[92976d9] | 1 | #include "kernel.hfa"
|
---|
| 2 |
|
---|
| 3 | #if !defined(HAVE_LINUX_IO_URING_H)
|
---|
| 4 | void __kernel_io_startup( cluster & this ) {
|
---|
| 5 | // Nothing to do without io_uring
|
---|
| 6 | }
|
---|
| 7 |
|
---|
| 8 | void __kernel_io_shutdown( cluster & this ) {
|
---|
| 9 | // Nothing to do without io_uring
|
---|
| 10 | }
|
---|
| 11 |
|
---|
| 12 | bool is_async( void (*)() ) {
|
---|
| 13 | return false;
|
---|
| 14 | }
|
---|
| 15 |
|
---|
| 16 | #else
|
---|
| 17 | extern "C" {
|
---|
| 18 | #define _GNU_SOURCE /* See feature_test_macros(7) */
|
---|
| 19 | #include <errno.h>
|
---|
| 20 | #include <stdint.h>
|
---|
| 21 | #include <string.h>
|
---|
| 22 | #include <unistd.h>
|
---|
| 23 | #include <sys/mman.h>
|
---|
| 24 | #include <sys/syscall.h>
|
---|
| 25 |
|
---|
| 26 | #include <linux/io_uring.h>
|
---|
| 27 | }
|
---|
| 28 |
|
---|
| 29 | #include "bits/signal.hfa"
|
---|
| 30 | #include "kernel_private.hfa"
|
---|
| 31 | #include "thread.hfa"
|
---|
| 32 |
|
---|
| 33 | uint32_t entries_per_cluster() {
|
---|
| 34 | return 256;
|
---|
| 35 | }
|
---|
| 36 |
|
---|
| 37 | static void * __io_poller( void * arg );
|
---|
| 38 |
|
---|
[185efe6] | 39 | // Weirdly, some systems that do support io_uring don't actually define these
|
---|
| 40 | #ifdef __alpha__
|
---|
| 41 | /*
|
---|
| 42 | * alpha is the only exception, all other architectures
|
---|
| 43 | * have common numbers for new system calls.
|
---|
| 44 | */
|
---|
| 45 | # ifndef __NR_io_uring_setup
|
---|
| 46 | # define __NR_io_uring_setup 535
|
---|
| 47 | # endif
|
---|
| 48 | # ifndef __NR_io_uring_enter
|
---|
| 49 | # define __NR_io_uring_enter 536
|
---|
| 50 | # endif
|
---|
| 51 | # ifndef __NR_io_uring_register
|
---|
| 52 | # define __NR_io_uring_register 537
|
---|
| 53 | # endif
|
---|
| 54 | #else /* !__alpha__ */
|
---|
| 55 | # ifndef __NR_io_uring_setup
|
---|
| 56 | # define __NR_io_uring_setup 425
|
---|
| 57 | # endif
|
---|
| 58 | # ifndef __NR_io_uring_enter
|
---|
| 59 | # define __NR_io_uring_enter 426
|
---|
| 60 | # endif
|
---|
| 61 | # ifndef __NR_io_uring_register
|
---|
| 62 | # define __NR_io_uring_register 427
|
---|
| 63 | # endif
|
---|
| 64 | #endif
|
---|
| 65 |
|
---|
| 66 |
|
---|
[92976d9] | 67 | //=============================================================================================
|
---|
| 68 | // I/O Startup / Shutdown logic
|
---|
| 69 | //=============================================================================================
|
---|
| 70 | void __kernel_io_startup( cluster & this ) {
|
---|
| 71 | // Step 1 : call to setup
|
---|
| 72 | struct io_uring_params params;
|
---|
| 73 | memset(¶ms, 0, sizeof(params));
|
---|
| 74 |
|
---|
[2d8f7b0] | 75 | uint32_t nentries = entries_per_cluster();
|
---|
| 76 |
|
---|
| 77 | int fd = syscall(__NR_io_uring_setup, nentries, ¶ms );
|
---|
[92976d9] | 78 | if(fd < 0) {
|
---|
| 79 | abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
|
---|
| 80 | }
|
---|
| 81 |
|
---|
| 82 | // Step 2 : mmap result
|
---|
| 83 | memset(&this.io, 0, sizeof(struct io_ring));
|
---|
[2d8f7b0] | 84 | struct io_uring_sq & sq = this.io.submit_q;
|
---|
| 85 | struct io_uring_cq & cq = this.io.completion_q;
|
---|
[92976d9] | 86 |
|
---|
| 87 | // calculate the right ring size
|
---|
[2d8f7b0] | 88 | sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) );
|
---|
| 89 | cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe));
|
---|
[92976d9] | 90 |
|
---|
| 91 | // Requires features
|
---|
| 92 | // // adjust the size according to the parameters
|
---|
| 93 | // if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
|
---|
| 94 | // cq->ring_sz = sq->ring_sz = max(cq->ring_sz, sq->ring_sz);
|
---|
| 95 | // }
|
---|
| 96 |
|
---|
| 97 | // mmap the Submit Queue into existence
|
---|
[2d8f7b0] | 98 | sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQ_RING);
|
---|
| 99 | if (sq.ring_ptr == (void*)MAP_FAILED) {
|
---|
[92976d9] | 100 | abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
|
---|
| 101 | }
|
---|
| 102 |
|
---|
| 103 | // mmap the Completion Queue into existence (may or may not be needed)
|
---|
| 104 | // Requires features
|
---|
| 105 | // if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
|
---|
| 106 | // cq->ring_ptr = sq->ring_ptr;
|
---|
| 107 | // }
|
---|
| 108 | // else {
|
---|
| 109 | // We need multiple call to MMAP
|
---|
[2d8f7b0] | 110 | cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_CQ_RING);
|
---|
| 111 | if (cq.ring_ptr == (void*)MAP_FAILED) {
|
---|
| 112 | munmap(sq.ring_ptr, sq.ring_sz);
|
---|
[92976d9] | 113 | abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
|
---|
| 114 | }
|
---|
| 115 | // }
|
---|
| 116 |
|
---|
| 117 | // mmap the submit queue entries
|
---|
| 118 | size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
|
---|
[2d8f7b0] | 119 | sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQES);
|
---|
| 120 | if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
|
---|
| 121 | munmap(sq.ring_ptr, sq.ring_sz);
|
---|
| 122 | if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
|
---|
[92976d9] | 123 | abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
|
---|
| 124 | }
|
---|
| 125 |
|
---|
| 126 | // Get the pointers from the kernel to fill the structure
|
---|
