[3e2b9c9] | 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 |
|
---|
[80444bb] | 19 | #if defined(__CFA_DEBUG__)
|
---|
| 20 | // #define __CFA_DEBUG_PRINT_IO__
|
---|
| 21 | // #define __CFA_DEBUG_PRINT_IO_CORE__
|
---|
| 22 | #endif
|
---|
| 23 |
|
---|
[3e2b9c9] | 24 | #include "io/types.hfa"
|
---|
[c44d652] | 25 | #include "kernel.hfa"
|
---|
[3e2b9c9] | 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 |
|
---|
[f277633e] | 36 | void ?{}(io_context_params & this) {}
|
---|
| 37 |
|
---|
[3e2b9c9] | 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 |
|
---|
[b7664a03] | 44 | void register_fixed_files( io_context &, int *, unsigned ) {}
|
---|
| 45 | void register_fixed_files( cluster &, int *, unsigned ) {}
|
---|
| 46 |
|
---|
[3e2b9c9] | 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>
|
---|
[426f60c] | 57 | #include <sys/eventfd.h>
|
---|
[3e2b9c9] | 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 {
|
---|
[4fc3343] | 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 | volatile bool stopped; // Whether the poller has finished running
|
---|
| 120 | volatile uint64_t epoch; // Epoch used for memory reclamation
|
---|
[3e2b9c9] | 121 | } iopoll;
|
---|
| 122 |
|
---|
| 123 | void __kernel_io_startup(void) {
|
---|
[80444bb] | 124 | __cfadbg_print_safe(io_core, "Kernel : Creating EPOLL instance\n" );
|
---|
[3e2b9c9] | 125 |
|
---|
| 126 | iopoll.epollfd = epoll_create1(0);
|
---|
| 127 | if (iopoll.epollfd == -1) {
|
---|
| 128 | abort( "internal error, epoll_create1\n");
|
---|
| 129 | }
|
---|
| 130 |
|
---|
[80444bb] | 131 | __cfadbg_print_safe(io_core, "Kernel : Starting io poller thread\n" );
|
---|
[3e2b9c9] | 132 |
|
---|
[4fc3343] | 133 | iopoll.stack = __create_pthread( &iopoll.thrd, iopoll_loop, 0p );
|
---|
| 134 | iopoll.run = true;
|
---|
| 135 | iopoll.stopped = false;
|
---|
| 136 | iopoll.epoch = 0;
|
---|
[3e2b9c9] | 137 | }
|
---|
| 138 |
|
---|
| 139 | void __kernel_io_shutdown(void) {
|
---|
| 140 | // Notify the io poller thread of the shutdown
|
---|
| 141 | iopoll.run = false;
|
---|
| 142 | sigval val = { 1 };
|
---|
| 143 | pthread_sigqueue( iopoll.thrd, SIGUSR1, val );
|
---|
| 144 |
|
---|
| 145 | // Wait for the io poller thread to finish
|
---|
| 146 |
|
---|
[bfcf6b9] | 147 | __destroy_pthread( iopoll.thrd, iopoll.stack, 0p );
|
---|
[3e2b9c9] | 148 |
|
---|
| 149 | int ret = close(iopoll.epollfd);
|
---|
| 150 | if (ret == -1) {
|
---|
| 151 | abort( "internal error, close epoll\n");
|
---|
| 152 | }
|
---|
| 153 |
|
---|
| 154 | // Io polling is now fully stopped
|
---|
| 155 |
|
---|
[80444bb] | 156 | __cfadbg_print_safe(io_core, "Kernel : IO poller stopped\n" );
|
---|
[3e2b9c9] | 157 | }
|
---|
| 158 |
|
---|
| 159 | static void * iopoll_loop( __attribute__((unused)) void * args ) {
|
---|
| 160 | __processor_id_t id;
|
---|
[58d64a4] | 161 | id.full_proc = false;
|
---|
[3e2b9c9] | 162 | id.id = doregister(&id);
|
---|
[8fc652e0] | 163 | __cfaabi_tls.this_proc_id = &id;
|
---|
[80444bb] | 164 | __cfadbg_print_safe(io_core, "Kernel : IO poller thread starting\n" );
|
---|
[3e2b9c9] | 165 |
|
---|
| 166 | // Block signals to control when they arrive
|
---|
