ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libev/ev_iouring.c
Revision: 1.17
Committed: Sat Dec 28 07:39:18 2019 UTC (4 years, 6 months ago) by root
Content type: text/plain
Branch: MAIN
Changes since 1.16: +1 -1 lines
Log Message:
*** empty log message ***

File Contents

# User Rev Content
1 root 1.1 /*
2     * libev linux io_uring fd activity backend
3     *
4     * Copyright (c) 2019 Marc Alexander Lehmann <libev@schmorp.de>
5     * All rights reserved.
6     *
7     * Redistribution and use in source and binary forms, with or without modifica-
8     * tion, are permitted provided that the following conditions are met:
9     *
10     * 1. Redistributions of source code must retain the above copyright notice,
11     * this list of conditions and the following disclaimer.
12     *
13     * 2. Redistributions in binary form must reproduce the above copyright
14     * notice, this list of conditions and the following disclaimer in the
15     * documentation and/or other materials provided with the distribution.
16     *
17     * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
18     * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MER-
19     * CHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO
20     * EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPE-
21     * CIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
22     * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
23     * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
24     * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTH-
25     * ERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
26     * OF THE POSSIBILITY OF SUCH DAMAGE.
27     *
28     * Alternatively, the contents of this file may be used under the terms of
29     * the GNU General Public License ("GPL") version 2 or any later version,
30     * in which case the provisions of the GPL are applicable instead of
31     * the above. If you wish to allow the use of your version of this file
32     * only under the terms of the GPL and not to allow others to use your
33     * version of this file under the BSD license, indicate your decision
34     * by deleting the provisions above and replace them with the notice
35     * and other provisions required by the GPL. If you do not delete the
36     * provisions above, a recipient may use your version of this file under
37     * either the BSD or the GPL.
38     */
39    
40     /*
41     * general notes about linux io_uring:
42     *
43     * a) it's the best interface I have seen so far. on linux.
44     * b) best is not necessarily very good.
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
47     * without any syscalls. what's not to like?
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,
50     * and I can't see what benefit three would have (other than being
51     * somehow resizable/relocatable, but that's apparently not possible).
52 root 1.7 * (FIXME: newer kernels can use 2 mmaps only, need to look into this).
53 root 1.1 * f) hmm, it's practiclaly undebuggable (gdb can't access the memory, and
54 root 1.7 * the bizarre way structure offsets are communicated makes it hard to
55 root 1.1 * just print the ring buffer heads, even *iff* the memory were visible
56     * in gdb. but then, that's also ok, really.
57     * g) well, you cannot specify a timeout when waiting for events. no,
58     * seriously, the interface doesn't support a timeout. never seen _that_
59     * before. sure, you can use a timerfd, but that's another syscall
60     * you could have avoided. overall, this bizarre omission smells
61     * like a ยต-optimisation by the io_uring author for his personal
62     * 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.
64 root 1.7 * (FIXME: jens mentioned timeout commands, need to investigate)
65 root 1.1 * h) there is a hardcoded limit of 4096 outstanding events. okay,
66     * at least there is no arbitrary low system-wide limit...
67 root 1.7 * (FIXME: apparently, this was increased to 32768 in later kernels(
68 root 1.1 * i) unlike linux aio, you *can* register more then the limit
69     * of fd events, and the kernel will "gracefully" signal an
70     * overflow, after which you could destroy and recreate the kernel
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 root 1.7 * (FIXME: iouring should no longer drop events, need to investigate)
76 root 1.1 * j) but, oh my! is has exactly the same bugs as the linux aio backend,
77     * where some undocumented poll combinations just fail.
78     * so we need epoll AGAIN as a fallback. AGAIN! epoll!! and of course,
79     * this is completely undocumented, have I mantioned this already?
80     * 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
82     * or might not get fixed (do I hold my breath?).
