Context Navigation

setup.cfa@ d5d3a90

Visit:

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

Last change on this file since d5d3a90 was 3e2b9c9, checked in by Thierry Delisle <tdelisle@…>, 5 years ago
More restructuring of translation units Unclear if it improves compilation time.
Property mode set to `100644`
File size: 15.8 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	#include "io/types.hfa"
20
21	#if !defined(CFA_HAVE_LINUX_IO_URING_H)
22	void __kernel_io_startup() {
23	// Nothing to do without io_uring
24	}
25
26	void __kernel_io_shutdown() {
27	// Nothing to do without io_uring
28	}
29
30	void ?{}(io_context & this, struct cluster & cl) {}
31	void ?{}(io_context & this, struct cluster & cl, const io_context_params & params) {}
32
33	void ^?{}(io_context & this) {}
34	void ^?{}(io_context & this, bool cluster_context) {}
35
36	#else
37	#include <errno.h>
38	#include <stdint.h>
39	#include <string.h>
40	#include <signal.h>
41	#include <unistd.h>
42
43	extern "C" {
44	#include <pthread.h>
45	#include <sys/epoll.h>
46	#include <sys/mman.h>
47	#include <sys/syscall.h>
48
49	#include <linux/io_uring.h>
50	}
51
52	#include "bitmanip.hfa"
53	#include "kernel_private.hfa"
54	#include "thread.hfa"
55
56	void ?{}(io_context_params & this) {
57	this.num_entries = 256;
58	this.num_ready = 256;
59	this.submit_aff = -1;
60	this.eager_submits = false;
61	this.poller_submits = false;
62	this.poll_submit = false;
63	this.poll_complete = false;
64	}
65
66	static void * __io_poller_slow( void * arg );
67
68	// Weirdly, some systems that do support io_uring don't actually define these
69	#ifdef __alpha__
70	/*
71	* alpha is the only exception, all other architectures
72	* have common numbers for new system calls.
73	*/
74	#ifndef __NR_io_uring_setup
75	#define __NR_io_uring_setup 535
76	#endif
77	#ifndef __NR_io_uring_enter
78	#define __NR_io_uring_enter 536
79	#endif
80	#ifndef __NR_io_uring_register
81	#define __NR_io_uring_register 537
82	#endif
83	#else /* !__alpha__ */
84	#ifndef __NR_io_uring_setup
85	#define __NR_io_uring_setup 425
86	#endif
87	#ifndef __NR_io_uring_enter
88	#define __NR_io_uring_enter 426
89	#endif
90	#ifndef __NR_io_uring_register
91	#define __NR_io_uring_register 427
92	#endif
93	#endif
94
95	//=============================================================================================
96	// I/O Startup / Shutdown logic + Master Poller
97	//=============================================================================================
98
99	// IO Master poller loop forward
100	static void * iopoll_loop( __attribute__((unused)) void * args );
101
102	static struct {
103	pthread_t thrd; // pthread handle to io poller thread
104	void * stack; // pthread stack for io poller thread
105	int epollfd; // file descriptor to the epoll instance
106	volatile bool run; // Whether or not to continue
107	} iopoll;
108
109	void __kernel_io_startup(void) {
110	__cfaabi_dbg_print_safe( "Kernel : Creating EPOLL instance\n" );
111
112	iopoll.epollfd = epoll_create1(0);
113	if (iopoll.epollfd == -1) {
114	abort( "internal error, epoll_create1\n");
115	}
116
117	__cfaabi_dbg_print_safe( "Kernel : Starting io poller thread\n" );
118
119	iopoll.run = true;
120	iopoll.stack = __create_pthread( &iopoll.thrd, iopoll_loop, 0p );
121	}
122
123	void __kernel_io_shutdown(void) {
124	// Notify the io poller thread of the shutdown
125	iopoll.run = false;
126	sigval val = { 1 };
127	pthread_sigqueue( iopoll.thrd, SIGUSR1, val );
128
129	// Wait for the io poller thread to finish
130
131	pthread_join( iopoll.thrd, 0p );
132	free( iopoll.stack );
133
134	int ret = close(iopoll.epollfd);
135	if (ret == -1) {
136	abort( "internal error, close epoll\n");
137	}
138
139	// Io polling is now fully stopped
140
141	__cfaabi_dbg_print_safe( "Kernel : IO poller stopped\n" );
142	}
143
144	static void * iopoll_loop( __attribute__((unused)) void * args ) {
145	__processor_id_t id;
146	id.id = doregister(&id);
147	__cfaabi_dbg_print_safe( "Kernel : IO poller thread starting\n" );
