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.cfa -- |
---|
8 | // |
---|
9 | // Author : Thierry Delisle |
---|
10 | // Created On : Thu Apr 23 17:31:00 2020 |
---|
11 | // Last Modified By : |
---|
12 | // Last Modified On : |
---|
13 | // Update Count : |
---|
14 | // |
---|
15 | |
---|
16 | #include "kernel.hfa" |
---|
17 | |
---|
18 | #if !defined(HAVE_LINUX_IO_URING_H) |
---|
19 | void __kernel_io_startup( cluster & this ) { |
---|
20 | // Nothing to do without io_uring |
---|
21 | } |
---|
22 | |
---|
23 | void __kernel_io_shutdown( cluster & this ) { |
---|
24 | // Nothing to do without io_uring |
---|
25 | } |
---|
26 | |
---|
27 | #else |
---|
28 | extern "C" { |
---|
29 | #define _GNU_SOURCE /* See feature_test_macros(7) */ |
---|
30 | #include <errno.h> |
---|
31 | #include <stdint.h> |
---|
32 | #include <string.h> |
---|
33 | #include <unistd.h> |
---|
34 | #include <sys/mman.h> |
---|
35 | #include <sys/syscall.h> |
---|
36 | |
---|
37 | #include <linux/io_uring.h> |
---|
38 | } |
---|
39 | |
---|
40 | #include "bits/signal.hfa" |
---|
41 | #include "kernel_private.hfa" |
---|
42 | #include "thread.hfa" |
---|
43 | |
---|
44 | uint32_t entries_per_cluster() { |
---|
45 | return 256; |
---|
46 | } |
---|
47 | |
---|
48 | static void * __io_poller( void * arg ); |
---|
49 | |
---|
50 | // Weirdly, some systems that do support io_uring don't actually define these |
---|
51 | #ifdef __alpha__ |
---|
52 | /* |
---|
53 | * alpha is the only exception, all other architectures |
---|
54 | * have common numbers for new system calls. |
---|
55 | */ |
---|
56 | # ifndef __NR_io_uring_setup |
---|
57 | # define __NR_io_uring_setup 535 |
---|
58 | # endif |
---|
59 | # ifndef __NR_io_uring_enter |
---|
60 | # define __NR_io_uring_enter 536 |
---|
61 | # endif |
---|
62 | # ifndef __NR_io_uring_register |
---|
63 | # define __NR_io_uring_register 537 |
---|
64 | # endif |
---|
65 | #else /* !__alpha__ */ |
---|
66 | # ifndef __NR_io_uring_setup |
---|
67 | # define __NR_io_uring_setup 425 |
---|
68 | # endif |
---|
69 | # ifndef __NR_io_uring_enter |
---|
70 | # define __NR_io_uring_enter 426 |
---|
71 | # endif |
---|
72 | # ifndef __NR_io_uring_register |
---|
73 | # define __NR_io_uring_register 427 |
---|
74 | # endif |
---|
75 | #endif |
---|
76 | |
---|
77 | |
---|
78 | //============================================================================================= |
---|
79 | // I/O Startup / Shutdown logic |
---|
80 | //============================================================================================= |
---|
81 | void __kernel_io_startup( cluster & this ) { |
---|
82 | // Step 1 : call to setup |
---|
83 | struct io_uring_params params; |
---|
84 | memset(¶ms, 0, sizeof(params)); |
---|
85 | |
---|
86 | uint32_t nentries = entries_per_cluster(); |
---|
87 | |
---|
88 | int fd = syscall(__NR_io_uring_setup, nentries, ¶ms ); |
---|
89 | if(fd < 0) { |
---|
90 | abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno)); |
---|
91 | } |
---|
92 | |
---|
93 | // Step 2 : mmap result |
---|
94 | memset(&this.io, 0, sizeof(struct io_ring)); |
---|
95 | struct io_uring_sq & sq = this.io.submit_q; |
---|
96 | struct io_uring_cq & cq = this.io.completion_q; |
---|
97 | |
---|
98 | // calculate the right ring size |
---|
99 | sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) ); |
---|
100 | cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe)); |
---|
101 | |
---|
102 | // Requires features |
---|
103 | #if defined(IORING_FEAT_SINGLE_MMAP) |
---|
104 | // adjust the size according to the parameters |
---|
105 | if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) { |
---|
106 | cq->ring_sz = sq->ring_sz = max(cq->ring_sz, sq->ring_sz); |
---|
107 | } |
---|
108 | #endif |
---|
109 | |
---|
110 | // mmap the Submit Queue into existence |
---|
111 | sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQ_RING); |
---|
112 | if (sq.ring_ptr == (void*)MAP_FAILED) { |
---|
113 | abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno)); |
---|
114 | } |
---|
115 | |
---|
116 | // Requires features |
---|
117 | #if defined(IORING_FEAT_SINGLE_MMAP) |
---|
118 | // mmap the Completion Queue into existence (may or may not be needed) |
---|
