Context Navigation

setup.cfa@ 58f99b3

Visit:

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

Last change on this file since 58f99b3 was 58f99b3, checked in by Thierry Delisle <tdelisle@…>, 5 years ago
Retry interruptable syscalls instead of blocking interrupts
Property mode set to `100644`
File size: 17.8 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	#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/eventfd.h>
55	#include <sys/mman.h>
56	#include <sys/syscall.h>
57
58	#include <linux/io_uring.h>
59	}
60
61	#include "bitmanip.hfa"
62	#include "kernel_private.hfa"
63	#include "thread.hfa"
64
65	void ?{}(io_context_params & this) {
66	this.num_entries = 256;
67	this.num_ready = 256;
68	this.submit_aff = -1;
69	this.eager_submits = false;
70	this.poller_submits = false;
71	this.poll_submit = false;
72	this.poll_complete = false;
73	}
74
75	static void * __io_poller_slow( void * arg );
76
77	// Weirdly, some systems that do support io_uring don't actually define these
78	#ifdef __alpha__
79	/*
80	* alpha is the only exception, all other architectures
81	* have common numbers for new system calls.
82	*/
83	#ifndef __NR_io_uring_setup
84	#define __NR_io_uring_setup 535
85	#endif
86	#ifndef __NR_io_uring_enter
87	#define __NR_io_uring_enter 536
88	#endif
89	#ifndef __NR_io_uring_register
90	#define __NR_io_uring_register 537
91	#endif
92	#else /* !__alpha__ */
93	#ifndef __NR_io_uring_setup
94	#define __NR_io_uring_setup 425
95	#endif
96	#ifndef __NR_io_uring_enter
97	#define __NR_io_uring_enter 426
98	#endif
99	#ifndef __NR_io_uring_register
100	#define __NR_io_uring_register 427
101	#endif
102	#endif
103
104	//=============================================================================================
105	// I/O Startup / Shutdown logic + Master Poller
106	//=============================================================================================
107
108	// IO Master poller loop forward
109	static void * iopoll_loop( __attribute__((unused)) void * args );
110
111	static struct {
112	pthread_t thrd; // pthread handle to io poller thread
113	void * stack; // pthread stack for io poller thread
114	int epollfd; // file descriptor to the epoll instance
115	volatile bool run; // Whether or not to continue
116	} iopoll;
117
118	void __kernel_io_startup(void) {
119	__cfadbg_print_safe(io_core, "Kernel : Creating EPOLL instance\n" );
120
121	iopoll.epollfd = epoll_create1(0);
122	if (iopoll.epollfd == -1) {
123	abort( "internal error, epoll_create1\n");
124	}
125
126	__cfadbg_print_safe(io_core, "Kernel : Starting io poller thread\n" );
127
128	iopoll.run = true;
129	iopoll.stack = __create_pthread( &iopoll.thrd, iopoll_loop, 0p );
130	}
131
132	void __kernel_io_shutdown(void) {
133	// Notify the io poller thread of the shutdown
134	iopoll.run = false;
135	sigval val = { 1 };
136	pthread_sigqueue( iopoll.thrd, SIGUSR1, val );
137
138	// Wait for the io poller thread to finish
139
140	__destroy_pthread( iopoll.thrd, iopoll.stack, 0p );
141
142	int ret = close(iopoll.epollfd);
143	if (ret == -1) {
144	abort( "internal error, close epoll\n");
145	}
146
147	// Io polling is now fully stopped
148
149	__cfadbg_print_safe(io_core, "Kernel : IO poller stopped\n" );
150	}
151
152	static void * iopoll_loop( __attribute__((unused)) void * args ) {
153	__processor_id_t id;
154	id.full_proc = false;
155	id.id = doregister(&id);
156	__cfaabi_tls.this_proc_id = &id;
157	__cfadbg_print_safe(io_core, "Kernel : IO poller thread starting\n" );
158
159	// Block signals to control when they arrive
160	sigset_t mask;
161	sigfillset(&mask);
162	if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
163	abort( "internal error, pthread_sigmask" );
164	}
165
166	sigdelset( &mask, SIGUSR1 );
167
168	// Create sufficient events
169	struct epoll_event events[10];
170	// Main loop
171	while( iopoll.run ) {
172	__cfadbg_print_safe(io_core, "Kernel I/O - epoll : waiting on io_uring contexts\n");
173
174	// Wait for events
175	int nfds = epoll_pwait( iopoll.epollfd, events, 10, -1, &mask );
176
177	__cfadbg_print_safe(io_core, "Kernel I/O - epoll : %d io contexts events, waking up\n", nfds);
178
179	// Check if an error occured
180	if (nfds == -1) {
181	if( errno == EINTR ) continue;
182	abort( "internal error, pthread_sigmask" );
183	}
184
185	for(i; nfds) {
