ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libev/ev_iouring.c
(Generate patch)

Comparing libev/ev_iouring.c (file contents):
Revision 1.10 by root, Fri Dec 27 21:17:11 2019 UTC vs.
Revision 1.14 by root, Sat Dec 28 05:20:17 2019 UTC

… …
87/* TODO: resize cq/sq size independently */ 87/* TODO: resize cq/sq size independently */
88 88
89#include <sys/timerfd.h> 89#include <sys/timerfd.h>
90#include <sys/mman.h> 90#include <sys/mman.h>
91#include <poll.h> 91#include <poll.h>
92#include <stdint.h>
92 93
93#define IOURING_INIT_ENTRIES 32 94#define IOURING_INIT_ENTRIES 32
94 95
95/*****************************************************************************/ 96/*****************************************************************************/
96/* syscall wrapdadoop - this section has the raw api/abi definitions */ 97/* syscall wrapdadoop - this section has the raw api/abi definitions */
… …
173 __u32 resv[4]; 174 __u32 resv[4];
174 struct io_sqring_offsets sq_off; 175 struct io_sqring_offsets sq_off;
175 struct io_cqring_offsets cq_off; 176 struct io_cqring_offsets cq_off;
176}; 177};
177 178
179#define IORING_SETUP_CQSIZE 0x00000008
180
178#define IORING_OP_POLL_ADD 6 181#define IORING_OP_POLL_ADD 6
179#define IORING_OP_POLL_REMOVE 7 182#define IORING_OP_POLL_REMOVE 7
183#define IORING_OP_TIMEOUT 11
184#define IORING_OP_TIMEOUT_REMOVE 12
185
186/* relative or absolute, reference clock is CLOCK_MONOTONIC */
187struct iouring_kernel_timespec
188{
189 int64_t tv_sec;
190 long long tv_nsec;
191};
192
193#define IORING_TIMEOUT_ABS 0x00000001
180 194
181#define IORING_ENTER_GETEVENTS 0x01 195#define IORING_ENTER_GETEVENTS 0x01
182 196
183#define IORING_OFF_SQ_RING 0x00000000ULL 197#define IORING_OFF_SQ_RING 0x00000000ULL
184#define IORING_OFF_CQ_RING 0x08000000ULL 198#define IORING_OFF_CQ_RING 0x08000000ULL
185#define IORING_OFF_SQES 0x10000000ULL 199#define IORING_OFF_SQES 0x10000000ULL
186 200
187#define IORING_FEAT_SINGLE_MMAP 0x1 201#define IORING_FEAT_SINGLE_MMAP 0x00000001
188#define IORING_FEAT_NODROP 0x2 202#define IORING_FEAT_NODROP 0x00000002
189#define IORING_FEAT_SUBMIT_STABLE 0x4 203#define IORING_FEAT_SUBMIT_STABLE 0x00000004
190 204
191inline_size 205inline_size
192int 206int
193evsys_io_uring_setup (unsigned entries, struct io_uring_params *params) 207evsys_io_uring_setup (unsigned entries, struct io_uring_params *params)
194{ 208{
… …
214 228
215/* the submit/completion queue entries */ 229/* the submit/completion queue entries */
216#define EV_SQES ((struct io_uring_sqe *) iouring_sqes) 230#define EV_SQES ((struct io_uring_sqe *) iouring_sqes)
217#define EV_CQES ((struct io_uring_cqe *)((char *)iouring_cq_ring + iouring_cq_cqes)) 231#define EV_CQES ((struct io_uring_cqe *)((char *)iouring_cq_ring + iouring_cq_cqes))
218 232
233/* TODO: this is not enough, we might have to reap events */
234/* TODO: but we can't, as that will re-arm events, causing */
235/* TODO: an endless loop in fd_reify */
236static int
237iouring_enter (EV_P_ ev_tstamp timeout)
238{
239 int res;
240
241 EV_RELEASE_CB;
242
243 res = evsys_io_uring_enter (iouring_fd, iouring_to_submit, 1,
244 timeout > EV_TS_CONST (0.) ? IORING_ENTER_GETEVENTS : 0, 0, 0);
245
246 assert (("libev: io_uring_enter did not consume all sqes", (res < 0 || res == iouring_to_submit)));
247
248 iouring_to_submit = 0;
249
250 EV_ACQUIRE_CB;
251
252 return res;
253}
254