119 | if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) { |
---|
120 | cq->ring_ptr = sq->ring_ptr; |
---|
121 | } |
---|
122 | else |
---|
123 | #endif |
---|
124 | { |
---|
125 | // We need multiple call to MMAP |
---|
126 | cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_CQ_RING); |
---|
127 | if (cq.ring_ptr == (void*)MAP_FAILED) { |
---|
128 | munmap(sq.ring_ptr, sq.ring_sz); |
---|
129 | abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno)); |
---|
130 | } |
---|
131 | } |
---|
132 | |
---|
133 | // mmap the submit queue entries |
---|
134 | size_t size = params.sq_entries * sizeof(struct io_uring_sqe); |
---|
135 | sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQES); |
---|
136 | if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) { |
---|
137 | munmap(sq.ring_ptr, sq.ring_sz); |
---|
138 | if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz); |
---|
139 | abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno)); |
---|
140 | } |
---|
141 | |
---|
142 | // Get the pointers from the kernel to fill the structure |
---|
143 | // submit queue |
---|
144 | sq.head = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.head); |
---|
145 | sq.tail = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail); |
---|
146 | sq.mask = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask); |
---|
147 | sq.num = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries); |
---|
148 | sq.flags = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags); |
---|
149 | sq.dropped = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped); |
---|
150 | sq.array = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.array); |
---|
151 | sq.alloc = *sq.tail; |
---|
152 | |
---|
153 | // completion queue |
---|
154 | cq.head = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.head); |
---|
155 | cq.tail = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail); |
---|
156 | cq.mask = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask); |
---|
157 | cq.num = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries); |
---|
158 | cq.overflow = ( uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow); |
---|
159 | cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes); |
---|
160 | |
---|
161 | // some paranoid checks |
---|
162 | /* 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 ); |
---|
163 | /* paranoid */ verifyf( (*cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num ); |
---|
164 | /* paranoid */ verifyf( (*cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head ); |
---|
165 | /* paranoid */ verifyf( (*cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail ); |
---|
166 | |
---|
167 | /* 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 ); |
---|
168 | /* paranoid */ verifyf( (*sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num ); |
---|
169 | /* paranoid */ verifyf( (*sq.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.head ); |
---|
170 | /* paranoid */ verifyf( (*sq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.tail ); |
---|
171 | |
---|
172 | // Update the global ring info |
---|
173 | this.io.flags = params.flags; |
---|
174 | this.io.fd = fd; |
---|
175 | this.io.done = false; |
---|
176 | (this.io.submit){ min(*sq.num, *cq.num) }; |
---|
177 | |
---|
178 | // Initialize statistics |
---|
179 | this.io.submit_q.stats.submit_avg.val = 0; |
---|
180 | this.io.submit_q.stats.submit_avg.cnt = 0; |
---|
181 | this.io.completion_q.stats.completed_avg.val = 0; |
---|
182 | this.io.completion_q.stats.completed_avg.cnt = 0; |
---|
183 | |
---|
184 | // Create the poller thread |
---|
185 | this.io.stack = __create_pthread( &this.io.poller, __io_poller, &this ); |
---|
186 | } |
---|
187 | |
---|
188 | void __kernel_io_shutdown( cluster & this ) { |
---|
189 | // Stop the IO Poller |
---|
190 | // Notify the poller thread of the shutdown |
---|
191 | __atomic_store_n(&this.io.done, true, __ATOMIC_SEQ_CST); |
---|
192 | sigval val = { 1 }; |
---|
193 | pthread_sigqueue( this.io.poller, SIGUSR1, val ); |
---|
194 | |
---|
195 | // Wait for the poller thread to finish |
---|
196 | pthread_join( this.io.poller, 0p ); |
---|
197 | free( this.io.stack ); |
---|
198 | |
---|
199 | // print statistics |
---|
200 | __cfaabi_bits_print_safe( STDERR_FILENO, |