| 167 | sigset_t mask;
|
---|
| 168 | sigfillset(&mask);
|
---|
| 169 | if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
|
---|
| 170 | abort( "internal error, pthread_sigmask" );
|
---|
| 171 | }
|
---|
| 172 |
|
---|
| 173 | sigdelset( &mask, SIGUSR1 );
|
---|
| 174 |
|
---|
| 175 | // Create sufficient events
|
---|
| 176 | struct epoll_event events[10];
|
---|
| 177 | // Main loop
|
---|
| 178 | while( iopoll.run ) {
|
---|
[ece0e80] | 179 | __cfadbg_print_safe(io_core, "Kernel I/O - epoll : waiting on io_uring contexts\n");
|
---|
| 180 |
|
---|
[d611995] | 181 | // increment the epoch to notify any deleters we are starting a new cycle
|
---|
| 182 | __atomic_fetch_add(&iopoll.epoch, 1, __ATOMIC_SEQ_CST);
|
---|
| 183 |
|
---|
[3e2b9c9] | 184 | // Wait for events
|
---|
| 185 | int nfds = epoll_pwait( iopoll.epollfd, events, 10, -1, &mask );
|
---|
| 186 |
|
---|
[ece0e80] | 187 | __cfadbg_print_safe(io_core, "Kernel I/O - epoll : %d io contexts events, waking up\n", nfds);
|
---|
| 188 |
|
---|
[3e2b9c9] | 189 | // Check if an error occured
|
---|
| 190 | if (nfds == -1) {
|
---|
| 191 | if( errno == EINTR ) continue;
|
---|
| 192 | abort( "internal error, pthread_sigmask" );
|
---|
| 193 | }
|
---|
| 194 |
|
---|
| 195 | for(i; nfds) {
|
---|
| 196 | $io_ctx_thread * io_ctx = ($io_ctx_thread *)(uintptr_t)events[i].data.u64;
|
---|
| 197 | /* paranoid */ verify( io_ctx );
|
---|
[426f60c] | 198 | __cfadbg_print_safe(io_core, "Kernel I/O - epoll : Unparking io poller %d (%p)\n", io_ctx->ring->fd, io_ctx);
|
---|
[3e2b9c9] | 199 | #if !defined( __CFA_NO_STATISTICS__ )
|
---|
[8fc652e0] | 200 | __cfaabi_tls.this_stats = io_ctx->self.curr_cluster->stats;
|
---|
[3e2b9c9] | 201 | #endif
|
---|
[d48b174] | 202 |
|
---|
| 203 | eventfd_t v;
|
---|
| 204 | eventfd_read(io_ctx->ring->efd, &v);
|
---|
| 205 |
|
---|
[e873838] | 206 | post( io_ctx->sem );
|
---|
[3e2b9c9] | 207 | }
|
---|
| 208 | }
|
---|
| 209 |
|
---|
[4fc3343] | 210 | __atomic_store_n(&iopoll.stopped, true, __ATOMIC_SEQ_CST);
|
---|
| 211 |
|
---|
[80444bb] | 212 | __cfadbg_print_safe(io_core, "Kernel : IO poller thread stopping\n" );
|
---|
[3e2b9c9] | 213 | unregister(&id);
|
---|
| 214 | return 0p;
|
---|
| 215 | }
|
---|
| 216 |
|
---|
| 217 | //=============================================================================================
|
---|
| 218 | // I/O Context Constrution/Destruction
|
---|
| 219 | //=============================================================================================
|
---|
| 220 |
|
---|
| 221 | void ?{}($io_ctx_thread & this, struct cluster & cl) { (this.self){ "IO Poller", cl }; }
|
---|
| 222 | void main( $io_ctx_thread & this );
|
---|
| 223 | static inline $thread * get_thread( $io_ctx_thread & this ) { return &this.self; }
|
---|
| 224 | void ^?{}( $io_ctx_thread & mutex this ) {}
|
---|
| 225 |
|
---|
| 226 | static void __io_create ( __io_data & this, const io_context_params & params_in );
|
---|
| 227 | static void __io_destroy( __io_data & this );
|
---|
| 228 |
|
---|
| 229 | void ?{}(io_context & this, struct cluster & cl, const io_context_params & params) {
|
---|
| 230 | (this.thrd){ cl };
|
---|
| 231 | this.thrd.ring = malloc();
