Context Navigation

setup.cfa @ 78a8440

ADTarm-ehast-experimentalenumforall-pointer-decayjacob/cs343-translationnew-ast-unique-exprpthread-emulationqualifiedEnum

Last change on this file since 78a8440 was d611995, checked in by Thierry Delisle <tdelisle@…>, 4 years ago
Fix crash on cluster exit. (Added epoch based blocking to iopoll.)
Property mode set to `100644`
File size: 18.7 KB

Line
1	//
2	// Cforall Version 1.0.0 Copyright (C) 2020 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	// io/setup.cfa --
8	//
9	// Author : Thierry Delisle
10	// Created On : Fri Jul 31 16:25:51 2020
11	// Last Modified By :
12	// Last Modified On :
13	// Update Count :
14	//
15
16	#define __cforall_thread__
17	#define _GNU_SOURCE /* See feature_test_macros(7) */
18
19	#if defined(__CFA_DEBUG__)
20	// #define __CFA_DEBUG_PRINT_IO__
21	// #define __CFA_DEBUG_PRINT_IO_CORE__
22	#endif
23
24	#include "io/types.hfa"
25	#include "kernel.hfa"
26
27	#if !defined(CFA_HAVE_LINUX_IO_URING_H)
28	void __kernel_io_startup() {
29	// Nothing to do without io_uring
30	}
31
32	void __kernel_io_shutdown() {
33	// Nothing to do without io_uring
34	}
35
36	void ?{}(io_context_params & this) {}
37
38	void ?{}(io_context & this, struct cluster & cl) {}
39	void ?{}(io_context & this, struct cluster & cl, const io_context_params & params) {}
40
41	void ^?{}(io_context & this) {}
42	void ^?{}(io_context & this, bool cluster_context) {}
43
44	void register_fixed_files( io_context &, int *, unsigned ) {}
45	void register_fixed_files( cluster &, int *, unsigned ) {}
46
47	#else
48	#include <errno.h>
49	#include <stdint.h>
50	#include <string.h>
51	#include <signal.h>
52	#include <unistd.h>
53
54	extern "C" {
55	#include <pthread.h>
56	#include <sys/epoll.h>
57	#include <sys/eventfd.h>
58	#include <sys/mman.h>
59	#include <sys/syscall.h>
60
61	#include <linux/io_uring.h>
62	}
63
64	#include "bitmanip.hfa"
65	#include "kernel_private.hfa"
66	#include "thread.hfa"
67
68	void ?{}(io_context_params & this) {
69	this.num_entries = 256;
70	this.num_ready = 256;
71	this.submit_aff = -1;
72	this.eager_submits = false;
73	this.poller_submits = false;
74	this.poll_submit = false;
75	this.poll_complete = false;
76	}
77
78	static void * __io_poller_slow( void * arg );
79
80	// Weirdly, some systems that do support io_uring don't actually define these
81	#ifdef __alpha__
82	/*
83	* alpha is the only exception, all other architectures
84	* have common numbers for new system calls.
85	*/
86	#ifndef __NR_io_uring_setup
87	#define __NR_io_uring_setup 535
88	#endif
89	#ifndef __NR_io_uring_enter
90	#define __NR_io_uring_enter 536
91	#endif
92	#ifndef __NR_io_uring_register
93	#define __NR_io_uring_register 537
94	#endif
95	#else /* !__alpha__ */
96	#ifndef __NR_io_uring_setup
97	#define __NR_io_uring_setup 425
98	#endif
99	#ifndef __NR_io_uring_enter
100	#define __NR_io_uring_enter 426
101	#endif
102	#ifndef __NR_io_uring_register
103	#define __NR_io_uring_register 427
104	#endif
105	#endif
106
107	//=============================================================================================
108	// I/O Startup / Shutdown logic + Master Poller
109	//=============================================================================================
110
111	// IO Master poller loop forward
112	static void * iopoll_loop( __attribute__((unused)) void * args );
113
114	static struct {
115	pthread_t thrd; // pthread handle to io poller thread