| 127 | // submit queue
|
---|
[2d8f7b0] | 128 | sq.head = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
|
---|
| 129 | sq.tail = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
|
---|
| 130 | sq.mask = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
|
---|
| 131 | sq.num = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
|
---|
| 132 | sq.flags = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
|
---|
| 133 | sq.dropped = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
|
---|
| 134 | sq.array = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
|
---|
| 135 | sq.alloc = *sq.tail;
|
---|
[92976d9] | 136 |
|
---|
| 137 | // completion queue
|
---|
[2d8f7b0] | 138 | cq.head = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
|
---|
| 139 | cq.tail = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
|
---|
| 140 | cq.mask = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
|
---|
| 141 | cq.num = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
|
---|
| 142 | cq.overflow = ( uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
|
---|
| 143 | cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
|
---|
| 144 |
|
---|
| 145 | // some paranoid checks
|
---|
| 146 | /* 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 );
|
---|
| 147 | /* paranoid */ verifyf( (*cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
|
---|
| 148 | /* paranoid */ verifyf( (*cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
|
---|
| 149 | /* paranoid */ verifyf( (*cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
|
---|
| 150 |
|
---|
| 151 | /* 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 );
|
---|
| 152 | /* paranoid */ verifyf( (*sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
|
---|
| 153 | /* paranoid */ verifyf( (*sq.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.head );
|
---|
| 154 | /* paranoid */ verifyf( (*sq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.tail );
|
---|
[92976d9] | 155 |
|
---|
| 156 | // Update the global ring info
|
---|
| 157 | this.io.flags = params.flags;
|
---|
| 158 | this.io.fd = fd;
|
---|
| 159 | this.io.done = false;
|
---|
[2d8f7b0] | 160 | (this.io.submit){ min(*sq.num, *cq.num) };
|
---|
[92976d9] | 161 |
|
---|
| 162 | // Create the poller thread
|
---|
| 163 | this.io.stack = __create_pthread( &this.io.poller, __io_poller, &this );
|
---|
| 164 | }
|
---|
| 165 |
|
---|
| 166 | void __kernel_io_shutdown( cluster & this ) {
|
---|
| 167 | // Stop the IO Poller
|
---|
| 168 | // Notify the poller thread of the shutdown
|
---|
| 169 | __atomic_store_n(&this.io.done, true, __ATOMIC_SEQ_CST);
|
---|
| 170 | sigval val = { 1 };
|
---|
| 171 | pthread_sigqueue( this.io.poller, SIGUSR1, val );
|
---|
| 172 |
|
---|
| 173 | // Wait for the poller thread to finish
|
---|
| 174 | pthread_join( this.io.poller, 0p );
|
---|
| 175 | free( this.io.stack );
|
---|
| 176 |
|
---|
| 177 | // Shutdown the io rings
|
---|
| 178 | struct io_uring_sq & sq = this.io.submit_q;
|
---|
| 179 | struct io_uring_cq & cq = this.io.completion_q;
|
---|
| 180 |
|
---|
| 181 | // unmap the submit queue entries
|
---|
[2d8f7b0] | 182 | munmap(sq.sqes, (*sq.num) * sizeof(struct io_uring_sqe));
|
---|
[92976d9] | 183 |
|
---|
| 184 | // unmap the Submit Queue ring
|
---|
| 185 | munmap(sq.ring_ptr, sq.ring_sz);
|
---|
| 186 |
|
---|
| 187 | // unmap the Completion Queue ring, if it is different
|
---|
| 188 | if (cq.ring_ptr != sq.ring_ptr) {
|
---|
| 189 | munmap(cq.ring_ptr, cq.ring_sz);
|
---|
| 190 | }
|
---|
| 191 |
|
---|
| 192 | // close the file descriptor
|
---|
| 193 | close(this.io.fd);
|
---|
| 194 | }
|
---|
| 195 |
|
---|
| 196 | //=============================================================================================
|
---|
| 197 | // I/O Polling
|
---|
| 198 | //=============================================================================================
|
---|
| 199 | struct io_user_data {
|
---|
| 200 | int32_t result;
|
---|
| 201 | $thread * thrd;
|
---|
| 202 | };
|
---|
| 203 |
|
---|
| 204 | // Process a single completion message from the io_uring
|
---|
| 205 | // This is NOT thread-safe
|
---|
| 206 | static bool __io_process(struct io_ring & ring) {
|
---|
| 207 | unsigned head = *ring.completion_q.head;
|
---|
| 208 | unsigned tail = __atomic_load_n(ring.completion_q.tail, __ATOMIC_ACQUIRE);
|
---|
| 209 |
|
---|
| 210 | if (head == tail) return false;
|
---|
| 211 |
|
---|
| 212 | unsigned idx = head & (*ring.completion_q.mask);
|
---|
| 213 | struct io_uring_cqe & cqe = ring.completion_q.cqes[idx];
|
---|
| 214 |
|
---|
| 215 | /* paranoid */ verify(&cqe);
|
---|
| 216 |
|
---|
| 217 | struct io_user_data * data = (struct io_user_data *)cqe.user_data;
|
---|
[185efe6] | 218 | // __cfaabi_bits_print_safe( STDERR_FILENO, "Performed reading io cqe %p, result %d for %p\n", data, cqe.res, data->thrd );
|
---|
[2d8f7b0] | 219 |
|
---|
[92976d9] | 220 | data->result = cqe.res;
|
---|
[2d8f7b0] | 221 | __unpark( data->thrd __cfaabi_dbg_ctx2 );
|
---|
| 222 |
|
---|
| 223 | // Allow new submissions to happen
|
---|
| 224 | V(ring.submit);
|
---|
[92976d9] | 225 |
|
---|
| 226 | // Mark to the kernel that the cqe has been seen
|
---|
| 227 | // Ensure that the kernel only sees the new value of the head index after the CQEs have been read.