|
---|
| 232 | __cfadbg_print_safe(io_core, "Kernel I/O : Creating ring for io_context %p\n", &this);
|
---|
| 233 | __io_create( *this.thrd.ring, params );
|
---|
| 234 |
|
---|
| 235 | __cfadbg_print_safe(io_core, "Kernel I/O : Starting poller thread for io_context %p\n", &this);
|
---|
| 236 | this.thrd.done = false;
|
---|
| 237 | __thrd_start( this.thrd, main );
|
---|
| 238 |
|
---|
| 239 | __cfadbg_print_safe(io_core, "Kernel I/O : io_context %p ready\n", &this);
|
---|
| 240 | }
|
---|
| 241 |
|
---|
| 242 | void ?{}(io_context & this, struct cluster & cl) {
|
---|
| 243 | io_context_params params;
|
---|
| 244 | (this){ cl, params };
|
---|
| 245 | }
|
---|
| 246 |
|
---|
| 247 | void ^?{}(io_context & this, bool cluster_context) {
|
---|
| 248 | __cfadbg_print_safe(io_core, "Kernel I/O : tearing down io_context %p\n", &this);
|
---|
| 249 |
|
---|
| 250 | // Notify the thread of the shutdown
|
---|
| 251 | __atomic_store_n(&this.thrd.done, true, __ATOMIC_SEQ_CST);
|
---|
| 252 |
|
---|
| 253 | // If this is an io_context within a cluster, things get trickier
|
---|
| 254 | $thread & thrd = this.thrd.self;
|
---|
| 255 | if( cluster_context ) {
|
---|
[ece0e80] | 256 | // We are about to do weird things with the threads
|
---|
| 257 | // we don't need interrupts to complicate everything
|
---|
| 258 | disable_interrupts();
|
---|
| 259 |
|
---|
| 260 | // Get cluster info
|
---|
[3e2b9c9] | 261 | cluster & cltr = *thrd.curr_cluster;
|
---|
[1eb239e4] | 262 | /* paranoid */ verify( cltr.idles.total == 0 || &cltr == mainCluster );
|
---|
[3e2b9c9] | 263 | /* paranoid */ verify( !ready_mutate_islocked() );
|
---|
| 264 |
|
---|
| 265 | // We need to adjust the clean-up based on where the thread is
|
---|
| 266 | if( thrd.state == Ready || thrd.preempted != __NO_PREEMPTION ) {
|
---|
[ece0e80] | 267 | // This is the tricky case
|
---|
| 268 | // The thread was preempted or ready to run and now it is on the ready queue
|
---|
| 269 | // but the cluster is shutting down, so there aren't any processors to run the ready queue
|
---|
| 270 | // the solution is to steal the thread from the ready-queue and pretend it was blocked all along
|
---|
[3e2b9c9] | 271 |
|
---|
[e873838] | 272 | ready_schedule_lock();
|
---|
[ece0e80] | 273 | // The thread should on the list
|
---|
[3e2b9c9] | 274 | /* paranoid */ verify( thrd.link.next != 0p );
|
---|
| 275 |
|
---|
| 276 | // Remove the thread from the ready queue of this cluster
|
---|
[ece0e80] | 277 | // The thread should be the last on the list
|
---|
[3e2b9c9] | 278 | __attribute__((unused)) bool removed = remove_head( &cltr, &thrd );
|
---|
| 279 | /* paranoid */ verify( removed );
|
---|
| 280 | thrd.link.next = 0p;
|
---|
| 281 | thrd.link.prev = 0p;
|
---|
| 282 |
|
---|
| 283 | // Fixup the thread state
|
---|
| 284 | thrd.state = Blocked;
|
---|
[6a77224] | 285 | thrd.ticket = TICKET_BLOCKED;
|
---|
[3e2b9c9] | 286 | thrd.preempted = __NO_PREEMPTION;
|
---|
| 287 |
|
---|
[e873838] | 288 | ready_schedule_unlock();
|
---|
[3e2b9c9] | 289 |
|
---|
| 290 | // Pretend like the thread was blocked all along
|
---|
| 291 | }
|
---|
| 292 | // !!! This is not an else if !!!
