Context Navigation

setup.cfa @ 54cc2bf

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

Last change on this file since 54cc2bf was 7222630, checked in by Thierry Delisle <tdelisle@…>, 3 years ago
Small refector so only preemption cares about signal
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
193	eventfd_t v;
194	eventfd_read(io_ctx->ring->efd, &v);
195
196	post( io_ctx->sem );
197	}
198	}
199
200	__cfadbg_print_safe(io_core, "Kernel : IO poller thread stopping\n" );
201	unregister(&id);
202	return 0p;
203	}
204
205	//=============================================================================================
206	// I/O Context Constrution/Destruction
207	//=============================================================================================
208
209	void ?{}($io_ctx_thread & this, struct cluster & cl) { (this.self){ "IO Poller", cl }; }
210	void main( $io_ctx_thread & this );
211	static inline $thread * get_thread( $io_ctx_thread & this ) { return &this.self; }
212	void ^?{}( $io_ctx_thread & mutex this ) {}
213
214	static void __io_create ( __io_data & this, const io_context_params & params_in );
215	static void __io_destroy( __io_data & this );
216
217	void ?{}(io_context & this, struct cluster & cl, const io_context_params & params) {
218	(this.thrd){ cl };
219	this.thrd.ring = malloc();
220	__cfadbg_print_safe(io_core, "Kernel I/O : Creating ring for io_context %p\n", &this);
221	__io_create( *this.thrd.ring, params );
222
223	__cfadbg_print_safe(io_core, "Kernel I/O : Starting poller thread for io_context %p\n", &this);
224	this.thrd.done = false;
225	__thrd_start( this.thrd, main );
226
227	__cfadbg_print_safe(io_core, "Kernel I/O : io_context %p ready\n", &this);
228	}
229
230	void ?{}(io_context & this, struct cluster & cl) {
231	io_context_params params;
232	(this){ cl, params };
233	}
234
235	void ^?{}(io_context & this, bool cluster_context) {
236	__cfadbg_print_safe(io_core, "Kernel I/O : tearing down io_context %p\n", &this);
237
238	// Notify the thread of the shutdown
239	__atomic_store_n(&this.thrd.done, true, __ATOMIC_SEQ_CST);
240
241	// If this is an io_context within a cluster, things get trickier
242	$thread & thrd = this.thrd.self;
243	if( cluster_context ) {
244	// We are about to do weird things with the threads
245	// we don't need interrupts to complicate everything
246	disable_interrupts();
247
248	// Get cluster info
249	cluster & cltr = *thrd.curr_cluster;
250	/* paranoid */ verify( cltr.idles.total == 0 \|\| &cltr == mainCluster );
251	/* paranoid */ verify( !ready_mutate_islocked() );
252
253	// We need to adjust the clean-up based on where the thread is
254	if( thrd.state == Ready \|\| thrd.preempted != __NO_PREEMPTION ) {
255	// This is the tricky case
256	// The thread was preempted or ready to run and now it is on the ready queue
257	// but the cluster is shutting down, so there aren't any processors to run the ready queue
258	// the solution is to steal the thread from the ready-queue and pretend it was blocked all along
259
260	ready_schedule_lock();
261	// The thread should on the list
262	/* paranoid */ verify( thrd.link.next != 0p );
263
264	// Remove the thread from the ready queue of this cluster
265	// The thread should be the last on the list
266	__attribute__((unused)) bool removed = remove_head( &cltr, &thrd );
267	/* paranoid */ verify( removed );
268	thrd.link.next = 0p;
269	thrd.link.prev = 0p;
270
271	// Fixup the thread state
272	thrd.state = Blocked;
273	thrd.ticket = TICKET_BLOCKED;
274	thrd.preempted = __NO_PREEMPTION;
275
276	ready_schedule_unlock();
277
278	// Pretend like the thread was blocked all along
279	}
280	// !!! This is not an else if !!!
281	// Ok, now the thread is blocked (whether we cheated to get here or not)
282	if( thrd.state == Blocked ) {
283	// This is the "easy case"
284	// The thread is parked and can easily be moved to active cluster
285	verify( thrd.curr_cluster != active_cluster() \|\| thrd.curr_cluster == mainCluster );
286	thrd.curr_cluster = active_cluster();
287
288	// unpark the fast io_poller
289	unpark( &thrd );
290	}
291	else {
292	// The thread is in a weird state
293	// I don't know what to do here
294	abort("io_context poller thread is in unexpected state, cannot clean-up correctly\n");
295	}
296
297	// The weird thread kidnapping stuff is over, restore interrupts.
