Context Navigation

setup.cfa@ 03f944f

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 03f944f was 8fc652e0, checked in by Thierry Delisle <tdelisle@…>, 5 years ago
Change usage of TLS to more strongly segregate in kernel and out of kernel usage.
Property mode set to `100644`
File size: 16.0 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	#include "io/types.hfa"
20	#include "kernel.hfa"
21
22	#if !defined(CFA_HAVE_LINUX_IO_URING_H)
23	void __kernel_io_startup() {
24	// Nothing to do without io_uring
25	}
26
27	void __kernel_io_shutdown() {
28	// Nothing to do without io_uring
29	}
30
31	void ?{}(io_context_params & this) {}
32
33	void ?{}(io_context & this, struct cluster & cl) {}
34	void ?{}(io_context & this, struct cluster & cl, const io_context_params & params) {}
35
36	void ^?{}(io_context & this) {}
37	void ^?{}(io_context & this, bool cluster_context) {}
38
39	#else
40	#include <errno.h>
41	#include <stdint.h>
42	#include <string.h>
43	#include <signal.h>
44	#include <unistd.h>
45
46	extern "C" {
47	#include <pthread.h>
48	#include <sys/epoll.h>
49	#include <sys/mman.h>
50	#include <sys/syscall.h>
51
52	#include <linux/io_uring.h>
53	}
54
55	#include "bitmanip.hfa"
56	#include "kernel_private.hfa"
57	#include "thread.hfa"
58
59	void ?{}(io_context_params & this) {
60	this.num_entries = 256;
61	this.num_ready = 256;
62	this.submit_aff = -1;
63	this.eager_submits = false;
64	this.poller_submits = false;
65	this.poll_submit = false;
66	this.poll_complete = false;
67	}
68
69	static void * __io_poller_slow( void * arg );
70
71	// Weirdly, some systems that do support io_uring don't actually define these
72	#ifdef __alpha__
73	/*
74	* alpha is the only exception, all other architectures
75	* have common numbers for new system calls.
76	*/
77	#ifndef __NR_io_uring_setup
78	#define __NR_io_uring_setup 535
79	#endif
80	#ifndef __NR_io_uring_enter
81	#define __NR_io_uring_enter 536
82	#endif
83	#ifndef __NR_io_uring_register
84	#define __NR_io_uring_register 537
85	#endif
86	#else /* !__alpha__ */
87	#ifndef __NR_io_uring_setup
88	#define __NR_io_uring_setup 425
89	#endif
90	#ifndef __NR_io_uring_enter
91	#define __NR_io_uring_enter 426
92	#endif
93	#ifndef __NR_io_uring_register
94	#define __NR_io_uring_register 427
95	#endif
96	#endif
97
98	//=============================================================================================
99	// I/O Startup / Shutdown logic + Master Poller
100	//=============================================================================================
101
102	// IO Master poller loop forward
103	static void * iopoll_loop( __attribute__((unused)) void * args );
104
105	static struct {
106	pthread_t thrd; // pthread handle to io poller thread
107	void * stack; // pthread stack for io poller thread
108	int epollfd; // file descriptor to the epoll instance
109	volatile bool run; // Whether or not to continue
110	} iopoll;
111
112	void __kernel_io_startup(void) {
113	__cfaabi_dbg_print_safe( "Kernel : Creating EPOLL instance\n" );
114
115	iopoll.epollfd = epoll_create1(0);
116	if (iopoll.epollfd == -1) {
117	abort( "internal error, epoll_create1\n");
118	}
119
120	__cfaabi_dbg_print_safe( "Kernel : Starting io poller thread\n" );
121
122	iopoll.run = true;
123	iopoll.stack = __create_pthread( &iopoll.thrd, iopoll_loop, 0p );
124	}
125
126	void __kernel_io_shutdown(void) {
127	// Notify the io poller thread of the shutdown
128	iopoll.run = false;
129	sigval val = { 1 };
130	pthread_sigqueue( iopoll.thrd, SIGUSR1, val );
131
132	// Wait for the io poller thread to finish
133
134	pthread_join( iopoll.thrd, 0p );
135	free( iopoll.stack );
136
137	int ret = close(iopoll.epollfd);
138	if (ret == -1) {
139	abort( "internal error, close epoll\n");
140	}
141
142	// Io polling is now fully stopped
143
144	__cfaabi_dbg_print_safe( "Kernel : IO poller stopped\n" );
145	}
146
147	static void * iopoll_loop( __attribute__((unused)) void * args ) {
148	__processor_id_t id;
149	id.full_proc = false;