|
---|
[ece0e80] | 293 | // Ok, now the thread is blocked (whether we cheated to get here or not)
|
---|
[3e2b9c9] | 294 | if( thrd.state == Blocked ) {
|
---|
| 295 | // This is the "easy case"
|
---|
| 296 | // The thread is parked and can easily be moved to active cluster
|
---|
| 297 | verify( thrd.curr_cluster != active_cluster() || thrd.curr_cluster == mainCluster );
|
---|
| 298 | thrd.curr_cluster = active_cluster();
|
---|
| 299 |
|
---|
| 300 | // unpark the fast io_poller
|
---|
[0d72d45] | 301 | unpark( &thrd );
|
---|
[3e2b9c9] | 302 | }
|
---|
| 303 | else {
|
---|
| 304 | // The thread is in a weird state
|
---|
| 305 | // I don't know what to do here
|
---|
| 306 | abort("io_context poller thread is in unexpected state, cannot clean-up correctly\n");
|
---|
| 307 | }
|
---|
[ece0e80] | 308 |
|
---|
| 309 | // The weird thread kidnapping stuff is over, restore interrupts.
|
---|
| 310 | enable_interrupts( __cfaabi_dbg_ctx );
|
---|
[3e2b9c9] | 311 | } else {
|
---|
[3c80ccc] | 312 | post( this.thrd.sem );
|
---|
[3e2b9c9] | 313 | }
|
---|
| 314 |
|
---|
| 315 | ^(this.thrd){};
|
---|
| 316 | __cfadbg_print_safe(io_core, "Kernel I/O : Stopped poller thread for io_context %p\n", &this);
|
---|
| 317 |
|
---|
| 318 | __io_destroy( *this.thrd.ring );
|
---|
| 319 | __cfadbg_print_safe(io_core, "Kernel I/O : Destroyed ring for io_context %p\n", &this);
|
---|
| 320 |
|
---|
| 321 | free(this.thrd.ring);
|
---|
| 322 | }
|
---|
| 323 |
|
---|
| 324 | void ^?{}(io_context & this) {
|
---|
| 325 | ^(this){ false };
|
---|
| 326 | }
|
---|
| 327 |
|
---|
[7222630] | 328 | extern void __disable_interrupts_hard();
|
---|
| 329 | extern void __enable_interrupts_hard();
|
---|
[bb58825] | 330 |
|
---|
[3e2b9c9] | 331 | static void __io_create( __io_data & this, const io_context_params & params_in ) {
|
---|
| 332 | // Step 1 : call to setup
|
---|
| 333 | struct io_uring_params params;
|
---|
| 334 | memset(¶ms, 0, sizeof(params));
|
---|
| 335 | if( params_in.poll_submit ) params.flags |= IORING_SETUP_SQPOLL;
|
---|
| 336 | if( params_in.poll_complete ) params.flags |= IORING_SETUP_IOPOLL;
|
---|
| 337 |
|
---|
[4998155] | 338 | __u32 nentries = params_in.num_entries != 0 ? params_in.num_entries : 256;
|
---|
[63fe427c] | 339 | if( !is_pow2(nentries) ) {
|
---|
| 340 | abort("ERROR: I/O setup 'num_entries' must be a power of 2\n");
|
---|
| 341 | }
|
---|
| 342 | if( params_in.poller_submits && params_in.eager_submits ) {
|
---|
| 343 | abort("ERROR: I/O setup 'poller_submits' and 'eager_submits' cannot be used together\n");
|
---|
| 344 | }
|
---|
[3e2b9c9] | 345 |
|
---|
| 346 | int fd = syscall(__NR_io_uring_setup, nentries, ¶ms );
|
---|
| 347 | if(fd < 0) {
|
---|
| 348 | abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
|
---|
| 349 | }
|
---|
| 350 |
|
---|
| 351 | // Step 2 : mmap result
|
---|
| 352 | memset( &this, 0, sizeof(struct __io_data) );
|
---|
| 353 | struct __submition_data & sq = this.submit_q;
|
---|
| 354 | struct __completion_data & cq = this.completion_q;
|
---|
| 355 |
|
---|
| 356 | // calculate the right ring size
|
---|
| 357 | sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
|
---|
| 358 | cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
|
---|
| 359 |
|
---|
| 360 | // Requires features
|
---|
| 361 | #if defined(IORING_FEAT_SINGLE_MMAP)
|
---|
| 362 | // adjust the size according to the parameters