---|
201 | "----- I/O uRing Stats -----\n" |
---|
202 | "- total submit calls : %llu\n" |
---|
203 | "- avg submit : %lf\n" |
---|
204 | "- total wait calls : %llu\n" |
---|
205 | "- avg completion/wait : %lf\n", |
---|
206 | this.io.submit_q.stats.submit_avg.cnt, |
---|
207 | ((double)this.io.submit_q.stats.submit_avg.val) / this.io.submit_q.stats.submit_avg.cnt, |
---|
208 | this.io.completion_q.stats.completed_avg.cnt, |
---|
209 | ((double)this.io.completion_q.stats.completed_avg.val) / this.io.completion_q.stats.completed_avg.cnt |
---|
210 | ); |
---|
211 | |
---|
212 | // Shutdown the io rings |
---|
213 | struct io_uring_sq & sq = this.io.submit_q; |
---|
214 | struct io_uring_cq & cq = this.io.completion_q; |
---|
215 | |
---|
216 | // unmap the submit queue entries |
---|
217 | munmap(sq.sqes, (*sq.num) * sizeof(struct io_uring_sqe)); |
---|
218 | |
---|
219 | // unmap the Submit Queue ring |
---|
220 | munmap(sq.ring_ptr, sq.ring_sz); |
---|
221 | |
---|
222 | // unmap the Completion Queue ring, if it is different |
---|
223 | if (cq.ring_ptr != sq.ring_ptr) { |
---|
224 | munmap(cq.ring_ptr, cq.ring_sz); |
---|
225 | } |
---|
226 | |
---|
227 | // close the file descriptor |
---|
228 | close(this.io.fd); |
---|
229 | } |
---|
230 | |
---|
231 | //============================================================================================= |
---|
232 | // I/O Polling |
---|
233 | //============================================================================================= |
---|
234 | struct io_user_data { |
---|
235 | int32_t result; |
---|
236 | $thread * thrd; |
---|
237 | }; |
---|
238 | |
---|
239 | // Process a single completion message from the io_uring |
---|
240 | // This is NOT thread-safe |
---|
241 | int __drain_io( struct io_ring & ring, sigset_t & mask, int waitcnt ) { |
---|
242 | int ret = syscall( __NR_io_uring_enter, ring.fd, 0, waitcnt, IORING_ENTER_GETEVENTS, &mask, _NSIG / 8); |
---|
243 | if( ret < 0 ) { |
---|
244 | switch((int)errno) { |
---|
245 | case EAGAIN: |
---|
246 | case EINTR: |
---|
247 | return -EAGAIN; |
---|
248 | default: |
---|
249 | abort( "KERNEL ERROR: IO_URING WAIT - %s\n", strerror(errno) ); |
---|
250 | } |
---|
251 | } |
---|
252 | |
---|
253 | // Drain the queue |
---|
254 | unsigned head = *ring.completion_q.head; |
---|
255 | unsigned tail = __atomic_load_n(ring.completion_q.tail, __ATOMIC_ACQUIRE); |
---|
256 | |
---|
257 | // Nothing was new return 0 |
---|
258 | if (head == tail) { |
---|
259 | ring.completion_q.stats.completed_avg.cnt += 1; |
---|
260 | return 0; |
---|
261 | } |
---|
262 | |
---|
263 | uint32_t count = tail - head; |
---|
264 | for(i; count) { |
---|
265 | unsigned idx = (head + i) & (*ring.completion_q.mask); |
---|
266 | struct io_uring_cqe & cqe = ring.completion_q.cqes[idx]; |
---|
267 | |
---|
268 | /* paranoid */ verify(&cqe); |
---|
269 | |
---|
270 | struct io_user_data * data = (struct io_user_data *)cqe.user_data; |
---|
271 | // __cfaabi_bits_print_safe( STDERR_FILENO, "Performed reading io cqe %p, result %d for %p\n", data, cqe.res, data->thrd ); |
---|
272 | |
---|
273 | data->result = cqe.res; |
---|
274 | __unpark( data->thrd __cfaabi_dbg_ctx2 ); |
---|
275 | } |
---|
276 | |
---|
277 | // Allow new submissions to happen |
---|
278 | V(ring.submit, count); |
---|
279 | |
---|
280 | // Mark to the kernel that the cqe has been seen |
---|
281 | // Ensure that the kernel only sees the new value of the head index after the CQEs have been read. |
---|
282 | __atomic_fetch_add( ring.completion_q.head, count, __ATOMIC_RELAXED ); |
---|
283 | |
---|
284 | ring.completion_q.stats.completed_avg.val += count; |
---|
285 | ring.completion_q.stats.completed_avg.cnt += 1; |
---|
286 | |
---|
287 | return count; |
---|
288 | } |
---|
289 | |
---|
290 | static void * __io_poller( void * arg ) { |
---|
291 | cluster * cltr = (cluster *)arg; |
---|
292 | struct io_ring & ring = cltr->io; |
---|
293 | |
---|
294 | sigset_t mask; |
---|
295 | sigfillset(&mask); |
---|
296 | if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) { |
---|
297 | abort( "KERNEL ERROR: IO_URING - pthread_sigmask" ); |
---|
298 | } |
---|
299 | |
---|
300 | sigdelset( &mask, SIGUSR1 ); |
---|
301 | |
---|
