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 1.6 by root, Fri Dec 20 05:20:23 2019 UTC vs.
Revision 1.11 by root, Fri Dec 27 21:56:29 2019 UTC

… …
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 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 50 * and I can't see what benefit three 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).
52 * f) hmm, it's practiclaly undebuggable (gdb can't access the memory, and 53 * f) hmm, it's practiclaly undebuggable (gdb can't access the memory, and
53 the bizarre way structure offsets are commuinicated makes it hard to 54 * the bizarre way structure offsets are communicated makes it hard to
54 * just print the ring buffer heads, even *iff* the memory were visible 55 * just print the ring buffer heads, even *iff* the memory were visible
55 * in gdb. but then, that's also ok, really. 56 * in gdb. but then, that's also ok, really.
56 * g) well, you cannot specify a timeout when waiting for events. no, 57 * g) well, you cannot specify a timeout when waiting for events. no,
57 * seriously, the interface doesn't support a timeout. never seen _that_ 58 * seriously, the interface doesn't support a timeout. never seen _that_
58 * before. sure, you can use a timerfd, but that's another syscall 59 * before. sure, you can use a timerfd, but that's another syscall
59 * you could have avoided. overall, this bizarre omission smells 60 * you could have avoided. overall, this bizarre omission smells
60 * like a µ-optimisation by the io_uring author for his personal 61 * like a µ-optimisation by the io_uring author for his personal
61 * applications, to the detriment of everybody else who just wants 62 * applications, to the detriment of everybody else who just wants
62 * an event loop. but, umm, ok, if that's all, it could be worse. 63 * an event loop. but, umm, ok, if that's all, it could be worse.
64 * (FIXME: jens mentioned timeout commands, need to investigate)
63 * h) there is a hardcoded limit of 4096 outstanding events. okay, 65 * h) there is a hardcoded limit of 4096 outstanding events. okay,
64 * at least there is no arbitrary low system-wide limit... 66 * at least there is no arbitrary low system-wide limit...
67 * (FIXME: apparently, this was increased to 32768 in later kernels(
65 * i) unlike linux aio, you *can* register more then the limit 68 * i) unlike linux aio, you *can* register more then the limit
66 * of fd events, and the kernel will "gracefully" signal an 69 * of fd events, and the kernel will "gracefully" signal an
67 * overflow, after which you could destroy and recreate the kernel 70 * overflow, after which you could destroy and recreate the kernel
68 * state, a bit bigger, or fall back to e.g. poll. thats not 71 * state, a bit bigger, or fall back to e.g. poll. thats not
69 * totally insane, but kind of questions the point a high 72 * totally insane, but kind of questions the point a high
70 * performance I/O framework when it doesn't really work 73 * performance I/O framework when it doesn't really work
71 * under stress. 74 * under stress.
75 * (FIXME: iouring should no longer drop events, need to investigate)
72 * j) but, oh my! is has exactly the same bugs as the linux aio backend, 76 * j) but, oh my! is has exactly the same bugs as the linux aio backend,
73 * where some undocumented poll combinations just fail. 77 * where some undocumented poll combinations just fail.
74 * so we need epoll AGAIN as a fallback. AGAIN! epoll!! and of course, 78 * so we need epoll AGAIN as a fallback. AGAIN! epoll!! and of course,
75 * this is completely undocumented, have I mantioned this already? 79 * this is completely undocumented, have I mantioned this already?
76 * k) overall, the *API* itself is, I dare to say, not a total trainwreck. 80 * k) overall, the *API* itself is, I dare to say, not a total trainwreck.
77 * the big isuess with it are the bugs requiring epoll, which might 81 * the big isuess with it are the bugs requiring epoll, which might
78 * or might not get fixed (do I hold my breath?). 82 * or might not get fixed (do I hold my breath?).
