source: libcfa/src/concurrency/clib/cfathread.cfa@ 7425720

ADT ast-experimental enum pthread-emulation qualifiedEnum
Last change on this file since 7425720 was 75965a6, checked in by Peter A. Buhr <pabuhr@…>, 4 years ago

replace thread_rand with prng, replace xorshift64 with xorshift_13_7_17
  • Property mode set to 100644
File size: 16.3 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// clib/cfathread.cfa --
8//
9// Author : Thierry Delisle
10// Created On : Tue Sep 22 15:31:20 2020
11// Last Modified By :
12// Last Modified On :
13// Update Count :
14//
15
16// #define EPOLL_FOR_SOCKETS
17
18#include "fstream.hfa"
19#include "locks.hfa"
20#include "kernel.hfa"
21#include "stats.hfa"
22#include "thread.hfa"
23#include "time.hfa"
24#include "stdlib.hfa"
25
26#include "cfathread.h"
27
28extern "C" {
29 #include <string.h>
30 #include <errno.h>
31}
32
33extern void ?{}(processor &, const char[], cluster &, thread$ *);
34extern "C" {
35 extern void __cfactx_invoke_thread(void (*main)(void *), void * this);
36 extern int accept4(int sockfd, struct sockaddr *addr, socklen_t *addrlen, int flags);
37}
38
39extern Time __kernel_get_time();
40extern unsigned register_proc_id( void );
41
42//================================================================================
43// Epoll support for sockets
44
45#if defined(EPOLL_FOR_SOCKETS)
46 extern "C" {
47 #include <sys/epoll.h>
48 #include <sys/resource.h>
49 }
50
51 static pthread_t master_poller;
52 static int master_epollfd = 0;
53 static size_t poller_cnt = 0;
54 static int * poller_fds = 0p;
55 static struct leaf_poller * pollers = 0p;
56
57 struct __attribute__((aligned)) fd_info_t {
58 int pollid;
59 size_t rearms;
60 };
61 rlim_t fd_limit = 0;
62 static fd_info_t * volatile * fd_map = 0p;
63
64 void * master_epoll( __attribute__((unused)) void * args ) {
65 unsigned id = register_proc_id();
66
67 enum { MAX_EVENTS = 5 };
68 struct epoll_event events[MAX_EVENTS];
69 for() {
70 int ret = epoll_wait(master_epollfd, events, MAX_EVENTS, -1);
71 if ( ret < 0 ) {
72 abort | "Master epoll error: " | strerror(errno);
73 }
74
75 for(i; ret) {
76 thread$ * thrd = (thread$ *)events[i].data.u64;
77 unpark( thrd );
78 }
79 }
80
81 return 0p;
82 }
83
84 static inline int epoll_rearm(int epollfd, int fd, uint32_t event) {
85 struct epoll_event eevent;
86 eevent.events = event | EPOLLET | EPOLLONESHOT;
87 eevent.data.u64 = (uint64_t)active_thread();
88
89 if(0 != epoll_ctl(epollfd, EPOLL_CTL_MOD, fd, &eevent))
90 {
91 if(errno == ENOENT) return -1;
92 abort | acquire | "epoll" | epollfd | "ctl rearm" | fd | "error: " | errno | strerror(errno);
93 }
94
95 park();
96 return 0;
97 }
98
99 thread leaf_poller {
100 int epollfd;
101 };
102
103 void ?{}(leaf_poller & this, int fd) { this.epollfd = fd; }
104
105 void main(leaf_poller & this) {
106 enum { MAX_EVENTS = 1024 };
107 struct epoll_event events[MAX_EVENTS];
108 const int max_retries = 5;
109 int retries = max_retries;
110
111 struct epoll_event event;
112 event.events = EPOLLIN | EPOLLET | EPOLLONESHOT;
113 event.data.u64 = (uint64_t)&(thread&)this;
114
115 if(0 != epoll_ctl(master_epollfd, EPOLL_CTL_ADD, this.epollfd, &event))
116 {
117 abort | "master epoll ctl add leaf: " | errno | strerror(errno);
118 }
119
120 park();
121
122 for() {
123 yield();
124 int ret = epoll_wait(this.epollfd, events, MAX_EVENTS, 0);
125 if ( ret < 0 ) {
126 abort | "Leaf epoll error: " | errno | strerror(errno);
127 }
128
129 if(ret) {
130 for(i; ret) {
131 thread$ * thrd = (thread$ *)events[i].data.u64;
132 unpark( thrd, UNPARK_REMOTE );
133 }
134 }
135 else if(0 >= --retries) {