83     */
84    
85 root 1.9 /* TODO: use internal TIMEOUT */
86     /* TODO: take advantage of single mmap, NODROP etc. */
87     /* TODO: resize cq/sq size independently */
88    
89 root 1.1 #include <sys/timerfd.h>
90     #include <sys/mman.h>
91     #include <poll.h>
92 root 1.12 #include <stdint.h>
93 root 1.1
94     #define IOURING_INIT_ENTRIES 32
95    
96     /*****************************************************************************/
97     /* syscall wrapdadoop - this section has the raw api/abi definitions */
98    
99     #include <linux/fs.h>
100     #include <linux/types.h>
101    
102     /* mostly directly taken from the kernel or documentation */
103    
104     struct io_uring_sqe
105     {
106     __u8 opcode;
107     __u8 flags;
108     __u16 ioprio;
109     __s32 fd;
110 root 1.8 union {
111     __u64 off;
112     __u64 addr2;
113     };
114 root 1.1 __u64 addr;
115     __u32 len;
116     union {
117     __kernel_rwf_t rw_flags;
118     __u32 fsync_flags;
119     __u16 poll_events;
120     __u32 sync_range_flags;
121     __u32 msg_flags;
122 root 1.8 __u32 timeout_flags;
123     __u32 accept_flags;
124     __u32 cancel_flags;
125     __u32 open_flags;
126     __u32 statx_flags;
127 root 1.1 };
128     __u64 user_data;
129     union {
130     __u16 buf_index;
131     __u64 __pad2[3];
132     };
133     };
134    
135     struct io_uring_cqe
136     {
137     __u64 user_data;
138     __s32 res;
139     __u32 flags;
140     };
141    
142     struct io_sqring_offsets
143     {
144     __u32 head;
145     __u32 tail;
146     __u32 ring_mask;
147     __u32 ring_entries;
148     __u32 flags;
149     __u32 dropped;
150     __u32 array;
151     __u32 resv1;
152     __u64 resv2;
153     };
154    
155     struct io_cqring_offsets
156     {
157     __u32 head;
158     __u32 tail;
159     __u32 ring_mask;
160     __u32 ring_entries;
161     __u32 overflow;
162     __u32 cqes;
163     __u64 resv[2];
164     };
165    
166     struct io_uring_params
167     {
168     __u32 sq_entries;
169     __u32 cq_entries;
170     __u32 flags;
171     __u32 sq_thread_cpu;
172     __u32 sq_thread_idle;
173 root 1.8 __u32 features;
174     __u32 resv[4];
175 root 1.1 struct io_sqring_offsets sq_off;
176     struct io_cqring_offsets cq_off;
177     };
178    
179 root 1.12 #define IORING_SETUP_CQSIZE 0x00000008
180    
181     #define IORING_OP_POLL_ADD 6
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 */
187     struct iouring_kernel_timespec
188     {
189     int64_t tv_sec;
190     long long tv_nsec;
191     };
192    
193     #define IORING_TIMEOUT_ABS 0x00000001
194 root 1.1
195     #define IORING_ENTER_GETEVENTS 0x01
196    
197     #define IORING_OFF_SQ_RING 0x00000000ULL
198     #define IORING_OFF_CQ_RING 0x08000000ULL
199     #define IORING_OFF_SQES 0x10000000ULL
200    
201 root 1.12 #define IORING_FEAT_SINGLE_MMAP 0x00000001
202     #define IORING_FEAT_NODROP 0x00000002
203     #define IORING_FEAT_SUBMIT_STABLE 0x00000004
204 root 1.8
205 root 1.1 inline_size
206     int
207     evsys_io_uring_setup (unsigned entries, struct io_uring_params *params)
208     {
209     return ev_syscall2 (SYS_io_uring_setup, entries, params);
210     }
211    
212     inline_size
213     int
214     evsys_io_uring_enter (int fd, unsigned to_submit, unsigned min_complete, unsigned flags, const sigset_t *sig, size_t sigsz)
215     {
216     return ev_syscall6 (SYS_io_uring_enter, fd, to_submit, min_complete, flags, sig, sigsz);
