source: libcfa/src/concurrency/kernel_private.hfa @ 24e321c

Last change on this file since 24e321c was 24e321c, checked in by Thierry Delisle <tdelisle@…>, 2 months ago

Unpark now takes a hint on locality.

  • Property mode set to 100644
File size: 10.9 KB
Line 
1//
2// Cforall Version 1.0.0 Copyright (C) 2016 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// kernel_private.hfa --
8//
9// Author           : Thierry Delisle
10// Created On       : Mon Feb 13 12:27:26 2017
11// Last Modified By : Peter A. Buhr
12// Last Modified On : Wed Aug 12 08:21:33 2020
13// Update Count     : 9
14//
15
16#pragma once
17
18#if !defined(__cforall_thread__)
19        #error kernel_private.hfa should only be included in libcfathread source
20#endif
21
22#include "kernel.hfa"
23#include "thread.hfa"
24
25#include "alarm.hfa"
26#include "stats.hfa"
27
28extern "C" {
29#if   defined(CFA_HAVE_LINUX_LIBRSEQ)
30        #include <rseq/rseq.h>
31#elif defined(CFA_HAVE_LINUX_RSEQ_H)
32        #include <linux/rseq.h>
33#else
34        #ifndef _GNU_SOURCE
35        #error kernel_private requires gnu_source
36        #endif
37        #include <sched.h>
38#endif
39}
40
41//-----------------------------------------------------------------------------
42// Scheduler
43extern "C" {
44        void disable_interrupts() OPTIONAL_THREAD;
45        void enable_interrupts( bool poll = true );
46}
47
48void schedule_thread$( thread$ *, unpark_hint hint ) __attribute__((nonnull (1)));
49
50extern bool __preemption_enabled();
51
52//release/wake-up the following resources
53void __thread_finish( thread$ * thrd );
54
55//-----------------------------------------------------------------------------
56// Hardware
57
58#if   defined(CFA_HAVE_LINUX_LIBRSEQ)
59        // No data needed
60#elif defined(CFA_HAVE_LINUX_RSEQ_H)
61        extern "Cforall" {
62                extern __attribute__((aligned(128))) thread_local volatile struct rseq __cfaabi_rseq;
63        }
64#else
65        // No data needed
66#endif
67
68static inline int __kernel_getcpu() {
69        /* paranoid */ verify( ! __preemption_enabled() );
70#if   defined(CFA_HAVE_LINUX_LIBRSEQ)
71        return rseq_current_cpu();
72#elif defined(CFA_HAVE_LINUX_RSEQ_H)
73        int r = __cfaabi_rseq.cpu_id;
74        /* paranoid */ verify( r >= 0 );
75        return r;
76#else
77        return sched_getcpu();
78#endif
79}
80
81//-----------------------------------------------------------------------------
82// Processor
83void main(processorCtx_t *);
84
85void * __create_pthread( pthread_t *, void * (*)(void *), void * );
86void __destroy_pthread( pthread_t pthread, void * stack, void ** retval );
87
88extern cluster * mainCluster;
89
90//-----------------------------------------------------------------------------
91// Threads
92extern "C" {
93      void __cfactx_invoke_thread(void (*main)(void *), void * this);
94}
95
96__cfaabi_dbg_debug_do(
97        extern void __cfaabi_dbg_thread_register  ( thread$ * thrd );
98        extern void __cfaabi_dbg_thread_unregister( thread$ * thrd );
99)
100
101#define TICKET_BLOCKED (-1) // thread is blocked
102#define TICKET_RUNNING ( 0) // thread is running
103#define TICKET_UNBLOCK ( 1) // thread should ignore next block
104
105//-----------------------------------------------------------------------------
106// Utils
107void doregister( struct cluster * cltr, struct thread$ & thrd );
108void unregister( struct cluster * cltr, struct thread$ & thrd );
109
110//-----------------------------------------------------------------------------
111// I/O
112$io_arbiter * create(void);
113void destroy($io_arbiter *);
114
115//=======================================================================
116// Cluster lock API
117//=======================================================================
118// Lock-Free registering/unregistering of threads
119// Register a processor to a given cluster and get its unique id in return
120unsigned register_proc_id( void );
121
122// Unregister a processor from a given cluster using its id, getting back the original pointer
123void unregister_proc_id( unsigned );
124
125//=======================================================================
126// Reader-writer lock implementation
127// Concurrent with doregister/unregister,
128//    i.e., threads can be added at any point during or between the entry/exit
129
130//-----------------------------------------------------------------------
