ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libev/ev_iouring.c
Revision: 1.8
Committed: Fri Dec 27 16:08:24 2019 UTC (4 years, 6 months ago) by root
Content type: text/plain
Branch: MAIN
Changes since 1.7: +25 -49 lines
Log Message:
rip out epoll from iouring

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     #include <sys/timerfd.h>
86     #include <sys/mman.h>
87     #include <poll.h>
88    
89     #define IOURING_INIT_ENTRIES 32
90    
91     /*****************************************************************************/
92     /* syscall wrapdadoop - this section has the raw api/abi definitions */
93    
94     #include <linux/fs.h>
95     #include <linux/types.h>
96    
97     /* mostly directly taken from the kernel or documentation */
98    
99     struct io_uring_sqe
100     {
101     __u8 opcode;
102     __u8 flags;
103     __u16 ioprio;
104     __s32 fd;
105 root 1.8 union {
106     __u64 off;
107     __u64 addr2;
108     };
109 root 1.1 __u64 addr;
110     __u32 len;
111     union {
112     __kernel_rwf_t rw_flags;
113     __u32 fsync_flags;
114     __u16 poll_events;
115     __u32 sync_range_flags;
116     __u32 msg_flags;
117 root 1.8 __u32 timeout_flags;
118     __u32 accept_flags;
119     __u32 cancel_flags;
120     __u32 open_flags;
121     __u32 statx_flags;
122 root 1.1 };
123     __u64 user_data;
124     union {
125     __u16 buf_index;
126     __u64 __pad2[3];
127     };
128     };
129    
130     struct io_uring_cqe
131     {
132     __u64 user_data;
133     __s32 res;
134     __u32 flags;
135     };
136    
137     struct io_sqring_offsets
138     {
139     __u32 head;
140     __u32 tail;
141     __u32 ring_mask;
142     __u32 ring_entries;
143     __u32 flags;
144     __u32 dropped;
145     __u32 array;
146     __u32 resv1;
147     __u64 resv2;
148     };
149    
150     struct io_cqring_offsets
151     {
152     __u32 head;
153     __u32 tail;
154     __u32 ring_mask;
155     __u32 ring_entries;
156     __u32 overflow;
157     __u32 cqes;
158     __u64 resv[2];
159     };
160    
161     struct io_uring_params
162     {
163     __u32 sq_entries;
164     __u32 cq_entries;
165     __u32 flags;
166     __u32 sq_thread_cpu;
167     __u32 sq_thread_idle;
168 root 1.8 __u32 features;
169     __u32 resv[4];
170 root 1.1 struct io_sqring_offsets sq_off;
171     struct io_cqring_offsets cq_off;
172     };
173    
174     #define IORING_OP_POLL_ADD 6
175     #define IORING_OP_POLL_REMOVE 7
176    
177     #define IORING_ENTER_GETEVENTS 0x01
178    
179     #define IORING_OFF_SQ_RING 0x00000000ULL
180     #define IORING_OFF_CQ_RING 0x08000000ULL
181     #define IORING_OFF_SQES 0x10000000ULL
182    
183 root 1.8 #define IORING_FEAT_SINGLE_MMAP 0x1
184     #define IORING_FEAT_NODROP 0x2
185     #define IORING_FEAT_SUBMIT_STABLE 0x4
186    
187 root 1.1 inline_size
188     int
189     evsys_io_uring_setup (unsigned entries, struct io_uring_params *params)
190     {
191     return ev_syscall2 (SYS_io_uring_setup, entries, params);
192     }
193    
194     inline_size
195     int
196     evsys_io_uring_enter (int fd, unsigned to_submit, unsigned min_complete, unsigned flags, const sigset_t *sig, size_t sigsz)
197     {
198     return ev_syscall6 (SYS_io_uring_enter, fd, to_submit, min_complete, flags, sig, sigsz);
199     }
200    
201     /*****************************************************************************/
202     /* actual backed implementation */
203    
204     /* we hope that volatile will make the compiler access this variables only once */
