Context Navigation

setup.cfa@ b57db73

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 b57db73 was b7664a03, checked in by Thierry Delisle <tdelisle@…>, 5 years ago
Removed some warnings and allowed io calls to compile without io_uring.
Property mode set to `100644`
File size: 18.0 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	} iopoll;
120
121	void __kernel_io_startup(void) {
122	__cfadbg_print_safe(io_core, "Kernel : Creating EPOLL instance\n" );
123
124	iopoll.epollfd = epoll_create1(0);
125	if (iopoll.epollfd == -1) {
126	abort( "internal error, epoll_create1\n");
127	}
128
129	__cfadbg_print_safe(io_core, "Kernel : Starting io poller thread\n" );
130
131	iopoll.run = true;
132	iopoll.stack = __create_pthread( &iopoll.thrd, iopoll_loop, 0p );
133	}
134
135	void __kernel_io_shutdown(void) {
136	// Notify the io poller thread of the shutdown
137	iopoll.run = false;
138	sigval val = { 1 };
139	pthread_sigqueue( iopoll.thrd, SIGUSR1, val );
140
141	// Wait for the io poller thread to finish
142
143	__destroy_pthread( iopoll.thrd, iopoll.stack, 0p );
144
145	int ret = close(iopoll.epollfd);
146	if (ret == -1) {
147	abort( "internal error, close epoll\n");
148	}
149
150	// Io polling is now fully stopped
151
152	__cfadbg_print_safe(io_core, "Kernel : IO poller stopped\n" );
153	}
154
155	static void * iopoll_loop( __attribute__((unused)) void * args ) {
156	__processor_id_t id;
157	id.full_proc = false;
158	id.id = doregister(&id);
159	__cfaabi_tls.this_proc_id = &id;
160	__cfadbg_print_safe(io_core, "Kernel : IO poller thread starting\n" );
161
162	// Block signals to control when they arrive
163	sigset_t mask;
164	sigfillset(&mask);
165	if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
166	abort( "internal error, pthread_sigmask" );
167	}
168
169	sigdelset( &mask, SIGUSR1 );
170
171	// Create sufficient events
172	struct epoll_event events[10];
173	// Main loop
174	while( iopoll.run ) {
175	__cfadbg_print_safe(io_core, "Kernel I/O - epoll : waiting on io_uring contexts\n");
176
177	// Wait for events
178	int nfds = epoll_pwait( iopoll.epollfd, events, 10, -1, &mask );
179
180	__cfadbg_print_safe(io_core, "Kernel I/O - epoll : %d io contexts events, waking up\n", nfds);
181
182	// Check if an error occured
183	if (nfds == -1) {
184	if( errno == EINTR ) continue;
185	abort( "internal error, pthread_sigmask" );
186	}
187
188	for(i; nfds) {
189	$io_ctx_thread * io_ctx = ($io_ctx_thread *)(uintptr_t)events[i].data.u64;
190	/* paranoid */ verify( io_ctx );
191	__cfadbg_print_safe(io_core, "Kernel I/O - epoll : Unparking io poller %d (%p)\n", io_ctx->ring->fd, io_ctx);
192	#if !defined( __CFA_NO_STATISTICS__ )
193	__cfaabi_tls.this_stats = io_ctx->self.curr_cluster->stats;
194	#endif
195
196	eventfd_t v;
197	eventfd_read(io_ctx->ring->efd, &v);
198
199	post( io_ctx->sem );
200	}
201	}
202
203	__cfadbg_print_safe(io_core, "Kernel : IO poller thread stopping\n" );
204	unregister(&id);
205	return 0p;
206	}
207
208	//=============================================================================================
209	// I/O Context Constrution/Destruction
210	//=============================================================================================
211
212	void ?{}($io_ctx_thread & this, struct cluster & cl) { (this.self){ "IO Poller", cl }; }
213	void main( $io_ctx_thread & this );
214	static inline $thread * get_thread( $io_ctx_thread & this ) { return &this.self; }