|
---|
[2d8f7b0] | 228 | __atomic_fetch_add( ring.completion_q.head, 1, __ATOMIC_RELAXED );
|
---|
[92976d9] | 229 |
|
---|
| 230 | return true;
|
---|
| 231 | }
|
---|
| 232 |
|
---|
| 233 | static void * __io_poller( void * arg ) {
|
---|
| 234 | cluster * cltr = (cluster *)arg;
|
---|
| 235 | struct io_ring & ring = cltr->io;
|
---|
| 236 |
|
---|
| 237 | sigset_t mask;
|
---|
| 238 | sigfillset(&mask);
|
---|
| 239 | if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
|
---|
| 240 | abort( "KERNEL ERROR: IO_URING - pthread_sigmask" );
|
---|
| 241 | }
|
---|
| 242 |
|
---|
| 243 | sigdelset( &mask, SIGUSR1 );
|
---|
| 244 |
|
---|
| 245 | verify( (*ring.submit_q.head) == (*ring.submit_q.tail) );
|
---|
| 246 | verify( (*ring.completion_q.head) == (*ring.completion_q.tail) );
|
---|
| 247 |
|
---|
| 248 | LOOP: while(!__atomic_load_n(&ring.done, __ATOMIC_SEQ_CST)) {
|
---|
| 249 | int ret = syscall( __NR_io_uring_enter, ring.fd, 0, 1, IORING_ENTER_GETEVENTS, &mask, _NSIG / 8);
|
---|
| 250 | if( ret < 0 ) {
|
---|
| 251 | switch((int)errno) {
|
---|
| 252 | case EAGAIN:
|
---|
| 253 | case EINTR:
|
---|
| 254 | continue LOOP;
|
---|
| 255 | default:
|
---|
| 256 | abort( "KERNEL ERROR: IO_URING WAIT - %s\n", strerror(errno) );
|
---|
| 257 | }
|
---|
| 258 | }
|
---|
| 259 |
|
---|
| 260 | // Drain the queue
|
---|
| 261 | while(__io_process(ring)) {}
|
---|
| 262 | }
|
---|
| 263 |
|
---|
| 264 | return 0p;
|
---|
| 265 | }
|
---|
| 266 |
|
---|
| 267 | //=============================================================================================
|
---|
| 268 | // I/O Submissions
|
---|
| 269 | //=============================================================================================
|
---|
| 270 |
|
---|
[2d8f7b0] | 271 | // Submition steps :
|
---|
| 272 | // 1 - We need to make sure we don't overflow any of the buffer, P(ring.submit) to make sure
|
---|
| 273 | // entries are available. The semaphore make sure that there is no more operations in
|
---|
| 274 | // progress then the number of entries in the buffer. This probably limits concurrency
|
---|
| 275 | // more than necessary since submitted but not completed operations don't need any
|
---|
| 276 | // entries in user space. However, I don't know what happens if we overflow the buffers
|
---|
| 277 | // because too many requests completed at once. This is a safe approach in all cases.
|
---|
| 278 | // Furthermore, with hundreds of entries, this may be okay.
|
---|
| 279 | //
|
---|
| 280 | // 2 - Allocate a queue entry. The ring already has memory for all entries but only the ones
|
---|
| 281 | // listed in sq.array are visible by the kernel. For those not listed, the kernel does not
|
---|
| 282 | // offer any assurance that an entry is not being filled by multiple flags. Therefore, we
|
---|
| 283 | // need to write an allocator that allows allocating concurrently.
|
---|
| 284 | //
|
---|
| 285 | // 3 - Actually fill the submit entry, this is the only simple and straightforward step.
