… | |
… | |
35 | * and other provisions required by the GPL. If you do not delete the |
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 |
36 | * provisions above, a recipient may use your version of this file under |
37 | * either the BSD or the GPL. |
37 | * either the BSD or the GPL. |
38 | */ |
38 | */ |
39 | |
39 | |
40 | #define EPOLL_FALLBACK 1 |
40 | /* |
|
|
41 | * general notes about linux aio: |
|
|
42 | * |
|
|
43 | * a) at first, the linux aio IOCB_CMD_POLL functionality introduced in |
|
|
44 | * 4.18 looks too good to be true: both watchers and events can be |
|
|
45 | * batched, and events can even be handled in userspace using |
|
|
46 | * a ring buffer shared with the kernel. watchers can be canceled |
|
|
47 | * regardless of whether the fd has been closed. no problems with fork. |
|
|
48 | * ok, the ring buffer is 200% undocumented (there isn't even a |
|
|
49 | * header file), but otherwise, it's pure bliss! |
|
|
50 | * b) ok, watchers are one-shot, so you have to re-arm active ones |
|
|
51 | * on every iteration. so much for syscall-less event handling, |
|
|
52 | * but at least these re-arms can be batched, no big deal, right? |
|
|
53 | * c) well, linux as usual: the documentation lies to you: io_submit |
|
|
54 | * sometimes returns EINVAL because the kernel doesn't feel like |
|
|
55 | * handling your poll mask - ttys can be polled for POLLOUT, |
|
|
56 | * POLLOUT|POLLIN, but polling for POLLIN fails. just great, |
|
|
57 | * so we have to fall back to something else (hello, epoll), |
|
|
58 | * but at least the fallback can be slow, because these are |
|
|
59 | * exceptional cases, right? |
|
|
60 | * d) hmm, you have to tell the kernel the maximum number of watchers |
|
|
61 | * you want to queue when initialising the aio context. but of |
|
|
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 |
|
|
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 |
|
|
66 | * and less gambling, but still ugly as hell). |
|
|
67 | * e) that's when you find out you can also hit an arbitrary system-wide |
|
|
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 |
|
|
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 |
|
|
72 | * have to use the normal epoll backend with all its quirks. |
|
|
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 |
|
|
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 |
|
|
77 | * your fds, batches watchers and events and doesn't require costly state |
|
|
78 | * recreates. well, until it does. |
|
|
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 that#s not counting |
|
|
81 | * the fact that this backend also depends on the epoll backend, making |
|
|
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 |
|
|
84 | * it's clear that whatever linux comes up with is far, far, far worse. |
|
|
85 | */ |
41 | |
86 | |
42 | #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 */ |
43 | #include <poll.h> |
88 | #include <poll.h> |
44 | #include <linux/aio_abi.h> |
89 | #include <linux/aio_abi.h> |
45 | |
90 | |
46 | #if EPOLL_FALLBACK |
|
|
47 | # include <sys/epoll.h> |
|
|
48 | #endif |
|
|
49 | |
|
|
50 | /* we try to fill 4kB pages exactly. |
|
|
51 | * the ring buffer header is 32 bytes, every io event is 32 bytes. |
|
|
52 | * the kernel takes the io event number, doubles it, adds 2, adds the ring buffer. |
|
|
53 | * therefore the calculation below will use "exactly" 4kB for the ring buffer |
|
|
54 | */ |
|
|
55 | #define EV_LINUXAIO_DEPTH (128 / 2 - 2 - 1) /* max. number of io events per batch */ |
|
|
56 | |
|
|
57 | /*****************************************************************************/ |
91 | /*****************************************************************************/ |
58 | /* syscall wrapdadoop - this section has the raw syscall definitions */ |
92 | /* syscall wrapdadoop - this section has the raw api/abi definitions */ |
59 | |
93 | |
60 | #include <sys/syscall.h> /* no glibc wrappers */ |
94 | #include <sys/syscall.h> /* no glibc wrappers */ |
61 | |
95 | |
62 | /* aio_abi.h is not versioned in any way, so we cannot test for its existance */ |
96 | /* aio_abi.h is not versioned in any way, so we cannot test for its existance */ |
63 | #define IOCB_CMD_POLL 5 |
97 | #define IOCB_CMD_POLL 5 |
64 | |
98 | |
65 | /* taken from linux/fs/aio.c */ |
99 | /* taken from linux/fs/aio.c. yup, that's a .c file. |
|
|
100 | * not only is this totally undocumented, not even the source code |
|
|
101 | * can tell you what the future semantics of compat_features and |
|
|
