out of BIAS chances
404	// Use all queues
405	i = rhigh;
406	local = false;
407	}
408	return [i, local];
409	}
410
411	__attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd, bool push_local) with (cltr->ready_queue) {
412	__cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
413
414	const bool external = !push_local \|\| (!kernelTLS().this_processor) \|\| (cltr != kernelTLS().this_processor->cltr);
415	/* paranoid */ verify(external \|\| kernelTLS().this_processor->rdq.id < lanes.count );
416
417	bool local;
418	int preferred = external ? -1 : kernelTLS().this_processor->rdq.id;
419
420	// Try to pick a lane and lock it
421	unsigned i;
422	do {
423	// Pick the index of a lane
424	unsigned r = __tls_rand_fwd();
425	[i, local] = idx_from_r(r, preferred);
426
427	i %= __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
428
429	#if !defined(__CFA_NO_STATISTICS__)
430	if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
431	else if(local) __tls_stats()->ready.push.local.attempt++;
432	else __tls_stats()->ready.push.share.attempt++;
433	#endif
434
435	// If we can't lock it retry
436	} while( !__atomic_try_acquire( &lanes.data[i].lock ) );
437
438	// Actually push it
439	push(lanes.data[i], thrd);
440
441	// Unlock and return
442	__atomic_unlock( &lanes.data[i].lock );
443
444	// Mark the current index in the tls rng instance as having an item
445	__tls_rand_advance_bck();
446
447	__cfadbg_print_safe(ready_queue, "Kernel : Pushed %p on cluster %p (idx: %u, mask %llu, first %d)\n", thrd, cltr, i, used.mask[0], lane_first);
448
449	// Update statistics
450	#if !defined(__CFA_NO_STATISTICS__)
451	if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
452	else if(local) __tls_stats()->ready.push.local.success++;
453	else __tls_stats()->ready.push.share.success++;
454	#endif
455	}
456
457	// Pop from the ready queue from a given cluster
458	__attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
459	/* paranoid */ verify( lanes.count > 0 );
460	/* paranoid */ verify( kernelTLS().this_processor );
461	/* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
462
463	unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
464	int preferred = kernelTLS().this_processor->rdq.id;
465
466
467	// As long as the list is not empty, try finding a lane that isn't empty and pop from it
468	for(25) {
469	// Pick two lists at random
470	unsigned ri = __tls_rand_bck();
471	unsigned rj = __tls_rand_bck();
472
473	unsigned i, j;
474	__attribute__((unused)) bool locali, localj;
475	[i, locali] = idx_from_r(ri, preferred);
476	[j, localj] = idx_from_r(rj, preferred);
477
478	i %= count;
479	j %= count;
480
481	// try popping from the 2 picked lists
482	struct $thread * thrd = try_pop(cltr, i, j __STATS(, *(locali \|\| localj ? &__tls_stats()->ready.pop.local : &__tls_stats()->ready.pop.help)));
483	if(thrd) {
484	return thrd;
485	}
486	}
487
488	// All lanes where empty return 0p
489	return 0p;
490	}
491
492	__attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) { return pop_fast(cltr); }
493	__attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) {
494	return search(cltr);
495	}
496	#endif
497	#if defined(USE_WORK_STEALING)
498	__attribute__((hot)) void push(struct cluster * cltr, struct $thread * thrd, bool push_local) with (cltr->ready_queue) {
499	__cfadbg_print_safe(ready_queue, "Kernel : Pushing %p on cluster %p\n", thrd, cltr);
500
501	// #define USE_PREFERRED
502	#if !defined(USE_PREFERRED)
503	const bool external = !push_local \|\| (!kernelTLS().this_processor) \|\| (cltr != kernelTLS().this_processor->cltr);
504	/* paranoid */ verify(external \|\| kernelTLS().this_processor->rdq.id < lanes.count );
