blob: 708a0954bf117fdd76d7490967ed3db406d19bd9 [file] [log] [blame]
Willy Tarreaude99e992007-04-16 00:53:59 +02001/*
2 * FD polling functions for Speculative I/O combined with Linux epoll()
3 *
Willy Tarreauf2e8ee22008-05-25 10:39:02 +02004 * Copyright 2000-2008 Willy Tarreau <w@1wt.eu>
Willy Tarreaude99e992007-04-16 00:53:59 +02005 *
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version
9 * 2 of the License, or (at your option) any later version.
10 *
Willy Tarreauf2e8ee22008-05-25 10:39:02 +020011 *
12 * This code implements "speculative I/O" under Linux. The principle is to
13 * try to perform expected I/O before registering the events in the poller.
14 * Each time this succeeds, it saves an expensive epoll_ctl(). It generally
15 * succeeds for all reads after an accept(), and for writes after a connect().
16 * It also improves performance for streaming connections because even if only
17 * one side is polled, the other one may react accordingly depending on the
18 * level of the buffer.
19 *
20 * It has a presents drawbacks though. If too many events are set for spec I/O,
21 * those ones can starve the polled events. Experiments show that when polled
22 * events starve, they quickly turn into spec I/O, making the situation even
23 * worse. While we can reduce the number of polled events processed at once,
24 * we cannot do this on speculative events because most of them are new ones
25 * (avg 2/3 new - 1/3 old from experiments).
26 *
27 * The solution against this problem relies on those two factors :
28 * 1) one FD registered as a spec event cannot be polled at the same time
29 * 2) even during very high loads, we will almost never be interested in
30 * simultaneous read and write streaming on the same FD.
31 *
32 * The first point implies that during starvation, we will not have more than
33 * half of our FDs in the poll list, otherwise it means there is less than that
34 * in the spec list, implying there is no starvation.
35 *
36 * The second point implies that we're statically only interested in half of
37 * the maximum number of file descriptors at once, because we will unlikely
38 * have simultaneous read and writes for a same buffer during long periods.
39 *
40 * So, if we make it possible to drain maxsock/2/2 during peak loads, then we
41 * can ensure that there will be no starvation effect. This means that we must
42 * always allocate maxsock/4 events for the poller.
43 *
44 *
Willy Tarreaude99e992007-04-16 00:53:59 +020045 */
46
47#include <unistd.h>
48#include <sys/time.h>
49#include <sys/types.h>
50
51#include <common/compat.h>
52#include <common/config.h>
Willy Tarreaud6f087e2008-01-18 17:20:13 +010053#include <common/debug.h>
Willy Tarreaude99e992007-04-16 00:53:59 +020054#include <common/standard.h>
Willy Tarreau0c303ee2008-07-07 00:09:58 +020055#include <common/ticks.h>
Willy Tarreaude99e992007-04-16 00:53:59 +020056#include <common/time.h>
Willy Tarreau1db37712007-06-03 17:16:49 +020057#include <common/tools.h>
Willy Tarreaude99e992007-04-16 00:53:59 +020058
Willy Tarreaude99e992007-04-16 00:53:59 +020059#include <types/global.h>
60
61#include <proto/fd.h>
62#include <proto/task.h>
63
64#if defined(USE_MY_EPOLL)
65#include <common/epoll.h>
66#include <errno.h>
67#include <sys/syscall.h>
68static _syscall1 (int, epoll_create, int, size);
69static _syscall4 (int, epoll_ctl, int, epfd, int, op, int, fd, struct epoll_event *, event);
70static _syscall4 (int, epoll_wait, int, epfd, struct epoll_event *, events, int, maxevents, int, timeout);
71#else
72#include <sys/epoll.h>
73#endif
74
75/*
76 * We define 4 states for each direction of a file descriptor, which we store
77 * as 2 bits :
78 *
79 * 00 = IDLE : we're not interested in this event
80 * 01 = SPEC : perform speculative I/O on this FD
81 * 10 = WAIT : really wait for an availability event on this FD (poll)
82 * 11 = STOP : was marked WAIT, but disabled. It can switch back to WAIT if
83 * the application changes its mind, otherwise disable FD polling
84 * and switch back to IDLE.