102 | * incompat_features are, or what header_length actually is for. |
|
|
103 | */ |
66 | #define AIO_RING_MAGIC 0xa10a10a1 |
104 | #define AIO_RING_MAGIC 0xa10a10a1 |
67 | #define AIO_RING_INCOMPAT_FEATURES 0 |
105 | #define EV_AIO_RING_INCOMPAT_FEATURES 0 |
68 | struct aio_ring |
106 | struct aio_ring |
69 | { |
107 | { |
70 | unsigned id; /* kernel internal index number */ |
108 | unsigned id; /* kernel internal index number */ |
71 | unsigned nr; /* number of io_events */ |
109 | unsigned nr; /* number of io_events */ |
72 | unsigned head; /* Written to by userland or by kernel. */ |
110 | unsigned head; /* Written to by userland or by kernel. */ |
… | |
… | |
78 | unsigned header_length; /* size of aio_ring */ |
116 | unsigned header_length; /* size of aio_ring */ |
79 | |
117 | |
80 | struct io_event io_events[0]; |
118 | struct io_event io_events[0]; |
81 | }; |
119 | }; |
82 | |
120 | |
|
|
121 | /* |
|
|
122 | * define some syscall wrappers for common architectures |
|
|
123 | * this is mostly for nice looks during debugging, not performance. |
|
|
124 | * our syscalls return < 0, not == -1, on error. which is good |
|
|
125 | * enough for linux aio. |
|
|
126 | * TODO: arm is also common nowadays, maybe even mips and x86 |
|
|
127 | * TODO: after implementing this, it suddenly looks like overkill, but its hard to remove... |
|
|
128 | */ |
|
|
129 | #if __GNUC__ && __linux && ECB_AMD64 && !defined __OPTIMIZE_SIZE__ |
|
|
130 | /* the costly errno access probably kills this for size optimisation */ |
|
|
131 | |
|
|
132 | #define ev_syscall(nr,narg,arg1,arg2,arg3,arg4,arg5) \ |
|
|
133 | ({ \ |
|
|
134 | long res; \ |
|
|
135 | register unsigned long r5 __asm__ ("r8" ); \ |
|
|
136 | register unsigned long r4 __asm__ ("r10"); \ |
|
|
137 | register unsigned long r3 __asm__ ("rdx"); \ |
|
|
138 | register unsigned long r2 __asm__ ("rsi"); \ |
|
|
139 | register unsigned long r1 __asm__ ("rdi"); \ |
|
|
140 | if (narg >= 5) r5 = (unsigned long)(arg5); \ |
|
|
141 | if (narg >= 4) r4 = (unsigned long)(arg4); \ |
|
|
142 | if (narg >= 3) r3 = (unsigned long)(arg3); \ |
|
|
143 | if (narg >= 2) r2 = (unsigned long)(arg2); \ |
|
|
144 | if (narg >= 1) r1 = (unsigned long)(arg1); \ |
|
|
145 | __asm__ __volatile__ ( \ |
|
|
146 | "syscall\n\t" \ |
|
|
147 | : "=a" (res) \ |
|
|
148 | : "0" (nr), "r" (r1), "r" (r2), "r" (r3), "r" (r4), "r" (r5) \ |
|
|
149 | : "cc", "r11", "cx", "memory"); \ |
|
|
150 | errno = -res; \ |
|
|
151 | res; \ |
|
|
152 | }) |
|
|
153 | |
|
|
154 | #endif |
|
|
155 | |
|
|
156 | #ifdef ev_syscall |
|
|
157 | #define ev_syscall0(nr) ev_syscall (nr, 0, 0, 0, 0, 0, 0 |
|
|
158 | #define ev_syscall1(nr,arg1) ev_syscall (nr, 1, arg1, 0, 0, 0, 0) |
|
|
159 | #define ev_syscall2(nr,arg1,arg2) ev_syscall (nr, 2, arg1, arg2, 0, 0, 0) |
|
|
160 | #define ev_syscall3(nr,arg1,arg2,arg3) ev_syscall (nr, 3, arg1, arg2, arg3, 0, 0) |
|
|
161 | #define ev_syscall4(nr,arg1,arg2,arg3,arg4) ev_syscall (nr, 3, arg1, arg2, arg3, arg4, 0) |
|
|
162 | #define ev_syscall5(nr,arg1,arg2,arg3,arg4,arg5) ev_syscall (nr, 5, arg1, arg2, arg3, arg4, arg5) |
|
|
163 | #else |
|
|
164 | #define ev_syscall0(nr) syscall (nr) |
|
|
165 | #define ev_syscall1(nr,arg1) syscall (nr, arg1) |
|
|
166 | #define ev_syscall2(nr,arg1,arg2) syscall (nr, arg1, arg2) |
|
|
167 | #define ev_syscall3(nr,arg1,arg2,arg3) syscall (nr, arg1, arg2, arg3) |
|
|
168 | #define ev_syscall4(nr,arg1,arg2,arg3,arg4) syscall (nr, arg1, arg2, arg3, arg4) |
|
|
169 | #define ev_syscall5(nr,arg1,arg2,arg3,arg4,arg5) syscall (nr, arg1, arg2, arg3, arg4, arg5) |
|
|
170 | #endif |
|
|
171 | |
83 | inline_size |
172 | inline_size |
84 | int |
173 | int |
85 | evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) |
174 | evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) |
86 | { |
175 | { |
87 | return syscall (SYS_io_setup, nr_events, ctx_idp); |
176 | return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp); |
88 | } |
177 | } |
89 | |
178 | |
90 | inline_size |
179 | inline_size |
91 | int |
180 | int |
92 | evsys_io_destroy (aio_context_t ctx_id) |
181 | evsys_io_destroy (aio_context_t ctx_id) |
93 | { |
182 | { |
94 | return syscall (SYS_io_destroy, ctx_id); |
183 | return ev_syscall1 (SYS_io_destroy, ctx_id); |
95 | } |
184 | } |
96 | |
185 | |
97 | inline_size |
186 | inline_size |
98 | int |
187 | int |
99 | evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) |
188 | evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) |
100 | { |
189 | { |
101 | return syscall (SYS_io_submit, ctx_id, nr, cbp); |