215	void ^?{}( $io_ctx_thread & mutex this ) {}
216
217	static void __io_create ( __io_data & this, const io_context_params & params_in );
218	static void __io_destroy( __io_data & this );
219
220	void ?{}(io_context & this, struct cluster & cl, const io_context_params & params) {
221	(this.thrd){ cl };
222	this.thrd.ring = malloc();
223	__cfadbg_print_safe(io_core, "Kernel I/O : Creating ring for io_context %p\n", &this);
224	__io_create( *this.thrd.ring, params );
225
226	__cfadbg_print_safe(io_core, "Kernel I/O : Starting poller thread for io_context %p\n", &this);
227	this.thrd.done = false;
228	__thrd_start( this.thrd, main );
229
230	__cfadbg_print_safe(io_core, "Kernel I/O : io_context %p ready\n", &this);
231	}
232
233	void ?{}(io_context & this, struct cluster & cl) {
234	io_context_params params;
235	(this){ cl, params };
236	}
237
238	void ^?{}(io_context & this, bool cluster_context) {
239	__cfadbg_print_safe(io_core, "Kernel I/O : tearing down io_context %p\n", &this);
240
241	// Notify the thread of the shutdown
242	__atomic_store_n(&this.thrd.done, true, __ATOMIC_SEQ_CST);
243
244	// If this is an io_context within a cluster, things get trickier
245	$thread & thrd = this.thrd.self;
246	if( cluster_context ) {
247	// We are about to do weird things with the threads
248	// we don't need interrupts to complicate everything
249	disable_interrupts();
250
251	// Get cluster info
252	cluster & cltr = *thrd.curr_cluster;
253	/* paranoid */ verify( cltr.idles.total == 0 \|\| &cltr == mainCluster );
254	/* paranoid */ verify( !ready_mutate_islocked() );
255
256	// We need to adjust the clean-up based on where the thread is
257	if( thrd.state == Ready \|\| thrd.preempted != __NO_PREEMPTION ) {
258	// This is the tricky case
259	// The thread was preempted or ready to run and now it is on the ready queue
260	// but the cluster is shutting down, so there aren't any processors to run the ready queue
261	// the solution is to steal the thread from the ready-queue and pretend it was blocked all along
262
263	ready_schedule_lock();
264	// The thread should on the list
265	/* paranoid */ verify( thrd.link.next != 0p );
266
267	// Remove the thread from the ready queue of this cluster
268	// The thread should be the last on the list
269	__attribute__((unused)) bool removed = remove_head( &cltr, &thrd );
270	/* paranoid */ verify( removed );
271	thrd.link.next = 0p;
272	thrd.link.prev = 0p;
273
274	// Fixup the thread state
275	thrd.state = Blocked;
276	thrd.ticket = TICKET_BLOCKED;
277	thrd.preempted = __NO_PREEMPTION;
278
279	ready_schedule_unlock();
280
281	// Pretend like the thread was blocked all along
282	}
283	// !!! This is not an else if !!!
284	// Ok, now the thread is blocked (whether we cheated to get here or not)
285	if( thrd.state == Blocked ) {
286	// This is the "easy case"
287	// The thread is parked and can easily be moved to active cluster
288	verify( thrd.curr_cluster != active_cluster() \|\| thrd.curr_cluster == mainCluster );
289	thrd.curr_cluster = active_cluster();
290
291	// unpark the fast io_poller
292	unpark( &thrd );
293	}
294	else {
295	// The thread is in a weird state
296	// I don't know what to do here
297	abort("io_context poller thread is in unexpected state, cannot clean-up correctly\n");
298	}
299
300	// The weird thread kidnapping stuff is over, restore interrupts.
