source: libcfa/src/concurrency/kernel/private.hfa @ 2af1943

ADTast-experimentalpthread-emulation
Last change on this file since 2af1943 was 2af1943, checked in by Thierry Delisle <tdelisle@…>, 19 months ago

Small improvements to some of the alignment requirements in CFA runtime.

  • Property mode set to 100644
File size: 12.1 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// Defines whether or not we *want* to use io_uring_enter as the idle_sleep blocking call
42// #define CFA_WANT_IO_URING_IDLE
43
44// Defines whether or not we *can* use io_uring_enter as the idle_sleep blocking call
45#if defined(CFA_WANT_IO_URING_IDLE) && defined(CFA_HAVE_LINUX_IO_URING_H)
46        #if defined(CFA_HAVE_IORING_OP_READ) || (defined(CFA_HAVE_READV) && defined(CFA_HAVE_IORING_OP_READV))
47                #define CFA_WITH_IO_URING_IDLE
48        #endif
49#endif
50
51//-----------------------------------------------------------------------------
52// Scheduler
53extern "C" {
54        void disable_interrupts() OPTIONAL_THREAD;
55        void enable_interrupts( bool poll = true );
56}
57
58void schedule_thread$( thread$ *, unpark_hint hint ) __attribute__((nonnull (1)));
59
60extern bool __preemption_enabled();
61
62enum {
63        PREEMPT_NORMAL    = 0,
64        PREEMPT_TERMINATE = 1,
65        PREEMPT_IO = 2,
66};
67
68static inline void __disable_interrupts_checked() {
69        /* paranoid */ verify( __preemption_enabled() );
70        disable_interrupts();
71        /* paranoid */ verify( ! __preemption_enabled() );
72}
73
74static inline void __enable_interrupts_checked( bool poll = true ) {
75        /* paranoid */ verify( ! __preemption_enabled() );
76        enable_interrupts( poll );
77        /* paranoid */ verify( __preemption_enabled() );
78}
79
80//release/wake-up the following resources
81void __thread_finish( thread$ * thrd );
82
83//-----------------------------------------------------------------------------
84// Hardware
85
86#if   defined(CFA_HAVE_LINUX_LIBRSEQ)
87        // No data needed
88#elif defined(CFA_HAVE_LINUX_RSEQ_H)
89        extern "Cforall" {
90                extern __attribute__((aligned(64))) thread_local volatile struct rseq __cfaabi_rseq;
91        }
92#else
93        // No data needed
94#endif
95
96static inline int __kernel_getcpu() {
97        /* paranoid */ verify( ! __preemption_enabled() );
98#if   defined(CFA_HAVE_LINUX_LIBRSEQ)
99        return rseq_current_cpu();
100#elif defined(CFA_HAVE_LINUX_RSEQ_H)
101        int r = __cfaabi_rseq.cpu_id;
102        /* paranoid */ verify( r >= 0 );
103        return r;
104#else
105        return sched_getcpu();
106#endif
107}
108
109//-----------------------------------------------------------------------------
110// Processor
111void main(processorCtx_t &);
112static inline coroutine$* get_coroutine(processorCtx_t & this) { return &this.self; }
113
114void * __create_pthread( pthread_t *, void * (*)(void *), void * );
115void __destroy_pthread( pthread_t pthread, void * stack, void ** retval );
116
117extern cluster * mainCluster;
118
119//-----------------------------------------------------------------------------
120// Threads
121extern "C" {
122      void __cfactx_invoke_thread(void (*main)(void *), void * this);
123}
124
125__cfaabi_dbg_debug_do(
126        extern void __cfaabi_dbg_thread_register  ( thread$ * thrd );
127        extern void __cfaabi_dbg_thread_unregister( thread$ * thrd );
128)
129
130#define TICKET_BLOCKED (-1) // thread is blocked
131#define TICKET_RUNNING ( 0) // thread is running
132#define TICKET_UNBLOCK ( 1) // thread should ignore next block
133
134//-----------------------------------------------------------------------------
135// Utils
136void doregister( struct cluster * cltr, struct thread$ & thrd );
137void unregister( struct cluster * cltr, struct thread$ & thrd );
138
139//-----------------------------------------------------------------------------
140// I/O
141$io_arbiter * create(void);
142void destroy($io_arbiter *);
143
144//=======================================================================
145// Cluster lock API
146//=======================================================================
147// Lock-Free registering/unregistering of threads
148// Register a processor to a given cluster and get its unique id in return
149unsigned register_proc_id( void );
150
151// Unregister a processor from a given cluster using its id, getting back the original pointer
152void unregister_proc_id( unsigned );
153
154//=======================================================================
155// Reader-writer lock implementation