505	#else
506	unsigned preferred = thrd->preferred;
507	const bool external = push_local \|\| (!kernelTLS().this_processor) \|\| preferred == -1u \|\| thrd->curr_cluster != cltr;
508	/* paranoid */ verifyf(external \|\| preferred < lanes.count, "Invalid preferred queue %u for %u lanes", preferred, lanes.count );
509
510	unsigned r = preferred % READYQ_SHARD_FACTOR;
511	const unsigned start = preferred - r;
512	#endif
513
514	// Try to pick a lane and lock it
515	unsigned i;
516	do {
517	#if !defined(__CFA_NO_STATISTICS__)
518	if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.attempt, 1, __ATOMIC_RELAXED);
519	else __tls_stats()->ready.push.local.attempt++;
520	#endif
521
522	if(unlikely(external)) {
523	i = __tls_rand() % lanes.count;
524	}
525	else {
526	#if !defined(USE_PREFERRED)
527	processor * proc = kernelTLS().this_processor;
528	unsigned r = proc->rdq.its++;
529	i = proc->rdq.id + (r % READYQ_SHARD_FACTOR);
530	#else
531	i = start + (r++ % READYQ_SHARD_FACTOR);
532	#endif
533	}
534	// If we can't lock it retry
535	} while( !__atomic_try_acquire( &lanes.data[i].lock ) );
536
537	// Actually push it
538	push(lanes.data[i], thrd);
539
540	// Unlock and return
541	__atomic_unlock( &lanes.data[i].lock );
542
543	#if !defined(__CFA_NO_STATISTICS__)
544	if(unlikely(external)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
545	else __tls_stats()->ready.push.local.success++;
546	#endif
547
548	__cfadbg_print_safe(ready_queue, "Kernel : Pushed %p on cluster %p (idx: %u, mask %llu, first %d)\n", thrd, cltr, i, used.mask[0], lane_first);
549	}
550
551	// Pop from the ready queue from a given cluster
552	__attribute__((hot)) $thread * pop_fast(struct cluster * cltr) with (cltr->ready_queue) {
553	/* paranoid */ verify( lanes.count > 0 );
554	/* paranoid */ verify( kernelTLS().this_processor );
555	/* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count );
556
557	processor * proc = kernelTLS().this_processor;
558
559	if(proc->rdq.target == -1u) {
560	unsigned long long min = ts(lanes.data[proc->rdq.id]);
561	for(int i = 0; i < READYQ_SHARD_FACTOR; i++) {
562	unsigned long long tsc = ts(lanes.data[proc->rdq.id + i]);
563	if(tsc < min) min = tsc;
564	}
565	proc->rdq.cutoff = min;
566	proc->rdq.target = __tls_rand() % lanes.count;
567	}
568	else {
569	unsigned target = proc->rdq.target;
570	proc->rdq.target = -1u;
571	const unsigned long long bias = 0; //2_500_000_000;
572	const unsigned long long cutoff = proc->rdq.cutoff > bias ? proc->rdq.cutoff - bias : proc->rdq.cutoff;
573	if(lanes.tscs[target].tv < cutoff && ts(lanes.data[target]) < cutoff) {
574	$thread * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
575	if(t) return t;
576	}
577	}
578
579	for(READYQ_SHARD_FACTOR) {
580	unsigned i = proc->rdq.id + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
581	if($thread * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
582	}
583	return 0p;
584	}
585
586	__attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr) with (cltr->ready_queue) {
587	unsigned i = __tls_rand() % lanes.count;
588	return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
589	}
590
591	__attribute__((hot)) struct $thread * pop_search(struct cluster * cltr) with (cltr->ready_queue) {
592	return search(cltr);
593	}
594	#endif
595
596	//=======================================================================
597	// Various Ready Queue utilities
598	//=======================================================================
599	// these function work the same or almost the same
600	// whether they are using work-stealing or relaxed fifo scheduling