217     }
218    
219     /*****************************************************************************/
220     /* actual backed implementation */
221    
222     /* we hope that volatile will make the compiler access this variables only once */
223     #define EV_SQ_VAR(name) *(volatile unsigned *)((char *)iouring_sq_ring + iouring_sq_ ## name)
224     #define EV_CQ_VAR(name) *(volatile unsigned *)((char *)iouring_cq_ring + iouring_cq_ ## name)
225    
226     /* the index array */
227     #define EV_SQ_ARRAY ((unsigned *)((char *)iouring_sq_ring + iouring_sq_array))
228    
229     /* the submit/completion queue entries */
230     #define EV_SQES ((struct io_uring_sqe *) iouring_sqes)
231     #define EV_CQES ((struct io_uring_cqe *)((char *)iouring_cq_ring + iouring_cq_cqes))
232    
233 root 1.16 inline_speed
234     int
235 root 1.14 iouring_enter (EV_P_ ev_tstamp timeout)
236     {
237     int res;
238    
239     EV_RELEASE_CB;
240    
241     res = evsys_io_uring_enter (iouring_fd, iouring_to_submit, 1,
242     timeout > EV_TS_CONST (0.) ? IORING_ENTER_GETEVENTS : 0, 0, 0);
243    
244     assert (("libev: io_uring_enter did not consume all sqes", (res < 0 || res == iouring_to_submit)));
245    
246     iouring_to_submit = 0;
247    
248     EV_ACQUIRE_CB;
249    
250     return res;
251     }
252    
253 root 1.16 /* TODO: can we move things around so we don't need this forward-reference? */
254     static void
255     iouring_poll (EV_P_ ev_tstamp timeout);
256    
257 root 1.1 static
258     struct io_uring_sqe *
259     iouring_sqe_get (EV_P)
260     {
261 root 1.16 unsigned tail;
262    
263     for (;;)
264     {
265     tail = EV_SQ_VAR (tail);
266 root 1.1
267 root 1.16 if (ecb_expect_true (tail + 1 - EV_SQ_VAR (head) <= EV_SQ_VAR (ring_entries)))
268     break; /* whats the problem, we have free sqes */
269    
270     /* queue full, need to flush and possibly handle some events */
271 root 1.14
272 root 1.16 #if EV_FEATURE_CODE
273     /* first we ask the kernel nicely, most often this frees up some sqes */
274 root 1.14 int res = iouring_enter (EV_A_ EV_TS_CONST (0.));
275    
276 root 1.16 ECB_MEMORY_FENCE_ACQUIRE; /* better safe than sorry */
277    
278     if (res >= 0)
279     continue; /* yes, it worked, try again */
280     #endif
281    
282     /* some problem, possibly EBUSY - do the full poll and let it handle any issues */
283 root 1.14
284 root 1.16 iouring_poll (EV_A_ EV_TS_CONST (0.));
285     /* iouring_poll should have done ECB_MEMORY_FENCE_ACQUIRE for us */
286 root 1.1 }
287    
288 root 1.14 /*assert (("libev: io_uring queue full after flush", tail + 1 - EV_SQ_VAR (head) <= EV_SQ_VAR (ring_entries)));*/
289 root 1.1
290     return EV_SQES + (tail & EV_SQ_VAR (ring_mask));
291     }
292    
293     inline_size
294     struct io_uring_sqe *
295     iouring_sqe_submit (EV_P_ struct io_uring_sqe *sqe)
296     {
297     unsigned idx = sqe - EV_SQES;
298    
299     EV_SQ_ARRAY [idx] = idx;
300     ECB_MEMORY_FENCE_RELEASE;
301     ++EV_SQ_VAR (tail);
302     /*ECB_MEMORY_FENCE_RELEASE; /* for the time being we assume this is not needed */
303     ++iouring_to_submit;
304     }
305    
306     /*****************************************************************************/
307    
308     /* when the timerfd expires we simply note the fact,
309     * as the purpose of the timerfd is to wake us up, nothing else.