190 | return ev_syscall3 (SYS_io_submit, ctx_id, nr, cbp); |
102 | } |
191 | } |
103 | |
192 | |
104 | inline_size |
193 | inline_size |
105 | int |
194 | int |
106 | evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) |
195 | evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) |
107 | { |
196 | { |
108 | return syscall (SYS_io_cancel, ctx_id, cbp, result); |
197 | return ev_syscall3 (SYS_io_cancel, ctx_id, cbp, result); |
109 | } |
198 | } |
110 | |
199 | |
111 | inline_size |
200 | inline_size |
112 | int |
201 | int |
113 | evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) |
202 | evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) |
114 | { |
203 | { |
115 | return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); |
204 | return ev_syscall5 (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); |
116 | } |
205 | } |
117 | |
206 | |
118 | /*****************************************************************************/ |
207 | /*****************************************************************************/ |
119 | /* actual backed implementation */ |
208 | /* actual backed implementation */ |
120 | |
209 | |
|
|
210 | ecb_cold |
|
|
211 | static int |
|
|
212 | linuxaio_nr_events (EV_P) |
|
|
213 | { |
|
|
214 | /* we start with 16 iocbs and incraese from there |
|
|
215 | * that's tiny, but the kernel has a rather low system-wide |
|
|
216 | * limit that can be reached quickly, so let's be parsimonious |
|
|
217 | * with this resource. |
|
|
218 | * Rest assured, the kernel generously rounds up small and big numbers |
|
|
219 | * in different ways (but doesn't seem to charge you for it). |
|
|
220 | * The 15 here is because the kernel usually has a power of two as aio-max-nr, |
|
|
221 | * and this helps to take advantage of that limit. |
|
|
222 | */ |
|
|
223 | |
|
|
224 | /* we try to fill 4kB pages exactly. |
|
|
225 | * the ring buffer header is 32 bytes, every io event is 32 bytes. |
|
|
226 | * the kernel takes the io requests number, doubles it, adds 2 |
|
|
227 | * and adds the ring buffer. |
|
|
228 | * the way we use this is by starting low, and then roughly doubling the |
|
|
229 | * size each time we hit a limit. |
|
|
230 | */ |
|
|
231 | |
|
|
232 | int requests = 15 << linuxaio_iteration; |
|
|
233 | int one_page = (4096 |
|
|
234 | / sizeof (struct io_event) ) / 2; /* how many fit into one page */ |
|
|
235 | int first_page = ((4096 - sizeof (struct aio_ring)) |
|
|
236 | / sizeof (struct io_event) - 2) / 2; /* how many fit into the first page */ |
|
|
237 | |
|
|
238 | /* if everything fits into one page, use count exactly */ |
|
|
239 | if (requests > first_page) |
|
|
240 | /* otherwise, round down to full pages and add the first page */ |
|
|
241 | requests = requests / one_page * one_page + first_page; |
|
|
242 | |
|
|
243 | return requests; |
|
|
244 | } |
|
|
245 | |
121 | /* we use out own wrapper structure in acse we ever want to do something "clever" */ |
246 | /* we use out own wrapper structure in case we ever want to do something "clever" */ |
122 | typedef struct aniocb |
247 | typedef struct aniocb |
123 | { |
248 | { |
124 | struct iocb io; |
249 | struct iocb io; |
125 | /*int inuse;*/ |
250 | /*int inuse;*/ |
126 | } *ANIOCBP; |
251 | } *ANIOCBP; |
127 | |
252 | |
128 | inline_size |
253 | inline_size |
129 | void |
254 | void |
130 | linuxaio_array_needsize_iocbp (ANIOCBP *base, int count) |
255 | linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count) |
131 | { |
256 | { |
132 | /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */ |
|
|
133 | while (count--) |
257 | while (count--) |
134 | { |
258 | { |
|
|
259 | /* TODO: quite the overhead to allocate every iocb separately, maybe use our own allocator? */ |
135 | *base = (ANIOCBP)ev_malloc (sizeof (**base)); |
260 | ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb)); |
136 | /* TODO: full zero initialize required? */ |
261 | |
|
|
262 | /* full zero initialise is probably not required at the moment, but |
|
|
263 | * this is not well documented, so we better do it. |
|
|
264 | */ |
137 | memset (*base, 0, sizeof (**base)); |
265 | memset (iocb, 0, sizeof (*iocb)); |
138 | /* would be nice to initialize fd/data as well, but array_needsize API doesn't support that */ |
266 | |
139 | (*base)->io.aio_lio_opcode = IOCB_CMD_POLL; |
267 | iocb->io.aio_lio_opcode = IOCB_CMD_POLL; |
140 | ++base; |
268 | iocb->io.aio_data = offset; |
|
|
269 | iocb->io.aio_fildes = offset; |
|
|
270 | |
|
|
271 | base [offset++] = iocb; |
141 | } |
272 | } |
142 | } |
273 | } |