601
602	//-----------------------------------------------------------------------
603	// try to pop from a lane given by index w
604	static inline struct $thread * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
605	__STATS( stats.attempt++; )
606
607	// Get relevant elements locally
608	__intrusive_lane_t & lane = lanes.data[w];
609
610	// If list looks empty retry
611	if( is_empty(lane) ) {
612	return 0p;
613	}
614
615	// If we can't get the lock retry
616	if( !__atomic_try_acquire(&lane.lock) ) {
617	return 0p;
618	}
619
620	// If list is empty, unlock and retry
621	if( is_empty(lane) ) {
622	__atomic_unlock(&lane.lock);
623	return 0p;
624	}
625
626	// Actually pop the list
627	struct $thread * thrd;
628	unsigned long long tsv;
629	[thrd, tsv] = pop(lane);
630
631	/* paranoid */ verify(thrd);
632	/* paranoid */ verify(tsv);
633	/* paranoid */ verify(lane.lock);
634
635	// Unlock and return
636	__atomic_unlock(&lane.lock);
637
638	// Update statistics
639	__STATS( stats.success++; )
640
641	#if defined(USE_WORK_STEALING)
642	lanes.tscs[w].tv = tsv;
643	#endif
644
645	thrd->preferred = w;
646
647	// return the popped thread
648	return thrd;
649	}
650
651	//-----------------------------------------------------------------------
652	// try to pop from any lanes making sure you don't miss any threads push
653	// before the start of the function
654	static inline struct $thread * search(struct cluster * cltr) with (cltr->ready_queue) {
655	/* paranoid */ verify( lanes.count > 0 );
656	unsigned count = __atomic_load_n( &lanes.count, __ATOMIC_RELAXED );
657	unsigned offset = __tls_rand();
658	for(i; count) {
659	unsigned idx = (offset + i) % count;
660	struct $thread * thrd = try_pop(cltr, idx __STATS(, __tls_stats()->ready.pop.search));
661	if(thrd) {
662	return thrd;
663	}
664	}
665
666	// All lanes where empty return 0p
667	return 0p;
668	}
669
670	//-----------------------------------------------------------------------
671	// Check that all the intrusive queues in the data structure are still consistent
672	static void check( __ready_queue_t & q ) with (q) {
673	#if defined(__CFA_WITH_VERIFY__)
674	{
675	for( idx ; lanes.count ) {
676	__intrusive_lane_t & sl = lanes.data[idx];
677	assert(!lanes.data[idx].lock);
678
679	if(is_empty(sl)) {
680	assert( sl.anchor.next == 0p );
681	assert( sl.anchor.ts == -1llu );
682	assert( mock_head(sl) == sl.prev );
683	} else {
684	assert( sl.anchor.next != 0p );
685	assert( sl.anchor.ts != -1llu );
686	assert( mock_head(sl) != sl.prev );
687	}
688	}
689	}
690	#endif
691	}
692
693	//-----------------------------------------------------------------------
694	// Given 2 indexes, pick the list with the oldest push an try to pop from it
695	static inline struct $thread * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) {
696	// Pick the bet list
697	int w = i;
698	if( __builtin_expect(!is_empty(lanes.data[j]), true) ) {
699	w = (ts(lanes.data[i]) < ts(lanes.data[j])) ? i : j;
700	}
701
702	return try_pop(cltr, w __STATS(, stats));
703	}
704
705	// Call this function of the intrusive list was moved using memcpy
706	// fixes the list so that the pointers back to anchors aren't left dangling
707	static inline void fix(__intrusive_lane_t & ll) {
708	if(is_empty(ll)) {
709	verify(ll.anchor.next == 0p);
710	ll.prev = mock_head(ll);
711	}
712	}
713
714	static void assign_list(unsigned & value, dlist(processor) & list, unsigned count) {
715	processor * it = &list`first;
716	for(unsigned i = 0; i < count; i++) {
717	/* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
718	it->rdq.id = value;
719	it->rdq.target = -1u;
720	value += READYQ_SHARD_FACTOR;