136 epoll_rearm(master_epollfd, this.epollfd, EPOLLIN);
137 }
138 }
139 }
140
141 void setup_epoll( void ) __attribute__(( constructor ));
142 void setup_epoll( void ) {
143 if(master_epollfd) abort | "Master epoll already setup";
144
145 master_epollfd = epoll_create1(0);
146 if(master_epollfd == -1) {
147 abort | "failed to create master epoll: " | errno | strerror(errno);
148 }
149
150 struct rlimit rlim;
151 if(int ret = getrlimit(RLIMIT_NOFILE, &rlim); 0 != ret) {
152 abort | "failed to get nofile limit: " | errno | strerror(errno);
153 }
154
155 fd_limit = rlim.rlim_cur;
156 fd_map = alloc(fd_limit);
157 for(i;fd_limit) {
158 fd_map[i] = 0p;
159 }
160
161 poller_cnt = 2;
162 poller_fds = alloc(poller_cnt);
163 pollers = alloc(poller_cnt);
164 for(i; poller_cnt) {
165 poller_fds[i] = epoll_create1(0);
166 if(poller_fds[i] == -1) {
167 abort | "failed to create leaf epoll [" | i | "]: " | errno | strerror(errno);
168 }
169
170 (pollers[i]){ poller_fds[i] };
171 }
172
173 pthread_attr_t attr;
174 if (int ret = pthread_attr_init(&attr); 0 != ret) {
175 abort | "failed to create master epoll thread attr: " | ret | strerror(ret);
176 }
177
178 if (int ret = pthread_create(&master_poller, &attr, master_epoll, 0p); 0 != ret) {
179 abort | "failed to create master epoll thread: " | ret | strerror(ret);
180 }
181 }
182
183 static inline int epoll_wait(int fd, uint32_t event) {
184 if(fd_map[fd] >= 1p) {
185 fd_map[fd]->rearms++;
186 epoll_rearm(poller_fds[fd_map[fd]->pollid], fd, event);
187 return 0;
188 }
189
190 for() {
191 fd_info_t * expected = 0p;
192 fd_info_t * sentinel = 1p;
193 if(__atomic_compare_exchange_n( &(fd_map[fd]), &expected, sentinel, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED)) {
194 struct epoll_event eevent;
195 eevent.events = event | EPOLLET | EPOLLONESHOT;
196 eevent.data.u64 = (uint64_t)active_thread();
197
198 int id = prng() % poller_cnt;
199 if(0 != epoll_ctl(poller_fds[id], EPOLL_CTL_ADD, fd, &eevent))
200 {
201 abort | "epoll ctl add" | poller_fds[id] | fd | fd_map[fd] | expected | "error: " | errno | strerror(errno);
202 }
203
204 fd_info_t * ninfo = alloc();
205 ninfo->pollid = id;
206 ninfo->rearms = 0;
207 __atomic_store_n( &fd_map[fd], ninfo, __ATOMIC_SEQ_CST);
208
209 park();
210 return 0;
211 }
212
213 if(expected >= 0) {
214 fd_map[fd]->rearms++;
215 epoll_rearm(poller_fds[fd_map[fd]->pollid], fd, event);
216 return 0;
217 }
218
219 Pause();
220 }
221 }
222#endif
223
224//================================================================================
225// Thread run by the C Interface
226
227struct cfathread_object {
228 thread$ self;
229 void * (*themain)( void * );
230 void * arg;
231 void * ret;
232};
233void main(cfathread_object & this);
234void ^?{}(cfathread_object & mutex this);
235
236static inline thread$ * get_thread( cfathread_object & this ) { return &this.self; }
237
238typedef ThreadCancelled(cfathread_object) cfathread_exception;
239typedef ThreadCancelled_vtable(cfathread_object) cfathread_vtable;
240
241void defaultResumptionHandler(ThreadCancelled(cfathread_object) & except) {
242 abort | "A thread was cancelled";
243}
244
245cfathread_vtable _cfathread_vtable_instance;
246
247cfathread_vtable & const _default_vtable = _cfathread_vtable_instance;
248
249cfathread_vtable const & get_exception_vtable(cfathread_exception *) {
250 return _cfathread_vtable_instance;
251}
252
253static void ?{}( cfathread_object & this, cluster & cl, void *(*themain)( void * ), void * arg ) {
254 this.themain = themain;
255 this.arg = arg;
256 (this.self){"C-thread", cl};
257 __thrd_start(this, main);
258}
259
260void ^?{}(cfathread_object & mutex this) {