186	$io_ctx_thread * io_ctx = ($io_ctx_thread *)(uintptr_t)events[i].data.u64;
187	/* paranoid */ verify( io_ctx );
188	__cfadbg_print_safe(io_core, "Kernel I/O - epoll : Unparking io poller %d (%p)\n", io_ctx->ring->fd, io_ctx);
189	#if !defined( __CFA_NO_STATISTICS__ )
190	__cfaabi_tls.this_stats = io_ctx->self.curr_cluster->stats;
191	#endif
192	post( io_ctx->sem );
193	}
194	}
195
196	__cfadbg_print_safe(io_core, "Kernel : IO poller thread stopping\n" );
197	unregister(&id);
198	return 0p;
199	}
200
201	//=============================================================================================
202	// I/O Context Constrution/Destruction
203	//=============================================================================================
204
205	void ?{}($io_ctx_thread & this, struct cluster & cl) { (this.self){ "IO Poller", cl }; }
206	void main( $io_ctx_thread & this );
207	static inline $thread * get_thread( $io_ctx_thread & this ) { return &this.self; }
208	void ^?{}( $io_ctx_thread & mutex this ) {}
209
210	static void __io_create ( __io_data & this, const io_context_params & params_in );
211	static void __io_destroy( __io_data & this );
212
213	void ?{}(io_context & this, struct cluster & cl, const io_context_params & params) {
214	(this.thrd){ cl };
215	this.thrd.ring = malloc();
216	__cfadbg_print_safe(io_core, "Kernel I/O : Creating ring for io_context %p\n", &this);
217	__io_create( *this.thrd.ring, params );
218
219	__cfadbg_print_safe(io_core, "Kernel I/O : Starting poller thread for io_context %p\n", &this);
220	this.thrd.done = false;
221	__thrd_start( this.thrd, main );
222
223	__cfadbg_print_safe(io_core, "Kernel I/O : io_context %p ready\n", &this);
224	}
225
226	void ?{}(io_context & this, struct cluster & cl) {
227	io_context_params params;
228	(this){ cl, params };
229	}
230
231	void ^?{}(io_context & this, bool cluster_context) {
232	__cfadbg_print_safe(io_core, "Kernel I/O : tearing down io_context %p\n", &this);
233
234	// Notify the thread of the shutdown
235	__atomic_store_n(&this.thrd.done, true, __ATOMIC_SEQ_CST);
236
237	// If this is an io_context within a cluster, things get trickier
238	$thread & thrd = this.thrd.self;
239	if( cluster_context ) {
240	// We are about to do weird things with the threads
241	// we don't need interrupts to complicate everything
242	disable_interrupts();
243
244	// Get cluster info
245	cluster & cltr = *thrd.curr_cluster;
246	/* paranoid */ verify( cltr.idles.total == 0 \|\| &cltr == mainCluster );
247	/* paranoid */ verify( !ready_mutate_islocked() );
248
249	// We need to adjust the clean-up based on where the thread is
250	if( thrd.state == Ready \|\| thrd.preempted != __NO_PREEMPTION ) {
251	// This is the tricky case
252	// The thread was preempted or ready to run and now it is on the ready queue
253	// but the cluster is shutting down, so there aren't any processors to run the ready queue
254	// the solution is to steal the thread from the ready-queue and pretend it was blocked all along
255
256	ready_schedule_lock();
257	// The thread should on the list
258	/* paranoid */ verify( thrd.link.next != 0p );
259
260	// Remove the thread from the ready queue of this cluster
261	// The thread should be the last on the list
262	__attribute__((unused)) bool removed = remove_head( &cltr, &thrd );
263	/* paranoid */ verify( removed );
264	thrd.link.next = 0p;
265	thrd.link.prev = 0p;
266
267	// Fixup the thread state
268	thrd.state = Blocked;
269	thrd.ticket = TICKET_BLOCKED;
270	thrd.preempted = __NO_PREEMPTION;
271
272	ready_schedule_unlock();
273
274	// Pretend like the thread was blocked all along
275	}
276	// !!! This is not an else if !!!
277	// Ok, now the thread is blocked (whether we cheated to get here or not)
278	if( thrd.state == Blocked ) {
279	// This is the "easy case"
280	// The thread is parked and can easily be moved to active cluster
281	verify( thrd.curr_cluster != active_cluster() \|\| thrd.curr_cluster == mainCluster );
282	thrd.curr_cluster = active_cluster();
283
284	// unpark the fast io_poller
285	unpark( &thrd );
286	}
287	else {
288	// The thread is in a weird state
289	// I don't know what to do here
290	abort("io_context poller thread is in unexpected state, cannot clean-up correctly\n");
291	}
292
293	// The weird thread kidnapping stuff is over, restore interrupts.
294	enable_interrupts( __cfaabi_dbg_ctx );