156// Concurrent with doregister/unregister,
157//    i.e., threads can be added at any point during or between the entry/exit
158
159//-----------------------------------------------------------------------
160// simple spinlock underlying the RWLock
161// Blocking acquire
162static inline void __atomic_acquire(volatile bool * ll) {
163        while( __builtin_expect(__atomic_exchange_n(ll, (bool)true, __ATOMIC_SEQ_CST), false) ) {
164                while(__atomic_load_n(ll, (int)__ATOMIC_RELAXED))
165                        Pause();
166        }
167        /* paranoid */ verify(*ll);
168}
169
170// Non-Blocking acquire
171static inline bool __atomic_try_acquire(volatile bool * ll) {
172        return !__atomic_exchange_n(ll, (bool)true, __ATOMIC_SEQ_CST);
173}
174
175// Release
176static inline void __atomic_unlock(volatile bool * ll) {
177        /* paranoid */ verify(*ll);
178        __atomic_store_n(ll, (bool)false, __ATOMIC_RELEASE);
179}
180
181//-----------------------------------------------------------------------
182// Reader-Writer lock protecting the ready-queues
183// while this lock is mostly generic some aspects
184// have been hard-coded to for the ready-queue for
185// simplicity and performance
186struct __scheduler_RWLock_t {
187        // total cachelines allocated
188        unsigned int max;
189
190        // cachelines currently in use
191        volatile unsigned int alloc;
192
193        // cachelines ready to itereate over
194        // (!= to alloc when thread is in second half of doregister)
195        volatile unsigned int ready;
196
197        // writer lock
198        volatile bool write_lock;
199
200        // data pointer
201        volatile bool * volatile * data;
202};
203
204void  ?{}(__scheduler_RWLock_t & this);
205void ^?{}(__scheduler_RWLock_t & this);
206
207extern __scheduler_RWLock_t * __scheduler_lock;
208
209//-----------------------------------------------------------------------
210// Reader side : acquire when using the ready queue to schedule but not
211//  creating/destroying queues
212static inline void ready_schedule_lock(void) with(*__scheduler_lock) {
213        /* paranoid */ verify( ! __preemption_enabled() );
214        /* paranoid */ verify( ! kernelTLS().in_sched_lock );
215        /* paranoid */ verify( data[kernelTLS().sched_id] == &kernelTLS().sched_lock );
216        /* paranoid */ verify( !kernelTLS().this_processor || kernelTLS().this_processor->unique_id == kernelTLS().sched_id );
217
218        // Step 1 : make sure no writer are in the middle of the critical section
219        while(__atomic_load_n(&write_lock, (int)__ATOMIC_RELAXED))
220                Pause();
221
222        // Fence needed because we don't want to start trying to acquire the lock
223        // before we read a false.
224        // Not needed on x86
225        // std::atomic_thread_fence(std::memory_order_seq_cst);
226
227        // Step 2 : acquire our local lock
228        __atomic_acquire( &kernelTLS().sched_lock );
229        /*paranoid*/ verify(kernelTLS().sched_lock);
230
231        #ifdef __CFA_WITH_VERIFY__
232                // Debug, check if this is owned for reading
233                kernelTLS().in_sched_lock = true;
234        #endif
235}
236
237static inline void ready_schedule_unlock(void) with(*__scheduler_lock) {
238        /* paranoid */ verify( ! __preemption_enabled() );
239        /* paranoid */ verify( data[kernelTLS().sched_id] == &kernelTLS().sched_lock );
240        /* paranoid */ verify( !kernelTLS().this_processor || kernelTLS().this_processor->unique_id == kernelTLS().sched_id );
241        /* paranoid */ verify( kernelTLS().sched_lock );
242        /* paranoid */ verify( kernelTLS().in_sched_lock );
243        #ifdef __CFA_WITH_VERIFY__
244                // Debug, check if this is owned for reading
245                kernelTLS().in_sched_lock = false;
246        #endif
247        __atomic_unlock(&kernelTLS().sched_lock);
248}
249
250#ifdef __CFA_WITH_VERIFY__
251        static inline bool ready_schedule_islocked(void) {
252                /* paranoid */ verify( ! __preemption_enabled() );
253                /* paranoid */ verify( (!kernelTLS().in_sched_lock) || kernelTLS().sched_lock );
254                return kernelTLS().sched_lock;
255        }
256
257        static inline bool ready_mutate_islocked() {
258                return __scheduler_lock->write_lock;
259        }
260#endif
261
262//-----------------------------------------------------------------------
263// Writer side : acquire when changing the ready queue, e.g. adding more
264//  queues or removing them.