310     * the next iteration should re-set it.
311     */
312     static void
313     iouring_tfd_cb (EV_P_ struct ev_io *w, int revents)
314     {
315     iouring_tfd_to = EV_TSTAMP_HUGE;
316     }
317    
318     /* called for full and partial cleanup */
319     ecb_cold
320     static int
321     iouring_internal_destroy (EV_P)
322     {
323     close (iouring_tfd);
324     close (iouring_fd);
325    
326     if (iouring_sq_ring != MAP_FAILED) munmap (iouring_sq_ring, iouring_sq_ring_size);
327     if (iouring_cq_ring != MAP_FAILED) munmap (iouring_cq_ring, iouring_cq_ring_size);
328     if (iouring_sqes != MAP_FAILED) munmap (iouring_sqes , iouring_sqes_size );
329    
330 root 1.8 if (ev_is_active (&iouring_tfd_w))
331     {
332     ev_ref (EV_A);
333     ev_io_stop (EV_A_ &iouring_tfd_w);
334     }
335 root 1.1 }
336    
337     ecb_cold
338     static int
339     iouring_internal_init (EV_P)
340     {
341     struct io_uring_params params = { 0 };
342    
343     iouring_to_submit = 0;
344    
345     iouring_tfd = -1;
346     iouring_sq_ring = MAP_FAILED;
347     iouring_cq_ring = MAP_FAILED;
348     iouring_sqes = MAP_FAILED;
349    
350 root 1.12 if (!have_monotonic) /* cannot really happen, but what if11 */
351     return -1;
352    
353 root 1.1 for (;;)
354     {
355     iouring_fd = evsys_io_uring_setup (iouring_entries, &params);
356    
357     if (iouring_fd >= 0)
358     break; /* yippie */
359    
360     if (errno != EINVAL)
361     return -1; /* we failed */
362    
363 root 1.11 #if TODO
364 root 1.17 if ((~params.features) & (IORING_FEAT_NODROP | IORING_FEATURE_SINGLE_MMAP | IORING_FEAT_SUBMIT_STABLE))
365 root 1.11 return -1; /* we require the above features */
366     #endif
367    
368 root 1.1 /* EINVAL: lots of possible reasons, but maybe
369     * it is because we hit the unqueryable hardcoded size limit
370     */
371    
372     /* we hit the limit already, give up */
373     if (iouring_max_entries)
374     return -1;
375    
376     /* first time we hit EINVAL? assume we hit the limit, so go back and retry */
377     iouring_entries >>= 1;
378     iouring_max_entries = iouring_entries;
379     }
380    
381     iouring_sq_ring_size = params.sq_off.array + params.sq_entries * sizeof (unsigned);
382     iouring_cq_ring_size = params.cq_off.cqes + params.cq_entries * sizeof (struct io_uring_cqe);
383     iouring_sqes_size = params.sq_entries * sizeof (struct io_uring_sqe);
384    
385     iouring_sq_ring = mmap (0, iouring_sq_ring_size, PROT_READ | PROT_WRITE,
386     MAP_SHARED | MAP_POPULATE, iouring_fd, IORING_OFF_SQ_RING);
387     iouring_cq_ring = mmap (0, iouring_cq_ring_size, PROT_READ | PROT_WRITE,
388     MAP_SHARED | MAP_POPULATE, iouring_fd, IORING_OFF_CQ_RING);
389     iouring_sqes = mmap (0, iouring_sqes_size, PROT_READ | PROT_WRITE,
390     MAP_SHARED | MAP_POPULATE, iouring_fd, IORING_OFF_SQES);
391    
392     if (iouring_sq_ring == MAP_FAILED || iouring_cq_ring == MAP_FAILED || iouring_sqes == MAP_FAILED)
393     return -1;
394    
395     iouring_sq_head = params.sq_off.head;