143 | |
274 | |
144 | ecb_cold |
275 | ecb_cold |
145 | static void |
276 | static void |
… | |
… | |
153 | |
284 | |
154 | static void |
285 | static void |
155 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
286 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
156 | { |
287 | { |
157 | array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); |
288 | array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); |
158 | struct aniocb *iocb = linuxaio_iocbps [fd]; |
289 | ANIOCBP iocb = linuxaio_iocbps [fd]; |
159 | |
290 | |
160 | #if EPOLL_FALLBACK |
|
|
161 | if (iocb->io.aio_reqprio < 0) |
291 | if (iocb->io.aio_reqprio < 0) |
162 | { |
292 | { |
|
|
293 | /* we handed this fd over to epoll, so undo this first */ |
|
|
294 | /* we do it manually because the optimisations on epoll_modify won't do us any good */ |
163 | epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); |
295 | epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); |
|
|
296 | anfds [fd].emask = 0; |
164 | iocb->io.aio_reqprio = 0; |
297 | iocb->io.aio_reqprio = 0; |
165 | } |
298 | } |
166 | #endif |
|
|
167 | |
299 | |
168 | if (iocb->io.aio_buf) |
300 | if (iocb->io.aio_buf) |
169 | evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ |
301 | { |
|
|
302 | for (;;) |
|
|
303 | { |
|
|
304 | /* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */ |
|
|
305 | if (expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 0)) |
|
|
306 | break; |
|
|
307 | |
|
|
308 | if (expect_true (errno == EINPROGRESS)) |
|
|
309 | break; |
|
|
310 | |
|
|
311 | /* the EINPROGRESS test is for nicer error message. clumsy. */ |
|
|
312 | assert (("libev: linuxaio unexpected io_cancel failed", errno != EINPROGRESS && errno != EINTR)); |
|
|
313 | } |
|
|
314 | } |
170 | |
315 | |
171 | if (nev) |
316 | if (nev) |
172 | { |
317 | { |
173 | iocb->io.aio_data = fd; |
|
|
174 | iocb->io.aio_fildes = fd; |
|
|
175 | iocb->io.aio_buf = |
318 | iocb->io.aio_buf = |
176 | (nev & EV_READ ? POLLIN : 0) |
319 | (nev & EV_READ ? POLLIN : 0) |
177 | | (nev & EV_WRITE ? POLLOUT : 0); |
320 | | (nev & EV_WRITE ? POLLOUT : 0); |
178 | |
321 | |
179 | /* queue iocb up for io_submit */ |
322 | /* queue iocb up for io_submit */ |
180 | /* this assumes we only ever get one call per fd per loop iteration */ |
323 | /* this assumes we only ever get one call per fd per loop iteration */ |
… | |
… | |
182 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
325 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
183 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
326 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
184 | } |
327 | } |
185 | } |
328 | } |
186 | |
329 | |
187 | #if EPOLL_FALLBACK |
|
|
188 | |
|
|
189 | static void |
|
|
190 | linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op) |
|
|
191 | { |
|
|
192 | struct epoll_event eev; |
|
|
193 | |
|
|
194 | eev.events = EPOLLONESHOT; |
|
|
195 | if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ; |
|
|
196 | if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT; |
|
|
197 | eev.data.fd = iocb->aio_fildes; |
|
|
198 | |
|
|
199 | if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0) |
|
|
200 | ev_syserr ("(libeio) linuxaio epoll_ctl"); |
|
|
201 | } |
|
|
202 | |
|
|
203 | static void |
330 | static void |
204 | linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
331 | linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
205 | { |
332 | { |
206 | struct epoll_event events[16]; |
333 | epoll_poll (EV_A_ 0); |
207 | |
|
|
208 | for (;;) |
|
|
209 | { |
|
|
210 | int idx; |
|
|
211 | int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0); |
|
|
212 | |
|
|
213 | if (expect_false (res < 0)) |
|
|
214 | ev_syserr ("(libev) linuxaio epoll_wait"); |
|
|
215 | else if (!res) |
|
|
216 | break; |
|
|
217 | |
|
|
218 | for (idx = res; idx--; ) |
|
|
219 | { |
|
|
220 | int fd = events [idx].data.fd; |
|
|
221 | uint32_t ev = events [idx].events; |
|
|
222 | |
|
|
223 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
|
|
224 | |
|
|
225 | linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD); |
|
|
226 | |
|
|
227 | fd_event (EV_A_ fd, |
|
|
228 | (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) |
|
|
229 | | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0)); |
|
|
230 | } |
|
|
231 | |
|
|
232 | if (res < sizeof (events) / sizeof (events [0])) |
|
|
233 | break; |
|
|
234 | } |
|
|
235 | } |
334 | } |
236 | |
335 | |
237 | #endif |
336 | inline_speed |
|
|
337 | void |
|
|
338 | linuxaio_fd_rearm (EV_P_ int fd) |