205     #define EV_SQ_VAR(name) *(volatile unsigned *)((char *)iouring_sq_ring + iouring_sq_ ## name)
206     #define EV_CQ_VAR(name) *(volatile unsigned *)((char *)iouring_cq_ring + iouring_cq_ ## name)
207    
208     /* the index array */
209     #define EV_SQ_ARRAY ((unsigned *)((char *)iouring_sq_ring + iouring_sq_array))
210    
211     /* the submit/completion queue entries */
212     #define EV_SQES ((struct io_uring_sqe *) iouring_sqes)
213     #define EV_CQES ((struct io_uring_cqe *)((char *)iouring_cq_ring + iouring_cq_cqes))
214    
215     static
216     struct io_uring_sqe *
217     iouring_sqe_get (EV_P)
218     {
219     unsigned tail = EV_SQ_VAR (tail);
220    
221     if (tail + 1 - EV_SQ_VAR (head) > EV_SQ_VAR (ring_entries))
222     {
223     /* queue full, flush */
224     evsys_io_uring_enter (iouring_fd, iouring_to_submit, 0, 0, 0, 0);
225     iouring_to_submit = 0;
226     }
227    
228     assert (("libev: io_uring queue full after flush", tail + 1 - EV_SQ_VAR (head) <= EV_SQ_VAR (ring_entries)));
229    
230     return EV_SQES + (tail & EV_SQ_VAR (ring_mask));
231     }
232    
233     inline_size
234     struct io_uring_sqe *
235     iouring_sqe_submit (EV_P_ struct io_uring_sqe *sqe)
236     {
237     unsigned idx = sqe - EV_SQES;
238    
239     EV_SQ_ARRAY [idx] = idx;
240     ECB_MEMORY_FENCE_RELEASE;
241     ++EV_SQ_VAR (tail);
242     /*ECB_MEMORY_FENCE_RELEASE; /* for the time being we assume this is not needed */
243     ++iouring_to_submit;
244     }
245    
246     /*****************************************************************************/
247    
248     /* when the timerfd expires we simply note the fact,
249     * as the purpose of the timerfd is to wake us up, nothing else.
250     * the next iteration should re-set it.
251     */
252     static void
253     iouring_tfd_cb (EV_P_ struct ev_io *w, int revents)
254     {
255     iouring_tfd_to = EV_TSTAMP_HUGE;
256     }
257    
258     /* called for full and partial cleanup */
259     ecb_cold
260     static int
261     iouring_internal_destroy (EV_P)
262     {
263     close (iouring_tfd);
264     close (iouring_fd);
265    
266     if (iouring_sq_ring != MAP_FAILED) munmap (iouring_sq_ring, iouring_sq_ring_size);
267     if (iouring_cq_ring != MAP_FAILED) munmap (iouring_cq_ring, iouring_cq_ring_size);
268     if (iouring_sqes != MAP_FAILED) munmap (iouring_sqes , iouring_sqes_size );
269    
270 root 1.8 if (ev_is_active (&iouring_tfd_w))
271     {
272     ev_ref (EV_A);
273     ev_io_stop (EV_A_ &iouring_tfd_w);
274     }
275 root 1.1 }
276    
277     ecb_cold
278     static int
279     iouring_internal_init (EV_P)
280     {
281     struct io_uring_params params = { 0 };
282    
283     iouring_to_submit = 0;
284    
285     iouring_tfd = -1;
286     iouring_sq_ring = MAP_FAILED;
287     iouring_cq_ring = MAP_FAILED;
288     iouring_sqes = MAP_FAILED;
289    
290     for (;;)
291     {
292     iouring_fd = evsys_io_uring_setup (iouring_entries, &params);
293    
294     if (iouring_fd >= 0)
295     break; /* yippie */
296    
297     if (errno != EINVAL)
298     return -1; /* we failed */
299    
300     /* EINVAL: lots of possible reasons, but maybe
301     * it is because we hit the unqueryable hardcoded size limit
302     */
303    
304     /* we hit the limit already, give up */
305     if (iouring_max_entries)
306     return -1;
307    
308     /* first time we hit EINVAL? assume we hit the limit, so go back and retry */
309     iouring_entries >>= 1;
310     iouring_max_entries = iouring_entries;
311     }