396     iouring_sq_tail = params.sq_off.tail;
397     iouring_sq_ring_mask = params.sq_off.ring_mask;
398     iouring_sq_ring_entries = params.sq_off.ring_entries;
399     iouring_sq_flags = params.sq_off.flags;
400     iouring_sq_dropped = params.sq_off.dropped;
401     iouring_sq_array = params.sq_off.array;
402    
403     iouring_cq_head = params.cq_off.head;
404     iouring_cq_tail = params.cq_off.tail;
405     iouring_cq_ring_mask = params.cq_off.ring_mask;
406     iouring_cq_ring_entries = params.cq_off.ring_entries;
407     iouring_cq_overflow = params.cq_off.overflow;
408     iouring_cq_cqes = params.cq_off.cqes;
409    
410     iouring_tfd = timerfd_create (CLOCK_MONOTONIC, TFD_CLOEXEC);
411    
412     if (iouring_tfd < 0)
413     return iouring_tfd;
414    
415     iouring_tfd_to = EV_TSTAMP_HUGE;
416    
417     return 0;
418     }
419    
420     ecb_cold
421     static void
422     iouring_fork (EV_P)
423     {
424     iouring_internal_destroy (EV_A);
425    
426     while (iouring_internal_init (EV_A) < 0)
427     ev_syserr ("(libev) io_uring_setup");
428    
429 root 1.8 fd_rearm_all (EV_A);
430 root 1.1
431     ev_io_stop (EV_A_ &iouring_tfd_w);
432     ev_io_set (EV_A_ &iouring_tfd_w, iouring_tfd, EV_READ);
433     ev_io_start (EV_A_ &iouring_tfd_w);
434     }
435    
436     /*****************************************************************************/
437    
438     static void
439     iouring_modify (EV_P_ int fd, int oev, int nev)
440     {
441     if (oev)
442     {
443     /* we assume the sqe's are all "properly" initialised */
444     struct io_uring_sqe *sqe = iouring_sqe_get (EV_A);
445     sqe->opcode = IORING_OP_POLL_REMOVE;
446     sqe->fd = fd;
447 root 1.13 /* Jens Axboe notified me that user_data is not what is documented, but is
448     * some kind of unique ID that has to match, otherwise the request cannot
449     * be removed. Since we don't *really* have that, we pass in the old
450     * generation counter - if that fails, too bad, it will hopefully be removed
451     * at close time and then be ignored. */
452 root 1.15 sqe->addr = (uint32_t)fd | ((__u64)(uint32_t)anfds [fd].egen << 32);
453     sqe->user_data = (uint64_t)-1;
454 root 1.1 iouring_sqe_submit (EV_A_ sqe);
455 root 1.2
456     /* increment generation counter to avoid handling old events */
457     ++anfds [fd].egen;
458 root 1.1 }
459    
460     if (nev)
461     {
462     struct io_uring_sqe *sqe = iouring_sqe_get (EV_A);
463     sqe->opcode = IORING_OP_POLL_ADD;
464     sqe->fd = fd;
465 root 1.2 sqe->user_data = (uint32_t)fd | ((__u64)(uint32_t)anfds [fd].egen << 32);
466 root 1.1 sqe->poll_events =
467     (nev & EV_READ ? POLLIN : 0)
468     | (nev & EV_WRITE ? POLLOUT : 0);
469     iouring_sqe_submit (EV_A_ sqe);
470     }
471     }
472    
473     inline_size
474     void
475     iouring_tfd_update (EV_P_ ev_tstamp timeout)
476     {
477     ev_tstamp tfd_to = mn_now + timeout;
478    
479     /* we assume there will be many iterations per timer change, so
480     * we only re-set the timerfd when we have to because its expiry
481     * is too late.