85 *
86 * Since we do not want to scan all the FD list to find speculative I/O events,
87 * we store them in a list consisting in a linear array holding only the FD
Willy Tarreau7a52a5c2008-08-16 16:06:02 +020088 * indexes right now. Note that a closed FD cannot exist in the spec list,
89 * because it is closed by fd_delete() which in turn calls __fd_clo() which
90 * always removes it from the list.
Willy Tarreaude99e992007-04-16 00:53:59 +020091 *
92 * The STOP state requires the event to be present in the spec list so that
93 * it can be detected and flushed upon next scan without having to scan the
94 * whole FD list.
95 *
96 * This translates like this :
97 *
98 * EVENT_IN_SPEC_LIST = 01
99 * EVENT_IN_POLL_LIST = 10
100 *
101 * IDLE = 0
102 * SPEC = (EVENT_IN_SPEC_LIST)
103 * WAIT = (EVENT_IN_POLL_LIST)
104 * STOP = (EVENT_IN_SPEC_LIST|EVENT_IN_POLL_LIST)
105 *
106 * fd_is_set() just consists in checking that the status is 01 or 10.
107 *
108 * For efficiency reasons, we will store the Read and Write bits interlaced to
109 * form a 4-bit field, so that we can simply shift the value right by 0/1 and
110 * get what we want :
111 * 3 2 1 0
112 * Wp Rp Ws Rs
113 *
114 * The FD array has to hold a back reference to the speculative list. This
115 * reference is only valid if at least one of the directions is marked SPEC.
116 *
117 */
118
119#define FD_EV_IN_SL 1
120#define FD_EV_IN_PL 4
121
122#define FD_EV_IDLE 0
123#define FD_EV_SPEC (FD_EV_IN_SL)
124#define FD_EV_WAIT (FD_EV_IN_PL)
125#define FD_EV_STOP (FD_EV_IN_SL|FD_EV_IN_PL)
126
127/* Those match any of R or W for Spec list or Poll list */
128#define FD_EV_RW_SL (FD_EV_IN_SL | (FD_EV_IN_SL << 1))
129#define FD_EV_RW_PL (FD_EV_IN_PL | (FD_EV_IN_PL << 1))
130#define FD_EV_MASK_DIR (FD_EV_IN_SL|FD_EV_IN_PL)
131
132#define FD_EV_IDLE_R 0
133#define FD_EV_SPEC_R (FD_EV_IN_SL)
134#define FD_EV_WAIT_R (FD_EV_IN_PL)
135#define FD_EV_STOP_R (FD_EV_IN_SL|FD_EV_IN_PL)
136#define FD_EV_MASK_R (FD_EV_IN_SL|FD_EV_IN_PL)
137
138#define FD_EV_IDLE_W (FD_EV_IDLE_R << 1)
139#define FD_EV_SPEC_W (FD_EV_SPEC_R << 1)
140#define FD_EV_WAIT_W (FD_EV_WAIT_R << 1)
141#define FD_EV_STOP_W (FD_EV_STOP_R << 1)
142#define FD_EV_MASK_W (FD_EV_MASK_R << 1)
143
144#define FD_EV_MASK (FD_EV_MASK_W | FD_EV_MASK_R)
145
Willy Tarreau6653d172007-05-13 01:52:05 +0200146/* This is the minimum number of events successfully processed in speculative
147 * mode above which we agree to return without checking epoll() (1/2 times).
148 */
149#define MIN_RETURN_EVENTS 25
Willy Tarreaude99e992007-04-16 00:53:59 +0200150
151/* descriptor of one FD.
152 * FIXME: should be a bit field */
153struct fd_status {
154 unsigned int e:4; // read and write events status.
Willy Tarreau4eac2092007-08-31 17:01:18 +0200155 unsigned int s1:28; // Position in spec list+1. 0=not in list. Should be last.
Willy Tarreaude99e992007-04-16 00:53:59 +0200156};
157
158static int nbspec = 0; // current size of the spec list
Willy Tarreauf2e8ee22008-05-25 10:39:02 +0200159static int absmaxevents = 0; // absolute maximum amounts of polled events
Willy Tarreaude99e992007-04-16 00:53:59 +0200160
161static struct fd_status *fd_list = NULL; // list of FDs
162static unsigned int *spec_list = NULL; // speculative I/O list
163
164/* private data */
165static struct epoll_event *epoll_events;
166static int epoll_fd;
167
168/* This structure may be used for any purpose. Warning! do not use it in
169 * recursive functions !