265uint_fast32_t ready_mutate_lock( void );
266
267void ready_mutate_unlock( uint_fast32_t /* value returned by lock */ );
268
269//-----------------------------------------------------------------------
270// Lock-Free registering/unregistering of threads
271// Register a processor to a given cluster and get its unique id in return
272// For convenience, also acquires the lock
273static inline [unsigned, uint_fast32_t] ready_mutate_register() {
274        unsigned id = register_proc_id();
275        uint_fast32_t last = ready_mutate_lock();
276        return [id, last];
277}
278
279// Unregister a processor from a given cluster using its id, getting back the original pointer
280// assumes the lock is acquired
281static inline void ready_mutate_unregister( unsigned id, uint_fast32_t last_s ) {
282        ready_mutate_unlock( last_s );
283        unregister_proc_id( id );
284}
285
286//-----------------------------------------------------------------------
287// Cluster idle lock/unlock
288static inline void lock(__cluster_proc_list & this) {
289        /* paranoid */ verify( ! __preemption_enabled() );
290
291        // Start by locking the global RWlock so that we know no-one is
292        // adding/removing processors while we mess with the idle lock
293        ready_schedule_lock();
294
295        lock( this.lock __cfaabi_dbg_ctx2 );
296
297        /* paranoid */ verify( ! __preemption_enabled() );
298}
299
300static inline bool try_lock(__cluster_proc_list & this) {
301        /* paranoid */ verify( ! __preemption_enabled() );
302
303        // Start by locking the global RWlock so that we know no-one is
304        // adding/removing processors while we mess with the idle lock
305        ready_schedule_lock();
306
307        if(try_lock( this.lock __cfaabi_dbg_ctx2 )) {
308                // success
309                /* paranoid */ verify( ! __preemption_enabled() );
310                return true;
311        }
312
313        // failed to lock
314        ready_schedule_unlock();
315
316        /* paranoid */ verify( ! __preemption_enabled() );
317        return false;
318}
319
320static inline void unlock(__cluster_proc_list & this) {
321        /* paranoid */ verify( ! __preemption_enabled() );
322
323        unlock(this.lock);
324
325        // Release the global lock, which we acquired when locking
326        ready_schedule_unlock();
327
328        /* paranoid */ verify( ! __preemption_enabled() );
329}
330
331//=======================================================================
332// Ready-Queue API
333//-----------------------------------------------------------------------
334// push thread onto a ready queue for a cluster
335// returns true if the list was previously empty, false otherwise
336__attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint);
337
338//-----------------------------------------------------------------------
339// pop thread from the local queues of a cluster
340// returns 0p if empty
341// May return 0p spuriously
342__attribute__((hot)) struct thread$ * pop_fast(struct cluster * cltr);
343
344//-----------------------------------------------------------------------
345// pop thread from any ready queue of a cluster
346// returns 0p if empty
347// May return 0p spuriously
348__attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr);
349
350//-----------------------------------------------------------------------
351// search all ready queues of a cluster for any thread
352// returns 0p if empty
353// guaranteed to find any threads added before this call
354__attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr);
355
356//-----------------------------------------------------------------------
357// get preferred ready for new thread
358unsigned ready_queue_new_preferred();
359
360//-----------------------------------------------------------------------
361// Increase the width of the ready queue (number of lanes) by 4
362void ready_queue_grow  (struct cluster * cltr);
363
364//-----------------------------------------------------------------------
365// Decrease the width of the ready queue (number of lanes) by 4
366void ready_queue_shrink(struct cluster * cltr);
367
368//-----------------------------------------------------------------------
369// Decrease the width of the ready queue (number of lanes) by 4
370void ready_queue_close(struct cluster * cltr);
371
372// Local Variables: //
373// mode: c //
374// tab-width: 4 //
375// End: //
Note: See TracBrowser for help on using the repository browser.