301	enable_interrupts( __cfaabi_dbg_ctx );
302	} else {
303	post( this.thrd.sem );
304	}
305
306	^(this.thrd){};
307	__cfadbg_print_safe(io_core, "Kernel I/O : Stopped poller thread for io_context %p\n", &this);
308
309	__io_destroy( *this.thrd.ring );
310	__cfadbg_print_safe(io_core, "Kernel I/O : Destroyed ring for io_context %p\n", &this);
311
312	free(this.thrd.ring);
313	}
314
315	void ^?{}(io_context & this) {
316	^(this){ false };
317	}
318
319	extern void __disable_interrupts_hard();
320	extern void __enable_interrupts_hard();
321
322	static void __io_create( __io_data & this, const io_context_params & params_in ) {
323	// Step 1 : call to setup
324	struct io_uring_params params;
325	memset(&params, 0, sizeof(params));
326	if( params_in.poll_submit ) params.flags \|= IORING_SETUP_SQPOLL;
327	if( params_in.poll_complete ) params.flags \|= IORING_SETUP_IOPOLL;
328
329	__u32 nentries = params_in.num_entries != 0 ? params_in.num_entries : 256;
330	if( !is_pow2(nentries) ) {
331	abort("ERROR: I/O setup 'num_entries' must be a power of 2\n");
332	}
333	if( params_in.poller_submits && params_in.eager_submits ) {
334	abort("ERROR: I/O setup 'poller_submits' and 'eager_submits' cannot be used together\n");
335	}
336
337	int fd = syscall(__NR_io_uring_setup, nentries, &params );
338	if(fd < 0) {
339	abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
340	}
341
342	// Step 2 : mmap result
343	memset( &this, 0, sizeof(struct __io_data) );
344	struct __submition_data & sq = this.submit_q;
345	struct __completion_data & cq = this.completion_q;
346
347	// calculate the right ring size
348	sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
349	cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
350
351	// Requires features
352	#if defined(IORING_FEAT_SINGLE_MMAP)
353	// adjust the size according to the parameters
354	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
355	cq.ring_sz = sq.ring_sz = max(cq.ring_sz, sq.ring_sz);
356	}
357	#endif
358
359	// mmap the Submit Queue into existence
360	sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQ_RING);
361	if (sq.ring_ptr == (void*)MAP_FAILED) {
362	abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
363	}
364
365	// Requires features
366	#if defined(IORING_FEAT_SINGLE_MMAP)
367	// mmap the Completion Queue into existence (may or may not be needed)
368	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
369	cq.ring_ptr = sq.ring_ptr;
370	}
371	else
372	#endif
373	{
374	// We need multiple call to MMAP
375	cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_CQ_RING);
376	if (cq.ring_ptr == (void*)MAP_FAILED) {
377	munmap(sq.ring_ptr, sq.ring_sz);
378	abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
379	}
380	}
381
382	// mmap the submit queue entries
383	size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
384	sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQES);
385	if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
386	munmap(sq.ring_ptr, sq.ring_sz);
387	if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
388	abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
389	}
390
391	// Step 3 : Initialize the data structure
392	// Get the pointers from the kernel to fill the structure
393	// submit queue
394	sq.head = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
395	sq.tail = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
396	sq.mask = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
397	sq.num = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
398	sq.flags = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
399	sq.dropped = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
400	sq.array = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
401	sq.prev_head = *sq.head;
402
403	{
404	const __u32 num = *sq.num;
405	for( i; num ) {
406	__sqe_clean( &sq.sqes[i] );
407	}
408	}
409
410	(sq.submit_lock){};
411	(sq.release_lock){};
412
413	if( params_in.poller_submits \|\| params_in.eager_submits ) {
414	/* paranoid */ verify( is_pow2( params_in.num_ready ) \|\| (params_in.num_ready < 8) );
415	sq.ready_cnt = max( params_in.num_ready, 8 );
416	sq.ready = alloc( sq.ready_cnt, 64`align );
417	for(i; sq.ready_cnt) {
418	sq.ready[i] = -1ul32;
419	}
420	sq.prev_ready = 0;
421	}
422	else {
423	sq.ready_cnt = 0;
424	sq.ready = 0p;
425	sq.prev_ready = 0;
426	}
427
428	// completion queue
