… | |
… | |
56 | * POLLOUT|POLLIN, but polling for POLLIN fails. just great, |
56 | * POLLOUT|POLLIN, but polling for POLLIN fails. just great, |
57 | * so we have to fall back to something else (hello, epoll), |
57 | * so we have to fall back to something else (hello, epoll), |
58 | * but at least the fallback can be slow, because these are |
58 | * but at least the fallback can be slow, because these are |
59 | * exceptional cases, right? |
59 | * exceptional cases, right? |
60 | * d) hmm, you have to tell the kernel the maximum number of watchers |
60 | * d) hmm, you have to tell the kernel the maximum number of watchers |
61 | * you want to queue when initialiasing the aio context. but of |
61 | * you want to queue when initialising the aio context. but of |
62 | * course the real limit is magically calculated in the kernel, and |
62 | * course the real limit is magically calculated in the kernel, and |
63 | * is often higher then we asked for. so we just have to destroy |
63 | * is often higher then we asked for. so we just have to destroy |
64 | * the aio context and re-create it a bit larger if we hit the limit. |
64 | * the aio context and re-create it a bit larger if we hit the limit. |
65 | * (starts to remind you of epoll? well, it's a bit more deterministic |
65 | * (starts to remind you of epoll? well, it's a bit more deterministic |
66 | * and less gambling, but still ugly as hell). |
66 | * and less gambling, but still ugly as hell). |
… | |
… | |
68 | * limit. or the kernel simply doesn't want to handle your watchers. |
68 | * limit. or the kernel simply doesn't want to handle your watchers. |
69 | * what the fuck do we do then? you guessed it, in the middle |
69 | * what the fuck do we do then? you guessed it, in the middle |
70 | * of event handling we have to switch to 100% epoll polling. and |
70 | * of event handling we have to switch to 100% epoll polling. and |
71 | * that better is as fast as normal epoll polling, so you practically |
71 | * that better is as fast as normal epoll polling, so you practically |
72 | * have to use the normal epoll backend with all its quirks. |
72 | * have to use the normal epoll backend with all its quirks. |
73 | * f) end result of this trainwreck: it inherits all the disadvantages |
73 | * f) end result of this train wreck: it inherits all the disadvantages |
74 | * from epoll, while adding a number on its own. why even bother to use |
74 | * from epoll, while adding a number on its own. why even bother to use |
75 | * it? because if conditions are right and your fds are supported and you |
75 | * it? because if conditions are right and your fds are supported and you |
76 | * don't hit a limit, this backend is actually faster, doesn't gamble with |
76 | * don't hit a limit, this backend is actually faster, doesn't gamble with |
77 | * your fds, batches watchers and events and doesn't require costly state |
77 | * your fds, batches watchers and events and doesn't require costly state |
78 | * recreates. well, until it does. |
78 | * recreates. well, until it does. |
79 | * g) all of this makes this backend use almost twice as much code as epoll. |
79 | * g) all of this makes this backend use almost twice as much code as epoll. |
80 | * which in turn uses twice as much code as poll. and thats not counting |
80 | * which in turn uses twice as much code as poll. and that#s not counting |
81 | * the fact that this backend also depends on the epoll backend, making |
81 | * the fact that this backend also depends on the epoll backend, making |
82 | * it three times as much code as poll, or kqueue. |
82 | * it three times as much code as poll, or kqueue. |
83 | * h) bleah. why can't linux just do kqueue. sure kqueue is ugly, but by now |
83 | * h) bleah. why can't linux just do kqueue. sure kqueue is ugly, but by now |
84 | * it's clear that whetaver linux comes up with is far, far, far worse. |
84 | * it's clear that whatever linux comes up with is far, far, far worse. |
85 | */ |
85 | */ |
86 | |
86 | |
87 | #include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */ |
87 | #include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */ |
88 | #include <poll.h> |
88 | #include <poll.h> |
89 | #include <linux/aio_abi.h> |
89 | #include <linux/aio_abi.h> |
… | |
… | |
100 | * not only is this totally undocumented, not even the source code |
100 | * not only is this totally undocumented, not even the source code |
101 | * can tell you what the future semantics of compat_features and |
101 | * can tell you what the future semantics of compat_features and |
102 | * incompat_features are, or what header_length actually is for. |
102 | * incompat_features are, or what header_length actually is for. |
103 | */ |
103 | */ |
104 | #define AIO_RING_MAGIC 0xa10a10a1 |
104 | #define AIO_RING_MAGIC 0xa10a10a1 |
105 | #define AIO_RING_INCOMPAT_FEATURES 0 |
105 | #define EV_AIO_RING_INCOMPAT_FEATURES 0 |
106 | struct aio_ring |
106 | struct aio_ring |
107 | { |
107 | { |
108 | unsigned id; /* kernel internal index number */ |
108 | unsigned id; /* kernel internal index number */ |
109 | unsigned nr; /* number of io_events */ |
109 | unsigned nr; /* number of io_events */ |
110 | unsigned head; /* Written to by userland or by kernel. */ |
110 | unsigned head; /* Written to by userland or by kernel. */ |
… | |
… | |
120 | |
120 | |
121 | inline_size |
121 | inline_size |
122 | int |
122 | int |
123 | evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) |
123 | evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) |
124 | { |
124 | { |
125 | return syscall (SYS_io_setup, nr_events, ctx_idp); |
125 | return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp); |
126 | } |
126 | } |
127 | |
127 | |
128 | inline_size |
128 | inline_size |
129 | int |
129 | int |
130 | evsys_io_destroy (aio_context_t ctx_id) |
130 | evsys_io_destroy (aio_context_t ctx_id) |
131 | { |
131 | { |
132 | return syscall (SYS_io_destroy, ctx_id); |
132 | return ev_syscall1 (SYS_io_destroy, ctx_id); |
133 | } |
133 | } |
134 | |
134 | |
135 | inline_size |
135 | inline_size |
136 | int |
136 | int |
137 | evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) |
137 | evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) |
138 | { |
138 | { |
139 | return syscall (SYS_io_submit, ctx_id, nr, cbp); |
139 | return ev_syscall3 (SYS_io_submit, ctx_id, nr, cbp); |
140 | } |
140 | } |
141 | |
141 | |
142 | inline_size |
142 | inline_size |
143 | int |
143 | int |
144 | evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) |
144 | evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) |
145 | { |
145 | { |
146 | return syscall (SYS_io_cancel, ctx_id, cbp, result); |
146 | return ev_syscall3 (SYS_io_cancel, ctx_id, cbp, result); |
147 | } |
147 | } |
148 | |
148 | |
149 | inline_size |
149 | inline_size |
150 | int |
150 | int |
151 | evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) |
151 | evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) |
152 | { |
152 | { |
153 | return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); |
153 | return ev_syscall5 (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); |
154 | } |
154 | } |
155 | |
155 | |
156 | /*****************************************************************************/ |
156 | /*****************************************************************************/ |
157 | /* actual backed implementation */ |
157 | /* actual backed implementation */ |
158 | |
158 | |
… | |
… | |
190 | requests = requests / one_page * one_page + first_page; |
190 | requests = requests / one_page * one_page + first_page; |
191 | |
191 | |
192 | return requests; |
192 | return requests; |
193 | } |
193 | } |
194 | |
194 | |
195 | /* we use out own wrapper structure in acse we ever want to do something "clever" */ |
195 | /* we use out own wrapper structure in case we ever want to do something "clever" */ |
196 | typedef struct aniocb |
196 | typedef struct aniocb |
197 | { |
197 | { |
198 | struct iocb io; |
198 | struct iocb io; |
199 | /*int inuse;*/ |
199 | /*int inuse;*/ |
200 | } *ANIOCBP; |
200 | } *ANIOCBP; |
… | |
… | |
203 | void |
203 | void |