219static 255static
220struct io_uring_sqe * 256struct io_uring_sqe *
221iouring_sqe_get (EV_P) 257iouring_sqe_get (EV_P)
222{ 258{
223 unsigned tail = EV_SQ_VAR (tail); 259 unsigned tail = EV_SQ_VAR (tail);
224 260
225 if (tail + 1 - EV_SQ_VAR (head) > EV_SQ_VAR (ring_entries)) 261 while (ecb_expect_false (tail + 1 - EV_SQ_VAR (head) > EV_SQ_VAR (ring_entries)))
226 { 262 {
227 /* queue full, flush */ 263 /* queue full, need to flush */
228 evsys_io_uring_enter (iouring_fd, iouring_to_submit, 0, 0, 0, 0);
229 iouring_to_submit = 0;
230 }
231 264
265 int res = iouring_enter (EV_A_ EV_TS_CONST (0.));
266
267 /* io_uring_enter might fail with EBUSY and won't submit anything */
268 /* unfortunately, we can't handle this at the moment */
269
270 if (res < 0 && errno == EBUSY)
271 //TODO
272 ev_syserr ("(libev) io_uring_enter could not clear sq");
273 else
274 break;
275
276 /* iouring_poll should have done ECB_MEMORY_FENCE_ACQUIRE */
277 }
278
232 assert (("libev: io_uring queue full after flush", tail + 1 - EV_SQ_VAR (head) <= EV_SQ_VAR (ring_entries))); 279 /*assert (("libev: io_uring queue full after flush", tail + 1 - EV_SQ_VAR (head) <= EV_SQ_VAR (ring_entries)));*/
233 280
234 return EV_SQES + (tail & EV_SQ_VAR (ring_mask)); 281 return EV_SQES + (tail & EV_SQ_VAR (ring_mask));
235} 282}
236 283
237inline_size 284inline_size
… …
289 iouring_tfd = -1; 336 iouring_tfd = -1;
290 iouring_sq_ring = MAP_FAILED; 337 iouring_sq_ring = MAP_FAILED;
291 iouring_cq_ring = MAP_FAILED; 338 iouring_cq_ring = MAP_FAILED;
292 iouring_sqes = MAP_FAILED; 339 iouring_sqes = MAP_FAILED;
293 340
341 if (!have_monotonic) /* cannot really happen, but what if11 */
342 return -1;
343
294 for (;;) 344 for (;;)
295 { 345 {
296 iouring_fd = evsys_io_uring_setup (iouring_entries, &params); 346 iouring_fd = evsys_io_uring_setup (iouring_entries, &params);
297 347
298 if (iouring_fd >= 0) 348 if (iouring_fd >= 0)
299 break; /* yippie */ 349 break; /* yippie */
300 350
301 if (errno != EINVAL) 351 if (errno != EINVAL)
302 return -1; /* we failed */ 352 return -1; /* we failed */
353
354#if TODO
355 if ((~params.features) & (IORING_FEAT_NODROP | IORING_FEATURE_SINGLE_MMAP))
356 return -1; /* we require the above features */
357#endif
303 358
304 /* EINVAL: lots of possible reasons, but maybe 359 /* EINVAL: lots of possible reasons, but maybe
305 * it is because we hit the unqueryable hardcoded size limit 360 * it is because we hit the unqueryable hardcoded size limit
306 */ 361 */
307 362
… …
378 { 433 {
379 /* we assume the sqe's are all "properly" initialised */ 434 /* we assume the sqe's are all "properly" initialised */
380 struct io_uring_sqe *sqe = iouring_sqe_get (EV_A); 435 struct io_uring_sqe *sqe = iouring_sqe_get (EV_A);
381 sqe->opcode = IORING_OP_POLL_REMOVE; 436 sqe->opcode = IORING_OP_POLL_REMOVE;
382 sqe->fd = fd; 437 sqe->fd = fd;
383 sqe->user_data = -1; 438 /* Jens Axboe notified me that user_data is not what is documented, but is
439 * some kind of unique ID that has to match, otherwise the request cannot
440 * be removed. Since we don't *really* have that, we pass in the old
441 * generation counter - if that fails, too bad, it will hopefully be removed
442 * at close time and then be ignored. */
443 sqe->user_data = (uint32_t)fd | ((__u64)(uint32_t)anfds [fd].egen << 32);
384 iouring_sqe_submit (EV_A_ sqe); 444 iouring_sqe_submit (EV_A_ sqe);
385 445