116	void * stack; // pthread stack for io poller thread
117	int epollfd; // file descriptor to the epoll instance
118	volatile bool run; // Whether or not to continue
119	volatile size_t epoch; // Epoch used for memory reclamation
120	} iopoll;
121
122	void __kernel_io_startup(void) {
123	__cfadbg_print_safe(io_core, "Kernel : Creating EPOLL instance\n" );
124
125	iopoll.epollfd = epoll_create1(0);
126	if (iopoll.epollfd == -1) {
127	abort( "internal error, epoll_create1\n");
128	}
129
130	__cfadbg_print_safe(io_core, "Kernel : Starting io poller thread\n" );
131
132	iopoll.run = true;
133	iopoll.stack = __create_pthread( &iopoll.thrd, iopoll_loop, 0p );
134	iopoll.epoch = 0;
135	}
136
137	void __kernel_io_shutdown(void) {
138	// Notify the io poller thread of the shutdown
139	iopoll.run = false;
140	sigval val = { 1 };
141	pthread_sigqueue( iopoll.thrd, SIGUSR1, val );
142
143	// Wait for the io poller thread to finish
144
145	__destroy_pthread( iopoll.thrd, iopoll.stack, 0p );
146
147	int ret = close(iopoll.epollfd);
148	if (ret == -1) {
149	abort( "internal error, close epoll\n");
150	}
151
152	// Io polling is now fully stopped
153
154	__cfadbg_print_safe(io_core, "Kernel : IO poller stopped\n" );
155	}
156
157	static void * iopoll_loop( __attribute__((unused)) void * args ) {
158	__processor_id_t id;
159	id.full_proc = false;
160	id.id = doregister(&id);
161	__cfaabi_tls.this_proc_id = &id;
162	__cfadbg_print_safe(io_core, "Kernel : IO poller thread starting\n" );
163
164	// Block signals to control when they arrive
165	sigset_t mask;
166	sigfillset(&mask);
167	if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
168	abort( "internal error, pthread_sigmask" );
169	}
170
171	sigdelset( &mask, SIGUSR1 );
172
173	// Create sufficient events
174	struct epoll_event events[10];
175	// Main loop
176	while( iopoll.run ) {
177	__cfadbg_print_safe(io_core, "Kernel I/O - epoll : waiting on io_uring contexts\n");
178
179	// increment the epoch to notify any deleters we are starting a new cycle
180	__atomic_fetch_add(&iopoll.epoch, 1, __ATOMIC_SEQ_CST);
181
182	// Wait for events
183	int nfds = epoll_pwait( iopoll.epollfd, events, 10, -1, &mask );
184
185	__cfadbg_print_safe(io_core, "Kernel I/O - epoll : %d io contexts events, waking up\n", nfds);
186
187	// Check if an error occured
188	if (nfds == -1) {
189	if( errno == EINTR ) continue;
190	abort( "internal error, pthread_sigmask" );
191	}
192
193	for(i; nfds) {
194	$io_ctx_thread * io_ctx = ($io_ctx_thread *)(uintptr_t)events[i].data.u64;
195	/* paranoid */ verify( io_ctx );
196	__cfadbg_print_safe(io_core, "Kernel I/O - epoll : Unparking io poller %d (%p)\n", io_ctx->ring->fd, io_ctx);
197	#if !defined( __CFA_NO_STATISTICS__ )
198	__cfaabi_tls.this_stats = io_ctx->self.curr_cluster->stats;
199	#endif
200
201	eventfd_t v;
202	eventfd_read(io_ctx->ring->efd, &v);
203
204	post( io_ctx->sem );
205	}
206	}
207
208	__cfadbg_print_safe(io_core, "Kernel : IO poller thread stopping\n" );
209	unregister(&id);
210	return 0p;
211	}
212
213	//=============================================================================================
214	// I/O Context Constrution/Destruction
215	//=============================================================================================
216
217	void ?{}($io_ctx_thread & this, struct cluster & cl) { (this.self){ "IO Poller", cl }; }
218	void main( $io_ctx_thread & this );