261 ^(this.self){};
262}
263
264void main( cfathread_object & this ) {
265 __attribute__((unused)) void * const thrd_obj = (void*)&this;
266 __attribute__((unused)) void * const thrd_hdl = (void*)active_thread();
267 /* paranoid */ verify( thrd_obj == thrd_hdl );
268
269 this.ret = this.themain( this.arg );
270}
271
272//================================================================================
273// Special Init Thread responsible for the initialization or processors
274struct __cfainit {
275 thread$ self;
276 void (*init)( void * );
277 void * arg;
278};
279void main(__cfainit & this);
280void ^?{}(__cfainit & mutex this);
281
282static inline thread$ * get_thread( __cfainit & this ) { return &this.self; }
283
284typedef ThreadCancelled(__cfainit) __cfainit_exception;
285typedef ThreadCancelled_vtable(__cfainit) __cfainit_vtable;
286
287void defaultResumptionHandler(ThreadCancelled(__cfainit) & except) {
288 abort | "The init thread was cancelled";
289}
290
291__cfainit_vtable ___cfainit_vtable_instance;
292
293__cfainit_vtable const & get_exception_vtable(__cfainit_exception *) {
294 return ___cfainit_vtable_instance;
295}
296
297static void ?{}( __cfainit & this, void (*init)( void * ), void * arg ) {
298 this.init = init;
299 this.arg = arg;
300 (this.self){"Processir Init"};
301
302 // Don't use __thrd_start! just prep the context manually
303 thread$ * this_thrd = get_thread(this);
304 void (*main_p)(__cfainit &) = main;
305
306 disable_interrupts();
307 __cfactx_start(main_p, get_coroutine(this), this, __cfactx_invoke_thread);
308
309 this_thrd->context.[SP, FP] = this_thrd->self_cor.context.[SP, FP];
310 /* paranoid */ verify( this_thrd->context.SP );
311
312 this_thrd->state = Ready;
313 enable_interrupts();
314}
315
316void ^?{}(__cfainit & mutex this) {
317 ^(this.self){};
318}
319
320void main( __cfainit & this ) {
321 __attribute__((unused)) void * const thrd_obj = (void*)&this;
322 __attribute__((unused)) void * const thrd_hdl = (void*)active_thread();
323 /* paranoid */ verify( thrd_obj == thrd_hdl );
324
325 this.init( this.arg );
326}
327
328//================================================================================
329// Main Api
330extern "C" {
331 int cfathread_cluster_create(cfathread_cluster_t * cl) __attribute__((nonnull(1))) {
332 *cl = new();
333 return 0;
334 }
335
336 cfathread_cluster_t cfathread_cluster_self(void) {
337 return active_cluster();
338 }
339
340 int cfathread_cluster_print_stats( cfathread_cluster_t cl ) {
341 #if !defined(__CFA_NO_STATISTICS__)
342 print_stats_at_exit( *cl, CFA_STATS_READY_Q | CFA_STATS_IO );
343 print_stats_now( *cl, CFA_STATS_READY_Q | CFA_STATS_IO );
344 #endif
345 return 0;
346 }
347
348 int cfathread_cluster_add_worker(cfathread_cluster_t cl, pthread_t* tid, void (*init_routine) (void *), void * arg) {
349 __cfainit * it = 0p;
350 if(init_routine) {
351 it = alloc();
352 (*it){init_routine, arg};
353 }
354 processor * proc = alloc();
355 (*proc){ "C-processor", *cl, get_thread(*it) };
356
357 // Wait for the init thread to return before continuing
358 if(it) {
359 ^(*it){};
360 free(it);
361 }
362
363 if(tid) *tid = proc->kernel_thread;
364 return 0;
365 }
366
367 int cfathread_cluster_pause (cfathread_cluster_t) {
368 abort | "Pausing clusters is not supported";
369 exit(1);
370 }
371
372 int cfathread_cluster_resume(cfathread_cluster_t) {
373 abort | "Resuming clusters is not supported";
374 exit(1);
375 }
376
377 //--------------------
378 // Thread attributes
379 int cfathread_attr_init(cfathread_attr_t *attr) __attribute__((nonnull (1))) {
380 attr->cl = active_cluster();
381 return 0;
