Context Navigation

setup.cfa @ 58fe85a

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

Last change on this file since 58fe85a was 80444bb, checked in by Thierry Delisle <tdelisle@…>, 4 years ago
Fixed prints in io setup.
Property mode set to `100644`
File size: 16.2 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	#else
45	#include <errno.h>
46	#include <stdint.h>
47	#include <string.h>
48	#include <signal.h>
49	#include <unistd.h>
50
51	extern "C" {
52	#include <pthread.h>
53	#include <sys/epoll.h>
54	#include <sys/mman.h>
55	#include <sys/syscall.h>
56
57	#include <linux/io_uring.h>
58	}
59
60	#include "bitmanip.hfa"
61	#include "kernel_private.hfa"
62	#include "thread.hfa"
63
64	void ?{}(io_context_params & this) {
65	this.num_entries = 256;
66	this.num_ready = 256;
67	this.submit_aff = -1;
68	this.eager_submits = false;
69	this.poller_submits = false;
70	this.poll_submit = false;
71	this.poll_complete = false;
72	}
73
74	static void * __io_poller_slow( void * arg );
75
76	// Weirdly, some systems that do support io_uring don't actually define these
77	#ifdef __alpha__
78	/*
79	* alpha is the only exception, all other architectures
80	* have common numbers for new system calls.
81	*/
82	#ifndef __NR_io_uring_setup
83	#define __NR_io_uring_setup 535
84	#endif
85	#ifndef __NR_io_uring_enter
86	#define __NR_io_uring_enter 536
87	#endif
88	#ifndef __NR_io_uring_register
89	#define __NR_io_uring_register 537
90	#endif
91	#else /* !__alpha__ */
92	#ifndef __NR_io_uring_setup
93	#define __NR_io_uring_setup 425
94	#endif
95	#ifndef __NR_io_uring_enter
96	#define __NR_io_uring_enter 426
97	#endif
98	#ifndef __NR_io_uring_register
99	#define __NR_io_uring_register 427
100	#endif
101	#endif
102
103	//=============================================================================================
104	// I/O Startup / Shutdown logic + Master Poller
105	//=============================================================================================
106
107	// IO Master poller loop forward
108	static void * iopoll_loop( __attribute__((unused)) void * args );
109
110	static struct {
111	pthread_t thrd; // pthread handle to io poller thread
112	void * stack; // pthread stack for io poller thread
113	int epollfd; // file descriptor to the epoll instance
114	volatile bool run; // Whether or not to continue
115	} iopoll;
116
117	void __kernel_io_startup(void) {
118	__cfadbg_print_safe(io_core, "Kernel : Creating EPOLL instance\n" );
119
120	iopoll.epollfd = epoll_create1(0);
121	if (iopoll.epollfd == -1) {
122	abort( "internal error, epoll_create1\n");
123	}
124
125	__cfadbg_print_safe(io_core, "Kernel : Starting io poller thread\n" );
126
127	iopoll.run = true;
128	iopoll.stack = __create_pthread( &iopoll.thrd, iopoll_loop, 0p );
129	}
130
131	void __kernel_io_shutdown(void) {
132	// Notify the io poller thread of the shutdown
133	iopoll.run = false;
134	sigval val = { 1 };
135	pthread_sigqueue( iopoll.thrd, SIGUSR1, val );
136
137	// Wait for the io poller thread to finish
138
139	__destroy_pthread( iopoll.thrd, iopoll.stack, 0p );
140
141	int ret = close(iopoll.epollfd);
142	if (ret == -1) {
143	abort( "internal error, close epoll\n");
144	}
145
146	// Io polling is now fully stopped
147
148	__cfadbg_print_safe(io_core, "Kernel : IO poller stopped\n" );
149	}
150
151	static void * iopoll_loop( __attribute__((unused)) void * args ) {
152	__processor_id_t id;
153	id.full_proc = false;
154	id.id = doregister(&id);
155	__cfaabi_tls.this_proc_id = &id;
156	__cfadbg_print_safe(io_core, "Kernel : IO poller thread starting\n" );
157
158	// Block signals to control when they arrive
159	sigset_t mask;
160	sigfillset(&mask);
161	if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