302 | verify( (*ring.submit_q.head) == (*ring.submit_q.tail) ); |
---|
303 | verify( (*ring.completion_q.head) == (*ring.completion_q.tail) ); |
---|
304 | |
---|
305 | while(!__atomic_load_n(&ring.done, __ATOMIC_SEQ_CST)) { |
---|
306 | __drain_io( ring, mask, 1 ); |
---|
307 | } |
---|
308 | |
---|
309 | return 0p; |
---|
310 | } |
---|
311 | |
---|
312 | //============================================================================================= |
---|
313 | // I/O Submissions |
---|
314 | //============================================================================================= |
---|
315 | |
---|
316 | // Submition steps : |
---|
317 | // 1 - We need to make sure we don't overflow any of the buffer, P(ring.submit) to make sure |
---|
318 | // entries are available. The semaphore make sure that there is no more operations in |
---|
319 | // progress then the number of entries in the buffer. This probably limits concurrency |
---|
320 | // more than necessary since submitted but not completed operations don't need any |
---|
321 | // entries in user space. However, I don't know what happens if we overflow the buffers |
---|
322 | // because too many requests completed at once. This is a safe approach in all cases. |
---|
323 | // Furthermore, with hundreds of entries, this may be okay. |
---|
324 | // |
---|
325 | // 2 - Allocate a queue entry. The ring already has memory for all entries but only the ones |
---|
326 | // listed in sq.array are visible by the kernel. For those not listed, the kernel does not |
---|
327 | // offer any assurance that an entry is not being filled by multiple flags. Therefore, we |
---|
328 | // need to write an allocator that allows allocating concurrently. |
---|
329 | // |
---|
330 | // 3 - Actually fill the submit entry, this is the only simple and straightforward step. |
---|
331 | // |
---|
332 | // 4 - Append the entry index to the array and adjust the tail accordingly. This operation |
---|
333 | // needs to arrive to two concensus at the same time: |
---|
334 | // A - The order in which entries are listed in the array: no two threads must pick the |
---|
335 | // same index for their entries |
---|
336 | // B - When can the tail be update for the kernel. EVERY entries in the array between |
---|
337 | // head and tail must be fully filled and shouldn't ever be touched again. |
---|
338 | // |
---|
339 | |
---|
340 | static inline [* struct io_uring_sqe, uint32_t] __submit_alloc( struct io_ring & ring ) { |
---|
341 | // Wait for a spot to be available |
---|
342 | P(ring.submit); |
---|
343 | |
---|
344 | // Allocate the sqe |
---|
345 | uint32_t idx = __atomic_fetch_add(&ring.submit_q.alloc, 1ul32, __ATOMIC_SEQ_CST); |
---|
346 | |
---|
347 | // Validate that we didn't overflow anything |
---|
348 | // Check that nothing overflowed |
---|
349 | /* paranoid */ verify( true ); |
---|
350 | |
---|
351 | // Check that it goes head -> tail -> alloc and never head -> alloc -> tail |
---|
352 | /* paranoid */ verify( true ); |
---|
353 | |
---|
354 | // Return the sqe |
---|
355 | return [&ring.submit_q.sqes[ idx & (*ring.submit_q.mask)], idx]; |
---|
356 | } |
---|
357 | |
---|
358 | static inline void __submit( struct io_ring & ring, uint32_t idx ) { |
---|
359 | // get mutual exclusion |
---|
360 | lock(ring.submit_q.lock __cfaabi_dbg_ctx2); |
---|
361 | |
---|
362 | // Append to the list of ready entries |
---|
363 | uint32_t * tail = ring.submit_q.tail; |
---|
364 | const uint32_t mask = *ring.submit_q.mask; |
---|
365 | |
---|
366 | ring.submit_q.array[ (*tail) & mask ] = idx & mask; |
---|
367 | __atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST); |
---|
368 | |
---|
369 | // Submit however, many entries need to be submitted |
---|
370 | int ret = syscall( __NR_io_uring_enter, ring.fd, 1, 0, 0, 0p, 0); |
---|
371 | // __cfaabi_bits_print_safe( STDERR_FILENO, "Performed io_submit, returned %d\n", ret ); |
---|
372 | if( ret < 0 ) { |
---|
373 | switch((int)errno) { |
---|
374 | default: |
---|
375 | abort( "KERNEL ERROR: IO_URING SUBMIT - %s\n", strerror(errno) ); |
---|
376 | } |
---|
377 | } |
---|
378 | |
---|
379 | ring.submit_q.stats.submit_avg.val += 1; |
---|
380 | ring.submit_q.stats.submit_avg.cnt += 1; |
---|
381 | |
---|
382 | unlock(ring.submit_q.lock); |