79 */ 83 */
80 84
85/* TODO: use internal TIMEOUT */
86/* TODO: take advantage of single mmap, NODROP etc. */
87/* TODO: resize cq/sq size independently */
88
81#include <sys/timerfd.h> 89#include <sys/timerfd.h>
82#include <sys/mman.h> 90#include <sys/mman.h>
83#include <poll.h> 91#include <poll.h>
84 92
85#define IOURING_INIT_ENTRIES 32 93#define IOURING_INIT_ENTRIES 32
… …
96{ 104{
97 __u8 opcode; 105 __u8 opcode;
98 __u8 flags; 106 __u8 flags;
99 __u16 ioprio; 107 __u16 ioprio;
100 __s32 fd; 108 __s32 fd;
109 union {
101 __u64 off; 110 __u64 off;
111 __u64 addr2;
112 };
102 __u64 addr; 113 __u64 addr;
103 __u32 len; 114 __u32 len;
104 union { 115 union {
105 __kernel_rwf_t rw_flags; 116 __kernel_rwf_t rw_flags;
106 __u32 fsync_flags; 117 __u32 fsync_flags;
107 __u16 poll_events; 118 __u16 poll_events;
108 __u32 sync_range_flags; 119 __u32 sync_range_flags;
109 __u32 msg_flags; 120 __u32 msg_flags;
121 __u32 timeout_flags;
122 __u32 accept_flags;
123 __u32 cancel_flags;
124 __u32 open_flags;
125 __u32 statx_flags;
110 }; 126 };
111 __u64 user_data; 127 __u64 user_data;
112 union { 128 union {
113 __u16 buf_index; 129 __u16 buf_index;
114 __u64 __pad2[3]; 130 __u64 __pad2[3];
… …
151 __u32 sq_entries; 167 __u32 sq_entries;
152 __u32 cq_entries; 168 __u32 cq_entries;
153 __u32 flags; 169 __u32 flags;
154 __u32 sq_thread_cpu; 170 __u32 sq_thread_cpu;
155 __u32 sq_thread_idle; 171 __u32 sq_thread_idle;
172 __u32 features;
156 __u32 resv[5]; 173 __u32 resv[4];
157 struct io_sqring_offsets sq_off; 174 struct io_sqring_offsets sq_off;
158 struct io_cqring_offsets cq_off; 175 struct io_cqring_offsets cq_off;
159}; 176};
160 177
161#define IORING_OP_POLL_ADD 6 178#define IORING_OP_POLL_ADD 6
… …
164#define IORING_ENTER_GETEVENTS 0x01 181#define IORING_ENTER_GETEVENTS 0x01
165 182
166#define IORING_OFF_SQ_RING 0x00000000ULL 183#define IORING_OFF_SQ_RING 0x00000000ULL
167#define IORING_OFF_CQ_RING 0x08000000ULL 184#define IORING_OFF_CQ_RING 0x08000000ULL
168#define IORING_OFF_SQES 0x10000000ULL 185#define IORING_OFF_SQES 0x10000000ULL
186
187#define IORING_FEAT_SINGLE_MMAP 0x1
188#define IORING_FEAT_NODROP 0x2
189#define IORING_FEAT_SUBMIT_STABLE 0x4
169 190
170inline_size 191inline_size
171int 192int
172evsys_io_uring_setup (unsigned entries, struct io_uring_params *params) 193evsys_io_uring_setup (unsigned entries, struct io_uring_params *params)
173{ 194{
… …
236iouring_tfd_cb (EV_P_ struct ev_io *w, int revents) 257iouring_tfd_cb (EV_P_ struct ev_io *w, int revents)
237{ 258{
238 iouring_tfd_to = EV_TSTAMP_HUGE; 259 iouring_tfd_to = EV_TSTAMP_HUGE;
239} 260}
240 261
241static void
242iouring_epoll_cb (EV_P_ struct ev_io *w, int revents)
243{
244 epoll_poll (EV_A_ 0);
245}
246
247/* called for full and partial cleanup */ 262/* called for full and partial cleanup */
248ecb_cold 263ecb_cold
249static int 264static int
250iouring_internal_destroy (EV_P) 265iouring_internal_destroy (EV_P)
251{ 266{
… …
254 269
255 if (iouring_sq_ring != MAP_FAILED) munmap (iouring_sq_ring, iouring_sq_ring_size); 270 if (iouring_sq_ring != MAP_FAILED) munmap (iouring_sq_ring, iouring_sq_ring_size);
256 if (iouring_cq_ring != MAP_FAILED) munmap (iouring_cq_ring, iouring_cq_ring_size); 271 if (iouring_cq_ring != MAP_FAILED) munmap (iouring_cq_ring, iouring_cq_ring_size);
257 if (iouring_sqes != MAP_FAILED) munmap (iouring_sqes , iouring_sqes_size ); 272 if (iouring_sqes != MAP_FAILED) munmap (iouring_sqes , iouring_sqes_size );
258 273
259 if (ev_is_active (&iouring_epoll_w)) ev_ref (EV_A); ev_io_stop (EV_A_ &iouring_epoll_w); 274 if (ev_is_active (&iouring_tfd_w))
260 if (ev_is_active (&iouring_tfd_w )) ev_ref (EV_A); ev_io_stop (EV_A_ &iouring_tfd_w ); 275 {
276 ev_ref (EV_A);
277 ev_io_stop (EV_A_ &iouring_tfd_w);
278 }
261} 279}
262 280
263ecb_cold 281ecb_cold
264static int 282static int
265iouring_internal_init (EV_P) 283iouring_internal_init (EV_P)
… …