162	abort( "internal error, pthread_sigmask" );
163	}
164
165	sigdelset( &mask, SIGUSR1 );
166
167	// Create sufficient events
168	struct epoll_event events[10];
169	// Main loop
170	while( iopoll.run ) {
171	// Wait for events
172	int nfds = epoll_pwait( iopoll.epollfd, events, 10, -1, &mask );
173
174	// Check if an error occured
175	if (nfds == -1) {
176	if( errno == EINTR ) continue;
177	abort( "internal error, pthread_sigmask" );
178	}
179
180	for(i; nfds) {
181	$io_ctx_thread * io_ctx = ($io_ctx_thread *)(uintptr_t)events[i].data.u64;
182	/* paranoid */ verify( io_ctx );
183	__cfadbg_print_safe(io_core, "Kernel I/O : Unparking io poller %p\n", io_ctx);
184	#if !defined( __CFA_NO_STATISTICS__ )
185	__cfaabi_tls.this_stats = io_ctx->self.curr_cluster->stats;
186	#endif
187	post( io_ctx->sem );
188	}
189	}
190
191	__cfadbg_print_safe(io_core, "Kernel : IO poller thread stopping\n" );
192	unregister(&id);
193	return 0p;
194	}
195
196	//=============================================================================================
197	// I/O Context Constrution/Destruction
198	//=============================================================================================
199
200	void ?{}($io_ctx_thread & this, struct cluster & cl) { (this.self){ "IO Poller", cl }; }
201	void main( $io_ctx_thread & this );
202	static inline $thread * get_thread( $io_ctx_thread & this ) { return &this.self; }
203	void ^?{}( $io_ctx_thread & mutex this ) {}
204
205	static void __io_create ( __io_data & this, const io_context_params & params_in );
206	static void __io_destroy( __io_data & this );
207
208	void ?{}(io_context & this, struct cluster & cl, const io_context_params & params) {
209	(this.thrd){ cl };
210	this.thrd.ring = malloc();
211	__cfadbg_print_safe(io_core, "Kernel I/O : Creating ring for io_context %p\n", &this);
212	__io_create( *this.thrd.ring, params );
213
214	__cfadbg_print_safe(io_core, "Kernel I/O : Starting poller thread for io_context %p\n", &this);
215	this.thrd.done = false;
216	__thrd_start( this.thrd, main );
217
218	__cfadbg_print_safe(io_core, "Kernel I/O : io_context %p ready\n", &this);
219	}
220
221	void ?{}(io_context & this, struct cluster & cl) {
222	io_context_params params;
223	(this){ cl, params };
224	}
225
226	void ^?{}(io_context & this, bool cluster_context) {
227	__cfadbg_print_safe(io_core, "Kernel I/O : tearing down io_context %p\n", &this);
228
229	// Notify the thread of the shutdown
230	__atomic_store_n(&this.thrd.done, true, __ATOMIC_SEQ_CST);
231
232	// If this is an io_context within a cluster, things get trickier
233	$thread & thrd = this.thrd.self;
234	if( cluster_context ) {
235	cluster & cltr = *thrd.curr_cluster;
236	/* paranoid */ verify( cltr.idles.total == 0 \|\| &cltr == mainCluster );
237	/* paranoid */ verify( !ready_mutate_islocked() );
238
239	// We need to adjust the clean-up based on where the thread is
240	if( thrd.state == Ready \|\| thrd.preempted != __NO_PREEMPTION ) {
241
242	ready_schedule_lock();
243
244	// This is the tricky case
245	// The thread was preempted and now it is on the ready queue
246	// The thread should be the last on the list
247	/* paranoid */ verify( thrd.link.next != 0p );
248
249	// Remove the thread from the ready queue of this cluster
250	__attribute__((unused)) bool removed = remove_head( &cltr, &thrd );
251	/* paranoid */ verify( removed );
252	thrd.link.next = 0p;
253	thrd.link.prev = 0p;
254
255	// Fixup the thread state
256	thrd.state = Blocked;
257	thrd.ticket = TICKET_BLOCKED;
258	thrd.preempted = __NO_PREEMPTION;
259
260	ready_schedule_unlock();
261
262	// Pretend like the thread was blocked all along
263	}
264	// !!! This is not an else if !!!
