… | |
… | |
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 */ |
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 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 */ |
… | |
… | |
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 | ev_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 | ev_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 | ev_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 | ev_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 | ev_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_modfy 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 | ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ |
301 | /* io_cancel always returns some error on relevant kernels, but works */ |
|
|
302 | evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); |
170 | |
303 | |
171 | if (nev) |
304 | if (nev) |
172 | { |
305 | { |
173 | iocb->io.aio_data = fd; |
|
|
174 | iocb->io.aio_fildes = fd; |
|
|
175 | iocb->io.aio_buf = |
306 | iocb->io.aio_buf = |
176 | (nev & EV_READ ? POLLIN : 0) |
307 | (nev & EV_READ ? POLLIN : 0) |
177 | | (nev & EV_WRITE ? POLLOUT : 0); |
308 | | (nev & EV_WRITE ? POLLOUT : 0); |
178 | |
309 | |
179 | /* queue iocb up for io_submit */ |
310 | /* queue iocb up for io_submit */ |
180 | /* this assumes we only ever get one call per fd per loop iteration */ |
311 | /* this assumes we only ever get one call per fd per loop iteration */ |
… | |
… | |
183 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
314 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
184 | } |
315 | } |
185 | } |
316 | } |
186 | |
317 | |
187 | static void |
318 | static void |
|
|
319 | linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
|
|
320 | { |
|
|
321 | epoll_poll (EV_A_ 0); |
|
|
322 | } |
|
|
323 | |
|
|
324 | static void |
|
|
325 | linuxaio_fd_rearm (EV_P_ int fd) |
|
|
326 | { |
|
|
327 | anfds [fd].events = 0; |
|
|
328 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
|
|
329 | fd_change (EV_A_ fd, EV_ANFD_REIFY); |
|
|
330 | } |
|
|
331 | |
|
|
332 | static void |
188 | linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) |
333 | linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) |
189 | { |
334 | { |
190 | while (nr) |
335 | while (nr) |
191 | { |
336 | { |
192 | int fd = ev->data; |
337 | int fd = ev->data; |
193 | int res = ev->res; |
338 | int res = ev->res; |
194 | |
339 | |
195 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
340 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
196 | |
341 | |
197 | /* linux aio is oneshot: rearm fd */ |
|
|
198 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
|
|
199 | anfds [fd].events = 0; |
|
|
200 | fd_change (EV_A_ fd, 0); |
|
|
201 | |
|
|
202 | /* feed events, we do not expect or handle POLLNVAL */ |
342 | /* feed events, we do not expect or handle POLLNVAL */ |
203 | if (ecb_expect_false (res & POLLNVAL)) |
|
|
204 | fd_kill (EV_A_ fd); |
|
|
205 | else |
|
|
206 | fd_event ( |
343 | fd_event ( |
207 | EV_A_ |
344 | EV_A_ |
208 | fd, |
345 | fd, |
209 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
346 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
210 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
347 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
211 | ); |
348 | ); |
|
|
349 | |
|
|
350 | /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */ |
|
|
351 | linuxaio_fd_rearm (EV_A_ fd); |
212 | |
352 | |
213 | --nr; |
353 | --nr; |
214 | ++ev; |
354 | ++ev; |
215 | } |
355 | } |
216 | } |
356 | } |
217 | |
357 | |
218 | /* get any events from ringbuffer, return true if any were handled */ |
358 | /* get any events from ring buffer, return true if any were handled */ |
219 | static int |
359 | static int |
220 | linuxaio_get_events_from_ring (EV_P) |
360 | linuxaio_get_events_from_ring (EV_P) |
221 | { |
361 | { |
222 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