295	} else {
296	post( this.thrd.sem );
297	}
298
299	^(this.thrd){};
300	__cfadbg_print_safe(io_core, "Kernel I/O : Stopped poller thread for io_context %p\n", &this);
301
302	__io_destroy( *this.thrd.ring );
303	__cfadbg_print_safe(io_core, "Kernel I/O : Destroyed ring for io_context %p\n", &this);
304
305	free(this.thrd.ring);
306	}
307
308	void ^?{}(io_context & this) {
309	^(this){ false };
310	}
311
312	static void __io_create( __io_data & this, const io_context_params & params_in ) {
313	// Step 1 : call to setup
314	struct io_uring_params params;
315	memset(&params, 0, sizeof(params));
316	if( params_in.poll_submit ) params.flags \|= IORING_SETUP_SQPOLL;
317	if( params_in.poll_complete ) params.flags \|= IORING_SETUP_IOPOLL;
318
319	__u32 nentries = params_in.num_entries != 0 ? params_in.num_entries : 256;
320	if( !is_pow2(nentries) ) {
321	abort("ERROR: I/O setup 'num_entries' must be a power of 2\n");
322	}
323	if( params_in.poller_submits && params_in.eager_submits ) {
324	abort("ERROR: I/O setup 'poller_submits' and 'eager_submits' cannot be used together\n");
325	}
326
327	int fd = syscall(__NR_io_uring_setup, nentries, &params );
328	if(fd < 0) {
329	abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
330	}
331
332	// Step 2 : mmap result
333	memset( &this, 0, sizeof(struct __io_data) );
334	struct __submition_data & sq = this.submit_q;
335	struct __completion_data & cq = this.completion_q;
336
337	// calculate the right ring size
338	sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
339	cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
340
341	// Requires features
342	#if defined(IORING_FEAT_SINGLE_MMAP)
343	// adjust the size according to the parameters
344	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
345	cq.ring_sz = sq.ring_sz = max(cq.ring_sz, sq.ring_sz);
346	}
347	#endif
348
349	// mmap the Submit Queue into existence
350	sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQ_RING);
351	if (sq.ring_ptr == (void*)MAP_FAILED) {
352	abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
353	}
354
355	// Requires features
356	#if defined(IORING_FEAT_SINGLE_MMAP)
357	// mmap the Completion Queue into existence (may or may not be needed)
358	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
359	cq.ring_ptr = sq.ring_ptr;
360	}
361	else
362	#endif
363	{
364	// We need multiple call to MMAP
365	cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_CQ_RING);
366	if (cq.ring_ptr == (void*)MAP_FAILED) {
367	munmap(sq.ring_ptr, sq.ring_sz);
368	abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
369	}
370	}
371
372	// mmap the submit queue entries
373	size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
374	sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQES);
375	if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
376	munmap(sq.ring_ptr, sq.ring_sz);
377	if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
378	abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
379	}
380	memset(sq.sqes, 0xde, size);
381
382	verify( 0 != (params.features & IORING_FEAT_NODROP) );
383
384	// Step 3 : Initialize the data structure
385	// Get the pointers from the kernel to fill the structure
386	// submit queue
387	sq.head = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
388	sq.tail = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
389	sq.mask = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
390	sq.num = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
391	sq.flags = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
392	sq.dropped = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
393	sq.array = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
394	sq.prev_head = *sq.head;
395
396	{
397	const __u32 num = *sq.num;
398	for( i; num ) {
399	sq.sqes[i].opcode = IORING_OP_LAST;
400	sq.sqes[i].user_data = 3ul64;
401	}
402	}
403
404	(sq.submit_lock){};
405	(sq.release_lock){};
406
407	if( params_in.poller_submits \|\| params_in.eager_submits ) {
408	/* paranoid */ verify( is_pow2( params_in.num_ready ) \|\| (params_in.num_ready < 8) );
409	sq.ready_cnt = max( params_in.num_ready, 8 );
410	sq.ready = alloc( sq.ready_cnt, 64`align );
411	for(i; sq.ready_cnt) {
412	sq.ready[i] = -1ul32;
413	}
414	sq.prev_ready = 0;
415	}
416	else {
417	sq.ready_cnt = 0;
418	sq.ready = 0p;
419	sq.prev_ready = 0;
420	}
421