|
|
339 | { |
|
|
340 | anfds [fd].events = 0; |
|
|
341 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
|
|
342 | fd_change (EV_A_ fd, EV_ANFD_REIFY); |
|
|
343 | } |
238 | |
344 | |
239 | static void |
345 | static void |
240 | linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) |
346 | linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) |
241 | { |
347 | { |
242 | while (nr) |
348 | while (nr) |
… | |
… | |
244 | int fd = ev->data; |
350 | int fd = ev->data; |
245 | int res = ev->res; |
351 | int res = ev->res; |
246 | |
352 | |
247 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
353 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
248 | |
354 | |
249 | /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */ |
|
|
250 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
|
|
251 | anfds [fd].events = 0; |
|
|
252 | fd_change (EV_A_ fd, 0); |
|
|
253 | |
|
|
254 | /* feed events, we do not expect or handle POLLNVAL */ |
355 | /* feed events, we do not expect or handle POLLNVAL */ |
255 | if (expect_false (res & POLLNVAL)) |
|
|
256 | fd_kill (EV_A_ fd); |
|
|
257 | else |
|
|
258 | fd_event ( |
356 | fd_event ( |
259 | EV_A_ |
357 | EV_A_ |
260 | fd, |
358 | fd, |
261 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
359 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
262 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
360 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
263 | ); |
361 | ); |
|
|
362 | |
|
|
363 | /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */ |
|
|
364 | linuxaio_fd_rearm (EV_A_ fd); |
264 | |
365 | |
265 | --nr; |
366 | --nr; |
266 | ++ev; |
367 | ++ev; |
267 | } |
368 | } |
268 | } |
369 | } |
269 | |
370 | |
270 | /* get any events from ringbuffer, return true if any were handled */ |
371 | /* get any events from ring buffer, return true if any were handled */ |
271 | static int |
372 | static int |
272 | linuxaio_get_events_from_ring (EV_P) |
373 | linuxaio_get_events_from_ring (EV_P) |
273 | { |
374 | { |
274 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
375 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
275 | |
376 | |
… | |
… | |
280 | unsigned tail = *(volatile unsigned *)&ring->tail; |
381 | unsigned tail = *(volatile unsigned *)&ring->tail; |
281 | |
382 | |
282 | if (head == tail) |
383 | if (head == tail) |
283 | return 0; |
384 | return 0; |
284 | |
385 | |
285 | /* bail out if the ring buffer doesn't match the expected layout */ |
|
|
286 | if (expect_false (ring->magic != AIO_RING_MAGIC) |
|
|
287 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
|
|
288 | || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ |
|
|
289 | return 0; |
|
|
290 | |
|
|
291 | /* make sure the events up to tail are visible */ |
386 | /* make sure the events up to tail are visible */ |
292 | ECB_MEMORY_FENCE_ACQUIRE; |
387 | ECB_MEMORY_FENCE_ACQUIRE; |
293 | |
388 | |
294 | /* parse all available events, but only once, to avoid starvation */ |
389 | /* parse all available events, but only once, to avoid starvation */ |
295 | if (tail > head) /* normal case around */ |
390 | if (tail > head) /* normal case around */ |
… | |
… | |
298 | { |
393 | { |
299 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
394 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
300 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
395 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
301 | } |
396 | } |
302 | |
397 | |
303 | /* TODO: we only need a compiler barrier here, not a read fence */ |
|
|
304 | ECB_MEMORY_FENCE_RELEASE; |
398 | ECB_MEMORY_FENCE_RELEASE; |
305 | /* as an extension to C, we hope that the volatile will make this atomic and once-only */ |
399 | /* as an extension to C, we hope that the volatile will make this atomic and once-only */ |
306 | *(volatile unsigned *)&ring->head = tail; |
400 | *(volatile unsigned *)&ring->head = tail; |
307 | /* make sure kernel can see our new head value - probably not required */ |
|
|
308 | ECB_MEMORY_FENCE_RELEASE; |
|
|
309 | |
401 | |
310 | return 1; |
402 | return 1; |
|
|
403 | } |
|
|
404 | |
|
|
405 | inline_size |
|
|
406 | int |
|
|
407 | linuxaio_ringbuf_valid (EV_P) |
|
|
408 | { |
|
|
409 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
|
|
410 | |
|
|
411 | return expect_true (ring->magic == AIO_RING_MAGIC) |
|
|
412 | && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES |
|
|
413 | && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */ |
311 | } |
414 | } |
312 | |
415 | |