131// simple spinlock underlying the RWLock
132// Blocking acquire
133static inline void __atomic_acquire(volatile bool * ll) {
134        while( __builtin_expect(__atomic_exchange_n(ll, (bool)true, __ATOMIC_SEQ_CST), false) ) {
135                while(__atomic_load_n(ll, (int)__ATOMIC_RELAXED))
136                        Pause();
137        }
138        /* paranoid */ verify(*ll);
139}
140
141// Non-Blocking acquire
142static inline bool __atomic_try_acquire(volatile bool * ll) {
143        return !__atomic_exchange_n(ll, (bool)true, __ATOMIC_SEQ_CST);
144}
145
146// Release
147static inline void __atomic_unlock(volatile bool * ll) {
148        /* paranoid */ verify(*ll);
149        __atomic_store_n(ll, (bool)false, __ATOMIC_RELEASE);
150}
151
152
153
154
155
156//-----------------------------------------------------------------------
157// Reader-Writer lock protecting the ready-queues
158// while this lock is mostly generic some aspects
159// have been hard-coded to for the ready-queue for
160// simplicity and performance
161struct __scheduler_RWLock_t {
162        // total cachelines allocated
163        unsigned int max;
164
165        // cachelines currently in use
166        volatile unsigned int alloc;
167
168        // cachelines ready to itereate over
169        // (!= to alloc when thread is in second half of doregister)
170        volatile unsigned int ready;
171
172        // writer lock
173        volatile bool write_lock;
174
175        // data pointer
176        volatile bool * volatile * data;
177};
178
179void  ?{}(__scheduler_RWLock_t & this);
180void ^?{}(__scheduler_RWLock_t & this);
181
182extern __scheduler_RWLock_t * __scheduler_lock;
183
184//-----------------------------------------------------------------------
185// Reader side : acquire when using the ready queue to schedule but not
186//  creating/destroying queues
187static inline void ready_schedule_lock(void) with(*__scheduler_lock) {
188        /* paranoid */ verify( ! __preemption_enabled() );
189        /* paranoid */ verify( ! kernelTLS().in_sched_lock );
190        /* paranoid */ verify( data[kernelTLS().sched_id] == &kernelTLS().sched_lock );
191        /* paranoid */ verify( !kernelTLS().this_processor || kernelTLS().this_processor->unique_id == kernelTLS().sched_id );
192
193        // Step 1 : make sure no writer are in the middle of the critical section
194        while(__atomic_load_n(&write_lock, (int)__ATOMIC_RELAXED))
195                Pause();
196
197        // Fence needed because we don't want to start trying to acquire the lock
198        // before we read a false.
199        // Not needed on x86
200        // std::atomic_thread_fence(std::memory_order_seq_cst);
201
202        // Step 2 : acquire our local lock
203        __atomic_acquire( &kernelTLS().sched_lock );
204        /*paranoid*/ verify(kernelTLS().sched_lock);
205
206        #ifdef __CFA_WITH_VERIFY__
207                // Debug, check if this is owned for reading
208                kernelTLS().in_sched_lock = true;
209        #endif
210}
211
212static inline void ready_schedule_unlock(void) with(*__scheduler_lock) {
213        /* paranoid */ verify( ! __preemption_enabled() );
214        /* paranoid */ verify( data[kernelTLS().sched_id] == &kernelTLS().sched_lock );
215        /* paranoid */ verify( !kernelTLS().this_processor || kernelTLS().this_processor->unique_id == kernelTLS().sched_id );
216        /* paranoid */ verify( kernelTLS().sched_lock );
217        /* paranoid */ verify( kernelTLS().in_sched_lock );
218        #ifdef __CFA_WITH_VERIFY__
219                // Debug, check if this is owned for reading
220                kernelTLS().in_sched_lock = false;
221        #endif
222        __atomic_unlock(&kernelTLS().sched_lock);
223}
224
225#ifdef __CFA_WITH_VERIFY__
226        static inline bool ready_schedule_islocked(void) {
227                /* paranoid */ verify( ! __preemption_enabled() );
228                /* paranoid */ verify( (!kernelTLS().in_sched_lock) || kernelTLS().sched_lock );
229                return kernelTLS().sched_lock;
230        }
231
232        static inline bool ready_mutate_islocked() {
233                return __scheduler_lock->write_lock;
234        }
235#endif
236
237//-----------------------------------------------------------------------
238// Writer side : acquire when changing the ready queue, e.g. adding more
239//  queues or removing them.