382 }
383
384 //--------------------
385 // Thread
386 int cfathread_create( cfathread_t * handle, const cfathread_attr_t * attr, void *(*main)( void * ), void * arg ) __attribute__((nonnull (1))) {
387 cluster * cl = attr ? attr->cl : active_cluster();
388 cfathread_t thrd = alloc();
389 (*thrd){ *cl, main, arg };
390 *handle = thrd;
391 return 0;
392 }
393
394 int cfathread_join( cfathread_t thrd, void ** retval ) {
395 void * ret = join( *thrd ).ret;
396 ^( *thrd ){};
397 free(thrd);
398 if(retval) {
399 *retval = ret;
400 }
401 return 0;
402 }
403
404 int cfathread_get_errno(void) {
405 return errno;
406 }
407
408 cfathread_t cfathread_self(void) {
409 return (cfathread_t)active_thread();
410 }
411
412 int cfathread_usleep(useconds_t usecs) {
413 sleep(usecs`us);
414 return 0;
415 }
416
417 int cfathread_sleep(unsigned int secs) {
418 sleep(secs`s);
419 return 0;
420 }
421
422 void cfathread_park( void ) {
423 park();
424 }
425
426 void cfathread_unpark( cfathread_t thrd ) {
427 unpark( *thrd );
428 }
429
430 void cfathread_yield( void ) {
431 yield();
432 }
433
434 typedef struct cfathread_mutex * cfathread_mutex_t;
435
436 //--------------------
437 // Mutex
438 struct cfathread_mutex {
439 linear_backoff_then_block_lock impl;
440 };
441 int cfathread_mutex_init(cfathread_mutex_t *restrict mut, const cfathread_mutexattr_t *restrict) __attribute__((nonnull (1))) { *mut = new(); return 0; }
442 int cfathread_mutex_destroy(cfathread_mutex_t *mut) __attribute__((nonnull (1))) { delete( *mut ); return 0; }
443 int cfathread_mutex_lock (cfathread_mutex_t *mut) __attribute__((nonnull (1))) { lock( (*mut)->impl ); return 0; }
444 int cfathread_mutex_unlock (cfathread_mutex_t *mut) __attribute__((nonnull (1))) { unlock( (*mut)->impl ); return 0; }
445 int cfathread_mutex_trylock(cfathread_mutex_t *mut) __attribute__((nonnull (1))) {
446 bool ret = try_lock( (*mut)->impl );
447 if( ret ) return 0;
448 else return EBUSY;
449 }
450
451 //--------------------
452 // Condition
453 struct cfathread_condition {
454 condition_variable(linear_backoff_then_block_lock) impl;
455 };
456 int cfathread_cond_init(cfathread_cond_t *restrict cond, const cfathread_condattr_t *restrict) __attribute__((nonnull (1))) { *cond = new(); return 0; }
457 int cfathread_cond_signal(cfathread_cond_t *cond) __attribute__((nonnull (1))) { notify_one( (*cond)->impl ); return 0; }
458 int cfathread_cond_wait(cfathread_cond_t *restrict cond, cfathread_mutex_t *restrict mut) __attribute__((nonnull (1,2))) { wait( (*cond)->impl, (*mut)->impl ); return 0; }
459 int cfathread_cond_timedwait(cfathread_cond_t *restrict cond, cfathread_mutex_t *restrict mut, const struct timespec *restrict abstime) __attribute__((nonnull (1,2,3))) {
460 Time t = { *abstime };
461 timespec curr;
462 clock_gettime( CLOCK_REALTIME, &curr );
463 Time c = { curr };
464 if( wait( (*cond)->impl, (*mut)->impl, t - c ) ) {
465 return 0;
466 }
467 errno = ETIMEDOUT;
468 return ETIMEDOUT;
469 }
470}
471
472#include <iofwd.hfa>
473
474extern "C" {
475 #include <unistd.h>
476 #include <sys/types.h>
477 #include <sys/socket.h>
478
479 //--------------------
480 // IO operations
481 int cfathread_socket(int domain, int type, int protocol) {
482 return socket(domain, type
483 #if defined(EPOLL_FOR_SOCKETS)
484 | SOCK_NONBLOCK
485 #endif
486 , protocol);
487 }
488 int cfathread_bind(int socket, const struct sockaddr *address, socklen_t address_len) {
489 return bind(socket, address, address_len);
490 }
491
492 int cfathread_listen(int socket, int backlog) {
493 return listen(socket, backlog);
494 }
495