362 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
223 | |
363 | |
224 | unsigned head = ring->head; |
364 | /* the kernel reads and writes both of these variables, */ |
|
|
365 | /* as a C extension, we assume that volatile use here */ |
|
|
366 | /* both makes reads atomic and once-only */ |
|
|
367 | unsigned head = *(volatile unsigned *)&ring->head; |
225 | unsigned tail = *(volatile unsigned *)&ring->tail; |
368 | unsigned tail = *(volatile unsigned *)&ring->tail; |
226 | |
369 | |
227 | if (head == tail) |
370 | if (head == tail) |
228 | return 0; |
371 | return 0; |
229 | |
372 | |
230 | /* bail out if the ring buffer doesn't match the expected layout */ |
373 | /* bail out if the ring buffer doesn't match the expected layout */ |
231 | if (ecb_expect_false (ring->magic != AIO_RING_MAGIC) |
374 | if (expect_false (ring->magic != AIO_RING_MAGIC) |
232 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
375 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
233 | || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ |
376 | || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ |
234 | return 0; |
377 | return 0; |
235 | |
378 | |
|
|
379 | /* make sure the events up to tail are visible */ |
236 | ECB_MEMORY_FENCE_ACQUIRE; |
380 | ECB_MEMORY_FENCE_ACQUIRE; |
237 | |
381 | |
238 | /* parse all available events, but only once, to avoid starvation */ |
382 | /* parse all available events, but only once, to avoid starvation */ |
239 | if (tail > head) /* normal case around */ |
383 | if (tail > head) /* normal case around */ |
240 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
384 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
… | |
… | |
242 | { |
386 | { |
243 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
387 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
244 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
388 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
245 | } |
389 | } |
246 | |
390 | |
247 | ring->head = tail; |
391 | ECB_MEMORY_FENCE_RELEASE; |
|
|
392 | /* as an extension to C, we hope that the volatile will make this atomic and once-only */ |
|
|
393 | *(volatile unsigned *)&ring->head = tail; |
248 | |
394 | |
249 | return 1; |
395 | return 1; |
250 | } |
396 | } |
251 | |
397 | |
252 | /* read at least one event from kernel, or timeout */ |
398 | /* read at least one event from kernel, or timeout */ |
253 | inline_size |
399 | inline_size |
254 | void |
400 | void |
255 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
401 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
256 | { |
402 | { |
257 | struct timespec ts; |
403 | struct timespec ts; |
258 | struct io_event ioev; |
404 | struct io_event ioev[1]; |
259 | int res; |
405 | int res; |
260 | |
406 | |
261 | if (linuxaio_get_events_from_ring (EV_A)) |
407 | if (linuxaio_get_events_from_ring (EV_A)) |
262 | return; |
408 | return; |
263 | |
409 | |
264 | /* no events, so wait for at least one, then poll ring buffer again */ |
410 | /* no events, so wait for at least one, then poll ring buffer again */ |
265 | /* this degrades to one event per loop iteration */ |
411 | /* this degrades to one event per loop iteration */ |
266 | /* if the ring buffer changes layout, but so be it */ |
412 | /* if the ring buffer changes layout, but so be it */ |
267 | |
413 | |
|
|
414 | EV_RELEASE_CB; |
|
|
415 | |
268 | ts.tv_sec = (long)timeout; |
416 | ts.tv_sec = (long)timeout; |
269 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
417 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
270 | |
418 | |
271 | res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); |
419 | res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts); |
|
|
420 | |
|
|
421 | EV_ACQUIRE_CB; |
272 | |
422 | |
273 | if (res < 0) |
423 | if (res < 0) |
274 | if (errno == EINTR) |
424 | if (errno == EINTR) |
275 | /* ignored */; |