148
149	// Block signals to control when they arrive
150	sigset_t mask;
151	sigfillset(&mask);
152	if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
153	abort( "internal error, pthread_sigmask" );
154	}
155
156	sigdelset( &mask, SIGUSR1 );
157
158	// Create sufficient events
159	struct epoll_event events[10];
160	// Main loop
161	while( iopoll.run ) {
162	// Wait for events
163	int nfds = epoll_pwait( iopoll.epollfd, events, 10, -1, &mask );
164
165	// Check if an error occured
166	if (nfds == -1) {
167	if( errno == EINTR ) continue;
168	abort( "internal error, pthread_sigmask" );
169	}
170
171	for(i; nfds) {
172	$io_ctx_thread * io_ctx = ($io_ctx_thread *)(uintptr_t)events[i].data.u64;
173	/* paranoid */ verify( io_ctx );
174	__cfadbg_print_safe(io_core, "Kernel I/O : Unparking io poller %p\n", io_ctx);
175	#if !defined( __CFA_NO_STATISTICS__ )
176	kernelTLS.this_stats = io_ctx->self.curr_cluster->stats;
177	#endif
178	__post( io_ctx->sem, &id );
179	}
180	}
181
182	__cfaabi_dbg_print_safe( "Kernel : IO poller thread stopping\n" );
183	unregister(&id);
184	return 0p;
185	}
186
187	//=============================================================================================
188	// I/O Context Constrution/Destruction
189	//=============================================================================================
190
191	void ?{}($io_ctx_thread & this, struct cluster & cl) { (this.self){ "IO Poller", cl }; }
192	void main( $io_ctx_thread & this );
193	static inline $thread * get_thread( $io_ctx_thread & this ) { return &this.self; }
194	void ^?{}( $io_ctx_thread & mutex this ) {}
195
196	static void __io_create ( __io_data & this, const io_context_params & params_in );
197	static void __io_destroy( __io_data & this );
198
199	void ?{}(io_context & this, struct cluster & cl, const io_context_params & params) {
200	(this.thrd){ cl };
201	this.thrd.ring = malloc();
202	__cfadbg_print_safe(io_core, "Kernel I/O : Creating ring for io_context %p\n", &this);
203	__io_create( *this.thrd.ring, params );
204
205	__cfadbg_print_safe(io_core, "Kernel I/O : Starting poller thread for io_context %p\n", &this);
206	this.thrd.done = false;
207	__thrd_start( this.thrd, main );
208
209	__cfadbg_print_safe(io_core, "Kernel I/O : io_context %p ready\n", &this);
210	}
211
212	void ?{}(io_context & this, struct cluster & cl) {
213	io_context_params params;
214	(this){ cl, params };
215	}
216
217	void ^?{}(io_context & this, bool cluster_context) {
218	__cfadbg_print_safe(io_core, "Kernel I/O : tearing down io_context %p\n", &this);
219
220	// Notify the thread of the shutdown
221	__atomic_store_n(&this.thrd.done, true, __ATOMIC_SEQ_CST);
222
223	// If this is an io_context within a cluster, things get trickier
224	$thread & thrd = this.thrd.self;
225	if( cluster_context ) {
226	cluster & cltr = *thrd.curr_cluster;
227	/* paranoid */ verify( cltr.nprocessors == 0 \|\| &cltr == mainCluster );
228	/* paranoid */ verify( !ready_mutate_islocked() );
229
230	// We need to adjust the clean-up based on where the thread is
231	if( thrd.state == Ready \|\| thrd.preempted != __NO_PREEMPTION ) {
232
233	ready_schedule_lock( (struct __processor_id_t *)active_processor() );
234
235	// This is the tricky case
236	// The thread was preempted and now it is on the ready queue
237	// The thread should be the last on the list
238	/* paranoid */ verify( thrd.link.next != 0p );
239
240	// Remove the thread from the ready queue of this cluster
241	__attribute__((unused)) bool removed = remove_head( &cltr, &thrd );
242	/* paranoid */ verify( removed );
243	thrd.link.next = 0p;
244	thrd.link.prev = 0p;
245	__cfaabi_dbg_debug_do( thrd.unpark_stale = true );
246
247	// Fixup the thread state
248	thrd.state = Blocked;
249	thrd.ticket = 0;
250	thrd.preempted = __NO_PREEMPTION;
251
252	ready_schedule_unlock( (struct __processor_id_t *)active_processor() );
253
254	// Pretend like the thread was blocked all along
255	}
256	// !!! This is not an else if !!!
257	if( thrd.state == Blocked ) {