312    
313     iouring_sq_ring_size = params.sq_off.array + params.sq_entries * sizeof (unsigned);
314     iouring_cq_ring_size = params.cq_off.cqes + params.cq_entries * sizeof (struct io_uring_cqe);
315     iouring_sqes_size = params.sq_entries * sizeof (struct io_uring_sqe);
316    
317     iouring_sq_ring = mmap (0, iouring_sq_ring_size, PROT_READ | PROT_WRITE,
318     MAP_SHARED | MAP_POPULATE, iouring_fd, IORING_OFF_SQ_RING);
319     iouring_cq_ring = mmap (0, iouring_cq_ring_size, PROT_READ | PROT_WRITE,
320     MAP_SHARED | MAP_POPULATE, iouring_fd, IORING_OFF_CQ_RING);
321     iouring_sqes = mmap (0, iouring_sqes_size, PROT_READ | PROT_WRITE,
322     MAP_SHARED | MAP_POPULATE, iouring_fd, IORING_OFF_SQES);
323    
324     if (iouring_sq_ring == MAP_FAILED || iouring_cq_ring == MAP_FAILED || iouring_sqes == MAP_FAILED)
325     return -1;
326    
327     iouring_sq_head = params.sq_off.head;
328     iouring_sq_tail = params.sq_off.tail;
329     iouring_sq_ring_mask = params.sq_off.ring_mask;
330     iouring_sq_ring_entries = params.sq_off.ring_entries;
331     iouring_sq_flags = params.sq_off.flags;
332     iouring_sq_dropped = params.sq_off.dropped;
333     iouring_sq_array = params.sq_off.array;
334    
335     iouring_cq_head = params.cq_off.head;
336     iouring_cq_tail = params.cq_off.tail;
337     iouring_cq_ring_mask = params.cq_off.ring_mask;
338     iouring_cq_ring_entries = params.cq_off.ring_entries;
339     iouring_cq_overflow = params.cq_off.overflow;
340     iouring_cq_cqes = params.cq_off.cqes;
341    
342     iouring_tfd = timerfd_create (CLOCK_MONOTONIC, TFD_CLOEXEC);
343    
344     if (iouring_tfd < 0)
345     return iouring_tfd;
346    
347     iouring_tfd_to = EV_TSTAMP_HUGE;
348    
349     return 0;
350     }
351    
352     ecb_cold
353     static void
354     iouring_fork (EV_P)
355     {
356     iouring_internal_destroy (EV_A);
357    
358     while (iouring_internal_init (EV_A) < 0)
359     ev_syserr ("(libev) io_uring_setup");
360    
361 root 1.8 fd_rearm_all (EV_A);
362 root 1.1
363     ev_io_stop (EV_A_ &iouring_tfd_w);
364     ev_io_set (EV_A_ &iouring_tfd_w, iouring_tfd, EV_READ);
365     ev_io_start (EV_A_ &iouring_tfd_w);
366     }
367    
368     /*****************************************************************************/
369    
370     static void
371     iouring_modify (EV_P_ int fd, int oev, int nev)
372     {
373     if (oev)
374     {
375     /* we assume the sqe's are all "properly" initialised */
376     struct io_uring_sqe *sqe = iouring_sqe_get (EV_A);
377     sqe->opcode = IORING_OP_POLL_REMOVE;
378     sqe->fd = fd;
379     sqe->user_data = -1;
380     iouring_sqe_submit (EV_A_ sqe);
381 root 1.2
382     /* increment generation counter to avoid handling old events */
383     ++anfds [fd].egen;
384 root 1.1 }
385    
386     if (nev)
387     {
388     struct io_uring_sqe *sqe = iouring_sqe_get (EV_A);
389     sqe->opcode = IORING_OP_POLL_ADD;
390     sqe->fd = fd;
391 root 1.2 sqe->user_data = (uint32_t)fd | ((__u64)(uint32_t)anfds [fd].egen << 32);
392 root 1.1 sqe->poll_events =
393     (nev & EV_READ ? POLLIN : 0)
394     | (nev & EV_WRITE ? POLLOUT : 0);
395     iouring_sqe_submit (EV_A_ sqe);
396     }
397     }
398    
399     inline_size
400     void
401     iouring_tfd_update (EV_P_ ev_tstamp timeout)
402     {
403     ev_tstamp tfd_to = mn_now + timeout;
404    
405     /* we assume there will be many iterations per timer change, so
406     * we only re-set the timerfd when we have to because its expiry
407     * is too late.