|
---|
| 286 | //
|
---|
| 287 | // 4 - Append the entry index to the array and adjust the tail accordingly. This operation
|
---|
| 288 | // needs to arrive to two concensus at the same time:
|
---|
| 289 | // A - The order in which entries are listed in the array: no two threads must pick the
|
---|
| 290 | // same index for their entries
|
---|
| 291 | // B - When can the tail be update for the kernel. EVERY entries in the array between
|
---|
| 292 | // head and tail must be fully filled and shouldn't ever be touched again.
|
---|
| 293 | //
|
---|
| 294 |
|
---|
| 295 | static inline [* struct io_uring_sqe, uint32_t] __submit_alloc( struct io_ring & ring ) {
|
---|
| 296 | // Wait for a spot to be available
|
---|
| 297 | P(ring.submit);
|
---|
| 298 |
|
---|
| 299 | // Allocate the sqe
|
---|
| 300 | uint32_t idx = __atomic_fetch_add(&ring.submit_q.alloc, 1ul32, __ATOMIC_SEQ_CST);
|
---|
| 301 |
|
---|
| 302 | // Validate that we didn't overflow anything
|
---|
| 303 | // Check that nothing overflowed
|
---|
| 304 | /* paranoid */ verify( true );
|
---|
| 305 |
|
---|
| 306 | // Check that it goes head -> tail -> alloc and never head -> alloc -> tail
|
---|
| 307 | /* paranoid */ verify( true );
|
---|
| 308 |
|
---|
| 309 | // Return the sqe
|
---|
| 310 | return [&ring.submit_q.sqes[ idx & (*ring.submit_q.mask)], idx];
|
---|
| 311 | }
|
---|
| 312 |
|
---|
| 313 | static inline void __submit( struct io_ring & ring, uint32_t idx ) {
|
---|
| 314 | // get mutual exclusion
|
---|
| 315 | lock(ring.submit_q.lock __cfaabi_dbg_ctx2);
|
---|
| 316 |
|
---|
| 317 | // Append to the list of ready entries
|
---|
| 318 | uint32_t * tail = ring.submit_q.tail;
|
---|
| 319 | const uint32_t mask = *ring.submit_q.mask;
|
---|
| 320 |
|
---|
| 321 | ring.submit_q.array[ (*tail) & mask ] = idx & mask;
|
---|
| 322 | __atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST);
|
---|
| 323 |
|
---|
| 324 | // Submit however, many entries need to be submitted
|
---|
| 325 | int ret = syscall( __NR_io_uring_enter, ring.fd, 1, 0, 0, 0p, 0);
|
---|
[185efe6] | 326 | // __cfaabi_bits_print_safe( STDERR_FILENO, "Performed io_submit, returned %d\n", ret );
|
---|
[2d8f7b0] | 327 | if( ret < 0 ) {
|
---|
| 328 | switch((int)errno) {
|
---|
| 329 | default:
|
---|
| 330 | abort( "KERNEL ERROR: IO_URING SUBMIT - %s\n", strerror(errno) );
|
---|
| 331 | }
|
---|
| 332 | }
|
---|
| 333 |
|
---|
| 334 | unlock(ring.submit_q.lock);
|
---|
| 335 | // Make sure that idx was submitted
|
---|
| 336 | // Be careful to not get false positive if we cycled the entire list or that someone else submitted for us
|
---|
| 337 | }
|
---|
| 338 |
|
---|
| 339 | static inline void ?{}(struct io_uring_sqe & this, uint8_t opcode, int fd) {
|
---|
| 340 | this.opcode = opcode;
|
---|
| 341 | #if !defined(IOSQE_ASYNC)
|
---|
| 342 | this.flags = 0;
|
---|
| 343 | #else
|
---|
| 344 | this.flags = IOSQE_ASYNC;
|
---|
| 345 | #endif
|
---|
| 346 | this.ioprio = 0;
|
---|
| 347 | this.fd = fd;
|
---|
| 348 | this.off = 0;
|
---|
| 349 | this.addr = 0;
|
---|
| 350 | this.len = 0;
|
---|
| 351 | this.rw_flags = 0;
|
---|
| 352 | this.__pad2[0] = this.__pad2[1] = this.__pad2[2] = 0;
|
---|
| 353 | }
|
---|
| 354 |
|
---|
| 355 | static inline void ?{}(struct io_uring_sqe & this, uint8_t opcode, int fd, void * addr, uint32_t len, uint64_t off ) {
|
---|
| 356 | (this){ opcode, fd };
|
---|
| 357 | this.off = off;
|
---|
| 358 | this.addr = (uint64_t)addr;
|
---|
| 359 | this.len = len;
|
---|
[92976d9] | 360 | }
|
---|
| 361 |
|
---|
| 362 | //=============================================================================================
|
---|
| 363 | // I/O Interface
|
---|