258
259	// This is the "easy case"
260	// The thread is parked and can easily be moved to active cluster
261	verify( thrd.curr_cluster != active_cluster() \|\| thrd.curr_cluster == mainCluster );
262	thrd.curr_cluster = active_cluster();
263
264	// unpark the fast io_poller
265	unpark( &thrd __cfaabi_dbg_ctx2 );
266	}
267	else {
268
269	// The thread is in a weird state
270	// I don't know what to do here
271	abort("io_context poller thread is in unexpected state, cannot clean-up correctly\n");
272	}
273	} else {
274	unpark( &thrd __cfaabi_dbg_ctx2 );
275	}
276
277	^(this.thrd){};
278	__cfadbg_print_safe(io_core, "Kernel I/O : Stopped poller thread for io_context %p\n", &this);
279
280	__io_destroy( *this.thrd.ring );
281	__cfadbg_print_safe(io_core, "Kernel I/O : Destroyed ring for io_context %p\n", &this);
282
283	free(this.thrd.ring);
284	}
285
286	void ^?{}(io_context & this) {
287	^(this){ false };
288	}
289
290	static void __io_create( __io_data & this, const io_context_params & params_in ) {
291	// Step 1 : call to setup
292	struct io_uring_params params;
293	memset(&params, 0, sizeof(params));
294	if( params_in.poll_submit ) params.flags \|= IORING_SETUP_SQPOLL;
295	if( params_in.poll_complete ) params.flags \|= IORING_SETUP_IOPOLL;
296
297	uint32_t nentries = params_in.num_entries;
298
299	int fd = syscall(__NR_io_uring_setup, nentries, &params );
300	if(fd < 0) {
301	abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
302	}
303
304	// Step 2 : mmap result
305	memset( &this, 0, sizeof(struct __io_data) );
306	struct __submition_data & sq = this.submit_q;
307	struct __completion_data & cq = this.completion_q;
308
309	// calculate the right ring size
310	sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
311	cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
312
313	// Requires features
314	#if defined(IORING_FEAT_SINGLE_MMAP)
315	// adjust the size according to the parameters
316	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
317	cq.ring_sz = sq.ring_sz = max(cq.ring_sz, sq.ring_sz);
318	}
319	#endif
320
321	// mmap the Submit Queue into existence
322	sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQ_RING);
323	if (sq.ring_ptr == (void*)MAP_FAILED) {
324	abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
325	}
326
327	// Requires features
328	#if defined(IORING_FEAT_SINGLE_MMAP)
329	// mmap the Completion Queue into existence (may or may not be needed)
330	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
331	cq.ring_ptr = sq.ring_ptr;
332	}
333	else
334	#endif
335	{
336	// We need multiple call to MMAP
337	cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_CQ_RING);
338	if (cq.ring_ptr == (void*)MAP_FAILED) {
339	munmap(sq.ring_ptr, sq.ring_sz);
340	abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
341	}
342	}
343
344	// mmap the submit queue entries
345	size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
346	sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQES);
347	if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
348	munmap(sq.ring_ptr, sq.ring_sz);
349	if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
350	abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
351	}
352
353	// Get the pointers from the kernel to fill the structure
354	// submit queue
355	sq.head = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
356	sq.tail = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
357	sq.mask = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
358	sq.num = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
359	sq.flags = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
360	sq.dropped = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
361	sq.array = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
362	sq.prev_head = *sq.head;
363
364	{
365	const uint32_t num = *sq.num;
366	for( i; num ) {
367	sq.sqes[i].user_data = 0ul64;
368	}
369	}
370
371	(sq.lock){};
372	(sq.release_lock){};
373
374	if( params_in.poller_submits \|\| params_in.eager_submits ) {