150	id.id = doregister(&id);
151	__cfaabi_tls.this_proc_id = &id;
152	__cfaabi_dbg_print_safe( "Kernel : IO poller thread starting\n" );
153
154	// Block signals to control when they arrive
155	sigset_t mask;
156	sigfillset(&mask);
157	if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
158	abort( "internal error, pthread_sigmask" );
159	}
160
161	sigdelset( &mask, SIGUSR1 );
162
163	// Create sufficient events
164	struct epoll_event events[10];
165	// Main loop
166	while( iopoll.run ) {
167	// Wait for events
168	int nfds = epoll_pwait( iopoll.epollfd, events, 10, -1, &mask );
169
170	// Check if an error occured
171	if (nfds == -1) {
172	if( errno == EINTR ) continue;
173	abort( "internal error, pthread_sigmask" );
174	}
175
176	for(i; nfds) {
177	$io_ctx_thread * io_ctx = ($io_ctx_thread *)(uintptr_t)events[i].data.u64;
178	/* paranoid */ verify( io_ctx );
179	__cfadbg_print_safe(io_core, "Kernel I/O : Unparking io poller %p\n", io_ctx);
180	#if !defined( __CFA_NO_STATISTICS__ )
181	__cfaabi_tls.this_stats = io_ctx->self.curr_cluster->stats;
182	#endif
183	post( io_ctx->sem );
184	}
185	}
186
187	__cfaabi_dbg_print_safe( "Kernel : IO poller thread stopping\n" );
188	unregister(&id);
189	return 0p;
190	}
191
192	//=============================================================================================
193	// I/O Context Constrution/Destruction
194	//=============================================================================================
195
196	void ?{}($io_ctx_thread & this, struct cluster & cl) { (this.self){ "IO Poller", cl }; }
197	void main( $io_ctx_thread & this );
198	static inline $thread * get_thread( $io_ctx_thread & this ) { return &this.self; }
199	void ^?{}( $io_ctx_thread & mutex this ) {}
200
201	static void __io_create ( __io_data & this, const io_context_params & params_in );
202	static void __io_destroy( __io_data & this );
203
204	void ?{}(io_context & this, struct cluster & cl, const io_context_params & params) {
205	(this.thrd){ cl };
206	this.thrd.ring = malloc();
207	__cfadbg_print_safe(io_core, "Kernel I/O : Creating ring for io_context %p\n", &this);
208	__io_create( *this.thrd.ring, params );
209
210	__cfadbg_print_safe(io_core, "Kernel I/O : Starting poller thread for io_context %p\n", &this);
211	this.thrd.done = false;
212	__thrd_start( this.thrd, main );
213
214	__cfadbg_print_safe(io_core, "Kernel I/O : io_context %p ready\n", &this);
215	}
216
217	void ?{}(io_context & this, struct cluster & cl) {
218	io_context_params params;
219	(this){ cl, params };
220	}
221
222	void ^?{}(io_context & this, bool cluster_context) {
223	__cfadbg_print_safe(io_core, "Kernel I/O : tearing down io_context %p\n", &this);
224
225	// Notify the thread of the shutdown
226	__atomic_store_n(&this.thrd.done, true, __ATOMIC_SEQ_CST);
227
228	// If this is an io_context within a cluster, things get trickier
229	$thread & thrd = this.thrd.self;
230	if( cluster_context ) {
231	cluster & cltr = *thrd.curr_cluster;
232	/* paranoid */ verify( cltr.idles.total == 0 \|\| &cltr == mainCluster );
233	/* paranoid */ verify( !ready_mutate_islocked() );
234
235	// We need to adjust the clean-up based on where the thread is
236	if( thrd.state == Ready \|\| thrd.preempted != __NO_PREEMPTION ) {
237
238	ready_schedule_lock();
239
240	// This is the tricky case
241	// The thread was preempted and now it is on the ready queue
242	// The thread should be the last on the list
243	/* paranoid */ verify( thrd.link.next != 0p );
244
245	// Remove the thread from the ready queue of this cluster
246	__attribute__((unused)) bool removed = remove_head( &cltr, &thrd );
247	/* paranoid */ verify( removed );
248	thrd.link.next = 0p;
249	thrd.link.prev = 0p;
250
251	// Fixup the thread state
252	thrd.state = Blocked;
253	thrd.ticket = TICKET_BLOCKED;
254	thrd.preempted = __NO_PREEMPTION;
255
256	ready_schedule_unlock();
257
258	// Pretend like the thread was blocked all along
259	}
260	// !!! This is not an else if !!!