---|
383 | // Make sure that idx was submitted |
---|
384 | // Be careful to not get false positive if we cycled the entire list or that someone else submitted for us |
---|
385 | } |
---|
386 | |
---|
387 | static inline void ?{}(struct io_uring_sqe & this, uint8_t opcode, int fd) { |
---|
388 | this.opcode = opcode; |
---|
389 | #if !defined(IOSQE_ASYNC) |
---|
390 | this.flags = 0; |
---|
391 | #else |
---|
392 | this.flags = IOSQE_ASYNC; |
---|
393 | #endif |
---|
394 | this.ioprio = 0; |
---|
395 | this.fd = fd; |
---|
396 | this.off = 0; |
---|
397 | this.addr = 0; |
---|
398 | this.len = 0; |
---|
399 | this.rw_flags = 0; |
---|
400 | this.__pad2[0] = this.__pad2[1] = this.__pad2[2] = 0; |
---|
401 | } |
---|
402 | |
---|
403 | static inline void ?{}(struct io_uring_sqe & this, uint8_t opcode, int fd, void * addr, uint32_t len, uint64_t off ) { |
---|
404 | (this){ opcode, fd }; |
---|
405 | this.off = off; |
---|
406 | this.addr = (uint64_t)addr; |
---|
407 | this.len = len; |
---|
408 | } |
---|
409 | #endif |
---|
410 | |
---|
411 | //============================================================================================= |
---|
412 | // I/O Interface |
---|
413 | //============================================================================================= |
---|
414 | #if defined(HAVE_LINUX_IO_URING_H) |
---|
415 | #define __submit_prelude \ |
---|
416 | struct io_ring & ring = active_cluster()->io; \ |
---|
417 | struct io_uring_sqe * sqe; \ |
---|
418 | uint32_t idx; \ |
---|
419 | [sqe, idx] = __submit_alloc( ring ); |
---|
420 | |
---|
421 | #define __submit_wait \ |
---|
422 | io_user_data data = { 0, active_thread() }; \ |
---|
423 | /*__cfaabi_bits_print_safe( STDERR_FILENO, "Preparing user data %p for %p\n", &data, data.thrd );*/ \ |
---|
424 | sqe->user_data = (uint64_t)&data; \ |
---|
425 | __submit( ring, idx ); \ |
---|
426 | park( __cfaabi_dbg_ctx ); \ |
---|
427 | return data.result; |
---|
428 | #endif |
---|
429 | |
---|
430 | // Some forward declarations |
---|
431 | extern "C" { |
---|
432 | #include <sys/types.h> |
---|
433 | struct iovec; |
---|
434 | extern ssize_t preadv2 (int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags); |
---|
435 | extern ssize_t pwritev2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags); |
---|
436 | |
---|
437 | extern int fsync(int fd); |
---|
438 | extern int sync_file_range(int fd, int64_t offset, int64_t nbytes, unsigned int flags); |
---|
439 | |
---|
440 | struct msghdr; |
---|
441 | struct sockaddr; |
---|
442 | extern ssize_t sendmsg(int sockfd, const struct msghdr *msg, int flags); |
---|
443 | extern ssize_t recvmsg(int sockfd, struct msghdr *msg, int flags); |
---|
444 | extern ssize_t send(int sockfd, const void *buf, size_t len, int flags); |
---|
445 | extern ssize_t recv(int sockfd, void *buf, size_t len, int flags); |
---|
446 | extern int accept4(int sockfd, struct sockaddr *addr, socklen_t *addrlen, int flags); |
---|
447 | extern int connect(int sockfd, const struct sockaddr *addr, socklen_t addrlen); |
---|
448 | |
---|
449 | extern int fallocate(int fd, int mode, uint64_t offset, uint64_t len); |
---|
450 | extern int posix_fadvise(int fd, uint64_t offset, uint64_t len, int advice); |
---|
451 | extern int madvise(void *addr, size_t length, int advice); |
---|
452 | |
---|
453 | extern int openat(int dirfd, const char *pathname, int flags, mode_t mode); |
---|
454 | extern int close(int fd); |
---|
455 | |
---|
456 | struct statx; |
---|
457 | extern int statx(int dirfd, const char *pathname, int flags, unsigned int mask, struct statx *statxbuf); |
---|
458 | |
---|
459 | extern ssize_t read (int fd, void *buf, size_t count); |
---|
460 | } |
---|
461 | |
---|
462 | //----------------------------------------------------------------------------- |
---|
463 | // Asynchronous operations |
---|
464 | ssize_t cfa_preadv2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags) { |
---|
465 | #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_READV) |
---|
466 | return preadv2(fd, iov, iovcnt, offset, flags); |
---|
467 | #else |
---|
468 | __submit_prelude |
---|
469 | |
---|
470 | (*sqe){ IORING_OP_READV, fd, iov, iovcnt, offset }; |
---|
471 | |
---|
472 | __submit_wait |
---|
473 | #endif |