204 | linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count) |
204 | linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count) |
205 | { |
205 | { |
206 | while (count--) |
206 | while (count--) |
207 | { |
207 | { |
208 | /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */ |
208 | /* TODO: quite the overhead to allocate every iocb separately, maybe use our own allocator? */ |
209 | ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb)); |
209 | ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb)); |
210 | |
210 | |
211 | /* full zero initialise is probably not required at the moment, but |
211 | /* full zero initialise is probably not required at the moment, but |
212 | * this is not well documented, so we better do it. |
212 | * this is not well documented, so we better do it. |
213 | */ |
213 | */ |
… | |
… | |
235 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
235 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
236 | { |
236 | { |
237 | array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); |
237 | array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); |
238 | ANIOCBP iocb = linuxaio_iocbps [fd]; |
238 | ANIOCBP iocb = linuxaio_iocbps [fd]; |
239 | |
239 | |
240 | if (iocb->io.aio_reqprio < 0) |
240 | if (ecb_expect_false (iocb->io.aio_reqprio < 0)) |
241 | { |
241 | { |
242 | /* we handed this fd over to epoll, so undo this first */ |
242 | /* we handed this fd over to epoll, so undo this first */ |
243 | /* we do it manually becvause the optimisations on epoll_modfy won't do us any good */ |
243 | /* we do it manually because the optimisations on epoll_modify won't do us any good */ |
244 | epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); |
244 | epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); |
|
|
245 | anfds [fd].emask = 0; |
245 | iocb->io.aio_reqprio = 0; |
246 | iocb->io.aio_reqprio = 0; |
246 | } |
247 | } |
247 | |
248 | |
|
|
249 | if (ecb_expect_false (iocb->io.aio_buf)) |
|
|
250 | { |
|
|
251 | /* iocb active, so cancel it first before resubmit */ |
|
|
252 | for (;;) |
|
|
253 | { |
|
|
254 | /* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */ |
|
|
255 | if (ecb_expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 0)) |
|
|
256 | break; |
|
|
257 | |
|
|
258 | if (ecb_expect_true (errno == EINPROGRESS)) |
|
|
259 | break; |
|
|
260 | |
|
|
261 | /* the EINPROGRESS test is for nicer error message. clumsy. */ |
|
|
262 | assert (("libev: linuxaio unexpected io_cancel failed", errno != EINPROGRESS && errno != EINTR)); |
|
|
263 | } |
|
|
264 | } |
|
|
265 | |
248 | if (iocb->io.aio_buf) |
266 | iocb->io.aio_buf = |
249 | /* io_cancel always returns some error on relevant kernels, but works */ |
267 | (nev & EV_READ ? POLLIN : 0) |
250 | evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); |
268 | | (nev & EV_WRITE ? POLLOUT : 0); |
251 | |
269 | |
252 | if (nev) |
270 | if (nev) |
253 | { |
271 | { |
254 | iocb->io.aio_buf = |
|
|
255 | (nev & EV_READ ? POLLIN : 0) |
|
|
256 | | (nev & EV_WRITE ? POLLOUT : 0); |
|
|
257 | |
|
|
258 | /* queue iocb up for io_submit */ |
272 | /* queue iocb up for io_submit */ |
259 | /* this assumes we only ever get one call per fd per loop iteration */ |
273 | /* this assumes we only ever get one call per fd per loop iteration */ |
260 | ++linuxaio_submitcnt; |
274 | ++linuxaio_submitcnt; |
261 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
275 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
262 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
276 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
… | |
… | |
267 | linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
281 | linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
268 | { |
282 | { |
269 | epoll_poll (EV_A_ 0); |
283 | epoll_poll (EV_A_ 0); |
270 | } |
284 | } |
271 | |
285 | |
272 | static void |
286 | inline_speed |
|
|
287 | void |
273 | linuxaio_fd_rearm (EV_P_ int fd) |
288 | linuxaio_fd_rearm (EV_P_ int fd) |
274 | { |
289 | { |
275 | anfds [fd].events = 0; |
290 | anfds [fd].events = 0; |