|
---|
| 363 | if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
|
---|
| 364 | cq.ring_sz = sq.ring_sz = max(cq.ring_sz, sq.ring_sz);
|
---|
| 365 | }
|
---|
| 366 | #endif
|
---|
| 367 |
|
---|
| 368 | // mmap the Submit Queue into existence
|
---|
| 369 | sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQ_RING);
|
---|
| 370 | if (sq.ring_ptr == (void*)MAP_FAILED) {
|
---|
| 371 | abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
|
---|
| 372 | }
|
---|
| 373 |
|
---|
| 374 | // Requires features
|
---|
| 375 | #if defined(IORING_FEAT_SINGLE_MMAP)
|
---|
| 376 | // mmap the Completion Queue into existence (may or may not be needed)
|
---|
| 377 | if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
|
---|
| 378 | cq.ring_ptr = sq.ring_ptr;
|
---|
| 379 | }
|
---|
| 380 | else
|
---|
| 381 | #endif
|
---|
| 382 | {
|
---|
| 383 | // We need multiple call to MMAP
|
---|
| 384 | cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_CQ_RING);
|
---|
| 385 | if (cq.ring_ptr == (void*)MAP_FAILED) {
|
---|
| 386 | munmap(sq.ring_ptr, sq.ring_sz);
|
---|
| 387 | abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
|
---|
| 388 | }
|
---|
| 389 | }
|
---|
| 390 |
|
---|
| 391 | // mmap the submit queue entries
|
---|
| 392 | size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
|
---|
| 393 | sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQES);
|
---|
| 394 | if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
|
---|
| 395 | munmap(sq.ring_ptr, sq.ring_sz);
|
---|
| 396 | if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
|
---|
| 397 | abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
|
---|
| 398 | }
|
---|
| 399 |
|
---|
[426f60c] | 400 | // Step 3 : Initialize the data structure
|
---|
[3e2b9c9] | 401 | // Get the pointers from the kernel to fill the structure
|
---|
| 402 | // submit queue
|
---|
[4998155] | 403 | sq.head = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
|
---|
| 404 | sq.tail = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
|
---|
| 405 | sq.mask = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
|
---|
| 406 | sq.num = ( const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
|
---|
| 407 | sq.flags = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
|
---|
| 408 | sq.dropped = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
|
---|
| 409 | sq.array = ( __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
|
---|
[3e2b9c9] | 410 | sq.prev_head = *sq.head;
|
---|
| 411 |
|
---|
| 412 | {
|
---|
[4998155] | 413 | const __u32 num = *sq.num;
|
---|
[3e2b9c9] | 414 | for( i; num ) {
|
---|
[ec19b21] | 415 | __sqe_clean( &sq.sqes[i] );
|
---|
[3e2b9c9] | 416 | }
|
---|
| 417 | }
|
---|
| 418 |
|
---|
[2fafe7e] | 419 | (sq.submit_lock){};
|
---|
[3e2b9c9] | 420 | (sq.release_lock){};
|
---|
| 421 |
|
---|
| 422 | if( params_in.poller_submits || params_in.eager_submits ) {
|
---|
| 423 | /* paranoid */ verify( is_pow2( params_in.num_ready ) || (params_in.num_ready < 8) );
|
---|
| 424 | sq.ready_cnt = max( params_in.num_ready, 8 );
|
---|
[ceb7db8] | 425 | sq.ready = alloc( sq.ready_cnt, 64`align );
|
---|
[3e2b9c9] | 426 | for(i; sq.ready_cnt) {
|
---|
| 427 | sq.ready[i] = -1ul32;
|
---|
| 428 | }
|
---|
[1095ccd] | 429 | sq.prev_ready = 0;
|
---|
[3e2b9c9] | 430 | }
|
---|
| 431 | else {
|
---|
| 432 | sq.ready_cnt = 0;
|
---|