280 if (iouring_fd >= 0) 298 if (iouring_fd >= 0)
281 break; /* yippie */ 299 break; /* yippie */
282 300
283 if (errno != EINVAL) 301 if (errno != EINVAL)
284 return -1; /* we failed */ 302 return -1; /* we failed */
303
304#if TODO
305 if ((~params.features) & (IORING_FEAT_NODROP | IORING_FEATURE_SINGLE_MMAP))
306 return -1; /* we require the above features */
307#endif
285 308
286 /* EINVAL: lots of possible reasons, but maybe 309 /* EINVAL: lots of possible reasons, but maybe
287 * it is because we hit the unqueryable hardcoded size limit 310 * it is because we hit the unqueryable hardcoded size limit
288 */ 311 */
289 312
… …
342 iouring_internal_destroy (EV_A); 365 iouring_internal_destroy (EV_A);
343 366
344 while (iouring_internal_init (EV_A) < 0) 367 while (iouring_internal_init (EV_A) < 0)
345 ev_syserr ("(libev) io_uring_setup"); 368 ev_syserr ("(libev) io_uring_setup");
346 369
347 /* forking epoll should also effectively unregister all fds from the backend */
348 epoll_fork (EV_A);
349 /* epoll_fork already did this. hopefully */
350 /*fd_rearm_all (EV_A);*/ 370 fd_rearm_all (EV_A);
351
352 ev_io_stop (EV_A_ &iouring_epoll_w);
353 ev_io_set (EV_A_ &iouring_epoll_w, backend_fd, EV_READ);
354 ev_io_start (EV_A_ &iouring_epoll_w);
355 371
356 ev_io_stop (EV_A_ &iouring_tfd_w); 372 ev_io_stop (EV_A_ &iouring_tfd_w);
357 ev_io_set (EV_A_ &iouring_tfd_w, iouring_tfd, EV_READ); 373 ev_io_set (EV_A_ &iouring_tfd_w, iouring_tfd, EV_READ);
358 ev_io_start (EV_A_ &iouring_tfd_w); 374 ev_io_start (EV_A_ &iouring_tfd_w);
359} 375}
… …
361/*****************************************************************************/ 377/*****************************************************************************/
362 378
363static void 379static void
364iouring_modify (EV_P_ int fd, int oev, int nev) 380iouring_modify (EV_P_ int fd, int oev, int nev)
365{ 381{
366 if (ecb_expect_false (anfds [fd].eflags))
367 {
368 /* we handed this fd over to epoll, so undo this first */
369 /* we do it manually because the optimisations on epoll_modify won't do us any good */
370 epoll_ctl (iouring_fd, EPOLL_CTL_DEL, fd, 0);
371 anfds [fd].eflags = 0;
372 oev = 0;
373 }
374
375 if (oev) 382 if (oev)
376 { 383 {
377 /* we assume the sqe's are all "properly" initialised */ 384 /* we assume the sqe's are all "properly" initialised */
378 struct io_uring_sqe *sqe = iouring_sqe_get (EV_A); 385 struct io_uring_sqe *sqe = iouring_sqe_get (EV_A);
379 sqe->opcode = IORING_OP_POLL_REMOVE; 386 sqe->opcode = IORING_OP_POLL_REMOVE;
… …
428 int fd = cqe->user_data & 0xffffffffU; 435 int fd = cqe->user_data & 0xffffffffU;
429 uint32_t gen = cqe->user_data >> 32; 436 uint32_t gen = cqe->user_data >> 32;
430 int res = cqe->res; 437 int res = cqe->res;
431 438
432 /* ignore fd removal events, if there are any. TODO: verify */ 439 /* ignore fd removal events, if there are any. TODO: verify */
440 /* TODO: yes, this triggers */
433 if (cqe->user_data == (__u64)-1) 441 if (cqe->user_data == (__u64)-1)
434 abort ();//D 442 return;
435 443
436 assert (("libev: io_uring fd must be in-bounds", fd >= 0 && fd < anfdmax)); 444 assert (("libev: io_uring fd must be in-bounds", fd >= 0 && fd < anfdmax));
437 445
438 /* documentation lies, of course. the result value is NOT like 446 /* documentation lies, of course. the result value is NOT like
439 * normal syscalls, but like linux raw syscalls, i.e. negative 447 * normal syscalls, but like linux raw syscalls, i.e. negative
… …
446 if (ecb_expect_false (gen != (uint32_t)anfds [fd].egen)) 454 if (ecb_expect_false (gen != (uint32_t)anfds [fd].egen))
447 return; 455 return;
448 456
449 if (ecb_expect_false (res < 0)) 457 if (ecb_expect_false (res < 0))
450 { 458 {
451 if (res == -EINVAL) 459 /*TODO: EINVAL handling (was something failed with this fd)*/
452 { 460 /*TODO: EBUSY happens when?*/
453 /* we assume this error code means the fd/poll combination is buggy
454 * and fall back to epoll.