| 364 | //=============================================================================================
|
---|
| 365 | extern "C" {
|
---|
| 366 | #define __USE_GNU
|
---|
[2d8f7b0] | 367 | #define _GNU_SOURCE
|
---|
| 368 | #include <fcntl.h>
|
---|
[92976d9] | 369 | #include <sys/uio.h>
|
---|
[2d8f7b0] | 370 | #include <sys/socket.h>
|
---|
| 371 | #include <sys/stat.h>
|
---|
[92976d9] | 372 | }
|
---|
| 373 |
|
---|
[2d8f7b0] | 374 | #define __submit_prelude \
|
---|
| 375 | struct io_ring & ring = active_cluster()->io; \
|
---|
| 376 | struct io_uring_sqe * sqe; \
|
---|
| 377 | uint32_t idx; \
|
---|
| 378 | [sqe, idx] = __submit_alloc( ring );
|
---|
| 379 |
|
---|
| 380 | #define __submit_wait \
|
---|
| 381 | io_user_data data = { 0, active_thread() }; \
|
---|
[185efe6] | 382 | /*__cfaabi_bits_print_safe( STDERR_FILENO, "Preparing user data %p for %p\n", &data, data.thrd );*/ \
|
---|
[2d8f7b0] | 383 | sqe->user_data = (uint64_t)&data; \
|
---|
| 384 | __submit( ring, idx ); \
|
---|
| 385 | park( __cfaabi_dbg_ctx ); \
|
---|
| 386 | return data.result;
|
---|
| 387 |
|
---|
| 388 | //-----------------------------------------------------------------------------
|
---|
| 389 | // Asynchronous operations
|
---|
[92976d9] | 390 | ssize_t async_preadv2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags) {
|
---|
| 391 | #if !defined(IORING_OP_READV)
|
---|
| 392 | return preadv2(fd, iov, iovcnt, offset, flags);
|
---|
| 393 | #else
|
---|
[2d8f7b0] | 394 | __submit_prelude
|
---|
| 395 |
|
---|
| 396 | (*sqe){ IORING_OP_READV, fd, iov, iovcnt, offset };
|
---|
| 397 |
|
---|
| 398 | __submit_wait
|
---|
[92976d9] | 399 | #endif
|
---|
| 400 | }
|
---|
| 401 |
|
---|
| 402 | ssize_t async_pwritev2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags) {
|
---|
| 403 | #if !defined(IORING_OP_WRITEV)
|
---|
| 404 | return pwritev2(fd, iov, iovcnt, offset, flags);
|
---|
| 405 | #else
|
---|
[2d8f7b0] | 406 | __submit_prelude
|
---|
| 407 |
|
---|
| 408 | (*sqe){ IORING_OP_WRITEV, fd, iov, iovcnt, offset };
|
---|
| 409 |
|
---|
| 410 | __submit_wait
|
---|
| 411 | #endif
|
---|
| 412 | }
|
---|
| 413 |
|
---|
| 414 | int async_fsync(int fd) {
|
---|
| 415 | #if !defined(IORING_OP_FSYNC)
|
---|
| 416 | return fsync(fd);
|
---|
| 417 | #else
|
---|
| 418 | __submit_prelude
|
---|
| 419 |
|
---|
| 420 | (*sqe){ IORING_OP_FSYNC, fd };
|
---|
| 421 |
|
---|
| 422 | __submit_wait
|
---|
[92976d9] | 423 | #endif
|
---|
| 424 | }
|
---|
| 425 |
|
---|
[2d8f7b0] | 426 | int async_sync_file_range(int fd, int64_t offset, int64_t nbytes, unsigned int flags) {
|
---|
| 427 | #if !defined(IORING_OP_SYNC_FILE_RANGE)
|
---|
| 428 | return sync_file_range(fd, offset, nbytes, flags);
|
---|
| 429 | #else
|
---|
| 430 | __submit_prelude
|
---|
| 431 |
|
---|
| 432 | (*sqe){ IORING_OP_SYNC_FILE_RANGE, fd };
|
---|
| 433 | sqe->off = offset;
|
---|
| 434 | sqe->len = nbytes;
|
---|
| 435 | sqe->sync_range_flags = flags;
|
---|
| 436 |
|
---|
| 437 | __submit_wait
|
---|
| 438 | #endif
|
---|
| 439 | }
|
---|
| 440 |
|
---|
| 441 |
|
---|
| 442 | ssize_t async_sendmsg(int sockfd, const struct msghdr *msg, int flags) {
|
---|
| 443 | #if !defined(IORING_OP_SENDMSG)
|
---|
| 444 | return recv(sockfd, msg, flags);
|
---|
| 445 | #else
|
---|
| 446 | __submit_prelude
|
---|
| 447 |
|
---|
| 448 | (*sqe){ IORING_OP_SENDMSG, sockfd, msg, 1, 0 };
|
---|
| 449 | sqe->msg_flags = flags;
|
---|
| 450 |
|
---|
| 451 | __submit_wait
|
---|
| 452 | #endif
|
---|
| 453 | }
|
---|
| 454 |
|
---|
| 455 | ssize_t async_recvmsg(int sockfd, struct msghdr *msg, int flags) {
|
---|
| 456 | #if !defined(IORING_OP_RECVMSG)
|
---|
| 457 | return recv(sockfd, msg, flags);
|
---|
| 458 | #else
|
---|