425 | /* ignored */; |
276 | else |
426 | else |
277 | ev_syserr ("(libev) linuxaio io_getevents"); |
427 | ev_syserr ("(libev) linuxaio io_getevents"); |
278 | else if (res) |
428 | else if (res) |
279 | { |
429 | { |
280 | /* at least one event received, handle it and any remaining ones in the ring buffer */ |
430 | /* at least one event available, handle it and any remaining ones in the ring buffer */ |
281 | linuxaio_parse_events (EV_A_ &ioev, 1); |
431 | linuxaio_parse_events (EV_A_ ioev, res); |
282 | linuxaio_get_events_from_ring (EV_A); |
432 | linuxaio_get_events_from_ring (EV_A); |
283 | } |
433 | } |
284 | } |
434 | } |
285 | |
435 | |
286 | #if EPOLL_FALLBACK |
436 | inline_size |
287 | static void |
437 | int |
288 | linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op) |
438 | linuxaio_io_setup (EV_P) |
289 | { |
439 | { |
290 | struct epoll_event eev; |
440 | linuxaio_ctx = 0; |
291 | |
441 | return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx); |
292 | eev.events = EPOLLONESHOT; |
|
|
293 | if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ; |
|
|
294 | if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT; |
|
|
295 | eev.data.fd = iocb->aio_fildes; |
|
|
296 | |
|
|
297 | if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0) |
|
|
298 | ev_syserr ("(libeio) linuxaio epoll_ctl"); |
|
|
299 | } |
442 | } |
300 | #endif |
|
|
301 | |
443 | |
302 | static void |
444 | static void |
303 | linuxaio_poll (EV_P_ ev_tstamp timeout) |
445 | linuxaio_poll (EV_P_ ev_tstamp timeout) |
304 | { |
446 | { |
305 | int submitted; |
447 | int submitted; |
306 | |
448 | |
307 | /* first phase: submit new iocbs */ |
449 | /* first phase: submit new iocbs */ |
308 | |
450 | |
309 | /* io_submit might return less than the requested number of iocbs */ |
451 | /* io_submit might return less than the requested number of iocbs */ |
310 | /* this is, afaics, only because of errors, but we go by the book and use a loop, */ |
452 | /* this is, afaics, only because of errors, but we go by the book and use a loop, */ |
311 | /* which allows us to pinpoint the errornous iocb */ |
453 | /* which allows us to pinpoint the erroneous iocb */ |
312 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
454 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
313 | { |
455 | { |
314 | int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
456 | int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
315 | |
457 | |
316 | if (ecb_expect_false (res < 0)) |
458 | if (expect_false (res < 0)) |
317 | if (errno == EAGAIN) |
459 | if (errno == EINVAL) |
318 | { |
460 | { |
319 | /* This happens when the ring buffer is full, at least. I assume this means |
|
|
320 | * that the event was queued synchronously during io_submit, and thus |
|
|
321 | * the buffer overflowed. |
|
|
322 | * In this case, we just try in next loop iteration. |
|
|
323 | * This should not result in a few fds taking priority, as the interface |
|
|
324 | * is one-shot, and we submit iocb's in a round-robin fashion. |
|
|
325 | */ |
|
|
326 | memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); |
|
|
327 | linuxaio_submitcnt -= submitted; |
|
|
328 | timeout = 0; |
|
|
329 | break; |
|
|
330 | } |
|
|
331 | #if EPOLL_FALLBACK |
|
|
332 | else if (errno == EINVAL) |
|
|
333 | { |
|
|
334 | /* This hapΓΌpens for unsupported fds, officially, but in my testing, |
461 | /* This happens for unsupported fds, officially, but in my testing, |
335 | * also randomly happens for supported fds. We fall back to good old |
462 | * also randomly happens for supported fds. We fall back to good old |
336 | * poll() here, under the assumption that this is a very rare case. |
463 | * poll() here, under the assumption that this is a very rare case. |
|
|
464 | * See https://lore.kernel.org/patchwork/patch/1047453/ to see |
|
|
465 | * discussion about such a case (ttys) where polling for POLLIN |