455 * this error code might also indicate a bug, but the kernel doesn't
456 * distinguish between those two conditions, so... sigh...
457 */
458 461
459 epoll_modify (EV_A_ fd, 0, anfds [fd].events);
460 }
461 else if (res == -EBADF) 462 if (res == -EBADF)
462 { 463 {
463 assert (("libev: event loop rejected bad fd", res != -EBADF)); 464 assert (("libev: event loop rejected bad fd", res != -EBADF));
464 fd_kill (EV_A_ fd); 465 fd_kill (EV_A_ fd);
465 } 466 }
466 else 467 else
… …
492iouring_overflow (EV_P) 493iouring_overflow (EV_P)
493{ 494{
494 /* we have two options, resize the queue (by tearing down 495 /* we have two options, resize the queue (by tearing down
495 * everything and recreating it, or living with it 496 * everything and recreating it, or living with it
496 * and polling. 497 * and polling.
497 * we implement this by resizing tghe queue, and, if that fails, 498 * we implement this by resizing the queue, and, if that fails,
498 * we just recreate the state on every failure, which 499 * we just recreate the state on every failure, which
499 * kind of is a very inefficient poll. 500 * kind of is a very inefficient poll.
500 * one danger is, due to the bios toward lower fds, 501 * one danger is, due to the bios toward lower fds,
501 * we will only really get events for those, so 502 * we will only really get events for those, so
502 * maybe we need a poll() fallback, after all. 503 * maybe we need a poll() fallback, after all.
… …
514 else 515 else
515 { 516 {
516 /* we hit the kernel limit, we should fall back to something else. 517 /* we hit the kernel limit, we should fall back to something else.
517 * we can either poll() a few times and hope for the best, 518 * we can either poll() a few times and hope for the best,
518 * poll always, or switch to epoll. 519 * poll always, or switch to epoll.
519 * since we use epoll anyways, go epoll. 520 * TODO: is this necessary with newer kernels?
520 */ 521 */
521 522
522 iouring_internal_destroy (EV_A); 523 iouring_internal_destroy (EV_A);
523 524
524 /* this should make it so that on return, we don'T call any uring functions */ 525 /* this should make it so that on return, we don't call any uring functions */
525 iouring_to_submit = 0; 526 iouring_to_submit = 0;
526 527
527 for (;;) 528 for (;;)
528 { 529 {
529 backend = epoll_init (EV_A_ 0); 530 backend = epoll_init (EV_A_ 0);
… …
603 604
604inline_size 605inline_size
605int 606int
606iouring_init (EV_P_ int flags) 607iouring_init (EV_P_ int flags)
607{ 608{
608 if (!epoll_init (EV_A_ 0))
609 return 0;
610
611 iouring_entries = IOURING_INIT_ENTRIES; 609 iouring_entries = IOURING_INIT_ENTRIES;
612 iouring_max_entries = 0; 610 iouring_max_entries = 0;
613 611
614 if (iouring_internal_init (EV_A) < 0) 612 if (iouring_internal_init (EV_A) < 0)
615 { 613 {
616 iouring_internal_destroy (EV_A); 614 iouring_internal_destroy (EV_A);
617 return 0; 615 return 0;
618 } 616 }
619 617
620 ev_io_init (&iouring_epoll_w, iouring_epoll_cb, backend_fd, EV_READ);
621 ev_set_priority (&iouring_epoll_w, EV_MAXPRI);
622
623 ev_io_init (&iouring_tfd_w, iouring_tfd_cb, iouring_tfd, EV_READ); 618 ev_io_init (&iouring_tfd_w, iouring_tfd_cb, iouring_tfd, EV_READ);
624 ev_set_priority (&iouring_tfd_w, EV_MAXPRI); 619 ev_set_priority (&iouring_tfd_w, EV_MINPRI);
625
626 ev_io_start (EV_A_ &iouring_epoll_w);
627 ev_unref (EV_A); /* watcher should not keep loop alive */
628
629 ev_io_start (EV_A_ &iouring_tfd_w); 620 ev_io_start (EV_A_ &iouring_tfd_w);
630 ev_unref (EV_A); /* watcher should not keep loop alive */ 621 ev_unref (EV_A); /* watcher should not keep loop alive */
631 622
632 backend_modify = iouring_modify; 623 backend_modify = iouring_modify;
633 backend_poll = iouring_poll; 624 backend_poll = iouring_poll;
… …
638inline_size 629inline_size
639void 630void
640iouring_destroy (EV_P) 631iouring_destroy (EV_P)
641{ 632{
642 iouring_internal_destroy (EV_A); 633 iouring_internal_destroy (EV_A);
643 epoll_destroy (EV_A);
644} 634}
645 635

Diff Legend

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