313 | /* read at least one event from kernel, or timeout */ |
416 | /* read at least one event from kernel, or timeout */ |
314 | inline_size |
417 | inline_size |
315 | void |
418 | void |
316 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
419 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
317 | { |
420 | { |
318 | struct timespec ts; |
421 | struct timespec ts; |
319 | struct io_event ioev[1]; |
422 | struct io_event ioev[8]; /* 256 octet stack space */ |
320 | int res; |
423 | int want = 1; /* how many events to request */ |
|
|
424 | int ringbuf_valid = linuxaio_ringbuf_valid (EV_A); |
321 | |
425 | |
|
|
426 | if (expect_true (ringbuf_valid)) |
|
|
427 | { |
|
|
428 | /* if the ring buffer has any events, we don't wait or call the kernel at all */ |
322 | if (linuxaio_get_events_from_ring (EV_A)) |
429 | if (linuxaio_get_events_from_ring (EV_A)) |
323 | return; |
430 | return; |
324 | |
431 | |
325 | /* no events, so wait for at least one, then poll ring buffer again */ |
432 | /* if the ring buffer is empty, and we don't have a timeout, then don't call the kernel */ |
326 | /* this degrades to one event per loop iteration */ |
433 | if (!timeout) |
327 | /* if the ring buffer changes layout, but so be it */ |
434 | return; |
|
|
435 | } |
|
|
436 | else |
|
|
437 | /* no ringbuffer, request slightly larger batch */ |
|
|
438 | want = sizeof (ioev) / sizeof (ioev [0]); |
328 | |
439 | |
|
|
440 | /* no events, so wait for some |
|
|
441 | * for fairness reasons, we do this in a loop, to fetch all events |
|
|
442 | */ |
|
|
443 | for (;;) |
|
|
444 | { |
|
|
445 | int res; |
|
|
446 | |
329 | EV_RELEASE_CB; |
447 | EV_RELEASE_CB; |
330 | |
448 | |
331 | ts.tv_sec = (long)timeout; |
449 | ts.tv_sec = (long)timeout; |
332 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
450 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
333 | |
451 | |
334 | res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts); |
452 | res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts); |
335 | |
453 | |
336 | EV_ACQUIRE_CB; |
454 | EV_ACQUIRE_CB; |
337 | |
455 | |
338 | if (res < 0) |
456 | if (res < 0) |
339 | if (errno == EINTR) |
457 | if (errno == EINTR) |
340 | /* ignored */; |
458 | /* ignored, retry */; |
341 | else |
459 | else |
342 | ev_syserr ("(libev) linuxaio io_getevents"); |
460 | ev_syserr ("(libev) linuxaio io_getevents"); |
343 | else if (res) |
461 | else if (res) |
344 | { |
462 | { |
345 | /* at least one event received, handle it and any remaining ones in the ring buffer */ |
463 | /* at least one event available, handle them */ |
346 | linuxaio_parse_events (EV_A_ ioev, res); |
464 | linuxaio_parse_events (EV_A_ ioev, res); |
|
|
465 | |
|
|
466 | if (expect_true (ringbuf_valid)) |
|
|
467 | { |
|
|
468 | /* if we have a ring buffer, handle any remaining events in it */ |
347 | linuxaio_get_events_from_ring (EV_A); |
469 | linuxaio_get_events_from_ring (EV_A); |
|
|
470 | |
|
|
471 | /* at this point, we should have handled all outstanding events */ |
|
|
472 | break; |
|
|
473 | } |
|
|
474 | else if (res < want) |
|
|
475 | /* otherwise, if there were fewere events than we wanted, we assume there are no more */ |
|
|
476 | break; |
|
|
477 | } |
|
|
478 | else |
|
|
479 | break; /* no events from the kernel, we are done */ |
|
|
480 | |
|
|
481 | timeout = 0; /* only wait in the first iteration */ |
348 | } |
482 | } |
|
|
483 | } |
|
|
484 | |
|
|
485 | inline_size |
|
|
486 | int |
|
|
487 | linuxaio_io_setup (EV_P) |
|
|
488 | { |
|
|
489 | linuxaio_ctx = 0; |
|
|
490 | return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx); |
349 | } |
491 | } |
350 | |
492 | |
351 | static void |
493 | static void |
352 | linuxaio_poll (EV_P_ ev_tstamp timeout) |
494 | linuxaio_poll (EV_P_ ev_tstamp timeout) |
353 | { |
495 | { |
… | |
… | |
355 | |
497 | |
356 | /* first phase: submit new iocbs */ |
498 | /* first phase: submit new iocbs */ |
357 | |
499 | |
358 | /* io_submit might return less than the requested number of iocbs */ |
500 | /* io_submit might return less than the requested number of iocbs */ |
359 | /* this is, afaics, only because of errors, but we go by the book and use a loop, */ |
501 | /* this is, afaics, only because of errors, but we go by the book and use a loop, */ |
360 | /* which allows us to pinpoint the errornous iocb */ |
502 | /* which allows us to pinpoint the erroneous iocb */ |
361 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
503 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
362 | { |
504 | { |
363 | #if 0 |
|
|