---|
474 | } |
---|
475 | |
---|
476 | ssize_t cfa_pwritev2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags) { |
---|
477 | #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_WRITEV) |
---|
478 | return pwritev2(fd, iov, iovcnt, offset, flags); |
---|
479 | #else |
---|
480 | __submit_prelude |
---|
481 | |
---|
482 | (*sqe){ IORING_OP_WRITEV, fd, iov, iovcnt, offset }; |
---|
483 | |
---|
484 | __submit_wait |
---|
485 | #endif |
---|
486 | } |
---|
487 | |
---|
488 | int cfa_fsync(int fd) { |
---|
489 | #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_FSYNC) |
---|
490 | return fsync(fd); |
---|
491 | #else |
---|
492 | __submit_prelude |
---|
493 | |
---|
494 | (*sqe){ IORING_OP_FSYNC, fd }; |
---|
495 | |
---|
496 | __submit_wait |
---|
497 | #endif |
---|
498 | } |
---|
499 | |
---|
500 | int cfa_sync_file_range(int fd, int64_t offset, int64_t nbytes, unsigned int flags) { |
---|
501 | #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_SYNC_FILE_RANGE) |
---|
502 | return sync_file_range(fd, offset, nbytes, flags); |
---|
503 | #else |
---|
504 | __submit_prelude |
---|
505 | |
---|
506 | (*sqe){ IORING_OP_SYNC_FILE_RANGE, fd }; |
---|
507 | sqe->off = offset; |
---|
508 | sqe->len = nbytes; |
---|
509 | sqe->sync_range_flags = flags; |
---|
510 | |
---|
511 | __submit_wait |
---|
512 | #endif |
---|
513 | } |
---|
514 | |
---|
515 | |
---|
516 | ssize_t cfa_sendmsg(int sockfd, const struct msghdr *msg, int flags) { |
---|
517 | #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_SENDMSG) |
---|
518 | return recv(sockfd, msg, flags); |
---|
519 | #else |
---|
520 | __submit_prelude |
---|
521 | |
---|
522 | (*sqe){ IORING_OP_SENDMSG, sockfd, msg, 1, 0 }; |
---|
523 | sqe->msg_flags = flags; |
---|
524 | |
---|
525 | __submit_wait |
---|
526 | #endif |
---|
527 | } |
---|
528 | |
---|
529 | ssize_t cfa_recvmsg(int sockfd, struct msghdr *msg, int flags) { |
---|
530 | #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_RECVMSG) |
---|
531 | return recv(sockfd, msg, flags); |
---|
532 | #else |
---|
533 | __submit_prelude |
---|
534 | |
---|
535 | (*sqe){ IORING_OP_RECVMSG, sockfd, msg, 1, 0 }; |
---|
536 | sqe->msg_flags = flags; |
---|
537 | |
---|
538 | __submit_wait |
---|
539 | #endif |
---|
540 | } |
---|
541 | |
---|
542 | ssize_t cfa_send(int sockfd, const void *buf, size_t len, int flags) { |
---|
543 | #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_SEND) |
---|
544 | return send( sockfd, buf, len, flags ); |
---|
545 | #else |
---|
546 | __submit_prelude |
---|
547 | |
---|
548 | (*sqe){ IORING_OP_SEND, sockfd }; |
---|
549 | sqe->addr = (uint64_t)buf; |
---|
550 | sqe->len = len; |
---|
551 | sqe->msg_flags = flags; |
---|
552 | |
---|
553 | __submit_wait |
---|
554 | #endif |
---|
555 | } |
---|
556 | |
---|
557 | ssize_t cfa_recv(int sockfd, void *buf, size_t len, int flags) { |
---|
558 | #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_RECV) |
---|
559 | return recv( sockfd, buf, len, flags ); |
---|
560 | #else |
---|
561 | __submit_prelude |
---|
562 | |
---|
563 | (*sqe){ IORING_OP_RECV, sockfd }; |
---|
564 | sqe->addr = (uint64_t)buf; |
---|
565 | sqe->len = len; |
---|
566 | sqe->msg_flags = flags; |
---|
567 | |
---|
568 | __submit_wait |
---|
569 | #endif |
---|
570 | } |
---|
571 | |
---|
572 | int cfa_accept4(int sockfd, struct sockaddr *addr, socklen_t *addrlen, int flags) { |
---|
573 | #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_ACCEPT) |
---|
574 | return accept4( sockfd, addr, addrlen, flags ); |
---|
575 | #else |
---|
576 | __submit_prelude |
---|
577 | |
---|
578 | (*sqe){ IORING_OP_ACCEPT, sockfd }; |
---|
579 | sqe->addr = addr; |
---|
580 | sqe->addr2 = addrlen; |
---|
581 | sqe->accept_flags = flags; |
---|
582 | |
---|
583 | __submit_wait |
---|
584 | #endif |
---|
585 | } |
---|
586 | |
---|
587 | int cfa_connect(int sockfd, const struct sockaddr *addr, socklen_t addrlen) { |
---|
588 | #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_CONNECT) |
---|
589 | return connect( sockfd, addr, addrlen ); |
---|
590 | #else |
---|
591 | __submit_prelude |
---|
592 | |
---|
593 | (*sqe){ IORING_OP_CONNECT, sockfd }; |
---|
594 | sqe->addr = (uint64_t)addr; |
---|
595 | sqe->off = addrlen; |
---|
596 | |