240uint_fast32_t ready_mutate_lock( void );
241
242void ready_mutate_unlock( uint_fast32_t /* value returned by lock */ );
243
244//-----------------------------------------------------------------------
245// Lock-Free registering/unregistering of threads
246// Register a processor to a given cluster and get its unique id in return
247// For convenience, also acquires the lock
248static inline [unsigned, uint_fast32_t] ready_mutate_register() {
249        unsigned id = register_proc_id();
250        uint_fast32_t last = ready_mutate_lock();
251        return [id, last];
252}
253
254// Unregister a processor from a given cluster using its id, getting back the original pointer
255// assumes the lock is acquired
256static inline void ready_mutate_unregister( unsigned id, uint_fast32_t last_s ) {
257        ready_mutate_unlock( last_s );
258        unregister_proc_id( id );
259}
260
261//-----------------------------------------------------------------------
262// Cluster idle lock/unlock
263static inline void lock(__cluster_proc_list & this) {
264        /* paranoid */ verify( ! __preemption_enabled() );
265
266        // Start by locking the global RWlock so that we know no-one is
267        // adding/removing processors while we mess with the idle lock
268        ready_schedule_lock();
269
270        // Simple counting lock, acquired, acquired by incrementing the counter
271        // to an odd number
272        for() {
273                uint64_t l = this.lock;
274                if(
275                        (0 == (l % 2))
276                        && __atomic_compare_exchange_n(&this.lock, &l, l + 1, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)
277                ) return;
278                Pause();
279        }
280
281        /* paranoid */ verify( ! __preemption_enabled() );
282}
283
284static inline void unlock(__cluster_proc_list & this) {
285        /* paranoid */ verify( ! __preemption_enabled() );
286
287        /* paranoid */ verify( 1 == (this.lock % 2) );
288        // Simple couting lock, release by incrementing to an even number
289        __atomic_fetch_add( &this.lock, 1, __ATOMIC_SEQ_CST );
290
291        // Release the global lock, which we acquired when locking
292        ready_schedule_unlock();
293
294        /* paranoid */ verify( ! __preemption_enabled() );
295}
296
297//=======================================================================
298// Ready-Queue API
299//-----------------------------------------------------------------------
300// push thread onto a ready queue for a cluster
301// returns true if the list was previously empty, false otherwise
302__attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint);
303
304//-----------------------------------------------------------------------
305// pop thread from the local queues of a cluster
306// returns 0p if empty
307// May return 0p spuriously
308__attribute__((hot)) struct thread$ * pop_fast(struct cluster * cltr);
309
310//-----------------------------------------------------------------------
311// pop thread from any ready queue of a cluster
312// returns 0p if empty
313// May return 0p spuriously
314__attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr);
315
316//-----------------------------------------------------------------------
317// search all ready queues of a cluster for any thread
318// returns 0p if empty
319// guaranteed to find any threads added before this call
320__attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr);
321
322//-----------------------------------------------------------------------
323// get preferred ready for new thread
324unsigned ready_queue_new_preferred();
325
326//-----------------------------------------------------------------------
327// Increase the width of the ready queue (number of lanes) by 4
328void ready_queue_grow  (struct cluster * cltr);
329
330//-----------------------------------------------------------------------
331// Decrease the width of the ready queue (number of lanes) by 4
332void ready_queue_shrink(struct cluster * cltr);
333
334
335// Local Variables: //
336// mode: c //
337// tab-width: 4 //
338// End: //
Note: See TracBrowser for help on using the repository browser.