721	it = &(*it)`next;
722	}
723	}
724
725	static void reassign_cltr_id(struct cluster * cltr) {
726	unsigned preferred = 0;
727	assign_list(preferred, cltr->procs.actives, cltr->procs.total - cltr->procs.idle);
728	assign_list(preferred, cltr->procs.idles , cltr->procs.idle );
729	}
730
731	static void fix_times( struct cluster * cltr ) with( cltr->ready_queue ) {
732	#if defined(USE_WORK_STEALING)
733	lanes.tscs = alloc(lanes.count, lanes.tscs`realloc);
734	for(i; lanes.count) {
735	unsigned long long tsc1 = ts(lanes.data[i]);
736	unsigned long long tsc2 = rdtscl()
737	lanes.tscs[i].tv = min(tsc1, tsc2);
738	}
739	#endif
740	}
741
742	#if defined(USE_CPU_WORK_STEALING)
743	// ready_queue size is fixed in this case
744	void ready_queue_grow(struct cluster * cltr) {}
745	void ready_queue_shrink(struct cluster * cltr) {}
746	#else
747	// Grow the ready queue
748	void ready_queue_grow(struct cluster * cltr) {
749	size_t ncount;
750	int target = cltr->procs.total;
751
752	/* paranoid */ verify( ready_mutate_islocked() );
753	__cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue\n");
754
755	// Make sure that everything is consistent
756	/* paranoid */ check( cltr->ready_queue );
757
758	// grow the ready queue
759	with( cltr->ready_queue ) {
760	// Find new count
761	// Make sure we always have atleast 1 list
762	if(target >= 2) {
763	ncount = target * READYQ_SHARD_FACTOR;
764	} else {
765	ncount = SEQUENTIAL_SHARD;
766	}
767
768	// Allocate new array (uses realloc and memcpies the data)
769	lanes.data = alloc( ncount, lanes.data`realloc );
770
771	// Fix the moved data
772	for( idx; (size_t)lanes.count ) {
773	fix(lanes.data[idx]);
774	}
775
776	// Construct new data
777	for( idx; (size_t)lanes.count ~ ncount) {
778	(lanes.data[idx]){};
779	}
780
781	// Update original
782	lanes.count = ncount;
783	}
784
785	fix_times(cltr);
786
787	reassign_cltr_id(cltr);
788
789	// Make sure that everything is consistent
790	/* paranoid */ check( cltr->ready_queue );
791
792	__cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue done\n");
793
794	/* paranoid */ verify( ready_mutate_islocked() );
795	}
796
797	// Shrink the ready queue
798	void ready_queue_shrink(struct cluster * cltr) {
799	/* paranoid */ verify( ready_mutate_islocked() );
800	__cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue\n");
801
802	// Make sure that everything is consistent
803	/* paranoid */ check( cltr->ready_queue );
804
805	int target = cltr->procs.total;
806
807	with( cltr->ready_queue ) {
808	// Remember old count
809	size_t ocount = lanes.count;
810
811	// Find new count
812	// Make sure we always have atleast 1 list
813	lanes.count = target >= 2 ? target * READYQ_SHARD_FACTOR: SEQUENTIAL_SHARD;
814	/* paranoid */ verify( ocount >= lanes.count );
815	/* paranoid / verify( lanes.count == target READYQ_SHARD_FACTOR \|\| target < 2 );
816
817	// for printing count the number of displaced threads
818	#if defined(__CFA_DEBUG_PRINT__) \|\| defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
819	__attribute__((unused)) size_t displaced = 0;
820	#endif
821
822	// redistribute old data
823	for( idx; (size_t)lanes.count ~ ocount) {
824	// Lock is not strictly needed but makes checking invariants much easier
825	__attribute__((unused)) bool locked = __atomic_try_acquire(&lanes.data[idx].lock);
826	verify(locked);
827
828	// As long as we can pop from this lane to push the threads somewhere else in the queue
829	while(!is_empty(lanes.data[idx])) {
830	struct $thread * thrd;
831	unsigned long long _;
832	[thrd, _] = pop(lanes.data[idx]);
833
834	push(cltr, thrd, true);
835