---|
597 | __submit_wait |
---|
598 | #endif |
---|
599 | } |
---|
600 | |
---|
601 | int cfa_fallocate(int fd, int mode, uint64_t offset, uint64_t len) { |
---|
602 | #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_FALLOCATE) |
---|
603 | return fallocate( fd, mode, offset, len ); |
---|
604 | #else |
---|
605 | __submit_prelude |
---|
606 | |
---|
607 | (*sqe){ IORING_OP_FALLOCATE, fd }; |
---|
608 | sqe->off = offset; |
---|
609 | sqe->len = length; |
---|
610 | sqe->mode = mode; |
---|
611 | |
---|
612 | __submit_wait |
---|
613 | #endif |
---|
614 | } |
---|
615 | |
---|
616 | int cfa_fadvise(int fd, uint64_t offset, uint64_t len, int advice) { |
---|
617 | #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_FADVISE) |
---|
618 | return posix_fadvise( fd, offset, len, advice ); |
---|
619 | #else |
---|
620 | __submit_prelude |
---|
621 | |
---|
622 | (*sqe){ IORING_OP_FADVISE, fd }; |
---|
623 | sqe->off = (uint64_t)offset; |
---|
624 | sqe->len = length; |
---|
625 | sqe->fadvise_advice = advice; |
---|
626 | |
---|
627 | __submit_wait |
---|
628 | #endif |
---|
629 | } |
---|
630 | |
---|
631 | int cfa_madvise(void *addr, size_t length, int advice) { |
---|
632 | #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_MADVISE) |
---|
633 | return madvise( addr, length, advice ); |
---|
634 | #else |
---|
635 | __submit_prelude |
---|
636 | |
---|
637 | (*sqe){ IORING_OP_MADVISE, 0 }; |
---|
638 | sqe->addr = (uint64_t)addr; |
---|
639 | sqe->len = length; |
---|
640 | sqe->fadvise_advice = advice; |
---|
641 | |
---|
642 | __submit_wait |
---|
643 | #endif |
---|
644 | } |
---|
645 | |
---|
646 | int cfa_openat(int dirfd, const char *pathname, int flags, mode_t mode) { |
---|
647 | #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_OPENAT) |
---|
648 | return openat( dirfd, pathname, flags, mode ); |
---|
649 | #else |
---|
650 | __submit_prelude |
---|
651 | |
---|
652 | (*sqe){ IORING_OP_OPENAT, dirfd }; |
---|
653 | sqe->addr = (uint64_t)pathname; |
---|
654 | sqe->open_flags = flags; |
---|
655 | sqe->mode = mode; |
---|
656 | |
---|
657 | __submit_wait |
---|
658 | #endif |
---|
659 | } |
---|
660 | |
---|
661 | int cfa_close(int fd) { |
---|
662 | #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_CLOSE) |
---|
663 | return close( fd ); |
---|
664 | #else |
---|
665 | __submit_prelude |
---|
666 | |
---|
667 | (*sqe){ IORING_OP_CLOSE, fd }; |
---|
668 | |
---|
669 | __submit_wait |
---|
670 | #endif |
---|
671 | } |
---|
672 | |
---|
673 | int cfa_statx(int dirfd, const char *pathname, int flags, unsigned int mask, struct statx *statxbuf) { |
---|
674 | #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_STATX) |
---|
675 | //return statx( dirfd, pathname, flags, mask, statxbuf ); |
---|
676 | return syscall( __NR_io_uring_setup, dirfd, pathname, flags, mask, statxbuf ); |
---|
677 | #else |
---|
678 | __submit_prelude |
---|
679 | |
---|
680 | (*sqe){ IORING_OP_STATX, dirfd }; |
---|
681 | sqe->addr = (uint64_t)pathname; |
---|
682 | sqe->statx_flags = flags; |
---|
683 | sqe->len = mask; |
---|
684 | sqe->off = (uint64_t)statxbuf; |
---|
685 | |
---|
686 | __submit_wait |
---|
687 | #endif |
---|
688 | } |
---|
689 | |
---|
690 | |
---|
691 | ssize_t cfa_read(int fd, void *buf, size_t count) { |
---|
692 | #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_READ) |
---|
693 | return read( fd, buf, count ); |
---|
694 | #else |
---|
695 | __submit_prelude |
---|
696 | |
---|
697 | (*sqe){ IORING_OP_READ, fd, buf, count, 0 }; |
---|
698 | |
---|
699 | __submit_wait |
---|
700 | #endif |
---|
701 | } |
---|
702 | |
---|
703 | ssize_t cfa_write(int fd, void *buf, size_t count) { |
---|
704 | #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_WRITE) |
---|
705 | return read( fd, buf, count ); |
---|
706 | #else |
---|
707 | __submit_prelude |
---|
708 | |
---|
709 | (*sqe){ IORING_OP_WRITE, fd, buf, count, 0 }; |
---|
710 | |
---|
711 | __submit_wait |
---|
712 | #endif |
---|
713 | } |
---|
714 | |
---|
715 | //----------------------------------------------------------------------------- |
---|
716 | // Check if a function is asynchronous |
---|
717 | |
---|
718 | // Macro magic to reduce the size of the following switch case |
---|