261	if( thrd.state == Blocked ) {
262
263	// This is the "easy case"
264	// The thread is parked and can easily be moved to active cluster
265	verify( thrd.curr_cluster != active_cluster() \|\| thrd.curr_cluster == mainCluster );
266	thrd.curr_cluster = active_cluster();
267
268	// unpark the fast io_poller
269	unpark( &thrd );
270	}
271	else {
272
273	// The thread is in a weird state
274	// I don't know what to do here
275	abort("io_context poller thread is in unexpected state, cannot clean-up correctly\n");
276	}
277	} else {
278	post( this.thrd.sem );
279	}
280
281	^(this.thrd){};
282	__cfadbg_print_safe(io_core, "Kernel I/O : Stopped poller thread for io_context %p\n", &this);
283
284	__io_destroy( *this.thrd.ring );
285	__cfadbg_print_safe(io_core, "Kernel I/O : Destroyed ring for io_context %p\n", &this);
286
287	free(this.thrd.ring);
288	}
289
290	void ^?{}(io_context & this) {
291	^(this){ false };
292	}
293
294	static void __io_create( __io_data & this, const io_context_params & params_in ) {
295	// Step 1 : call to setup
296	struct io_uring_params params;
297	memset(&params, 0, sizeof(params));
298	if( params_in.poll_submit ) params.flags \|= IORING_SETUP_SQPOLL;
299	if( params_in.poll_complete ) params.flags \|= IORING_SETUP_IOPOLL;
300
301	__u32 nentries = params_in.num_entries != 0 ? params_in.num_entries : 256;
302	if( !is_pow2(nentries) ) {
303	abort("ERROR: I/O setup 'num_entries' must be a power of 2\n");
304	}
305	if( params_in.poller_submits && params_in.eager_submits ) {
306	abort("ERROR: I/O setup 'poller_submits' and 'eager_submits' cannot be used together\n");
307	}
308
309	int fd = syscall(__NR_io_uring_setup, nentries, &params );
310	if(fd < 0) {
311	abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
312	}
313
314	// Step 2 : mmap result
315	memset( &this, 0, sizeof(struct __io_data) );
316	struct __submition_data & sq = this.submit_q;
317	struct __completion_data & cq = this.completion_q;
318
319	// calculate the right ring size
320	sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
321	cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
322
323	// Requires features
324	#if defined(IORING_FEAT_SINGLE_MMAP)
325	// adjust the size according to the parameters
326	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
327	cq.ring_sz = sq.ring_sz = max(cq.ring_sz, sq.ring_sz);
328	}
329	#endif
330
331	// mmap the Submit Queue into existence
332	sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQ_RING);
333	if (sq.ring_ptr == (void*)MAP_FAILED) {
334	abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
335	}
336
337	// Requires features
338	#if defined(IORING_FEAT_SINGLE_MMAP)
339	// mmap the Completion Queue into existence (may or may not be needed)
340	if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
341	cq.ring_ptr = sq.ring_ptr;
342	}
343	else
344	#endif
345	{
346	// We need multiple call to MMAP
347	cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_CQ_RING);
348	if (cq.ring_ptr == (void*)MAP_FAILED) {
349	munmap(sq.ring_ptr, sq.ring_sz);
350	abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
351	}
352	}
353
354	// mmap the submit queue entries
355	size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
356	sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ \| PROT_WRITE, MAP_SHARED \| MAP_POPULATE, fd, IORING_OFF_SQES);
357	if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
358	munmap(sq.ring_ptr, sq.ring_sz);
359	if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
360	abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
361	}
362
363	// Get the pointers from the kernel to fill the structure
364	// submit queue
365	sq.head = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
366	sq.tail = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
367	sq.mask = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
368	sq.num = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
369	sq.flags = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
370	sq.dropped = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
371	sq.array = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
372	sq.prev_head = *sq.head;
373
374	{
375	const __u32 num = *sq.num;
376	for( i; num ) {