422	// completion queue
423	cq.head = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
424	cq.tail = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
425	cq.mask = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
426	cq.num = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
427	cq.overflow = ( __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
428	cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
429
430	// Step 4 : eventfd
431	int efd;
432	for() {
433	efd = eventfd(0, 0);
434	if (efd < 0) {
435	if (errno == EINTR) continue;
436	abort("KERNEL ERROR: IO_URING EVENTFD - %s\n", strerror(errno));
437	}
438	break;
439	}
440
441	int ret;
442	for() {
443	ret = syscall( __NR_io_uring_register, fd, IORING_REGISTER_EVENTFD, &efd, 1);
444	if (ret < 0) {
445	if (errno == EINTR) continue;
446	abort("KERNEL ERROR: IO_URING EVENTFD REGISTER - %s\n", strerror(errno));
447	}
448	break;
449	}
450
451	// some paranoid checks
452	/* 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 );
453	/* paranoid / verifyf( (cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
454	/* paranoid / verifyf( (cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
455	/* paranoid / verifyf( (cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
456
457	/* 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 );
458	/* paranoid / verifyf( (sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
459	/* paranoid / verifyf( (sq.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.head );
460	/* paranoid / verifyf( (sq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.tail );
461
462	// Update the global ring info
463	this.ring_flags = params.flags;
464	this.fd = fd;
465	this.efd = efd;
466	this.eager_submits = params_in.eager_submits;
467	this.poller_submits = params_in.poller_submits;
468	}
469
470	static void __io_destroy( __io_data & this ) {
471	// Shutdown the io rings
472	struct __submition_data & sq = this.submit_q;
473	struct __completion_data & cq = this.completion_q;
474
475	// unmap the submit queue entries
476	munmap(sq.sqes, (sq.num) sizeof(struct io_uring_sqe));
477
478	// unmap the Submit Queue ring
479	munmap(sq.ring_ptr, sq.ring_sz);
480
481	// unmap the Completion Queue ring, if it is different
482	if (cq.ring_ptr != sq.ring_ptr) {
483	munmap(cq.ring_ptr, cq.ring_sz);
484	}
485
486	// close the file descriptor
487	close(this.fd);
488	close(this.efd);
489
490	free( this.submit_q.ready ); // Maybe null, doesn't matter
491	}
492
493	//=============================================================================================
494	// I/O Context Sleep
495	//=============================================================================================
496
497	void __ioctx_register($io_ctx_thread & ctx, struct epoll_event & ev) {
498	ev.events = EPOLLIN \| EPOLLET \| EPOLLONESHOT;
499	ev.data.u64 = (__u64)&ctx;
500	int ret = epoll_ctl(iopoll.epollfd, EPOLL_CTL_ADD, ctx.ring->efd, &ev);
501	if (ret < 0) {
502	abort( "KERNEL ERROR: EPOLL ADD - (%d) %s\n", (int)errno, strerror(errno) );
503	}
504	}
505
506	void __ioctx_prepare_block($io_ctx_thread & ctx, struct epoll_event & ev) {
507	__cfadbg_print_safe(io_core, "Kernel I/O - epoll : Re-arming io poller %d (%p)\n", ctx.ring->fd, &ctx);
508	int ret = epoll_ctl(iopoll.epollfd, EPOLL_CTL_MOD, ctx.ring->efd, &ev);
509	if (ret < 0) {
510	abort( "KERNEL ERROR: EPOLL REARM - (%d) %s\n", (int)errno, strerror(errno) );
511	}
512	}
513
514	//=============================================================================================
515	// I/O Context Misc Setup
516	//=============================================================================================
517	void register_fixed_files( io_context & ctx, int * files, unsigned count ) {
518	int ret = syscall( __NR_io_uring_register, ctx.thrd.ring->fd, IORING_REGISTER_FILES, files, count );
519	if( ret < 0 ) {
520	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
521	}
522
523	__cfadbg_print_safe( io_core, "Kernel I/O : Performed io_register for %p, returned %d\n", active_thread(), ret );
524	}
525
526	void register_fixed_files( cluster & cltr, int * files, unsigned count ) {
527	for(i; cltr.io.cnt) {
528	register_fixed_files( cltr.io.ctxs[i], files, count );
529	}
530	}
531	#endif

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

Download in other formats:

Original Format