496 int cfathread_accept(int socket, struct sockaddr *restrict address, socklen_t *restrict address_len) {
497 #if defined(EPOLL_FOR_SOCKETS)
498 int ret;
499 for() {
500 yield();
501 ret = accept4(socket, address, address_len, SOCK_NONBLOCK);
502 if(ret >= 0) break;
503 if(errno != EAGAIN && errno != EWOULDBLOCK) break;
504
505 epoll_wait(socket, EPOLLIN);
506 }
507 return ret;
508 #else
509 return cfa_accept4(socket, address, address_len, 0, CFA_IO_LAZY);
510 #endif
511 }
512
513 int cfathread_connect(int socket, const struct sockaddr *address, socklen_t address_len) {
514 #if defined(EPOLL_FOR_SOCKETS)
515 int ret;
516 for() {
517 ret = connect(socket, address, address_len);
518 if(ret >= 0) break;
519 if(errno != EAGAIN && errno != EWOULDBLOCK) break;
520
521 epoll_wait(socket, EPOLLIN);
522 }
523 return ret;
524 #else
525 return cfa_connect(socket, address, address_len, CFA_IO_LAZY);
526 #endif
527 }
528
529 int cfathread_dup(int fildes) {
530 return dup(fildes);
531 }
532
533 int cfathread_close(int fildes) {
534 return cfa_close(fildes, CFA_IO_LAZY);
535 }
536
537 ssize_t cfathread_sendmsg(int socket, const struct msghdr *message, int flags) {
538 #if defined(EPOLL_FOR_SOCKETS)
539 ssize_t ret;
540 __STATS__( false, io.ops.sockwrite++; )
541 for() {
542 ret = sendmsg(socket, message, flags);
543 if(ret >= 0) break;
544 if(errno != EAGAIN && errno != EWOULDBLOCK) break;
545
546 __STATS__( false, io.ops.epllwrite++; )
547 epoll_wait(socket, EPOLLOUT);
548 }
549 #else
550 ssize_t ret = cfa_sendmsg(socket, message, flags, CFA_IO_LAZY);
551 #endif
552 return ret;
553 }
554
555 ssize_t cfathread_write(int fildes, const void *buf, size_t nbyte) {
556 // Use send rather then write for socket since it's faster
557 #if defined(EPOLL_FOR_SOCKETS)
558 ssize_t ret;
559 // __STATS__( false, io.ops.sockwrite++; )
560 for() {
561 ret = send(fildes, buf, nbyte, 0);
562 if(ret >= 0) break;
563 if(errno != EAGAIN && errno != EWOULDBLOCK) break;
564
565 // __STATS__( false, io.ops.epllwrite++; )
566 epoll_wait(fildes, EPOLLOUT);
567 }
568 #else
569 ssize_t ret = cfa_send(fildes, buf, nbyte, 0, CFA_IO_LAZY);
570 #endif
571 return ret;
572 }
573
574 ssize_t cfathread_recvfrom(int socket, void *restrict buffer, size_t length, int flags, struct sockaddr *restrict address, socklen_t *restrict address_len) {
575 struct iovec iov;
576 iov.iov_base = buffer;
577 iov.iov_len = length;
578
579 struct msghdr msg;
580 msg.msg_name = address;
581 msg.msg_namelen = address_len ? (socklen_t)*address_len : (socklen_t)0;
582 msg.msg_iov = &iov;
583 msg.msg_iovlen = 1;
584 msg.msg_control = 0p;
585 msg.msg_controllen = 0;
586
587 #if defined(EPOLL_FOR_SOCKETS)
588 ssize_t ret;
589 yield();
590 for() {
591 ret = recvmsg(socket, &msg, flags);
592 if(ret >= 0) break;
593 if(errno != EAGAIN && errno != EWOULDBLOCK) break;
594
595 epoll_wait(socket, EPOLLIN);
596 }
597 #else
598 ssize_t ret = cfa_recvmsg(socket, &msg, flags, CFA_IO_LAZY);
599 #endif
600
601 if(address_len) *address_len = msg.msg_namelen;
602 return ret;
603 }
604
605 ssize_t cfathread_read(int fildes, void *buf, size_t nbyte) {
606 // Use recv rather then read for socket since it's faster
607 #if defined(EPOLL_FOR_SOCKETS)
608 ssize_t ret;
609 __STATS__( false, io.ops.sockread++; )
610 yield();
611 for() {
612 ret = recv(fildes, buf, nbyte, 0);
613 if(ret >= 0) break;
614 if(errno != EAGAIN && errno != EWOULDBLOCK) break;
615
616 __STATS__( false, io.ops.epllread++; )
617 epoll_wait(fildes, EPOLLIN);
618 }
619 #else
620 ssize_t ret = cfa_recv(fildes, buf, nbyte, 0, CFA_IO_LAZY);
621 #endif
622 return ret;
623 }
624
625}
Note: See TracBrowser for help on using the repository browser.