276 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
291 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
277 | fd_change (EV_A_ fd, EV_ANFD_REIFY); |
292 | fd_change (EV_A_ fd, EV_ANFD_REIFY); |
… | |
… | |
293 | fd, |
308 | fd, |
294 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
309 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
295 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
310 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
296 | ); |
311 | ); |
297 | |
312 | |
298 | /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */ |
313 | /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */ |
299 | linuxaio_fd_rearm (EV_A_ fd); |
314 | linuxaio_fd_rearm (EV_A_ fd); |
300 | |
315 | |
301 | --nr; |
316 | --nr; |
302 | ++ev; |
317 | ++ev; |
303 | } |
318 | } |
304 | } |
319 | } |
305 | |
320 | |
306 | /* get any events from ringbuffer, return true if any were handled */ |
321 | /* get any events from ring buffer, return true if any were handled */ |
307 | static int |
322 | static int |
308 | linuxaio_get_events_from_ring (EV_P) |
323 | linuxaio_get_events_from_ring (EV_P) |
309 | { |
324 | { |
310 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
325 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
|
|
326 | unsigned head, tail; |
311 | |
327 | |
312 | /* the kernel reads and writes both of these variables, */ |
328 | /* the kernel reads and writes both of these variables, */ |
313 | /* as a C extension, we assume that volatile use here */ |
329 | /* as a C extension, we assume that volatile use here */ |
314 | /* both makes reads atomic and once-only */ |
330 | /* both makes reads atomic and once-only */ |
315 | unsigned head = *(volatile unsigned *)&ring->head; |
331 | head = *(volatile unsigned *)&ring->head; |
|
|
332 | ECB_MEMORY_FENCE_ACQUIRE; |
316 | unsigned tail = *(volatile unsigned *)&ring->tail; |
333 | tail = *(volatile unsigned *)&ring->tail; |
317 | |
334 | |
318 | if (head == tail) |
335 | if (head == tail) |
319 | return 0; |
336 | return 0; |
320 | |
337 | |
321 | /* bail out if the ring buffer doesn't match the expected layout */ |
|
|
322 | if (expect_false (ring->magic != AIO_RING_MAGIC) |
|
|
323 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
|
|
324 | || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ |
|
|
325 | return 0; |
|
|
326 | |
|
|
327 | /* make sure the events up to tail are visible */ |
|
|
328 | ECB_MEMORY_FENCE_ACQUIRE; |
|
|
329 | |
|
|
330 | /* parse all available events, but only once, to avoid starvation */ |
338 | /* parse all available events, but only once, to avoid starvation */ |
331 | if (tail > head) /* normal case around */ |
339 | if (ecb_expect_true (tail > head)) /* normal case around */ |
332 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
340 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
333 | else /* wrapped around */ |
341 | else /* wrapped around */ |
334 | { |
342 | { |
335 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
343 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
336 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
344 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
337 | } |
345 | } |
338 | |
346 | |
339 | ECB_MEMORY_FENCE_RELAXED; |
347 | ECB_MEMORY_FENCE_RELEASE; |
340 | /* as an extension to C, we hope that the volatile will make this atomic and once-only */ |
348 | /* as an extension to C, we hope that the volatile will make this atomic and once-only */ |
341 | *(volatile unsigned *)&ring->head = tail; |
349 | *(volatile unsigned *)&ring->head = tail; |
342 | /* make sure kernel can see our new head value - probably not required */ |
|
|
343 | ECB_MEMORY_FENCE_RELEASE; |
|
|
344 | |
350 | |
345 | return 1; |
351 | return 1; |
|
|
352 | } |
|
|
353 | |
|
|
354 | inline_size |
|
|
355 | int |
|
|
356 | linuxaio_ringbuf_valid (EV_P) |
|
|
357 | { |
|
|
358 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
|
|
359 | |
|
|
360 | return ecb_expect_true (ring->magic == AIO_RING_MAGIC) |
|
|
361 | && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES |
|
|