375	/* paranoid */ verify( is_pow2( params_in.num_ready ) \|\| (params_in.num_ready < 8) );
376	sq.ready_cnt = max( params_in.num_ready, 8 );
377	sq.ready = alloc_align( 64, sq.ready_cnt );
378	for(i; sq.ready_cnt) {
379	sq.ready[i] = -1ul32;
380	}
381	}
382	else {
383	sq.ready_cnt = 0;
384	sq.ready = 0p;
385	}
386
387	// completion queue
388	cq.head = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
389	cq.tail = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
390	cq.mask = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
391	cq.num = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
392	cq.overflow = ( uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
393	cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
394
395	// some paranoid checks
396	/* 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 );
397	/* paranoid / verifyf( (cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
398	/* paranoid / verifyf( (cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
399	/* paranoid / verifyf( (cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
400
401	/* 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 );
402	/* paranoid / verifyf( (sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
403	/* paranoid / verifyf( (sq.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.head );
404	/* paranoid / verifyf( (sq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.tail );
405
406	// Update the global ring info
407	this.ring_flags = params.flags;
408	this.fd = fd;
409	this.eager_submits = params_in.eager_submits;
410	this.poller_submits = params_in.poller_submits;
411	}
412
413	static void __io_destroy( __io_data & this ) {
414	// Shutdown the io rings
415	struct __submition_data & sq = this.submit_q;
416	struct __completion_data & cq = this.completion_q;
417
418	// unmap the submit queue entries
419	munmap(sq.sqes, (sq.num) sizeof(struct io_uring_sqe));
420
421	// unmap the Submit Queue ring
422	munmap(sq.ring_ptr, sq.ring_sz);
423
424	// unmap the Completion Queue ring, if it is different
425	if (cq.ring_ptr != sq.ring_ptr) {
426	munmap(cq.ring_ptr, cq.ring_sz);
427	}
428
429	// close the file descriptor
430	close(this.fd);
431
432	free( this.submit_q.ready ); // Maybe null, doesn't matter
433	}
434
435	//=============================================================================================
436	// I/O Context Sleep
437	//=============================================================================================
438
439	void __ioctx_register($io_ctx_thread & ctx, struct epoll_event & ev) {
440	ev.events = EPOLLIN \| EPOLLONESHOT;
441	ev.data.u64 = (uint64_t)&ctx;
442	int ret = epoll_ctl(iopoll.epollfd, EPOLL_CTL_ADD, ctx.ring->fd, &ev);
443	if (ret < 0) {
444	abort( "KERNEL ERROR: EPOLL ADD - (%d) %s\n", (int)errno, strerror(errno) );
445	}
446	}
447
448	void __ioctx_prepare_block($io_ctx_thread & ctx, struct epoll_event & ev) {
449	int ret = epoll_ctl(iopoll.epollfd, EPOLL_CTL_MOD, ctx.ring->fd, &ev);
450	if (ret < 0) {
451	abort( "KERNEL ERROR: EPOLL REARM - (%d) %s\n", (int)errno, strerror(errno) );
452	}
453	}
454
455	//=============================================================================================
456	// I/O Context Misc Setup
457	//=============================================================================================
458	void register_fixed_files( io_context & ctx, int * files, unsigned count ) {
459	int ret = syscall( __NR_io_uring_register, ctx.thrd.ring->fd, IORING_REGISTER_FILES, files, count );
460	if( ret < 0 ) {
461	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
462	}
463
464	__cfadbg_print_safe( io_core, "Kernel I/O : Performed io_register for %p, returned %d\n", active_thread(), ret );
465	}
466
467	void register_fixed_files( cluster & cltr, int * files, unsigned count ) {
468	for(i; cltr.io.cnt) {
469	register_fixed_files( cltr.io.ctxs[i], files, count );
470	}
471	}
472	#endif

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

Download in other formats:

Original Format