408     */
409     if (ecb_expect_false (tfd_to < iouring_tfd_to))
410     {
411     struct itimerspec its;
412    
413     iouring_tfd_to = tfd_to;
414     EV_TS_SET (its.it_interval, 0.);
415     EV_TS_SET (its.it_value, tfd_to);
416    
417     if (timerfd_settime (iouring_tfd, TFD_TIMER_ABSTIME, &its, 0) < 0)
418     assert (("libev: iouring timerfd_settime failed", 0));
419     }
420     }
421    
422     inline_size
423     void
424     iouring_process_cqe (EV_P_ struct io_uring_cqe *cqe)
425     {
426     int fd = cqe->user_data & 0xffffffffU;
427     uint32_t gen = cqe->user_data >> 32;
428     int res = cqe->res;
429    
430     /* ignore fd removal events, if there are any. TODO: verify */
431     if (cqe->user_data == (__u64)-1)
432     abort ();//D
433    
434     assert (("libev: io_uring fd must be in-bounds", fd >= 0 && fd < anfdmax));
435    
436     /* documentation lies, of course. the result value is NOT like
437     * normal syscalls, but like linux raw syscalls, i.e. negative
438     * error numbers. fortunate, as otherwise there would be no way
439     * to get error codes at all. still, why not document this?
440     */
441    
442     /* ignore event if generation doesn't match */
443     /* this should actually be very rare */
444 root 1.2 if (ecb_expect_false (gen != (uint32_t)anfds [fd].egen))
445 root 1.1 return;
446    
447     if (ecb_expect_false (res < 0))
448     {
449 root 1.8 //TODO: EINVAL handling (was something failed with this fd)
450     //TODO: EBUSY happens when?
451 root 1.1
452 root 1.8 if (res == -EBADF)
453 root 1.1 {
454     assert (("libev: event loop rejected bad fd", res != -EBADF));
455     fd_kill (EV_A_ fd);
456     }
457     else
458     {
459     errno = -res;
460     ev_syserr ("(libev) IORING_OP_POLL_ADD");
461     }
462    
463     return;
464     }
465    
466     /* feed events, we do not expect or handle POLLNVAL */
467     fd_event (
468     EV_A_
469     fd,
470     (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
471     | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
472     );
473    
474     /* io_uring is oneshot, so we need to re-arm the fd next iteration */
475     /* this also means we usually have to do at least one syscall per iteration */
476     anfds [fd].events = 0;
477     fd_change (EV_A_ fd, EV_ANFD_REIFY);
478     }
479    
480     /* called when the event queue overflows */
481     ecb_cold
482     static void
483     iouring_overflow (EV_P)
484     {
485     /* we have two options, resize the queue (by tearing down
486     * everything and recreating it, or living with it
487     * and polling.
488     * we implement this by resizing tghe queue, and, if that fails,
489     * we just recreate the state on every failure, which
490     * kind of is a very inefficient poll.
491     * one danger is, due to the bios toward lower fds,
492     * we will only really get events for those, so
493     * maybe we need a poll() fallback, after all.