219	static inline $thread * get_thread( $io_ctx_thread & this ) { return &this.self; }
220	void ^?{}( $io_ctx_thread & mutex this ) {}
221
222	static void __io_create ( __io_data & this, const io_context_params & params_in );
223	static void __io_destroy( __io_data & this );
224
225	void ?{}(io_context & this, struct cluster & cl, const io_context_params & params) {
226	(this.thrd){ cl };
227	this.thrd.ring = malloc();
228	__cfadbg_print_safe(io_core, "Kernel I/O : Creating ring for io_context %p\n", &this);
229	__io_create( *this.thrd.ring, params );
230
231	__cfadbg_print_safe(io_core, "Kernel I/O : Starting poller thread for io_context %p\n", &this);
232	this.thrd.done = false;
233	__thrd_start( this.thrd, main );
234
235	__cfadbg_print_safe(io_core, "Kernel I/O : io_context %p ready\n", &this);
236	}
237
238	void ?{}(io_context & this, struct cluster & cl) {
239	io_context_params params;
240	(this){ cl, params };
241	}
242
243	void ^?{}(io_context & this, bool cluster_context) {
244	__cfadbg_print_safe(io_core, "Kernel I/O : tearing down io_context %p\n", &this);
245
246	// Notify the thread of the shutdown
247	__atomic_store_n(&this.thrd.done, true, __ATOMIC_SEQ_CST);
248
249	// If this is an io_context within a cluster, things get trickier
250	$thread & thrd = this.thrd.self;
251	if( cluster_context ) {
252	// We are about to do weird things with the threads
253	// we don't need interrupts to complicate everything
254	disable_interrupts();
255
256	// Get cluster info
257	cluster & cltr = *thrd.curr_cluster;
258	/* paranoid */ verify( cltr.idles.total == 0 \|\| &cltr == mainCluster );
259	/* paranoid */ verify( !ready_mutate_islocked() );
260
261	// We need to adjust the clean-up based on where the thread is
262	if( thrd.state == Ready \|\| thrd.preempted != __NO_PREEMPTION ) {
263	// This is the tricky case
264	// The thread was preempted or ready to run and now it is on the ready queue
265	// but the cluster is shutting down, so there aren't any processors to run the ready queue
266	// the solution is to steal the thread from the ready-queue and pretend it was blocked all along
267
268	ready_schedule_lock();
269	// The thread should on the list
270	/* paranoid */ verify( thrd.link.next != 0p );
271
272	// Remove the thread from the ready queue of this cluster
273	// The thread should be the last on the list
274	__attribute__((unused)) bool removed = remove_head( &cltr, &thrd );
275	/* paranoid */ verify( removed );
276	thrd.link.next = 0p;
277	thrd.link.prev = 0p;
278
279	// Fixup the thread state
280	thrd.state = Blocked;
281	thrd.ticket = TICKET_BLOCKED;
282	thrd.preempted = __NO_PREEMPTION;
283
284	ready_schedule_unlock();
285
286	// Pretend like the thread was blocked all along
287	}
288	// !!! This is not an else if !!!
289	// Ok, now the thread is blocked (whether we cheated to get here or not)
290	if( thrd.state == Blocked ) {
291	// This is the "easy case"
292	// The thread is parked and can easily be moved to active cluster
293	verify( thrd.curr_cluster != active_cluster() \|\| thrd.curr_cluster == mainCluster );
294	thrd.curr_cluster = active_cluster();
295
296	// unpark the fast io_poller
297	unpark( &thrd );
298	}
299	else {
300	// The thread is in a weird state
301	// I don't know what to do here
302	abort("io_context poller thread is in unexpected state, cannot clean-up correctly\n");
303	}
304
305	// The weird thread kidnapping stuff is over, restore interrupts.