| 433 | sq.ready = 0p;
|
---|
[1095ccd] | 434 | sq.prev_ready = 0;
|
---|
[3e2b9c9] | 435 | }
|
---|
| 436 |
|
---|
| 437 | // completion queue
|
---|
[4998155] | 438 | cq.head = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
|
---|
| 439 | cq.tail = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
|
---|
| 440 | cq.mask = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
|
---|
| 441 | cq.num = ( const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
|
---|
| 442 | cq.overflow = ( __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
|
---|
| 443 | cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
|
---|
[3e2b9c9] | 444 |
|
---|
[426f60c] | 445 | // Step 4 : eventfd
|
---|
[bb58825] | 446 | // io_uring_register is so f*cking slow on some machine that it
|
---|
| 447 | // will never succeed if preemption isn't hard blocked
|
---|
[7222630] | 448 | __disable_interrupts_hard();
|
---|
[bb58825] | 449 |
|
---|
| 450 | int efd = eventfd(0, 0);
|
---|
| 451 | if (efd < 0) {
|
---|
| 452 | abort("KERNEL ERROR: IO_URING EVENTFD - %s\n", strerror(errno));
|
---|
[426f60c] | 453 | }
|
---|
| 454 |
|
---|
[bb58825] | 455 | int ret = syscall( __NR_io_uring_register, fd, IORING_REGISTER_EVENTFD, &efd, 1);
|
---|
| 456 | if (ret < 0) {
|
---|
| 457 | abort("KERNEL ERROR: IO_URING EVENTFD REGISTER - %s\n", strerror(errno));
|
---|
[426f60c] | 458 | }
|
---|
| 459 |
|
---|
[7222630] | 460 | __enable_interrupts_hard();
|
---|
[bb58825] | 461 |
|
---|
[3e2b9c9] | 462 | // some paranoid checks
|
---|
| 463 | /* 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 );
|
---|
| 464 | /* paranoid */ verifyf( (*cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
|
---|
| 465 | /* paranoid */ verifyf( (*cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
|
---|
| 466 | /* paranoid */ verifyf( (*cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
|
---|
| 467 |
|
---|
| 468 | /* 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 );
|
---|
| 469 | /* paranoid */ verifyf( (*sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
|
---|
| 470 | /* paranoid */ verifyf( (*sq.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.head );
|
---|
| 471 | /* paranoid */ verifyf( (*sq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.tail );
|
---|
| 472 |
|
---|
| 473 | // Update the global ring info
|
---|
| 474 | this.ring_flags = params.flags;
|
---|
| 475 | this.fd = fd;
|
---|
[426f60c] | 476 | this.efd = efd;
|
---|
[3e2b9c9] | 477 | this.eager_submits = params_in.eager_submits;
|
---|
| 478 | this.poller_submits = params_in.poller_submits;
|
---|
| 479 | }
|
---|
| 480 |
|
---|
| 481 | static void __io_destroy( __io_data & this ) {
|
---|
| 482 | // Shutdown the io rings
|
---|
| 483 | struct __submition_data & sq = this.submit_q;
|
---|
| 484 | struct __completion_data & cq = this.completion_q;
|
---|
| 485 |
|
---|
| 486 | // unmap the submit queue entries
|
---|
| 487 | munmap(sq.sqes, (*sq.num) * sizeof(struct io_uring_sqe));
|
---|
| 488 |
|
---|
| 489 | // unmap the Submit Queue ring
|
---|
| 490 | munmap(sq.ring_ptr, sq.ring_sz);
|
---|
| 491 |
|
---|
| 492 | // unmap the Completion Queue ring, if it is different
|
---|
| 493 | if (cq.ring_ptr != sq.ring_ptr) {
|
---|
| 494 | munmap(cq.ring_ptr, cq.ring_sz);
|
---|
| 495 | }
|
---|
| 496 |
|
---|
| 497 | // close the file descriptor
|
---|
| 498 | close(this.fd);
|
---|