482     */
483     if (ecb_expect_false (tfd_to < iouring_tfd_to))
484     {
485     struct itimerspec its;
486    
487     iouring_tfd_to = tfd_to;
488     EV_TS_SET (its.it_interval, 0.);
489     EV_TS_SET (its.it_value, tfd_to);
490    
491     if (timerfd_settime (iouring_tfd, TFD_TIMER_ABSTIME, &its, 0) < 0)
492     assert (("libev: iouring timerfd_settime failed", 0));
493     }
494     }
495    
496     inline_size
497     void
498     iouring_process_cqe (EV_P_ struct io_uring_cqe *cqe)
499     {
500     int fd = cqe->user_data & 0xffffffffU;
501     uint32_t gen = cqe->user_data >> 32;
502     int res = cqe->res;
503    
504 root 1.15 /* user_data -1 is a remove that we are not atm. interested in */
505     if (cqe->user_data == (uint64_t)-1)
506     return;
507    
508 root 1.1 assert (("libev: io_uring fd must be in-bounds", fd >= 0 && fd < anfdmax));
509    
510     /* documentation lies, of course. the result value is NOT like
511     * normal syscalls, but like linux raw syscalls, i.e. negative
512     * error numbers. fortunate, as otherwise there would be no way
513     * to get error codes at all. still, why not document this?
514     */
515    
516     /* ignore event if generation doesn't match */
517 root 1.13 /* other than skipping removal events, */
518 root 1.1 /* this should actually be very rare */
519 root 1.2 if (ecb_expect_false (gen != (uint32_t)anfds [fd].egen))
520 root 1.1 return;
521    
522     if (ecb_expect_false (res < 0))
523     {
524 root 1.11 /*TODO: EINVAL handling (was something failed with this fd)*/
525     /*TODO: EBUSY happens when?*/
526 root 1.1
527 root 1.8 if (res == -EBADF)
528 root 1.1 {
529     assert (("libev: event loop rejected bad fd", res != -EBADF));
530     fd_kill (EV_A_ fd);
531     }
532     else
533     {
534     errno = -res;
535     ev_syserr ("(libev) IORING_OP_POLL_ADD");
536     }
537    
538     return;
539     }
540    
541     /* feed events, we do not expect or handle POLLNVAL */
542     fd_event (
543     EV_A_
544     fd,
545     (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
546     | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
547     );
548    
549     /* io_uring is oneshot, so we need to re-arm the fd next iteration */
550     /* this also means we usually have to do at least one syscall per iteration */
551     anfds [fd].events = 0;
552     fd_change (EV_A_ fd, EV_ANFD_REIFY);
553     }
554    
555     /* called when the event queue overflows */
556     ecb_cold
557     static void
558     iouring_overflow (EV_P)
559     {
560     /* we have two options, resize the queue (by tearing down
561     * everything and recreating it, or living with it
562     * and polling.
563 root 1.10 * we implement this by resizing the queue, and, if that fails,
564 root 1.1 * we just recreate the state on every failure, which
565     * kind of is a very inefficient poll.
566     * one danger is, due to the bios toward lower fds,
567     * we will only really get events for those, so
568     * maybe we need a poll() fallback, after all.
569     */
570     /*EV_CQ_VAR (overflow) = 0;*/ /* need to do this if we keep the state and poll manually */
571    
572     fd_rearm_all (EV_A);
573    
574     /* we double the size until we hit the hard-to-probe maximum */
575     if (!iouring_max_entries)
576     {
577     iouring_entries <<= 1;
578     iouring_fork (EV_A);
579     }
580     else
581     {
582     /* we hit the kernel limit, we should fall back to something else.
583     * we can either poll() a few times and hope for the best,
584     * poll always, or switch to epoll.