362 | && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */ |
346 | } |
363 | } |
347 | |
364 | |
348 | /* read at least one event from kernel, or timeout */ |
365 | /* read at least one event from kernel, or timeout */ |
349 | inline_size |
366 | inline_size |
350 | void |
367 | void |
351 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
368 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
352 | { |
369 | { |
353 | struct timespec ts; |
370 | struct timespec ts; |
354 | struct io_event ioev[1]; |
371 | struct io_event ioev[8]; /* 256 octet stack space */ |
355 | int res; |
372 | int want = 1; /* how many events to request */ |
|
|
373 | int ringbuf_valid = linuxaio_ringbuf_valid (EV_A); |
356 | |
374 | |
|
|
375 | if (ecb_expect_true (ringbuf_valid)) |
|
|
376 | { |
|
|
377 | /* if the ring buffer has any events, we don't wait or call the kernel at all */ |
357 | if (linuxaio_get_events_from_ring (EV_A)) |
378 | if (linuxaio_get_events_from_ring (EV_A)) |
358 | return; |
379 | return; |
359 | |
380 | |
360 | /* no events, so wait for at least one, then poll ring buffer again */ |
381 | /* if the ring buffer is empty, and we don't have a timeout, then don't call the kernel */ |
361 | /* this degrades to one event per loop iteration */ |
382 | if (!timeout) |
362 | /* if the ring buffer changes layout, but so be it */ |
383 | return; |
|
|
384 | } |
|
|
385 | else |
|
|
386 | /* no ringbuffer, request slightly larger batch */ |
|
|
387 | want = sizeof (ioev) / sizeof (ioev [0]); |
363 | |
388 | |
|
|
389 | /* no events, so wait for some |
|
|
390 | * for fairness reasons, we do this in a loop, to fetch all events |
|
|
391 | */ |
|
|
392 | for (;;) |
|
|
393 | { |
|
|
394 | int res; |
|
|
395 | |
364 | EV_RELEASE_CB; |
396 | EV_RELEASE_CB; |
365 | |
397 | |
366 | ts.tv_sec = (long)timeout; |
398 | EV_TS_SET (ts, timeout); |
367 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
399 | res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts); |
368 | |
400 | |
369 | res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts); |
|
|
370 | |
|
|
371 | EV_ACQUIRE_CB; |
401 | EV_ACQUIRE_CB; |
372 | |
402 | |
373 | if (res < 0) |
403 | if (res < 0) |
374 | if (errno == EINTR) |
404 | if (errno == EINTR) |
375 | /* ignored */; |
405 | /* ignored, retry */; |
376 | else |
406 | else |
377 | ev_syserr ("(libev) linuxaio io_getevents"); |
407 | ev_syserr ("(libev) linuxaio io_getevents"); |
378 | else if (res) |
408 | else if (res) |
379 | { |
409 | { |
380 | /* at least one event received, handle it and any remaining ones in the ring buffer */ |
410 | /* at least one event available, handle them */ |
381 | linuxaio_parse_events (EV_A_ ioev, res); |
411 | linuxaio_parse_events (EV_A_ ioev, res); |
|
|
412 | |
|
|
413 | if (ecb_expect_true (ringbuf_valid)) |
|
|
414 | { |
|
|
415 | /* if we have a ring buffer, handle any remaining events in it */ |
382 | linuxaio_get_events_from_ring (EV_A); |
416 | linuxaio_get_events_from_ring (EV_A); |
383 | } |
|
|
384 | } |
|
|
385 | |
417 | |
386 | static int |
418 | /* at this point, we should have handled all outstanding events */ |
|
|
419 | break; |
|
|
420 | } |
|
|
421 | else if (res < want) |
|
|
422 | /* otherwise, if there were fewere events than we wanted, we assume there are no more */ |
|
|
423 | break; |
|
|
424 | } |
|
|
425 | else |
|
|
426 | break; /* no events from the kernel, we are done */ |
|
|
427 | |
|
|
428 | timeout = 0; /* only wait in the first iteration */ |
|
|
429 | } |
|
|
430 | } |
|
|
431 | |
|
|
432 | inline_size |
|
|
433 | int |
387 | linuxaio_io_setup (EV_P) |
434 | linuxaio_io_setup (EV_P) |
388 | { |
435 | { |
389 | linuxaio_ctx = 0; |
436 | linuxaio_ctx = 0; |
390 | return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx); |
437 | return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx); |
391 | } |
438 | } |
… | |
… | |
397 | |
444 | |
398 | /* first phase: submit new iocbs */ |
445 | /* first phase: submit new iocbs */ |
399 | |
446 | |
400 | /* io_submit might return less than the requested number of iocbs */ |