| 459 | __submit_prelude
|
---|
| 460 |
|
---|
| 461 | (*sqe){ IORING_OP_RECVMSG, sockfd, msg, 1, 0 };
|
---|
| 462 | sqe->msg_flags = flags;
|
---|
| 463 |
|
---|
| 464 | __submit_wait
|
---|
| 465 | #endif
|
---|
| 466 | }
|
---|
| 467 |
|
---|
| 468 | ssize_t async_send(int sockfd, const void *buf, size_t len, int flags) {
|
---|
| 469 | #if !defined(IORING_OP_SEND)
|
---|
| 470 | return send( sockfd, buf, len, flags );
|
---|
| 471 | #else
|
---|
| 472 | __submit_prelude
|
---|
| 473 |
|
---|
| 474 | (*sqe){ IORING_OP_SEND, sockfd };
|
---|
| 475 | sqe->addr = (uint64_t)buf;
|
---|
| 476 | sqe->len = len;
|
---|
| 477 | sqe->msg_flags = flags;
|
---|
| 478 |
|
---|
| 479 | __submit_wait
|
---|
| 480 | #endif
|
---|
| 481 | }
|
---|
| 482 |
|
---|
| 483 | ssize_t async_recv(int sockfd, void *buf, size_t len, int flags) {
|
---|
| 484 | #if !defined(IORING_OP_RECV)
|
---|
| 485 | return recv( sockfd, buf, len, flags );
|
---|
| 486 | #else
|
---|
| 487 | __submit_prelude
|
---|
| 488 |
|
---|
| 489 | (*sqe){ IORING_OP_RECV, sockfd };
|
---|
| 490 | sqe->addr = (uint64_t)buf;
|
---|
| 491 | sqe->len = len;
|
---|
| 492 | sqe->msg_flags = flags;
|
---|
| 493 |
|
---|
| 494 | __submit_wait
|
---|
| 495 | #endif
|
---|
| 496 | }
|
---|
| 497 |
|
---|
| 498 | int async_accept4(int sockfd, struct sockaddr *addr, socklen_t *addrlen, int flags) {
|
---|
| 499 | #if !defined(IORING_OP_ACCEPT)
|
---|
| 500 | __SOCKADDR_ARG _addr;
|
---|
| 501 | _addr.__sockaddr__ = addr;
|
---|
| 502 | return accept4( sockfd, _addr, addrlen, flags );
|
---|
| 503 | #else
|
---|
| 504 | __submit_prelude
|
---|
| 505 |
|
---|
| 506 | (*sqe){ IORING_OP_ACCEPT, sockfd };
|
---|
| 507 | sqe->addr = addr;
|
---|
| 508 | sqe->addr2 = addrlen;
|
---|
| 509 | sqe->accept_flags = flags;
|
---|
| 510 |
|
---|
| 511 | __submit_wait
|
---|
| 512 | #endif
|
---|
| 513 | }
|
---|
| 514 |
|
---|
| 515 | int async_connect(int sockfd, const struct sockaddr *addr, socklen_t addrlen) {
|
---|
| 516 | #if !defined(IORING_OP_CONNECT)
|
---|
| 517 | __CONST_SOCKADDR_ARG _addr;
|
---|
| 518 | _addr.__sockaddr__ = addr;
|
---|
| 519 | return connect( sockfd, _addr, addrlen );
|
---|
| 520 | #else
|
---|
| 521 | __submit_prelude
|
---|
| 522 |
|
---|
| 523 | (*sqe){ IORING_OP_CONNECT, sockfd };
|
---|
| 524 | sqe->addr = (uint64_t)addr;
|
---|
| 525 | sqe->off = addrlen;
|
---|
| 526 |
|
---|
| 527 | __submit_wait
|
---|
| 528 | #endif
|
---|
| 529 | }
|
---|
| 530 |
|
---|
| 531 | int async_fallocate(int fd, int mode, uint64_t offset, uint64_t len) {
|
---|
| 532 | #if !defined(IORING_OP_FALLOCATE)
|
---|
| 533 | return fallocate( fd, mode, offset, len );
|
---|
| 534 | #else
|
---|
| 535 | __submit_prelude
|
---|
| 536 |
|
---|
| 537 | (*sqe){ IORING_OP_FALLOCATE, fd };
|
---|
| 538 | sqe->off = offset;
|
---|
| 539 | sqe->len = length;
|
---|
| 540 | sqe->mode = mode;
|
---|
| 541 |
|
---|
| 542 | __submit_wait
|
---|
| 543 | #endif
|
---|
| 544 | }
|
---|
| 545 |
|
---|
| 546 | int async_fadvise(int fd, uint64_t offset, uint64_t len, int advice) {
|
---|
| 547 | #if !defined(IORING_OP_FADVISE)
|
---|
| 548 | return posix_fadvise( fd, offset, len, advice );
|
---|
| 549 | #else
|
---|
| 550 | __submit_prelude
|
---|
| 551 |
|
---|
| 552 | (*sqe){ IORING_OP_FADVISE, fd };
|
---|
| 553 | sqe->off = (uint64_t)offset;
|
---|
| 554 | sqe->len = length;
|
---|
| 555 | sqe->fadvise_advice = advice;
|
---|
| 556 |
|
---|
| 557 | __submit_wait
|
---|
| 558 | #endif
|
---|
| 559 | }
|
---|
| 560 |
|
---|
| 561 | int async_madvise(void *addr, size_t length, int advice) {
|
---|
| 562 | #if !defined(IORING_OP_MADVISE)
|
---|
| 563 | return madvise( addr, length, advice );
|
---|
| 564 | #else
|
---|