364 | int res; |
|
|
365 | if (linuxaio_submits[submitted]->aio_fildes == backend_fd) |
|
|
366 | res = evsys_io_submit (linuxaio_ctx, 1, linuxaio_submits + submitted); |
|
|
367 | else |
|
|
368 | { res = -1; errno = EINVAL; }; |
|
|
369 | #else |
|
|
370 | int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
505 | int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
371 | #endif |
|
|
372 | |
506 | |
373 | if (expect_false (res < 0)) |
507 | if (expect_false (res < 0)) |
374 | if (errno == EAGAIN) |
|
|
375 | { |
|
|
376 | /* This happens when the ring buffer is full, at least. I assume this means |
|
|
377 | * that the event was queued synchronously during io_submit, and thus |
|
|
378 | * the buffer overflowed. |
|
|
379 | * In this case, we just try in next loop iteration. |
|
|
380 | * This should not result in a few fds taking priority, as the interface |
|
|
381 | * is one-shot, and we submit iocb's in a round-robin fashion. |
|
|
382 | * TODO: maybe make "submitted" persistent, so we don't have to memmove? |
|
|
383 | */ |
|
|
384 | if (ecb_expect_false (submitted)) |
|
|
385 | { |
|
|
386 | memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); |
|
|
387 | linuxaio_submitcnt -= submitted; |
|
|
388 | } |
|
|
389 | |
|
|
390 | timeout = 0; |
|
|
391 | break; |
|
|
392 | } |
|
|
393 | #if EPOLL_FALLBACK |
|
|
394 | else if (errno == EINVAL) |
508 | if (errno == EINVAL) |
395 | { |
509 | { |
396 | /* This happens for unsupported fds, officially, but in my testing, |
510 | /* This happens for unsupported fds, officially, but in my testing, |
397 | * also randomly happens for supported fds. We fall back to good old |
511 | * also randomly happens for supported fds. We fall back to good old |
398 | * poll() here, under the assumption that this is a very rare case. |
512 | * poll() here, under the assumption that this is a very rare case. |
399 | * See https://lore.kernel.org/patchwork/patch/1047453/ to see |
513 | * See https://lore.kernel.org/patchwork/patch/1047453/ to see |
400 | * discussion about such a case (ttys) where polling for POLLIN |
514 | * discussion about such a case (ttys) where polling for POLLIN |
401 | * fails but POLLIN|POLLOUT works. |
515 | * fails but POLLIN|POLLOUT works. |
402 | */ |
516 | */ |
403 | struct iocb *iocb = linuxaio_submits [submitted]; |
517 | struct iocb *iocb = linuxaio_submits [submitted]; |
|
|
518 | epoll_modify (EV_A_ iocb->aio_fildes, 0, anfds [iocb->aio_fildes].events); |
|
|
519 | iocb->aio_reqprio = -1; /* mark iocb as epoll */ |
|
|
520 | |
|
|
521 | res = 1; /* skip this iocb - another iocb, another chance */ |
|
|
522 | } |
|
|
523 | else if (errno == EAGAIN) |
|
|
524 | { |
|
|
525 | /* This happens when the ring buffer is full, or some other shit we |
|
|
526 | * don't know and isn't documented. Most likely because we have too |
|
|
527 | * many requests and linux aio can't be assed to handle them. |
|
|
528 | * In this case, we try to allocate a larger ring buffer, freeing |
|
|
529 | * ours first. This might fail, in which case we have to fall back to 100% |
|
|
530 | * epoll. |
|
|
531 | * God, how I hate linux not getting its act together. Ever. |
|
|
532 | */ |
|
|
533 | evsys_io_destroy (linuxaio_ctx); |
|
|
534 | linuxaio_submitcnt = 0; |
|
|
535 | |
|
|
536 | /* rearm all fds with active iocbs */ |
|
|
537 | { |
|
|
538 | int fd; |
|
|
539 | for (fd = 0; fd < linuxaio_iocbpmax; ++fd) |
|
|
540 | if (linuxaio_iocbps [fd]->io.aio_buf) |
|
|
541 | linuxaio_fd_rearm (EV_A_ fd); |
|
|
542 | } |
|
|
543 | |
|
|
544 | ++linuxaio_iteration; |
|
|
545 | if (linuxaio_io_setup (EV_A) < 0) |
|
|
546 | { |
|
|
547 | /* to bad, we can't get a new aio context, go 100% epoll */ |
|
|
548 | linuxaio_free_iocbp (EV_A); |
|
|
549 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
|
|
550 | ev_ref (EV_A); |
|
|
551 | linuxaio_ctx = 0; |
|
|
552 | backend_modify = epoll_modify; |
|
|
553 | backend_poll = epoll_poll; |
|
|
554 | } |
|
|
555 | |
|
|
556 | timeout = 0; |
|
|
557 | /* it's easiest to handle this mess in another iteration */ |
|
|
558 | return; |
|
|
559 | } |
|
|
560 | else if (errno == EBADF) |
|
|
561 | { |
|
|
562 | assert (("libev: event loop rejected bad fd", errno != EBADF)); |
|
|
563 | fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); |
|
|
564 | |
404 | res = 1; /* skip this iocb */ |
565 | res = 1; /* skip this iocb */ |
405 | |
|
|
406 | linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD); |
|
|
407 | iocb->aio_reqprio = -1; /* mark iocb as epoll */ |
|
|
408 | } |
566 | } |