494     */
495     /*EV_CQ_VAR (overflow) = 0;*/ /* need to do this if we keep the state and poll manually */
496    
497     fd_rearm_all (EV_A);
498    
499     /* we double the size until we hit the hard-to-probe maximum */
500     if (!iouring_max_entries)
501     {
502     iouring_entries <<= 1;
503     iouring_fork (EV_A);
504     }
505     else
506     {
507     /* we hit the kernel limit, we should fall back to something else.
508     * we can either poll() a few times and hope for the best,
509     * poll always, or switch to epoll.
510     * since we use epoll anyways, go epoll.
511     */
512    
513     iouring_internal_destroy (EV_A);
514    
515     /* this should make it so that on return, we don'T call any uring functions */
516     iouring_to_submit = 0;
517    
518     for (;;)
519     {
520     backend = epoll_init (EV_A_ 0);
521    
522     if (backend)
523     break;
524    
525     ev_syserr ("(libev) iouring switch to epoll");
526     }
527     }
528     }
529    
530     /* handle any events in the completion queue, return true if there were any */
531     static int
532     iouring_handle_cq (EV_P)
533     {
534     unsigned head, tail, mask;
535    
536     head = EV_CQ_VAR (head);
537     ECB_MEMORY_FENCE_ACQUIRE;
538     tail = EV_CQ_VAR (tail);
539    
540     if (head == tail)
541     return 0;
542    
543     /* it can only overflow if we have events, yes, yes? */
544     if (ecb_expect_false (EV_CQ_VAR (overflow)))
545     {
546     iouring_overflow (EV_A);
547     return 1;
548     }
549    
550     mask = EV_CQ_VAR (ring_mask);
551    
552     do
553     iouring_process_cqe (EV_A_ &EV_CQES [head++ & mask]);
554     while (head != tail);
555    
556     EV_CQ_VAR (head) = head;
557     ECB_MEMORY_FENCE_RELEASE;
558    
559     return 1;
560     }
561    
562     static void
563     iouring_poll (EV_P_ ev_tstamp timeout)
564     {
565     /* if we have events, no need for extra syscalls, but we might have to queue events */
566     if (iouring_handle_cq (EV_A))
567 root 1.3 timeout = EV_TS_CONST (0.);
568 root 1.1 else
569     /* no events, so maybe wait for some */
570     iouring_tfd_update (EV_A_ timeout);
571    
572 root 1.6 /* only enter the kernel if we have something to submit, or we need to wait */
573 root 1.1 if (timeout || iouring_to_submit)
574     {
575     int res;
576    
577     EV_RELEASE_CB;
578    
579     res = evsys_io_uring_enter (iouring_fd, iouring_to_submit, 1,
580 root 1.3 timeout > EV_TS_CONST (0.) ? IORING_ENTER_GETEVENTS : 0, 0, 0);
581 root 1.1 iouring_to_submit = 0;
582    
583     EV_ACQUIRE_CB;
584    
585     if (ecb_expect_false (res < 0))
586     if (errno == EINTR)
587     /* ignore */;
588     else
589     ev_syserr ("(libev) iouring setup");
590     else
591     iouring_handle_cq (EV_A);
592     }
593     }
594    
595     inline_size
596     int
597     iouring_init (EV_P_ int flags)
598     {
599     iouring_entries = IOURING_INIT_ENTRIES;
600     iouring_max_entries = 0;
601    
602     if (iouring_internal_init (EV_A) < 0)
603     {
604     iouring_internal_destroy (EV_A);
605     return 0;
606     }
607    
608 root 1.6 ev_io_init (&iouring_tfd_w, iouring_tfd_cb, iouring_tfd, EV_READ);
609 root 1.8 ev_set_priority (&iouring_tfd_w, EV_MINPRI);
610 root 1.1 ev_io_start (EV_A_ &iouring_tfd_w);
611     ev_unref (EV_A); /* watcher should not keep loop alive */
612    
613     backend_modify = iouring_modify;
614     backend_poll = iouring_poll;
615    
616     return EVBACKEND_IOURING;
617     }
618    
619     inline_size
620     void
621     iouring_destroy (EV_P)
622     {
623     iouring_internal_destroy (EV_A);
624     }
625