| 565 | __submit_prelude
|
---|
| 566 |
|
---|
| 567 | (*sqe){ IORING_OP_MADVISE, 0 };
|
---|
| 568 | sqe->addr = (uint64_t)addr;
|
---|
| 569 | sqe->len = length;
|
---|
| 570 | sqe->fadvise_advice = advice;
|
---|
| 571 |
|
---|
| 572 | __submit_wait
|
---|
| 573 | #endif
|
---|
| 574 | }
|
---|
| 575 |
|
---|
| 576 | int async_openat(int dirfd, const char *pathname, int flags, mode_t mode) {
|
---|
| 577 | #if !defined(IORING_OP_OPENAT)
|
---|
| 578 | return openat( dirfd, pathname, flags, mode );
|
---|
| 579 | #else
|
---|
| 580 | __submit_prelude
|
---|
| 581 |
|
---|
| 582 | (*sqe){ IORING_OP_OPENAT, dirfd };
|
---|
| 583 | sqe->addr = (uint64_t)pathname;
|
---|
| 584 | sqe->open_flags = flags;
|
---|
| 585 | sqe->mode = mode;
|
---|
| 586 |
|
---|
| 587 | __submit_wait
|
---|
| 588 | #endif
|
---|
| 589 | }
|
---|
| 590 |
|
---|
| 591 | int async_close(int fd) {
|
---|
| 592 | #if !defined(IORING_OP_CLOSE)
|
---|
| 593 | return close( fd );
|
---|
| 594 | #else
|
---|
| 595 | __submit_prelude
|
---|
| 596 |
|
---|
| 597 | (*sqe){ IORING_OP_CLOSE, fd };
|
---|
| 598 |
|
---|
| 599 | __submit_wait
|
---|
| 600 | #endif
|
---|
| 601 | }
|
---|
| 602 |
|
---|
| 603 | int async_statx(int dirfd, const char *pathname, int flags, unsigned int mask, struct statx *statxbuf) {
|
---|
| 604 | #if !defined(IORING_OP_STATX)
|
---|
[8962722] | 605 | //return statx( dirfd, pathname, flags, mask, statxbuf );
|
---|
| 606 | return syscall( __NR_io_uring_setup, dirfd, pathname, flags, mask, statxbuf );
|
---|
[2d8f7b0] | 607 | #else
|
---|
| 608 | __submit_prelude
|
---|
| 609 |
|
---|
| 610 | (*sqe){ IORING_OP_STATX, dirfd };
|
---|
| 611 | sqe->addr = (uint64_t)pathname;
|
---|
| 612 | sqe->statx_flags = flags;
|
---|
| 613 | sqe->len = mask;
|
---|
| 614 | sqe->off = (uint64_t)statxbuf;
|
---|
| 615 |
|
---|
| 616 | __submit_wait
|
---|
| 617 | #endif
|
---|
| 618 | }
|
---|
| 619 |
|
---|
| 620 |
|
---|
| 621 | ssize_t async_read(int fd, void *buf, size_t count) {
|
---|
| 622 | #if !defined(IORING_OP_READ)
|
---|
| 623 | return read( fd, buf, count );
|
---|
| 624 | #else
|
---|
| 625 | __submit_prelude
|
---|
| 626 |
|
---|
| 627 | (*sqe){ IORING_OP_READ, fd, buf, count, 0 };
|
---|
| 628 |
|
---|
| 629 | __submit_wait
|
---|
| 630 | #endif
|
---|
| 631 | }
|
---|
| 632 |
|
---|
| 633 | ssize_t async_write(int fd, void *buf, size_t count) {
|
---|
| 634 | #if !defined(IORING_OP_WRITE)
|
---|
| 635 | return read( fd, buf, count );
|
---|
| 636 | #else
|
---|
| 637 | __submit_prelude
|
---|
| 638 |
|
---|
| 639 | (*sqe){ IORING_OP_WRITE, fd, buf, count, 0 };
|
---|
| 640 |
|
---|
| 641 | __submit_wait
|
---|
| 642 | #endif
|
---|
| 643 | }
|
---|
| 644 |
|
---|
| 645 | //-----------------------------------------------------------------------------
|
---|
| 646 | // Check if a function is asynchronous
|
---|
| 647 |
|
---|
| 648 | // Macro magic to reduce the size of the following switch case
|
---|
| 649 | #define IS_DEFINED_APPLY(f, ...) f(__VA_ARGS__)
|
---|
| 650 | #define IS_DEFINED_SECOND(first, second, ...) second
|
---|
| 651 | #define IS_DEFINED_TEST(expansion) _CFA_IO_FEATURE_##expansion
|
---|
| 652 | #define IS_DEFINED(macro) IS_DEFINED_APPLY( IS_DEFINED_SECOND,IS_DEFINED_TEST(macro) false, true)
|
---|
| 653 |
|
---|
| 654 | bool is_async( fptr_t func ) {
|
---|
| 655 |
|
---|
| 656 | if( /*func == (fptr_t)preadv2 || */
|
---|
| 657 | func == (fptr_t)async_preadv2 )
|
---|
| 658 | #define _CFA_IO_FEATURE_IORING_OP_READV ,
|
---|
| 659 | return IS_DEFINED(IORING_OP_READV);
|
---|
| 660 |
|
---|
| 661 | if( /*func == (fptr_t)pwritev2 || */
|
---|
| 662 | func == (fptr_t)async_pwritev2 )
|
---|
| 663 | #define _CFA_IO_FEATURE_IORING_OP_WRITEV ,
|
---|
