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 1.14 by root, Sat Dec 28 05:20:17 2019 UTC vs.
Revision 1.19 by root, Sat Dec 28 07:58:51 2019 UTC

… …
44 * b) best is not necessarily very good. 44 * b) best is not necessarily very good.
45 * c) it's better than the aio mess, doesn't suffer from the fork problems 45 * c) it's better than the aio mess, doesn't suffer from the fork problems
46 * of linux aio or epoll and so on and so on. and you could do event stuff 46 * of linux aio or epoll and so on and so on. and you could do event stuff
47 * without any syscalls. what's not to like? 47 * without any syscalls. what's not to like?
48 * d) ok, it's vastly more complex, but that's ok, really. 48 * d) ok, it's vastly more complex, but that's ok, really.
49 * e) why 3 mmaps instead of one? one would be more space-efficient, 49 * e) why two mmaps instead of one? one would be more space-efficient,
50 * and I can't see what benefit three would have (other than being 50 * and I can't see what benefit two would have (other than being
51 * somehow resizable/relocatable, but that's apparently not possible). 51 * somehow resizable/relocatable, but that's apparently not possible).
52 * (FIXME: newer kernels can use 2 mmaps only, need to look into this).
53 * f) hmm, it's practiclaly undebuggable (gdb can't access the memory, and 52 * f) hmm, it's practically undebuggable (gdb can't access the memory, and
54 * the bizarre way structure offsets are communicated makes it hard to 53 * the bizarre way structure offsets are communicated makes it hard to
55 * just print the ring buffer heads, even *iff* the memory were visible 54 * just print the ring buffer heads, even *iff* the memory were visible
56 * in gdb. but then, that's also ok, really. 55 * in gdb. but then, that's also ok, really.
57 * g) well, you cannot specify a timeout when waiting for events. no, 56 * g) well, you cannot specify a timeout when waiting for events. no,
58 * seriously, the interface doesn't support a timeout. never seen _that_ 57 * seriously, the interface doesn't support a timeout. never seen _that_
59 * before. sure, you can use a timerfd, but that's another syscall 58 * before. sure, you can use a timerfd, but that's another syscall
60 * you could have avoided. overall, this bizarre omission smells 59 * you could have avoided. overall, this bizarre omission smells
61 * like a µ-optimisation by the io_uring author for his personal 60 * like a µ-optimisation by the io_uring author for his personal
62 * applications, to the detriment of everybody else who just wants 61 * applications, to the detriment of everybody else who just wants
63 * an event loop. but, umm, ok, if that's all, it could be worse. 62 * an event loop. but, umm, ok, if that's all, it could be worse.
64 * (FIXME: jens mentioned timeout commands, need to investigate) 63 * (from what I gather from the author Jens Axboe, it simply didn't
64 * occur to him, and he made good on it by adding an unlimited nuber
65 * of timeouts later :).
65 * h) there is a hardcoded limit of 4096 outstanding events. okay, 66 * h) initially there was a hardcoded limit of 4096 outstanding events.
66 * at least there is no arbitrary low system-wide limit... 67 * later versions not only bump this to 32k, but also can handle
67 * (FIXME: apparently, this was increased to 32768 in later kernels( 68 * an unlimited amount of events, so this only affects the batch size.
68 * i) unlike linux aio, you *can* register more then the limit 69 * i) unlike linux aio, you *can* register more then the limit
69 * of fd events, and the kernel will "gracefully" signal an 70 * of fd events. while early verisons of io_uring signalled an overflow
70 * overflow, after which you could destroy and recreate the kernel 71 * and you ended up getting wet. 5.5+ does not do this anymore.
71 * state, a bit bigger, or fall back to e.g. poll. thats not
72 * totally insane, but kind of questions the point a high
73 * performance I/O framework when it doesn't really work
74 * under stress.
75 * (FIXME: iouring should no longer drop events, need to investigate)
76 * j) but, oh my! is has exactly the same bugs as the linux aio backend, 72 * j) but, oh my! it had exactly the same bugs as the linux aio backend,
77 * where some undocumented poll combinations just fail. 73 * where some undocumented poll combinations just fail. fortunately,
78 * so we need epoll AGAIN as a fallback. AGAIN! epoll!! and of course, 74 * after finally reaching the author, he was more than willing to fix
79 * this is completely undocumented, have I mantioned this already? 75 * this probably in 5.6+.