429	cq.head = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
430	cq.tail = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
431	cq.mask = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
432	cq.num = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
433	cq.overflow = ( __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
434	cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
435
436	// Step 4 : eventfd
437	// io_uring_register is so f*cking slow on some machine that it
438	// will never succeed if preemption isn't hard blocked
439	__disable_interrupts_hard();
440
441	int efd = eventfd(0, 0);
442	if (efd < 0) {
443	abort("KERNEL ERROR: IO_URING EVENTFD - %s\n", strerror(errno));
444	}
445
446	int ret = syscall( __NR_io_uring_register, fd, IORING_REGISTER_EVENTFD, &efd, 1);
447	if (ret < 0) {
448	abort("KERNEL ERROR: IO_URING EVENTFD REGISTER - %s\n", strerror(errno));
449	}
450
451	__enable_interrupts_hard();
452
453	// some paranoid checks
454	/* 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 );
455	/* paranoid / verifyf( (cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
456	/* paranoid / verifyf( (cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
457	/* paranoid / verifyf( (cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
458
459	/* 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 );
460	/* paranoid / verifyf( (sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
461	/* paranoid / verifyf( (sq.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.head );
462	/* paranoid / verifyf( (sq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.tail );
463
464	// Update the global ring info
465	this.ring_flags = params.flags;
466	this.fd = fd;
467	this.efd = efd;
468	this.eager_submits = params_in.eager_submits;
469	this.poller_submits = params_in.poller_submits;
470	}
471
472	static void __io_destroy( __io_data & this ) {
473	// Shutdown the io rings
474	struct __submition_data & sq = this.submit_q;
475	struct __completion_data & cq = this.completion_q;
476
477	// unmap the submit queue entries
478	munmap(sq.sqes, (sq.num) sizeof(struct io_uring_sqe));
479
480	// unmap the Submit Queue ring
481	munmap(sq.ring_ptr, sq.ring_sz);
482
483	// unmap the Completion Queue ring, if it is different
484	if (cq.ring_ptr != sq.ring_ptr) {
485	munmap(cq.ring_ptr, cq.ring_sz);
486	}
487
488	// close the file descriptor
489	close(this.fd);
490	close(this.efd);
491
492	free( this.submit_q.ready ); // Maybe null, doesn't matter
493	}
494
495	//=============================================================================================
496	// I/O Context Sleep
497	//=============================================================================================
498	#define IOEVENTS EPOLLIN \| EPOLLONESHOT
499
500	static inline void __ioctx_epoll_ctl($io_ctx_thread & ctx, int op, const char * error) {
501	struct epoll_event ev;
502	ev.events = IOEVENTS;
503	ev.data.u64 = (__u64)&ctx;
504	int ret = epoll_ctl(iopoll.epollfd, op, ctx.ring->efd, &ev);
505	if (ret < 0) {
506	abort( "KERNEL ERROR: EPOLL %s - (%d) %s\n", error, (int)errno, strerror(errno) );
507	}
508	}
509
510	void __ioctx_register($io_ctx_thread & ctx) {
511	__ioctx_epoll_ctl(ctx, EPOLL_CTL_ADD, "ADD");
512	}
513
514	void __ioctx_prepare_block($io_ctx_thread & ctx) {
515	__cfadbg_print_safe(io_core, "Kernel I/O - epoll : Re-arming io poller %d (%p)\n", ctx.ring->fd, &ctx);
516	__ioctx_epoll_ctl(ctx, EPOLL_CTL_MOD, "REARM");
517	}
518
519	//=============================================================================================
520	// I/O Context Misc Setup
521	//=============================================================================================
522	void register_fixed_files( io_context & ctx, int * files, unsigned count ) {
523	int ret = syscall( __NR_io_uring_register, ctx.thrd.ring->fd, IORING_REGISTER_FILES, files, count );
524	if( ret < 0 ) {
525	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
526	}
527
528	__cfadbg_print_safe( io_core, "Kernel I/O : Performed io_register for %p, returned %d\n", active_thread(), ret );
529	}
530
531	void register_fixed_files( cluster & cltr, int * files, unsigned count ) {
532	for(i; cltr.io.cnt) {
533	register_fixed_files( cltr.io.ctxs[i], files, count );
534	}
535	}
536	#endif

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

Download in other formats:

Original Format