265	if( thrd.state == Blocked ) {
266
267	// This is the "easy case"
268	// The thread is parked and can easily be moved to active cluster
269	verify( thrd.curr_cluster != active_cluster() \|\| thrd.curr_cluster == mainCluster );
270	thrd.curr_cluster = active_cluster();
271
272	// unpark the fast io_poller
273	unpark( &thrd );
274	}
275	else {
276
277	// The thread is in a weird state
278	// I don't know what to do here
279	abort("io_context poller thread is in unexpected state, cannot clean-up correctly\n");
280	}
281	} else {
282	post( this.thrd.sem );
283	}
284
285	^(this.thrd){};
286	__cfadbg_print_safe(io_core, "Kernel I/O : Stopped poller thread for io_context %p\n", &this);
287
288	__io_destroy( *this.thrd.ring );
289	__cfadbg_print_safe(io_core, "Kernel I/O : Destroyed ring for io_context %p\n", &this);
290
291	free(this.thrd.ring);
292	}
293
294	void ^?{}(io_context & this) {
295	^(this){ false };
296	}
297
298	static void __io_create( __io_data & this, const io_context_params & params_in ) {
299	// Step 1 : call to setup
300	struct io_uring_params params;
301	memset(&params, 0, sizeof(params));
302	if( params_in.poll_submit ) params.flags \|= IORING_SETUP_SQPOLL;
303	if( params_in.poll_complete ) params.flags \|= IORING_SETUP_IOPOLL;
304
305	__u32 nentries = params_in.num_entries != 0 ? params_in.num_entries : 256;
306	if( !is_pow2(nentries) ) {
307	abort("ERROR: I/O setup 'num_entries' must be a power of 2\n");
308	}
309	if( params_in.poller_submits && params_in.eager_submits ) {
310	abort("ERROR: I/O setup 'poller_submits' and 'eager_submits' cannot be used together\n");
311	}
312
313	int fd = syscall(__NR_io_uring_setup, nentries, &params );
314	if(fd < 0) {
315	abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
316	}
317
318	// Step 2 : mmap result
319	memset( &this, 0, sizeof(struct __io_data) );
320	struct __submition_data & sq = this.submit_q;
321	struct __completion_data & cq = this.completion_q;
322
323	// calculate the right ring size
324	sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
325	cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
326
327	// Requires features
328	#if defined(IORING_FEAT_SINGLE_MMAP)
329	// adjust the size according to the parameters
330	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
331	cq.ring_sz = sq.ring_sz = max(cq.ring_sz, sq.ring_sz);
332	}
333	#endif
334
335	// mmap the Submit Queue into existence
336	sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQ_RING);
337	if (sq.ring_ptr == (void*)MAP_FAILED) {
338	abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
339	}
340
341	// Requires features
342	#if defined(IORING_FEAT_SINGLE_MMAP)
343	// mmap the Completion Queue into existence (may or may not be needed)
344	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
345	cq.ring_ptr = sq.ring_ptr;
346	}
347	else
348	#endif
349	{
350	// We need multiple call to MMAP
351	cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_CQ_RING);
352	if (cq.ring_ptr == (void*)MAP_FAILED) {
353	munmap(sq.ring_ptr, sq.ring_sz);
354	abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
355	}
356	}
357
358	// mmap the submit queue entries
359	size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
360	sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQES);
361	if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
362	munmap(sq.ring_ptr, sq.ring_sz);
363	if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
364	abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
365	}
366
367	// Get the pointers from the kernel to fill the structure
368	// submit queue
369	sq.head = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
370	sq.tail = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
371	sq.mask = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
372	sq.num = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
373	sq.flags = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
374	sq.dropped = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
375	sq.array = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
376	sq.prev_head = *sq.head;
377
378	{
379	const __u32 num = *sq.num;
380	for( i; num ) {
381	sq.sqes[i].user_data = 0ul64;
382	}