80 * k) overall, the *API* itself is, I dare to say, not a total trainwreck. 76 * k) overall, the *API* itself is, I dare to say, not a total trainwreck.
81 * the big isuess with it are the bugs requiring epoll, which might 77 * once the bugs ae fixed (probably in 5.6+), it will be without
82 * or might not get fixed (do I hold my breath?). 78 * competition.
83 */ 79 */
84 80
85/* TODO: use internal TIMEOUT */ 81/* TODO: use internal TIMEOUT */
86/* TODO: take advantage of single mmap, NODROP etc. */ 82/* TODO: take advantage of single mmap, NODROP etc. */
87/* TODO: resize cq/sq size independently */ 83/* TODO: resize cq/sq size independently */
… …
228 224
229/* the submit/completion queue entries */ 225/* the submit/completion queue entries */
230#define EV_SQES ((struct io_uring_sqe *) iouring_sqes) 226#define EV_SQES ((struct io_uring_sqe *) iouring_sqes)
231#define EV_CQES ((struct io_uring_cqe *)((char *)iouring_cq_ring + iouring_cq_cqes)) 227#define EV_CQES ((struct io_uring_cqe *)((char *)iouring_cq_ring + iouring_cq_cqes))
232 228
233/* TODO: this is not enough, we might have to reap events */ 229inline_speed
234/* TODO: but we can't, as that will re-arm events, causing */ 230int
235/* TODO: an endless loop in fd_reify */
236static int
237iouring_enter (EV_P_ ev_tstamp timeout) 231iouring_enter (EV_P_ ev_tstamp timeout)
238{ 232{
239 int res; 233 int res;
240 234
241 EV_RELEASE_CB; 235 EV_RELEASE_CB;
… …
249 243
250 EV_ACQUIRE_CB; 244 EV_ACQUIRE_CB;
251 245
252 return res; 246 return res;
253} 247}
248
249/* TODO: can we move things around so we don't need this forward-reference? */
250static void
251iouring_poll (EV_P_ ev_tstamp timeout);
254 252
255static 253static
256struct io_uring_sqe * 254struct io_uring_sqe *
257iouring_sqe_get (EV_P) 255iouring_sqe_get (EV_P)
258{ 256{
257 unsigned tail;
258
259 for (;;)
260 {
259 unsigned tail = EV_SQ_VAR (tail); 261 tail = EV_SQ_VAR (tail);
260 262
261 while (ecb_expect_false (tail + 1 - EV_SQ_VAR (head) > EV_SQ_VAR (ring_entries))) 263 if (ecb_expect_true (tail + 1 - EV_SQ_VAR (head) <= EV_SQ_VAR (ring_entries)))
262 { 264 break; /* whats the problem, we have free sqes */
263 /* queue full, need to flush */
264 265
266 /* queue full, need to flush and possibly handle some events */
267
268#if EV_FEATURE_CODE
269 /* first we ask the kernel nicely, most often this frees up some sqes */
265 int res = iouring_enter (EV_A_ EV_TS_CONST (0.)); 270 int res = iouring_enter (EV_A_ EV_TS_CONST (0.));
266 271
267 /* io_uring_enter might fail with EBUSY and won't submit anything */ 272 ECB_MEMORY_FENCE_ACQUIRE; /* better safe than sorry */
268 /* unfortunately, we can't handle this at the moment */
269 273
270 if (res < 0 && errno == EBUSY) 274 if (res >= 0)
271 //TODO 275 continue; /* yes, it worked, try again */
272 ev_syserr ("(libev) io_uring_enter could not clear sq"); 276#endif
273 else 277
274 break; 278 /* some problem, possibly EBUSY - do the full poll and let it handle any issues */
275 279
280 iouring_poll (EV_A_ EV_TS_CONST (0.));
276 /* iouring_poll should have done ECB_MEMORY_FENCE_ACQUIRE */ 281 /* iouring_poll should have done ECB_MEMORY_FENCE_ACQUIRE for us */
277 } 282 }
278 283
279 /*assert (("libev: io_uring queue full after flush", tail + 1 - EV_SQ_VAR (head) <= EV_SQ_VAR (ring_entries)));*/ 284 /*assert (("libev: io_uring queue full after flush", tail + 1 - EV_SQ_VAR (head) <= EV_SQ_VAR (ring_entries)));*/
280 285
281 return EV_SQES + (tail & EV_SQ_VAR (ring_mask)); 286 return EV_SQES + (tail & EV_SQ_VAR (ring_mask));