|
|
466 | * fails but POLLIN|POLLOUT works. |
337 | */ |
467 | */ |
338 | struct iocb *iocb = linuxaio_submits [submitted]; |
468 | struct iocb *iocb = linuxaio_submits [submitted]; |
|
|
469 | epoll_modify (EV_A_ iocb->aio_fildes, 0, anfds [iocb->aio_fildes].events); |
|
|
470 | iocb->aio_reqprio = -1; /* mark iocb as epoll */ |
|
|
471 | |
|
|
472 | res = 1; /* skip this iocb - another iocb, another chance */ |
|
|
473 | } |
|
|
474 | else if (errno == EAGAIN) |
|
|
475 | { |
|
|
476 | /* This happens when the ring buffer is full, or some other shit we |
|
|
477 | * don't know and isn't documented. Most likely because we have too |
|
|
478 | * many requests and linux aio can't be assed to handle them. |
|
|
479 | * In this case, we try to allocate a larger ring buffer, freeing |
|
|
480 | * ours first. This might fail, in which case we have to fall back to 100% |
|
|
481 | * epoll. |
|
|
482 | * God, how I hate linux not getting its act together. Ever. |
|
|
483 | */ |
|
|
484 | evsys_io_destroy (linuxaio_ctx); |
|
|
485 | linuxaio_submitcnt = 0; |
|
|
486 | |
|
|
487 | /* rearm all fds with active iocbs */ |
|
|
488 | { |
|
|
489 | int fd; |
|
|
490 | for (fd = 0; fd < linuxaio_iocbpmax; ++fd) |
|
|
491 | if (linuxaio_iocbps [fd]->io.aio_buf) |
|
|
492 | linuxaio_fd_rearm (EV_A_ fd); |
|
|
493 | } |
|
|
494 | |
|
|
495 | ++linuxaio_iteration; |
|
|
496 | if (linuxaio_io_setup (EV_A) < 0) |
|
|
497 | { |
|
|
498 | /* to bad, we can't get a new aio context, go 100% epoll */ |
|
|
499 | linuxaio_free_iocbp (EV_A); |
|
|
500 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
|
|
501 | ev_ref (EV_A); |
|
|
502 | linuxaio_ctx = 0; |
|
|
503 | backend_modify = epoll_modify; |
|
|
504 | backend_poll = epoll_poll; |
|
|
505 | } |
|
|
506 | |
|
|
507 | timeout = 0; |
|
|
508 | /* it's easiest to handle this mess in another iteration */ |
|
|
509 | return; |
|
|
510 | } |
|
|
511 | else if (errno == EBADF) |
|
|
512 | { |
|
|
513 | fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); |
|
|
514 | |
339 | res = 1; /* skip this iocb */ |
515 | res = 1; /* skip this iocb */ |
340 | |
|
|
341 | linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD); |
|
|
342 | iocb->aio_reqprio = -1; /* mark iocb as epoll */ |
|
|
343 | } |
516 | } |
344 | #endif |
|
|
345 | else |
517 | else |
346 | ev_syserr ("(libev) linuxaio io_submit"); |
518 | ev_syserr ("(libev) linuxaio io_submit"); |
347 | |
519 | |
348 | submitted += res; |
520 | submitted += res; |
349 | } |
521 | } |
… | |
… | |
353 | /* second phase: fetch and parse events */ |
525 | /* second phase: fetch and parse events */ |
354 | |
526 | |
355 | linuxaio_get_events (EV_A_ timeout); |
527 | linuxaio_get_events (EV_A_ timeout); |
356 | } |
528 | } |
357 | |
529 | |
358 | #if EPOLL_FALLBACK |
|
|
359 | |
|
|
360 | static void |
|
|
361 | linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
|
|
362 | { |
|
|
363 | struct epoll_event events[16]; |
|
|
364 | |
|
|
365 | for (;;) |
|
|
366 | { |
|
|
367 | int idx; |
|
|
368 | int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0); |
|
|
369 | |
|
|
370 | if (ecb_expect_false (res < 0)) |
|
|
371 | ev_syserr ("(libev) linuxaio epoll_wait"); |
|
|
372 | else if (!res) |
|
|
373 | break; |
|
|
374 | |
|
|
375 | for (idx = res; idx--; ) |
|
|
376 | { |
|
|
377 | int fd = events [idx].data.fd; |
|
|
378 | uint32_t ev = events [idx].events; |
|
|
379 | |
|
|
380 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
|
|
381 | |
|
|
382 | linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD); |
|
|
383 | |
|
|
384 | fd_event (EV_A_ fd, |
|
|
385 | (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) |
|
|
386 | | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0)); |
|
|
387 | } |
|
|
388 | |
|
|
389 | if (res < sizeof (events) / sizeof (events [0])) |
|
|
390 | break; |
|
|
391 | } |
|
|
392 | } |
|
|
393 | |
|
|