409 | #endif |
567 | else if (errno == EINTR) /* not seen in reality, not documented */ |
|
|
568 | res = 0; /* silently ignore and retry */ |
410 | else |
569 | else |
411 | ev_syserr ("(libev) linuxaio io_submit"); |
570 | ev_syserr ("(libev) linuxaio io_submit"); |
412 | |
571 | |
413 | submitted += res; |
572 | submitted += res; |
414 | } |
573 | } |
… | |
… | |
424 | int |
583 | int |
425 | linuxaio_init (EV_P_ int flags) |
584 | linuxaio_init (EV_P_ int flags) |
426 | { |
585 | { |
427 | /* would be great to have a nice test for IOCB_CMD_POLL instead */ |
586 | /* would be great to have a nice test for IOCB_CMD_POLL instead */ |
428 | /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ |
587 | /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ |
429 | #if EPOLL_FALLBACK |
588 | /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */ |
430 | /* 4.19 made epoll work */ |
|
|
431 | if (ev_linux_version () < 0x041300) |
589 | if (ev_linux_version () < 0x041300) |
432 | return 0; |
590 | return 0; |
433 | #else |
591 | |
434 | /* 4.18 introduced IOCB_CMD_POLL */ |
592 | if (!epoll_init (EV_A_ 0)) |
435 | if (ev_linux_version () < 0x041200) |
|
|
436 | return 0; |
593 | return 0; |
437 | #endif |
|
|
438 | |
594 | |
439 | linuxaio_ctx = 0; |
595 | linuxaio_iteration = 0; |
440 | if (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
|
|
441 | return 0; |
|
|
442 | |
596 | |
443 | #if EPOLL_FALLBACK |
597 | if (linuxaio_io_setup (EV_A) < 0) |
444 | backend_fd = ev_epoll_create (); |
|
|
445 | if (backend_fd < 0) |
|
|
446 | { |
598 | { |
447 | evsys_io_destroy (linuxaio_ctx); |
599 | epoll_destroy (EV_A); |
448 | return 0; |
600 | return 0; |
449 | } |
601 | } |
450 | |
602 | |
451 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
603 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
452 | ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); |
604 | ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); |
453 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
605 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
454 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
606 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
455 | #endif |
|
|
456 | |
607 | |
457 | backend_modify = linuxaio_modify; |
608 | backend_modify = linuxaio_modify; |
458 | backend_poll = linuxaio_poll; |
609 | backend_poll = linuxaio_poll; |
459 | |
610 | |
460 | linuxaio_iocbpmax = 0; |
611 | linuxaio_iocbpmax = 0; |
… | |
… | |
469 | |
620 | |
470 | inline_size |
621 | inline_size |
471 | void |
622 | void |
472 | linuxaio_destroy (EV_P) |
623 | linuxaio_destroy (EV_P) |
473 | { |
624 | { |
474 | #if EPOLL_FALLBACK |
625 | epoll_destroy (EV_A); |
475 | close (backend_fd); |
|
|
476 | #endif |
|
|
477 | linuxaio_free_iocbp (EV_A); |
626 | linuxaio_free_iocbp (EV_A); |
478 | ev_io_destroy (linuxaio_ctx); |
627 | evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ |
479 | } |
628 | } |
480 | |
629 | |
481 | inline_size |
630 | inline_size |
482 | void |
631 | void |
483 | linuxaio_fork (EV_P) |
632 | linuxaio_fork (EV_P) |
484 | { |
633 | { |
485 | /* this frees all iocbs, which is very heavy-handed */ |
634 | /* this frees all iocbs, which is very heavy-handed */ |
486 | linuxaio_destroy (EV_A); |
635 | linuxaio_destroy (EV_A); |
487 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
636 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
488 | |
637 | |
489 | linuxaio_ctx = 0; |
638 | linuxaio_iteration = 0; /* we start over in the child */ |
490 | while (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
639 | |
|
|
640 | while (linuxaio_io_setup (EV_A) < 0) |
491 | ev_syserr ("(libev) linuxaio io_setup"); |
641 | ev_syserr ("(libev) linuxaio io_setup"); |
492 | |
642 | |
493 | #if EPOLL_FALLBACK |
643 | /* forking epoll should also effectively unregister all fds from the backend */ |
494 | while ((backend_fd = ev_epoll_create ()) < 0) |
644 | epoll_fork (EV_A); |
495 | ev_syserr ("(libev) linuxaio epoll_create"); |
|
|
496 | |
645 | |
497 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
646 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
498 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
647 | ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
499 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
648 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
500 | #endif |
|
|
501 | |
649 | |
|
|
650 | /* epoll_fork already did this. hopefully */ |
502 | fd_rearm_all (EV_A); |
651 | /*fd_rearm_all (EV_A);*/ |
503 | } |
652 | } |
504 | |
653 | |