[426f60c] | 499 | close(this.efd);
|
---|
[3e2b9c9] | 500 |
|
---|
| 501 | free( this.submit_q.ready ); // Maybe null, doesn't matter
|
---|
| 502 | }
|
---|
| 503 |
|
---|
| 504 | //=============================================================================================
|
---|
| 505 | // I/O Context Sleep
|
---|
| 506 | //=============================================================================================
|
---|
[d48b174] | 507 | static inline void __ioctx_epoll_ctl($io_ctx_thread & ctx, int op, const char * error) {
|
---|
| 508 | struct epoll_event ev;
|
---|
[d611995] | 509 | ev.events = EPOLLIN | EPOLLONESHOT;
|
---|
[4998155] | 510 | ev.data.u64 = (__u64)&ctx;
|
---|
[d48b174] | 511 | int ret = epoll_ctl(iopoll.epollfd, op, ctx.ring->efd, &ev);
|
---|
[3e2b9c9] | 512 | if (ret < 0) {
|
---|
[d48b174] | 513 | abort( "KERNEL ERROR: EPOLL %s - (%d) %s\n", error, (int)errno, strerror(errno) );
|
---|
[3e2b9c9] | 514 | }
|
---|
| 515 | }
|
---|
| 516 |
|
---|
[d48b174] | 517 | void __ioctx_register($io_ctx_thread & ctx) {
|
---|
| 518 | __ioctx_epoll_ctl(ctx, EPOLL_CTL_ADD, "ADD");
|
---|
| 519 | }
|
---|
| 520 |
|
---|
| 521 | void __ioctx_prepare_block($io_ctx_thread & ctx) {
|
---|
[426f60c] | 522 | __cfadbg_print_safe(io_core, "Kernel I/O - epoll : Re-arming io poller %d (%p)\n", ctx.ring->fd, &ctx);
|
---|
[d48b174] | 523 | __ioctx_epoll_ctl(ctx, EPOLL_CTL_MOD, "REARM");
|
---|
[3e2b9c9] | 524 | }
|
---|
| 525 |
|
---|
[d611995] | 526 | void __ioctx_unregister($io_ctx_thread & ctx) {
|
---|
| 527 | // Read the current epoch so we know when to stop
|
---|
| 528 | size_t curr = __atomic_load_n(&iopoll.epoch, __ATOMIC_SEQ_CST);
|
---|
| 529 |
|
---|
| 530 | // Remove the fd from the iopoller
|
---|
| 531 | __ioctx_epoll_ctl(ctx, EPOLL_CTL_DEL, "REMOVE");
|
---|
| 532 |
|
---|
| 533 | // Notify the io poller thread of the shutdown
|
---|
| 534 | iopoll.run = false;
|
---|
| 535 | sigval val = { 1 };
|
---|
| 536 | pthread_sigqueue( iopoll.thrd, SIGUSR1, val );
|
---|
| 537 |
|
---|
| 538 | // Make sure all this is done
|
---|
| 539 | __atomic_thread_fence(__ATOMIC_SEQ_CST);
|
---|
| 540 |
|
---|
| 541 | // Wait for the next epoch
|
---|
[4fc3343] | 542 | while(curr == iopoll.epoch && !iopoll.stopped) Pause();
|
---|
[d611995] | 543 | }
|
---|
| 544 |
|
---|
[3e2b9c9] | 545 | //=============================================================================================
|
---|
| 546 | // I/O Context Misc Setup
|
---|
| 547 | //=============================================================================================
|
---|
| 548 | void register_fixed_files( io_context & ctx, int * files, unsigned count ) {
|
---|
| 549 | int ret = syscall( __NR_io_uring_register, ctx.thrd.ring->fd, IORING_REGISTER_FILES, files, count );
|
---|
| 550 | if( ret < 0 ) {
|
---|
[2fab24e3] | 551 | abort( "KERNEL ERROR: IO_URING REGISTER - (%d) %s\n", (int)errno, strerror(errno) );
|
---|
[3e2b9c9] | 552 | }
|
---|
| 553 |
|
---|
| 554 | __cfadbg_print_safe( io_core, "Kernel I/O : Performed io_register for %p, returned %d\n", active_thread(), ret );
|
---|
| 555 | }
|
---|
| 556 |
|
---|
| 557 | void register_fixed_files( cluster & cltr, int * files, unsigned count ) {
|
---|
| 558 | for(i; cltr.io.cnt) {
|
---|
| 559 | register_fixed_files( cltr.io.ctxs[i], files, count );
|
---|
| 560 | }
|
---|
| 561 | }
|
---|
[c44d652] | 562 | #endif
|
---|