585 root 1.10 * TODO: is this necessary with newer kernels?
586 root 1.1 */
587    
588     iouring_internal_destroy (EV_A);
589    
590 root 1.10 /* this should make it so that on return, we don't call any uring functions */
591 root 1.1 iouring_to_submit = 0;
592    
593     for (;;)
594     {
595     backend = epoll_init (EV_A_ 0);
596    
597     if (backend)
598     break;
599    
600     ev_syserr ("(libev) iouring switch to epoll");
601     }
602     }
603     }
604    
605     /* handle any events in the completion queue, return true if there were any */
606     static int
607     iouring_handle_cq (EV_P)
608     {
609     unsigned head, tail, mask;
610    
611     head = EV_CQ_VAR (head);
612     ECB_MEMORY_FENCE_ACQUIRE;
613     tail = EV_CQ_VAR (tail);
614    
615     if (head == tail)
616     return 0;
617    
618     /* it can only overflow if we have events, yes, yes? */
619     if (ecb_expect_false (EV_CQ_VAR (overflow)))
620     {
621     iouring_overflow (EV_A);
622     return 1;
623     }
624    
625     mask = EV_CQ_VAR (ring_mask);
626    
627     do
628     iouring_process_cqe (EV_A_ &EV_CQES [head++ & mask]);
629     while (head != tail);
630    
631     EV_CQ_VAR (head) = head;
632     ECB_MEMORY_FENCE_RELEASE;
633    
634     return 1;
635     }
636    
637     static void
638     iouring_poll (EV_P_ ev_tstamp timeout)
639     {
640     /* if we have events, no need for extra syscalls, but we might have to queue events */
641 root 1.16 /* we also clar the timeout if there are outstanding fdchanges */
642     /* the latter should only happen if both the sq and cq are full, most likely */
643     /* because we have a lot of event sources that immediately complete */
644     /* TODO: fdchacngecnt is always 0 because fd_reify does not have two buffers yet */
645     if (iouring_handle_cq (EV_A) || fdchangecnt)
646 root 1.3 timeout = EV_TS_CONST (0.);
647 root 1.1 else
648     /* no events, so maybe wait for some */
649     iouring_tfd_update (EV_A_ timeout);
650    
651 root 1.6 /* only enter the kernel if we have something to submit, or we need to wait */
652 root 1.1 if (timeout || iouring_to_submit)
653     {
654 root 1.14 int res = iouring_enter (EV_A_ timeout);
655 root 1.1
656     if (ecb_expect_false (res < 0))
657     if (errno == EINTR)
658     /* ignore */;
659 root 1.14 else if (errno == EBUSY)
660     /* cq full, cannot submit - should be rare because we flush the cq first, so simply ignore */;
661 root 1.1 else
662     ev_syserr ("(libev) iouring setup");
663     else
664     iouring_handle_cq (EV_A);
665     }
666     }
667    
668     inline_size
669     int
670     iouring_init (EV_P_ int flags)
671     {
672     iouring_entries = IOURING_INIT_ENTRIES;
673     iouring_max_entries = 0;
674    
675     if (iouring_internal_init (EV_A) < 0)
676     {
677     iouring_internal_destroy (EV_A);
678     return 0;
679     }
680    
681 root 1.6 ev_io_init (&iouring_tfd_w, iouring_tfd_cb, iouring_tfd, EV_READ);
682 root 1.8 ev_set_priority (&iouring_tfd_w, EV_MINPRI);
683 root 1.1 ev_io_start (EV_A_ &iouring_tfd_w);
684     ev_unref (EV_A); /* watcher should not keep loop alive */
685    
686     backend_modify = iouring_modify;
687     backend_poll = iouring_poll;
688    
689     return EVBACKEND_IOURING;
690     }
691    
692     inline_size
693     void
694     iouring_destroy (EV_P)
695     {
696     iouring_internal_destroy (EV_A);
697     }
698