447 | /* io_submit might return less than the requested number of iocbs */ |
401 | /* this is, afaics, only because of errors, but we go by the book and use a loop, */ |
448 | /* this is, afaics, only because of errors, but we go by the book and use a loop, */ |
402 | /* which allows us to pinpoint the errornous iocb */ |
449 | /* which allows us to pinpoint the erroneous iocb */ |
403 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
450 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
404 | { |
451 | { |
405 | int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
452 | int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
406 | |
453 | |
407 | if (expect_false (res < 0)) |
454 | if (ecb_expect_false (res < 0)) |
408 | if (errno == EINVAL) |
455 | if (errno == EINVAL) |
409 | { |
456 | { |
410 | /* This happens for unsupported fds, officially, but in my testing, |
457 | /* This happens for unsupported fds, officially, but in my testing, |
411 | * also randomly happens for supported fds. We fall back to good old |
458 | * also randomly happens for supported fds. We fall back to good old |
412 | * poll() here, under the assumption that this is a very rare case. |
459 | * poll() here, under the assumption that this is a very rare case. |
… | |
… | |
421 | res = 1; /* skip this iocb - another iocb, another chance */ |
468 | res = 1; /* skip this iocb - another iocb, another chance */ |
422 | } |
469 | } |
423 | else if (errno == EAGAIN) |
470 | else if (errno == EAGAIN) |
424 | { |
471 | { |
425 | /* This happens when the ring buffer is full, or some other shit we |
472 | /* This happens when the ring buffer is full, or some other shit we |
426 | * dont' know and isn't documented. Most likely because we have too |
473 | * don't know and isn't documented. Most likely because we have too |
427 | * many requests and linux aio can't be assed to handle them. |
474 | * many requests and linux aio can't be assed to handle them. |
428 | * In this case, we try to allocate a larger ring buffer, freeing |
475 | * In this case, we try to allocate a larger ring buffer, freeing |
429 | * ours first. This might fail, in which case we have to fall back to 100% |
476 | * ours first. This might fail, in which case we have to fall back to 100% |
430 | * epoll. |
477 | * epoll. |
431 | * God, how I hate linux not getting its act together. Ever. |
478 | * God, how I hate linux not getting its act together. Ever. |
… | |
… | |
442 | } |
489 | } |
443 | |
490 | |
444 | ++linuxaio_iteration; |
491 | ++linuxaio_iteration; |
445 | if (linuxaio_io_setup (EV_A) < 0) |
492 | if (linuxaio_io_setup (EV_A) < 0) |
446 | { |
493 | { |
|
|
494 | /* TODO: rearm all and recreate epoll backend from scratch */ |
|
|
495 | /* TODO: might be more prudent? */ |
|
|
496 | |
447 | /* to bad, we can't get a new aio context, go 100% epoll */ |
497 | /* to bad, we can't get a new aio context, go 100% epoll */ |
448 | linuxaio_free_iocbp (EV_A); |
498 | linuxaio_free_iocbp (EV_A); |
449 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
499 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
450 | ev_ref (EV_A); |
500 | ev_ref (EV_A); |
451 | linuxaio_ctx = 0; |
501 | linuxaio_ctx = 0; |
|
|
502 | |
|
|
503 | backend = EVBACKEND_EPOLL; |
452 | backend_modify = epoll_modify; |
504 | backend_modify = epoll_modify; |
453 | backend_poll = epoll_poll; |
505 | backend_poll = epoll_poll; |
454 | } |
506 | } |
455 | |
507 | |
456 | timeout = 0; |
508 | timeout = 0; |
457 | /* it's easiest to handle this mess in another iteration */ |
509 | /* it's easiest to handle this mess in another iteration */ |
458 | return; |
510 | return; |
459 | } |
511 | } |
460 | else if (errno == EBADF) |
512 | else if (errno == EBADF) |
461 | { |
513 | { |
|
|
514 | assert (("libev: event loop rejected bad fd", errno != EBADF)); |
462 | fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); |
515 | fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); |
463 | |
516 | |
464 | res = 1; /* skip this iocb */ |
517 | res = 1; /* skip this iocb */ |
465 | } |
518 | } |
|
|