… …
350 355
351 if (errno != EINVAL) 356 if (errno != EINVAL)
352 return -1; /* we failed */ 357 return -1; /* we failed */
353 358
354#if TODO 359#if TODO
355 if ((~params.features) & (IORING_FEAT_NODROP | IORING_FEATURE_SINGLE_MMAP)) 360 if ((~params.features) & (IORING_FEAT_NODROP | IORING_FEATURE_SINGLE_MMAP | IORING_FEAT_SUBMIT_STABLE))
356 return -1; /* we require the above features */ 361 return -1; /* we require the above features */
357#endif 362#endif
358 363
359 /* EINVAL: lots of possible reasons, but maybe 364 /* EINVAL: lots of possible reasons, but maybe
360 * it is because we hit the unqueryable hardcoded size limit 365 * it is because we hit the unqueryable hardcoded size limit
… …
438 /* Jens Axboe notified me that user_data is not what is documented, but is 443 /* 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 444 * 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 445 * 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 446 * generation counter - if that fails, too bad, it will hopefully be removed
442 * at close time and then be ignored. */ 447 * at close time and then be ignored. */
443 sqe->user_data = (uint32_t)fd | ((__u64)(uint32_t)anfds [fd].egen << 32); 448 sqe->addr = (uint32_t)fd | ((__u64)(uint32_t)anfds [fd].egen << 32);
449 sqe->user_data = (uint64_t)-1;
444 iouring_sqe_submit (EV_A_ sqe); 450 iouring_sqe_submit (EV_A_ sqe);
445 451
446 /* increment generation counter to avoid handling old events */ 452 /* increment generation counter to avoid handling old events */
447 ++anfds [fd].egen; 453 ++anfds [fd].egen;
448 } 454 }
… …
488iouring_process_cqe (EV_P_ struct io_uring_cqe *cqe) 494iouring_process_cqe (EV_P_ struct io_uring_cqe *cqe)
489{ 495{
490 int fd = cqe->user_data & 0xffffffffU; 496 int fd = cqe->user_data & 0xffffffffU;
491 uint32_t gen = cqe->user_data >> 32; 497 uint32_t gen = cqe->user_data >> 32;
492 int res = cqe->res; 498 int res = cqe->res;
499
500 /* user_data -1 is a remove that we are not atm. interested in */
501 if (cqe->user_data == (uint64_t)-1)
502 return;
493 503
494 assert (("libev: io_uring fd must be in-bounds", fd >= 0 && fd < anfdmax)); 504 assert (("libev: io_uring fd must be in-bounds", fd >= 0 && fd < anfdmax));
495 505
496 /* documentation lies, of course. the result value is NOT like 506 /* documentation lies, of course. the result value is NOT like
497 * normal syscalls, but like linux raw syscalls, i.e. negative 507 * normal syscalls, but like linux raw syscalls, i.e. negative
… …
622 632
623static void 633static void
624iouring_poll (EV_P_ ev_tstamp timeout) 634iouring_poll (EV_P_ ev_tstamp timeout)
625{ 635{
626 /* if we have events, no need for extra syscalls, but we might have to queue events */ 636 /* if we have events, no need for extra syscalls, but we might have to queue events */
637 /* we also clar the timeout if there are outstanding fdchanges */
638 /* the latter should only happen if both the sq and cq are full, most likely */
639 /* because we have a lot of event sources that immediately complete */
640 /* TODO: fdchacngecnt is always 0 because fd_reify does not have two buffers yet */
627 if (iouring_handle_cq (EV_A)) 641 if (iouring_handle_cq (EV_A) || fdchangecnt)
628 timeout = EV_TS_CONST (0.); 642 timeout = EV_TS_CONST (0.);
629 else 643 else
630 /* no events, so maybe wait for some */ 644 /* no events, so maybe wait for some */
631 iouring_tfd_update (EV_A_ timeout); 645 iouring_tfd_update (EV_A_ timeout);
632 646

Diff Legend

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