383	}
384
385	(sq.submit_lock){};
386	(sq.release_lock){};
387
388	if( params_in.poller_submits \|\| params_in.eager_submits ) {
389	/* paranoid */ verify( is_pow2( params_in.num_ready ) \|\| (params_in.num_ready < 8) );
390	sq.ready_cnt = max( params_in.num_ready, 8 );
391	sq.ready = alloc( sq.ready_cnt, 64`align );
392	for(i; sq.ready_cnt) {
393	sq.ready[i] = -1ul32;
394	}
395	sq.prev_ready = 0;
396	}
397	else {
398	sq.ready_cnt = 0;
399	sq.ready = 0p;
400	sq.prev_ready = 0;
401	}
402
403	// completion queue
404	cq.head = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
405	cq.tail = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
406	cq.mask = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
407	cq.num = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
408	cq.overflow = ( __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
409	cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
410
411	// some paranoid checks
412	/* 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 );
413	/* paranoid / verifyf( (cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
414	/* paranoid / verifyf( (cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
415	/* paranoid / verifyf( (cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
416
417	/* 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 );
418	/* paranoid / verifyf( (sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
419	/* paranoid / verifyf( (sq.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.head );
420	/* paranoid / verifyf( (sq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.tail );
421
422	// Update the global ring info
423	this.ring_flags = params.flags;
424	this.fd = fd;
425	this.eager_submits = params_in.eager_submits;
426	this.poller_submits = params_in.poller_submits;
427	}
428
429	static void __io_destroy( __io_data & this ) {
430	// Shutdown the io rings
431	struct __submition_data & sq = this.submit_q;
432	struct __completion_data & cq = this.completion_q;
433
434	// unmap the submit queue entries
435	munmap(sq.sqes, (sq.num) sizeof(struct io_uring_sqe));
436
437	// unmap the Submit Queue ring
438	munmap(sq.ring_ptr, sq.ring_sz);
439
440	// unmap the Completion Queue ring, if it is different
441	if (cq.ring_ptr != sq.ring_ptr) {
442	munmap(cq.ring_ptr, cq.ring_sz);
443	}
444
445	// close the file descriptor
446	close(this.fd);
447
448	free( this.submit_q.ready ); // Maybe null, doesn't matter
449	}
450
451	//=============================================================================================
452	// I/O Context Sleep
453	//=============================================================================================
454
455	void __ioctx_register($io_ctx_thread & ctx, struct epoll_event & ev) {
456	ev.events = EPOLLIN \| EPOLLONESHOT;
457	ev.data.u64 = (__u64)&ctx;
458	int ret = epoll_ctl(iopoll.epollfd, EPOLL_CTL_ADD, ctx.ring->fd, &ev);
459	if (ret < 0) {
460	abort( "KERNEL ERROR: EPOLL ADD - (%d) %s\n", (int)errno, strerror(errno) );
461	}
462	}
463
464	void __ioctx_prepare_block($io_ctx_thread & ctx, struct epoll_event & ev) {
465	int ret = epoll_ctl(iopoll.epollfd, EPOLL_CTL_MOD, ctx.ring->fd, &ev);
466	if (ret < 0) {
467	abort( "KERNEL ERROR: EPOLL REARM - (%d) %s\n", (int)errno, strerror(errno) );
468	}
469	}
470
471	//=============================================================================================
472	// I/O Context Misc Setup
473	//=============================================================================================
474	void register_fixed_files( io_context & ctx, int * files, unsigned count ) {
475	int ret = syscall( __NR_io_uring_register, ctx.thrd.ring->fd, IORING_REGISTER_FILES, files, count );
476	if( ret < 0 ) {
477	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
478	}
479
480	__cfadbg_print_safe( io_core, "Kernel I/O : Performed io_register for %p, returned %d\n", active_thread(), ret );
481	}
482
483	void register_fixed_files( cluster & cltr, int * files, unsigned count ) {
484	for(i; cltr.io.cnt) {
485	register_fixed_files( cltr.io.ctxs[i], files, count );
486	}
487	}
488	#endif

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

Download in other formats:

Original Format