394 | #endif |
|
|
395 | |
|
|
396 | inline_size |
530 | inline_size |
397 | int |
531 | int |
398 | linuxaio_init (EV_P_ int flags) |
532 | linuxaio_init (EV_P_ int flags) |
399 | { |
533 | { |
400 | /* would be great to have a nice test for IOCB_CMD_POLL instead */ |
534 | /* would be great to have a nice test for IOCB_CMD_POLL instead */ |
401 | /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ |
535 | /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ |
402 | if (ev_linux_version () < 0x041200) /* 4.18 introduced IOCB_CMD_POLL */ |
536 | /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */ |
|
|
537 | if (ev_linux_version () < 0x041300) |
403 | return 0; |
538 | return 0; |
404 | |
539 | |
405 | linuxaio_ctx = 0; |
540 | if (!epoll_init (EV_A_ 0)) |
406 | if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
|
|
407 | return 0; |
541 | return 0; |
408 | |
542 | |
409 | #if EPOLL_FALLBACK |
543 | linuxaio_iteration = 0; |
410 | backend_fd = ev_epoll_create (); |
544 | |
411 | if (backend_fd < 0) |
545 | if (linuxaio_io_setup (EV_A) < 0) |
412 | { |
546 | { |
413 | ev_io_destroy (linuxaio_ctx); |
547 | epoll_destroy (EV_A); |
414 | return 0; |
548 | return 0; |
415 | } |
549 | } |
416 | |
550 | |
417 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
551 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
|
|
552 | ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); |
418 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
553 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
419 | #endif |
554 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
420 | |
555 | |
421 | backend_modify = linuxaio_modify; |
556 | backend_modify = linuxaio_modify; |
422 | backend_poll = linuxaio_poll; |
557 | backend_poll = linuxaio_poll; |
423 | |
558 | |
424 | linuxaio_iocbpmax = 0; |
559 | linuxaio_iocbpmax = 0; |
… | |
… | |
433 | |
568 | |
434 | inline_size |
569 | inline_size |
435 | void |
570 | void |
436 | linuxaio_destroy (EV_P) |
571 | linuxaio_destroy (EV_P) |
437 | { |
572 | { |
438 | #if EPOLL_FALLBACK |
573 | epoll_destroy (EV_A); |
439 | close (backend_fd); |
|
|
440 | #endif |
|
|
441 | linuxaio_free_iocbp (EV_A); |
574 | linuxaio_free_iocbp (EV_A); |
442 | ev_io_destroy (linuxaio_ctx); |
575 | evsys_io_destroy (linuxaio_ctx); |
443 | } |
576 | } |
444 | |
577 | |
445 | inline_size |
578 | inline_size |
446 | void |
579 | void |
447 | linuxaio_fork (EV_P) |
580 | linuxaio_fork (EV_P) |
448 | { |
581 | { |
449 | /* this frees all iocbs, which is very heavy-handed */ |
582 | /* this frees all iocbs, which is very heavy-handed */ |
450 | linuxaio_destroy (EV_A); |
583 | linuxaio_destroy (EV_A); |
451 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
584 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
452 | |
585 | |
453 | linuxaio_ctx = 0; |
586 | linuxaio_iteration = 0; /* we start over in the child */ |
454 | while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
587 | |
|
|
588 | while (linuxaio_io_setup (EV_A) < 0) |
455 | ev_syserr ("(libev) linuxaio io_setup"); |
589 | ev_syserr ("(libev) linuxaio io_setup"); |
456 | |
590 | |
457 | #if EPOLL_FALLBACK |
591 | epoll_fork (EV_A); |
458 | while ((backend_fd = ev_epoll_create ()) < 0) |
|
|
459 | ev_syserr ("(libev) linuxaio epoll_create"); |
|
|
460 | |
592 | |
461 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
593 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
462 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
594 | ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
463 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
595 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
464 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
|
|
465 | #endif |
|
|
466 | |
596 | |
|
|
597 | /* epoll_fork already did this. hopefully */ |
467 | fd_rearm_all (EV_A); |
598 | /*fd_rearm_all (EV_A);*/ |
468 | } |
599 | } |
469 | |
600 | |