519 | else if (errno == EINTR) /* not seen in reality, not documented */ |
|
|
520 | res = 0; /* silently ignore and retry */ |
466 | else |
521 | else |
|
|
522 | { |
467 | ev_syserr ("(libev) linuxaio io_submit"); |
523 | ev_syserr ("(libev) linuxaio io_submit"); |
|
|
524 | res = 0; |
|
|
525 | } |
468 | |
526 | |
469 | submitted += res; |
527 | submitted += res; |
470 | } |
528 | } |
471 | |
529 | |
472 | linuxaio_submitcnt = 0; |
530 | linuxaio_submitcnt = 0; |
… | |
… | |
480 | int |
538 | int |
481 | linuxaio_init (EV_P_ int flags) |
539 | linuxaio_init (EV_P_ int flags) |
482 | { |
540 | { |
483 | /* would be great to have a nice test for IOCB_CMD_POLL instead */ |
541 | /* would be great to have a nice test for IOCB_CMD_POLL instead */ |
484 | /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ |
542 | /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ |
485 | /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work */ |
543 | /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */ |
486 | if (ev_linux_version () < 0x041300) |
544 | if (ev_linux_version () < 0x041300) |
487 | return 0; |
545 | return 0; |
488 | |
546 | |
489 | if (!epoll_init (EV_A_ 0)) |
547 | if (!epoll_init (EV_A_ 0)) |
490 | return 0; |
548 | return 0; |
… | |
… | |
500 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
558 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
501 | ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); |
559 | ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); |
502 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
560 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
503 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
561 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
504 | |
562 | |
505 | backend_modify = linuxaio_modify; |
563 | backend_modify = linuxaio_modify; |
506 | backend_poll = linuxaio_poll; |
564 | backend_poll = linuxaio_poll; |
507 | |
565 | |
508 | linuxaio_iocbpmax = 0; |
566 | linuxaio_iocbpmax = 0; |
509 | linuxaio_iocbps = 0; |
567 | linuxaio_iocbps = 0; |
510 | |
568 | |
511 | linuxaio_submits = 0; |
569 | linuxaio_submits = 0; |
… | |
… | |
519 | void |
577 | void |
520 | linuxaio_destroy (EV_P) |
578 | linuxaio_destroy (EV_P) |
521 | { |
579 | { |
522 | epoll_destroy (EV_A); |
580 | epoll_destroy (EV_A); |
523 | linuxaio_free_iocbp (EV_A); |
581 | linuxaio_free_iocbp (EV_A); |
524 | evsys_io_destroy (linuxaio_ctx); |
582 | evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ |
525 | } |
583 | } |
526 | |
584 | |
527 | inline_size |
585 | ecb_cold |
528 | void |
586 | static void |
529 | linuxaio_fork (EV_P) |
587 | linuxaio_fork (EV_P) |
530 | { |
588 | { |
531 | /* this frees all iocbs, which is very heavy-handed */ |
589 | /* this frees all iocbs, which is very heavy-handed */ |
532 | linuxaio_destroy (EV_A); |
590 | linuxaio_destroy (EV_A); |
533 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
591 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
… | |
… | |
535 | linuxaio_iteration = 0; /* we start over in the child */ |
593 | linuxaio_iteration = 0; /* we start over in the child */ |
536 | |
594 | |
537 | while (linuxaio_io_setup (EV_A) < 0) |
595 | while (linuxaio_io_setup (EV_A) < 0) |
538 | ev_syserr ("(libev) linuxaio io_setup"); |
596 | ev_syserr ("(libev) linuxaio io_setup"); |
539 | |
597 | |
|
|
598 | /* forking epoll should also effectively unregister all fds from the backend */ |
540 | epoll_fork (EV_A); |
599 | epoll_fork (EV_A); |
|
|
600 | /* epoll_fork already did this. hopefully */ |
|
|
601 | /*fd_rearm_all (EV_A);*/ |
541 | |
602 | |
542 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
603 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
543 | ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
604 | ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
544 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
605 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
545 | |
|
|
546 | /* epoll_fork already did this. hopefully */ |
|
|
547 | /*fd_rearm_all (EV_A);*/ |
|
|
548 | } |
606 | } |
549 | |
607 | |