306	enable_interrupts( __cfaabi_dbg_ctx );
307	} else {
308	post( this.thrd.sem );
309	}
310
311	^(this.thrd){};
312	__cfadbg_print_safe(io_core, "Kernel I/O : Stopped poller thread for io_context %p\n", &this);
313
314	__io_destroy( *this.thrd.ring );
315	__cfadbg_print_safe(io_core, "Kernel I/O : Destroyed ring for io_context %p\n", &this);
316
317	free(this.thrd.ring);
318	}
319
320	void ^?{}(io_context & this) {
321	^(this){ false };
322	}
323
324	extern void __disable_interrupts_hard();
325	extern void __enable_interrupts_hard();
326
327	static void __io_create( __io_data & this, const io_context_params & params_in ) {
328	// Step 1 : call to setup
329	struct io_uring_params params;
330	memset(&params, 0, sizeof(params));
331	if( params_in.poll_submit ) params.flags \|= IORING_SETUP_SQPOLL;
332	if( params_in.poll_complete ) params.flags \|= IORING_SETUP_IOPOLL;
333
334	__u32 nentries = params_in.num_entries != 0 ? params_in.num_entries : 256;
335	if( !is_pow2(nentries) ) {
336	abort("ERROR: I/O setup 'num_entries' must be a power of 2\n");
337	}
338	if( params_in.poller_submits && params_in.eager_submits ) {
339	abort("ERROR: I/O setup 'poller_submits' and 'eager_submits' cannot be used together\n");
340	}
341
342	int fd = syscall(__NR_io_uring_setup, nentries, &params );
343	if(fd < 0) {
344	abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
345	}
346
347	// Step 2 : mmap result
348	memset( &this, 0, sizeof(struct __io_data) );
349	struct __submition_data & sq = this.submit_q;
350	struct __completion_data & cq = this.completion_q;
351
352	// calculate the right ring size
353	sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
354	cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
355
356	// Requires features
357	#if defined(IORING_FEAT_SINGLE_MMAP)
358	// adjust the size according to the parameters
359	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
360	cq.ring_sz = sq.ring_sz = max(cq.ring_sz, sq.ring_sz);
361	}
362	#endif
363
364	// mmap the Submit Queue into existence
365	sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQ_RING);
366	if (sq.ring_ptr == (void*)MAP_FAILED) {
367	abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
368	}
369
370	// Requires features
371	#if defined(IORING_FEAT_SINGLE_MMAP)
372	// mmap the Completion Queue into existence (may or may not be needed)
373	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
374	cq.ring_ptr = sq.ring_ptr;
375	}
376	else
377	#endif
378	{
379	// We need multiple call to MMAP
380	cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_CQ_RING);
381	if (cq.ring_ptr == (void*)MAP_FAILED) {
382	munmap(sq.ring_ptr, sq.ring_sz);
383	abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
384	}
385	}
386
387	// mmap the submit queue entries
388	size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
389	sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQES);
390	if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
391	munmap(sq.ring_ptr, sq.ring_sz);
392	if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
393	abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
394	}
395
396	// Step 3 : Initialize the data structure
397	// Get the pointers from the kernel to fill the structure
398	// submit queue
399	sq.head = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
400	sq.tail = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
401	sq.mask = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
402	sq.num = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
403	sq.flags = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
404	sq.dropped = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
405	sq.array = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
406	sq.prev_head = *sq.head;
407
408	{
409	const __u32 num = *sq.num;
410	for( i; num ) {
411	__sqe_clean( &sq.sqes[i] );
412	}
413	}
414
415	(sq.submit_lock){};
416	(sq.release_lock){};
417
418	if( params_in.poller_submits \|\| params_in.eager_submits ) {
419	/* paranoid */ verify( is_pow2( params_in.num_ready ) \|\| (params_in.num_ready < 8) );
420	sq.ready_cnt = max( params_in.num_ready, 8 );
421	sq.ready = alloc( sq.ready_cnt, 64`align );
422	for(i; sq.ready_cnt) {
423	sq.ready[i] = -1ul32;
424	}
425	sq.prev_ready = 0;
426	}
427	else {
428	sq.ready_cnt = 0;
429	sq.ready = 0p;
430	sq.prev_ready = 0;
431	}
432
433	// completion queue
434	cq.head = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
435	cq.tail = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
436	cq.mask = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
437	cq.num = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
438	cq.overflow = ( __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
439	cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