719 | #define IS_DEFINED_APPLY(f, ...) f(__VA_ARGS__) |
---|
720 | #define IS_DEFINED_SECOND(first, second, ...) second |
---|
721 | #define IS_DEFINED_TEST(expansion) _CFA_IO_FEATURE_##expansion |
---|
722 | #define IS_DEFINED(macro) IS_DEFINED_APPLY( IS_DEFINED_SECOND,IS_DEFINED_TEST(macro) false, true) |
---|
723 | |
---|
724 | bool has_user_level_blocking( fptr_t func ) { |
---|
725 | #if defined(HAVE_LINUX_IO_URING_H) |
---|
726 | if( /*func == (fptr_t)preadv2 || */ |
---|
727 | func == (fptr_t)cfa_preadv2 ) |
---|
728 | #define _CFA_IO_FEATURE_IORING_OP_READV , |
---|
729 | return IS_DEFINED(IORING_OP_READV); |
---|
730 | |
---|
731 | if( /*func == (fptr_t)pwritev2 || */ |
---|
732 | func == (fptr_t)cfa_pwritev2 ) |
---|
733 | #define _CFA_IO_FEATURE_IORING_OP_WRITEV , |
---|
734 | return IS_DEFINED(IORING_OP_WRITEV); |
---|
735 | |
---|
736 | if( /*func == (fptr_t)fsync || */ |
---|
737 | func == (fptr_t)cfa_fsync ) |
---|
738 | #define _CFA_IO_FEATURE_IORING_OP_FSYNC , |
---|
739 | return IS_DEFINED(IORING_OP_FSYNC); |
---|
740 | |
---|
741 | if( /*func == (fptr_t)ync_file_range || */ |
---|
742 | func == (fptr_t)cfa_sync_file_range ) |
---|
743 | #define _CFA_IO_FEATURE_IORING_OP_SYNC_FILE_RANGE , |
---|
744 | return IS_DEFINED(IORING_OP_SYNC_FILE_RANGE); |
---|
745 | |
---|
746 | if( /*func == (fptr_t)sendmsg || */ |
---|
747 | func == (fptr_t)cfa_sendmsg ) |
---|
748 | #define _CFA_IO_FEATURE_IORING_OP_SENDMSG , |
---|
749 | return IS_DEFINED(IORING_OP_SENDMSG); |
---|
750 | |
---|
751 | if( /*func == (fptr_t)recvmsg || */ |
---|
752 | func == (fptr_t)cfa_recvmsg ) |
---|
753 | #define _CFA_IO_FEATURE_IORING_OP_RECVMSG , |
---|
754 | return IS_DEFINED(IORING_OP_RECVMSG); |
---|
755 | |
---|
756 | if( /*func == (fptr_t)send || */ |
---|
757 | func == (fptr_t)cfa_send ) |
---|
758 | #define _CFA_IO_FEATURE_IORING_OP_SEND , |
---|
759 | return IS_DEFINED(IORING_OP_SEND); |
---|
760 | |
---|
761 | if( /*func == (fptr_t)recv || */ |
---|
762 | func == (fptr_t)cfa_recv ) |
---|
763 | #define _CFA_IO_FEATURE_IORING_OP_RECV , |
---|
764 | return IS_DEFINED(IORING_OP_RECV); |
---|
765 | |
---|
766 | if( /*func == (fptr_t)accept4 || */ |
---|
767 | func == (fptr_t)cfa_accept4 ) |
---|
768 | #define _CFA_IO_FEATURE_IORING_OP_ACCEPT , |
---|
769 | return IS_DEFINED(IORING_OP_ACCEPT); |
---|
770 | |
---|
771 | if( /*func == (fptr_t)connect || */ |
---|
772 | func == (fptr_t)cfa_connect ) |
---|
773 | #define _CFA_IO_FEATURE_IORING_OP_CONNECT , |
---|
774 | return IS_DEFINED(IORING_OP_CONNECT); |
---|
775 | |
---|
776 | if( /*func == (fptr_t)fallocate || */ |
---|
777 | func == (fptr_t)cfa_fallocate ) |
---|
778 | #define _CFA_IO_FEATURE_IORING_OP_FALLOCATE , |
---|
779 | return IS_DEFINED(IORING_OP_FALLOCATE); |
---|
780 | |
---|
781 | if( /*func == (fptr_t)posix_fadvise || */ |
---|
782 | func == (fptr_t)cfa_fadvise ) |
---|
783 | #define _CFA_IO_FEATURE_IORING_OP_FADVISE , |
---|
784 | return IS_DEFINED(IORING_OP_FADVISE); |
---|
785 | |
---|
786 | if( /*func == (fptr_t)madvise || */ |
---|
787 | func == (fptr_t)cfa_madvise ) |
---|
788 | #define _CFA_IO_FEATURE_IORING_OP_MADVISE , |
---|
789 | return IS_DEFINED(IORING_OP_MADVISE); |
---|
790 | |
---|
791 | if( /*func == (fptr_t)openat || */ |
---|
792 | func == (fptr_t)cfa_openat ) |
---|
793 | #define _CFA_IO_FEATURE_IORING_OP_OPENAT , |
---|
794 | return IS_DEFINED(IORING_OP_OPENAT); |
---|
795 | |
---|
796 | if( /*func == (fptr_t)close || */ |
---|
797 | func == (fptr_t)cfa_close ) |
---|
798 | #define _CFA_IO_FEATURE_IORING_OP_CLOSE , |
---|
799 | return IS_DEFINED(IORING_OP_CLOSE); |
---|
800 | |
---|
801 | if( /*func == (fptr_t)statx || */ |
---|
802 | func == (fptr_t)cfa_statx ) |
---|
803 | #define _CFA_IO_FEATURE_IORING_OP_STATX , |
---|
804 | return IS_DEFINED(IORING_OP_STATX); |
---|
805 | |
---|
806 | if( /*func == (fptr_t)read || */ |
---|
807 | func == (fptr_t)cfa_read ) |
---|
808 | #define _CFA_IO_FEATURE_IORING_OP_READ , |
---|
809 | return IS_DEFINED(IORING_OP_READ); |
---|
810 | |
---|
811 | if( /*func == (fptr_t)write || */ |
---|
812 | func == (fptr_t)cfa_write ) |
---|
813 | #define _CFA_IO_FEATURE_IORING_OP_WRITE , |
---|
814 | return IS_DEFINED(IORING_OP_WRITE); |
---|
815 | #endif |
---|
816 | |
---|
817 | return false; |
---|
818 | } |
---|