298	enable_interrupts( __cfaabi_dbg_ctx );
299	} else {
300	post( this.thrd.sem );
301	}
302
303	^(this.thrd){};
304	__cfadbg_print_safe(io_core, "Kernel I/O : Stopped poller thread for io_context %p\n", &this);
305
306	__io_destroy( *this.thrd.ring );
307	__cfadbg_print_safe(io_core, "Kernel I/O : Destroyed ring for io_context %p\n", &this);
308
309	free(this.thrd.ring);
310	}
311
312	void ^?{}(io_context & this) {
313	^(this){ false };
314	}
315
316	extern void __disable_interrupts_hard();
317	extern void __enable_interrupts_hard();
318
319	static void __io_create( __io_data & this, const io_context_params & params_in ) {
320	// Step 1 : call to setup
321	struct io_uring_params params;
322	memset(&params, 0, sizeof(params));
323	if( params_in.poll_submit ) params.flags \|= IORING_SETUP_SQPOLL;
324	if( params_in.poll_complete ) params.flags \|= IORING_SETUP_IOPOLL;
325
326	__u32 nentries = params_in.num_entries != 0 ? params_in.num_entries : 256;
327	if( !is_pow2(nentries) ) {
328	abort("ERROR: I/O setup 'num_entries' must be a power of 2\n");
329	}
330	if( params_in.poller_submits && params_in.eager_submits ) {
331	abort("ERROR: I/O setup 'poller_submits' and 'eager_submits' cannot be used together\n");
332	}
333
334	int fd = syscall(__NR_io_uring_setup, nentries, &params );
335	if(fd < 0) {
336	abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
337	}
338
339	// Step 2 : mmap result
340	memset( &this, 0, sizeof(struct __io_data) );
341	struct __submition_data & sq = this.submit_q;
342	struct __completion_data & cq = this.completion_q;
343
344	// calculate the right ring size
345	sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
346	cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
347
348	// Requires features
349	#if defined(IORING_FEAT_SINGLE_MMAP)
350	// adjust the size according to the parameters
351	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
352	cq.ring_sz = sq.ring_sz = max(cq.ring_sz, sq.ring_sz);
353	}
354	#endif
355
356	// mmap the Submit Queue into existence
357	sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQ_RING);
358	if (sq.ring_ptr == (void*)MAP_FAILED) {
359	abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
360	}
361
362	// Requires features
363	#if defined(IORING_FEAT_SINGLE_MMAP)
364	// mmap the Completion Queue into existence (may or may not be needed)
365	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
366	cq.ring_ptr = sq.ring_ptr;
367	}
368	else
369	#endif
370	{
371	// We need multiple call to MMAP
372	cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_CQ_RING);
373	if (cq.ring_ptr == (void*)MAP_FAILED) {
374	munmap(sq.ring_ptr, sq.ring_sz);
375	abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
376	}
377	}
378
379	// mmap the submit queue entries
380	size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
381	sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQES);
382	if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
383	munmap(sq.ring_ptr, sq.ring_sz);
384	if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
385	abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
386	}
387
388	// Step 3 : Initialize the data structure
389	// Get the pointers from the kernel to fill the structure
390	// submit queue
391	sq.head = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
392	sq.tail = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
393	sq.mask = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
394	sq.num = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
395	sq.flags = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
396	sq.dropped = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
397	sq.array = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
398	sq.prev_head = *sq.head;
399
400	{
401	const __u32 num = *sq.num;
402	for( i; num ) {
403	__sqe_clean( &sq.sqes[i] );
404	}
405	}
406
407	(sq.submit_lock){};
408	(sq.release_lock){};
409
410	if( params_in.poller_submits \|\| params_in.eager_submits ) {
411	/* paranoid */ verify( is_pow2( params_in.num_ready ) \|\| (params_in.num_ready < 8) );
412	sq.ready_cnt = max( params_in.num_ready, 8 );
413	sq.ready = alloc( sq.ready_cnt, 64`align );
414	for(i; sq.ready_cnt) {
415	sq.ready[i] = -1ul32;
416	}
417	sq.prev_ready = 0;
418	}
419	else {
420	sq.ready_cnt = 0;
421	sq.ready = 0p;
422	sq.prev_ready = 0;
423	}
424
425	// completion queue