386 /* increment generation counter to avoid handling old events */ 446 /* increment generation counter to avoid handling old events */
387 ++anfds [fd].egen; 447 ++anfds [fd].egen;
388 } 448 }
… …
429{ 489{
430 int fd = cqe->user_data & 0xffffffffU; 490 int fd = cqe->user_data & 0xffffffffU;
431 uint32_t gen = cqe->user_data >> 32; 491 uint32_t gen = cqe->user_data >> 32;
432 int res = cqe->res; 492 int res = cqe->res;
433 493
434 /* ignore fd removal events, if there are any. TODO: verify */
435 /* TODO: yes, this triggers */
436 if (cqe->user_data == (__u64)-1)
437 return;
438
439 assert (("libev: io_uring fd must be in-bounds", fd >= 0 && fd < anfdmax)); 494 assert (("libev: io_uring fd must be in-bounds", fd >= 0 && fd < anfdmax));
440 495
441 /* documentation lies, of course. the result value is NOT like 496 /* documentation lies, of course. the result value is NOT like
442 * normal syscalls, but like linux raw syscalls, i.e. negative 497 * normal syscalls, but like linux raw syscalls, i.e. negative
443 * error numbers. fortunate, as otherwise there would be no way 498 * error numbers. fortunate, as otherwise there would be no way
444 * to get error codes at all. still, why not document this? 499 * to get error codes at all. still, why not document this?
445 */ 500 */
446 501
447 /* ignore event if generation doesn't match */ 502 /* ignore event if generation doesn't match */
503 /* other than skipping removal events, */
448 /* this should actually be very rare */ 504 /* this should actually be very rare */
449 if (ecb_expect_false (gen != (uint32_t)anfds [fd].egen)) 505 if (ecb_expect_false (gen != (uint32_t)anfds [fd].egen))
450 return; 506 return;
451 507
452 if (ecb_expect_false (res < 0)) 508 if (ecb_expect_false (res < 0))
453 { 509 {
454 //TODO: EINVAL handling (was something failed with this fd) 510 /*TODO: EINVAL handling (was something failed with this fd)*/
455 //TODO: EBUSY happens when? 511 /*TODO: EBUSY happens when?*/
456 512
457 if (res == -EBADF) 513 if (res == -EBADF)
458 { 514 {
459 assert (("libev: event loop rejected bad fd", res != -EBADF)); 515 assert (("libev: event loop rejected bad fd", res != -EBADF));
460 fd_kill (EV_A_ fd); 516 fd_kill (EV_A_ fd);
… …
575 iouring_tfd_update (EV_A_ timeout); 631 iouring_tfd_update (EV_A_ timeout);
576 632
577 /* only enter the kernel if we have something to submit, or we need to wait */ 633 /* only enter the kernel if we have something to submit, or we need to wait */
578 if (timeout || iouring_to_submit) 634 if (timeout || iouring_to_submit)
579 { 635 {
580 int res; 636 int res = iouring_enter (EV_A_ timeout);
581
582 EV_RELEASE_CB;
583
584 res = evsys_io_uring_enter (iouring_fd, iouring_to_submit, 1,
585 timeout > EV_TS_CONST (0.) ? IORING_ENTER_GETEVENTS : 0, 0, 0);
586 iouring_to_submit = 0;
587
588 EV_ACQUIRE_CB;
589 637
590 if (ecb_expect_false (res < 0)) 638 if (ecb_expect_false (res < 0))
591 if (errno == EINTR) 639 if (errno == EINTR)
592 /* ignore */; 640 /* ignore */;
641 else if (errno == EBUSY)
642 /* cq full, cannot submit - should be rare because we flush the cq first, so simply ignore */;
593 else 643 else
594 ev_syserr ("(libev) iouring setup"); 644 ev_syserr ("(libev) iouring setup");
595 else 645 else
596 iouring_handle_cq (EV_A); 646 iouring_handle_cq (EV_A);
597 } 647 }

Diff Legend

– Removed lines
+ Added lines
< Changed lines
> Changed lines