440
441	// Step 4 : eventfd
442	// io_uring_register is so f*cking slow on some machine that it
443	// will never succeed if preemption isn't hard blocked
444	__disable_interrupts_hard();
445
446	int efd = eventfd(0, 0);
447	if (efd < 0) {
448	abort("KERNEL ERROR: IO_URING EVENTFD - %s\n", strerror(errno));
449	}
450
451	int ret = syscall( __NR_io_uring_register, fd, IORING_REGISTER_EVENTFD, &efd, 1);
452	if (ret < 0) {
453	abort("KERNEL ERROR: IO_URING EVENTFD REGISTER - %s\n", strerror(errno));
454	}
455
456	__enable_interrupts_hard();
457
458	// some paranoid checks
459	/* paranoid / verifyf( (cq.mask) == ((cq.num) - 1ul32), "IO_URING Expected mask to be %u (%u entries), was %u", (cq.num) - 1ul32, cq.num, cq.mask );
460	/* paranoid / verifyf( (cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
461	/* paranoid / verifyf( (cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
462	/* paranoid / verifyf( (cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
463
464	/* paranoid / verifyf( (sq.mask) == ((sq.num) - 1ul32), "IO_URING Expected mask to be %u (%u entries), was %u", (sq.num) - 1ul32, sq.num, sq.mask );
465	/* paranoid / verifyf( (sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
466	/* paranoid / verifyf( (sq.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.head );
467	/* paranoid / verifyf( (sq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.tail );
468
469	// Update the global ring info
470	this.ring_flags = params.flags;
471	this.fd = fd;
472	this.efd = efd;
473	this.eager_submits = params_in.eager_submits;
474	this.poller_submits = params_in.poller_submits;
475	}
476
477	static void __io_destroy( __io_data & this ) {
478	// Shutdown the io rings
479	struct __submition_data & sq = this.submit_q;
480	struct __completion_data & cq = this.completion_q;
481
482	// unmap the submit queue entries
483	munmap(sq.sqes, (sq.num) sizeof(struct io_uring_sqe));
484
485	// unmap the Submit Queue ring
486	munmap(sq.ring_ptr, sq.ring_sz);
487
488	// unmap the Completion Queue ring, if it is different
489	if (cq.ring_ptr != sq.ring_ptr) {
490	munmap(cq.ring_ptr, cq.ring_sz);
491	}
492
493	// close the file descriptor
494	close(this.fd);
495	close(this.efd);
496
497	free( this.submit_q.ready ); // Maybe null, doesn't matter
498	}
499
500	//=============================================================================================
501	// I/O Context Sleep
502	//=============================================================================================
503	static inline void __ioctx_epoll_ctl($io_ctx_thread & ctx, int op, const char * error) {
504	struct epoll_event ev;
505	ev.events = EPOLLIN \| EPOLLONESHOT;
506	ev.data.u64 = (__u64)&ctx;
507	int ret = epoll_ctl(iopoll.epollfd, op, ctx.ring->efd, &ev);
508	if (ret < 0) {
509	abort( "KERNEL ERROR: EPOLL %s - (%d) %s\n", error, (int)errno, strerror(errno) );
510	}
511	}
512
513	void __ioctx_register($io_ctx_thread & ctx) {
514	__ioctx_epoll_ctl(ctx, EPOLL_CTL_ADD, "ADD");
515	}
516
517	void __ioctx_prepare_block($io_ctx_thread & ctx) {
518	__cfadbg_print_safe(io_core, "Kernel I/O - epoll : Re-arming io poller %d (%p)\n", ctx.ring->fd, &ctx);
519	__ioctx_epoll_ctl(ctx, EPOLL_CTL_MOD, "REARM");
520	}
521
522	void __ioctx_unregister($io_ctx_thread & ctx) {
523	// Read the current epoch so we know when to stop
524	size_t curr = __atomic_load_n(&iopoll.epoch, __ATOMIC_SEQ_CST);
525
526	// Remove the fd from the iopoller
527	__ioctx_epoll_ctl(ctx, EPOLL_CTL_DEL, "REMOVE");
528
529	// Notify the io poller thread of the shutdown
530	iopoll.run = false;
531	sigval val = { 1 };
532	pthread_sigqueue( iopoll.thrd, SIGUSR1, val );
533
534	// Make sure all this is done
535	__atomic_thread_fence(__ATOMIC_SEQ_CST);
536
537	// Wait for the next epoch
538	while(curr == __atomic_load_n(&iopoll.epoch, __ATOMIC_RELAXED)) yield();
539	}
540
541	//=============================================================================================
542	// I/O Context Misc Setup
543	//=============================================================================================
544	void register_fixed_files( io_context & ctx, int * files, unsigned count ) {
545	int ret = syscall( __NR_io_uring_register, ctx.thrd.ring->fd, IORING_REGISTER_FILES, files, count );
546	if( ret < 0 ) {
547	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
548	}
549
550	__cfadbg_print_safe( io_core, "Kernel I/O : Performed io_register for %p, returned %d\n", active_thread(), ret );
551	}
552
553	void register_fixed_files( cluster & cltr, int * files, unsigned count ) {
554	for(i; cltr.io.cnt) {
555	register_fixed_files( cltr.io.ctxs[i], files, count );
556	}
557	}
558	#endif

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

Download in other formats:

Original Format