426	cq.head = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
427	cq.tail = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
428	cq.mask = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
429	cq.num = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
430	cq.overflow = ( __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
431	cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
432
433	// Step 4 : eventfd
434	// io_uring_register is so f*cking slow on some machine that it
435	// will never succeed if preemption isn't hard blocked
436	__disable_interrupts_hard();
437
438	int efd = eventfd(0, 0);
439	if (efd < 0) {
440	abort("KERNEL ERROR: IO_URING EVENTFD - %s\n", strerror(errno));
441	}
442
443	int ret = syscall( __NR_io_uring_register, fd, IORING_REGISTER_EVENTFD, &efd, 1);
444	if (ret < 0) {
445	abort("KERNEL ERROR: IO_URING EVENTFD REGISTER - %s\n", strerror(errno));
446	}
447
448	__enable_interrupts_hard();
449
450	// some paranoid checks
451	/* 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 );
452	/* paranoid / verifyf( (cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
453	/* paranoid / verifyf( (cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
454	/* paranoid / verifyf( (cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
455
456	/* 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 );
457	/* paranoid / verifyf( (sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
458	/* paranoid / verifyf( (sq.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.head );
459	/* paranoid / verifyf( (sq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.tail );
460
461	// Update the global ring info
462	this.ring_flags = params.flags;
463	this.fd = fd;
464	this.efd = efd;
465	this.eager_submits = params_in.eager_submits;
466	this.poller_submits = params_in.poller_submits;
467	}
468
469	static void __io_destroy( __io_data & this ) {
470	// Shutdown the io rings
471	struct __submition_data & sq = this.submit_q;
472	struct __completion_data & cq = this.completion_q;
473
474	// unmap the submit queue entries
475	munmap(sq.sqes, (sq.num) sizeof(struct io_uring_sqe));
476
477	// unmap the Submit Queue ring
478	munmap(sq.ring_ptr, sq.ring_sz);
479
480	// unmap the Completion Queue ring, if it is different
481	if (cq.ring_ptr != sq.ring_ptr) {
482	munmap(cq.ring_ptr, cq.ring_sz);
483	}
484
485	// close the file descriptor
486	close(this.fd);
487	close(this.efd);
488
489	free( this.submit_q.ready ); // Maybe null, doesn't matter
490	}
491
492	//=============================================================================================
493	// I/O Context Sleep
494	//=============================================================================================
495	#define IOEVENTS EPOLLIN \| EPOLLONESHOT
496
497	static inline void __ioctx_epoll_ctl($io_ctx_thread & ctx, int op, const char * error) {
498	struct epoll_event ev;
499	ev.events = IOEVENTS;
500	ev.data.u64 = (__u64)&ctx;
501	int ret = epoll_ctl(iopoll.epollfd, op, ctx.ring->efd, &ev);
502	if (ret < 0) {
503	abort( "KERNEL ERROR: EPOLL %s - (%d) %s\n", error, (int)errno, strerror(errno) );
504	}
505	}
506
507	void __ioctx_register($io_ctx_thread & ctx) {
508	__ioctx_epoll_ctl(ctx, EPOLL_CTL_ADD, "ADD");
509	}
510
511	void __ioctx_prepare_block($io_ctx_thread & ctx) {
512	__cfadbg_print_safe(io_core, "Kernel I/O - epoll : Re-arming io poller %d (%p)\n", ctx.ring->fd, &ctx);
513	__ioctx_epoll_ctl(ctx, EPOLL_CTL_MOD, "REARM");
514	}
515
516	//=============================================================================================
517	// I/O Context Misc Setup
518	//=============================================================================================
519	void register_fixed_files( io_context & ctx, int * files, unsigned count ) {
520	int ret = syscall( __NR_io_uring_register, ctx.thrd.ring->fd, IORING_REGISTER_FILES, files, count );
521	if( ret < 0 ) {
522	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
523	}
524
525	__cfadbg_print_safe( io_core, "Kernel I/O : Performed io_register for %p, returned %d\n", active_thread(), ret );
526	}
527
528	void register_fixed_files( cluster & cltr, int * files, unsigned count ) {
529	for(i; cltr.io.cnt) {
530	register_fixed_files( cltr.io.ctxs[i], files, count );
531	}
532	}
533	#endif

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

Download in other formats:

Original Format