377	sq.sqes[i].user_data = 0ul64;
378	}
379	}
380
381	(sq.submit_lock){};
382	(sq.release_lock){};
383
384	if( params_in.poller_submits \|\| params_in.eager_submits ) {
385	/* paranoid */ verify( is_pow2( params_in.num_ready ) \|\| (params_in.num_ready < 8) );
386	sq.ready_cnt = max( params_in.num_ready, 8 );
387	sq.ready = alloc( sq.ready_cnt, 64`align );
388	for(i; sq.ready_cnt) {
389	sq.ready[i] = -1ul32;
390	}
391	sq.prev_ready = 0;
392	}
393	else {
394	sq.ready_cnt = 0;
395	sq.ready = 0p;
396	sq.prev_ready = 0;
397	}
398
399	// completion queue
400	cq.head = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
401	cq.tail = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
402	cq.mask = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
403	cq.num = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
404	cq.overflow = ( __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
405	cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
406
407	// some paranoid checks
408	/* 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 );
409	/* paranoid / verifyf( (cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
410	/* paranoid / verifyf( (cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
411	/* paranoid / verifyf( (cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
412
413	/* 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 );
414	/* paranoid / verifyf( (sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
415	/* paranoid / verifyf( (sq.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.head );
416	/* paranoid / verifyf( (sq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.tail );
417
418	// Update the global ring info
419	this.ring_flags = params.flags;
420	this.fd = fd;
421	this.eager_submits = params_in.eager_submits;
422	this.poller_submits = params_in.poller_submits;
423	}
424
425	static void __io_destroy( __io_data & this ) {
426	// Shutdown the io rings
427	struct __submition_data & sq = this.submit_q;
428	struct __completion_data & cq = this.completion_q;
429
430	// unmap the submit queue entries
431	munmap(sq.sqes, (sq.num) sizeof(struct io_uring_sqe));
432
433	// unmap the Submit Queue ring
434	munmap(sq.ring_ptr, sq.ring_sz);
435
436	// unmap the Completion Queue ring, if it is different
437	if (cq.ring_ptr != sq.ring_ptr) {
438	munmap(cq.ring_ptr, cq.ring_sz);
439	}
440
441	// close the file descriptor
442	close(this.fd);
443
444	free( this.submit_q.ready ); // Maybe null, doesn't matter
445	}
446
447	//=============================================================================================
448	// I/O Context Sleep
449	//=============================================================================================
450
451	void __ioctx_register($io_ctx_thread & ctx, struct epoll_event & ev) {
452	ev.events = EPOLLIN \| EPOLLONESHOT;
453	ev.data.u64 = (__u64)&ctx;
454	int ret = epoll_ctl(iopoll.epollfd, EPOLL_CTL_ADD, ctx.ring->fd, &ev);
455	if (ret < 0) {
456	abort( "KERNEL ERROR: EPOLL ADD - (%d) %s\n", (int)errno, strerror(errno) );
457	}
458	}
459
460	void __ioctx_prepare_block($io_ctx_thread & ctx, struct epoll_event & ev) {
461	int ret = epoll_ctl(iopoll.epollfd, EPOLL_CTL_MOD, ctx.ring->fd, &ev);
462	if (ret < 0) {
463	abort( "KERNEL ERROR: EPOLL REARM - (%d) %s\n", (int)errno, strerror(errno) );
464	}
465	}
466
467	//=============================================================================================
468	// I/O Context Misc Setup
469	//=============================================================================================
470	void register_fixed_files( io_context & ctx, int * files, unsigned count ) {
471	int ret = syscall( __NR_io_uring_register, ctx.thrd.ring->fd, IORING_REGISTER_FILES, files, count );
472	if( ret < 0 ) {
473	abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
474	}
475
476	__cfadbg_print_safe( io_core, "Kernel I/O : Performed io_register for %p, returned %d\n", active_thread(), ret );
477	}
478
479	void register_fixed_files( cluster & cltr, int * files, unsigned count ) {
480	for(i; cltr.io.cnt) {
481	register_fixed_files( cltr.io.ctxs[i], files, count );
482	}
483	}
484	#endif

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

Download in other formats:

Original Format