836	// for printing count the number of displaced threads
837	#if defined(__CFA_DEBUG_PRINT__) \|\| defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
838	displaced++;
839	#endif
840	}
841
842	// Unlock the lane
843	__atomic_unlock(&lanes.data[idx].lock);
844
845	// TODO print the queue statistics here
846
847	^(lanes.data[idx]){};
848	}
849
850	__cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue displaced %zu threads\n", displaced);
851
852	// Allocate new array (uses realloc and memcpies the data)
853	lanes.data = alloc( lanes.count, lanes.data`realloc );
854
855	// Fix the moved data
856	for( idx; (size_t)lanes.count ) {
857	fix(lanes.data[idx]);
858	}
859	}
860
861	fix_times(cltr);
862
863	reassign_cltr_id(cltr);
864
865	// Make sure that everything is consistent
866	/* paranoid */ check( cltr->ready_queue );
867
868	__cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue done\n");
869	/* paranoid */ verify( ready_mutate_islocked() );
870	}
871	#endif
872
873	#if !defined(__CFA_NO_STATISTICS__)
874	unsigned cnt(const __ready_queue_t & this, unsigned idx) {
875	/* paranoid */ verify(this.lanes.count > idx);
876	return this.lanes.data[idx].cnt;
877	}
878	#endif
879
880
881	#if defined(CFA_HAVE_LINUX_LIBRSEQ)
882	// No definition needed
883	#elif defined(CFA_HAVE_LINUX_RSEQ_H)
884
885	#if defined( __x86_64 ) \|\| defined( __i386 )
886	#define RSEQ_SIG 0x53053053
887	#elif defined( __ARM_ARCH )
888	#ifdef __ARMEB__
889	#define RSEQ_SIG 0xf3def5e7 /* udf #24035 ; 0x5de3 (ARMv6+) */
890	#else
891	#define RSEQ_SIG 0xe7f5def3 /* udf #24035 ; 0x5de3 */
892	#endif
893	#endif
894
895	extern void __disable_interrupts_hard();
896	extern void __enable_interrupts_hard();
897
898	void __kernel_raw_rseq_register (void) {
899	/* paranoid */ verify( __cfaabi_rseq.cpu_id == RSEQ_CPU_ID_UNINITIALIZED );
900
901	// int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), 0, (sigset_t *)0p, _NSIG / 8);
902	int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), 0, RSEQ_SIG);
903	if(ret != 0) {
904	int e = errno;
905	switch(e) {
906	case EINVAL: abort("KERNEL ERROR: rseq register invalid argument");
907	case ENOSYS: abort("KERNEL ERROR: rseq register no supported");
908	case EFAULT: abort("KERNEL ERROR: rseq register with invalid argument");
909	case EBUSY : abort("KERNEL ERROR: rseq register already registered");
910	case EPERM : abort("KERNEL ERROR: rseq register sig argument on unregistration does not match the signature received on registration");
911	default: abort("KERNEL ERROR: rseq register unexpected return %d", e);
912	}
913	}
914	}
915
916	void __kernel_raw_rseq_unregister(void) {
917	/* paranoid */ verify( __cfaabi_rseq.cpu_id >= 0 );
918
919	// int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), RSEQ_FLAG_UNREGISTER, (sigset_t *)0p, _NSIG / 8);
920	int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), RSEQ_FLAG_UNREGISTER, RSEQ_SIG);
921	if(ret != 0) {
922	int e = errno;
923	switch(e) {
924	case EINVAL: abort("KERNEL ERROR: rseq unregister invalid argument");
925	case ENOSYS: abort("KERNEL ERROR: rseq unregister no supported");
926	case EFAULT: abort("KERNEL ERROR: rseq unregister with invalid argument");
927	case EBUSY : abort("KERNEL ERROR: rseq unregister already registered");
928	case EPERM : abort("KERNEL ERROR: rseq unregister sig argument on unregistration does not match the signature received on registration");
929	default: abort("KERNEL ERROR: rseq unregisteunexpected return %d", e);
930	}
931	}
932	}
933	#else
934	// No definition needed
935	#endif

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

Download in other formats:

Original Format