| 664 | return IS_DEFINED(IORING_OP_WRITEV);
|
---|
| 665 |
|
---|
| 666 | if( /*func == (fptr_t)fsync || */
|
---|
| 667 | func == (fptr_t)async_fsync )
|
---|
| 668 | #define _CFA_IO_FEATURE_IORING_OP_FSYNC ,
|
---|
| 669 | return IS_DEFINED(IORING_OP_FSYNC);
|
---|
| 670 |
|
---|
| 671 | if( /*func == (fptr_t)ync_file_range || */
|
---|
| 672 | func == (fptr_t)async_sync_file_range )
|
---|
| 673 | #define _CFA_IO_FEATURE_IORING_OP_SYNC_FILE_RANGE ,
|
---|
| 674 | return IS_DEFINED(IORING_OP_SYNC_FILE_RANGE);
|
---|
| 675 |
|
---|
| 676 | if( /*func == (fptr_t)sendmsg || */
|
---|
| 677 | func == (fptr_t)async_sendmsg )
|
---|
| 678 | #define _CFA_IO_FEATURE_IORING_OP_SENDMSG ,
|
---|
| 679 | return IS_DEFINED(IORING_OP_SENDMSG);
|
---|
| 680 |
|
---|
| 681 | if( /*func == (fptr_t)recvmsg || */
|
---|
| 682 | func == (fptr_t)async_recvmsg )
|
---|
| 683 | #define _CFA_IO_FEATURE_IORING_OP_RECVMSG ,
|
---|
| 684 | return IS_DEFINED(IORING_OP_RECVMSG);
|
---|
| 685 |
|
---|
| 686 | if( /*func == (fptr_t)send || */
|
---|
| 687 | func == (fptr_t)async_send )
|
---|
| 688 | #define _CFA_IO_FEATURE_IORING_OP_SEND ,
|
---|
| 689 | return IS_DEFINED(IORING_OP_SEND);
|
---|
| 690 |
|
---|
| 691 | if( /*func == (fptr_t)recv || */
|
---|
| 692 | func == (fptr_t)async_recv )
|
---|
| 693 | #define _CFA_IO_FEATURE_IORING_OP_RECV ,
|
---|
| 694 | return IS_DEFINED(IORING_OP_RECV);
|
---|
| 695 |
|
---|
| 696 | if( /*func == (fptr_t)accept4 || */
|
---|
| 697 | func == (fptr_t)async_accept4 )
|
---|
| 698 | #define _CFA_IO_FEATURE_IORING_OP_ACCEPT ,
|
---|
| 699 | return IS_DEFINED(IORING_OP_ACCEPT);
|
---|
| 700 |
|
---|
| 701 | if( /*func == (fptr_t)connect || */
|
---|
| 702 | func == (fptr_t)async_connect )
|
---|
| 703 | #define _CFA_IO_FEATURE_IORING_OP_CONNECT ,
|
---|
| 704 | return IS_DEFINED(IORING_OP_CONNECT);
|
---|
| 705 |
|
---|
| 706 | if( /*func == (fptr_t)fallocate || */
|
---|
| 707 | func == (fptr_t)async_fallocate )
|
---|
| 708 | #define _CFA_IO_FEATURE_IORING_OP_FALLOCATE ,
|
---|
| 709 | return IS_DEFINED(IORING_OP_FALLOCATE);
|
---|
| 710 |
|
---|
| 711 | if( /*func == (fptr_t)fadvise || */
|
---|
| 712 | func == (fptr_t)async_fadvise )
|
---|
| 713 | #define _CFA_IO_FEATURE_IORING_OP_FADVISE ,
|
---|
| 714 | return IS_DEFINED(IORING_OP_FADVISE);
|
---|
| 715 |
|
---|
| 716 | if( /*func == (fptr_t)madvise || */
|
---|
| 717 | func == (fptr_t)async_madvise )
|
---|
| 718 | #define _CFA_IO_FEATURE_IORING_OP_MADVISE ,
|
---|
| 719 | return IS_DEFINED(IORING_OP_MADVISE);
|
---|
| 720 |
|
---|
| 721 | if( /*func == (fptr_t)openat || */
|
---|
| 722 | func == (fptr_t)async_openat )
|
---|
| 723 | #define _CFA_IO_FEATURE_IORING_OP_OPENAT ,
|
---|
| 724 | return IS_DEFINED(IORING_OP_OPENAT);
|
---|
| 725 |
|
---|
| 726 | if( /*func == (fptr_t)close || */
|
---|
| 727 | func == (fptr_t)async_close )
|
---|
| 728 | #define _CFA_IO_FEATURE_IORING_OP_CLOSE ,
|
---|
| 729 | return IS_DEFINED(IORING_OP_CLOSE);
|
---|
| 730 |
|
---|
| 731 | if( /*func == (fptr_t)statx || */
|
---|
| 732 | func == (fptr_t)async_statx )
|
---|
| 733 | #define _CFA_IO_FEATURE_IORING_OP_STATX ,
|
---|
| 734 | return IS_DEFINED(IORING_OP_STATX);
|
---|
| 735 |
|
---|
| 736 | if( /*func == (fptr_t)read || */
|
---|
| 737 | func == (fptr_t)async_read )
|
---|
| 738 | #define _CFA_IO_FEATURE_IORING_OP_READ ,
|
---|
| 739 | return IS_DEFINED(IORING_OP_READ);
|
---|
| 740 |
|
---|
| 741 | if( /*func == (fptr_t)write || */
|
---|
| 742 | func == (fptr_t)async_write )
|
---|
| 743 | #define _CFA_IO_FEATURE_IORING_OP_WRITE ,
|
---|
| 744 | return IS_DEFINED(IORING_OP_WRITE);
|
---|
| 745 |
|
---|
| 746 | return false;
|
---|
[92976d9] | 747 | }
|
---|
| 748 |
|
---|
| 749 | #endif
|
---|