170 */
171static struct epoll_event ev;
172
173
174REGPRM1 static void alloc_spec_entry(const int fd)
175{
Willy Tarreau4eac2092007-08-31 17:01:18 +0200176 if (fd_list[fd].s1)
Willy Tarreaude99e992007-04-16 00:53:59 +0200177 return;
Willy Tarreau4eac2092007-08-31 17:01:18 +0200178 fd_list[fd].s1 = nbspec + 1;
179 spec_list[nbspec] = fd;
180 nbspec++;
Willy Tarreaude99e992007-04-16 00:53:59 +0200181}
182
Willy Tarreau4eac2092007-08-31 17:01:18 +0200183/* Removes entry used by fd <fd> from the spec list and replaces it with the
184 * last one. The fd_list is adjusted to match the back reference if needed.
185 * If the fd has no entry assigned, return immediately.
Willy Tarreaude99e992007-04-16 00:53:59 +0200186 */
Willy Tarreau4eac2092007-08-31 17:01:18 +0200187REGPRM1 static void release_spec_entry(int fd)
Willy Tarreaude99e992007-04-16 00:53:59 +0200188{
Willy Tarreau4eac2092007-08-31 17:01:18 +0200189 unsigned int pos;
190
191 pos = fd_list[fd].s1;
192 if (!pos)
193 return;
194
195 fd_list[fd].s1 = 0;
196 pos--;
197 /* we have spec_list[pos]==fd */
Willy Tarreaude99e992007-04-16 00:53:59 +0200198
199 nbspec--;
200 if (pos == nbspec)
201 return;
202
Willy Tarreau4eac2092007-08-31 17:01:18 +0200203 /* we replace current FD by the highest one, which may sometimes be the same */
Willy Tarreaude99e992007-04-16 00:53:59 +0200204 fd = spec_list[nbspec];
Willy Tarreau4eac2092007-08-31 17:01:18 +0200205 fd_list[fd].s1 = pos + 1;
Willy Tarreaude99e992007-04-16 00:53:59 +0200206 spec_list[pos] = fd;
Willy Tarreaude99e992007-04-16 00:53:59 +0200207}
208
209/*
210 * Returns non-zero if <fd> is already monitored for events in direction <dir>.
211 */
212REGPRM2 static int __fd_is_set(const int fd, int dir)
213{
214 int ret;
215
Willy Tarreau7a52a5c2008-08-16 16:06:02 +0200216#if DEBUG_DEV
217 if (fdtab[fd].state == FD_STCLOSE) {
218 fprintf(stderr, "sepoll.fd_isset called on closed fd #%d.\n", fd);
219 ABORT_NOW();
220 }
221#endif
Willy Tarreaude99e992007-04-16 00:53:59 +0200222 ret = ((unsigned)fd_list[fd].e >> dir) & FD_EV_MASK_DIR;
223 return (ret == FD_EV_SPEC || ret == FD_EV_WAIT);
224}
225
226/*
227 * Don't worry about the strange constructs in __fd_set/__fd_clr, they are
228 * designed like this in order to reduce the number of jumps (verified).
229 */
230REGPRM2 static int __fd_set(const int fd, int dir)
231{
232 __label__ switch_state;
233 unsigned int i;
234
Willy Tarreau7a52a5c2008-08-16 16:06:02 +0200235#if DEBUG_DEV
236 if (fdtab[fd].state == FD_STCLOSE) {
237 fprintf(stderr, "sepoll.fd_set called on closed fd #%d.\n", fd);
238 ABORT_NOW();
239 }
240#endif
Willy Tarreaude99e992007-04-16 00:53:59 +0200241 i = ((unsigned)fd_list[fd].e >> dir) & FD_EV_MASK_DIR;
242
243 if (i == FD_EV_IDLE) {
244 // switch to SPEC state and allocate a SPEC entry.
245 alloc_spec_entry(fd);
246 switch_state:
247 fd_list[fd].e ^= (unsigned int)(FD_EV_IN_SL << dir);
248 return 1;
249 }
250 else if (i == FD_EV_STOP) {
251 // switch to WAIT state
252 goto switch_state;
253 }
254 else
255 return 0;
256}
257
258REGPRM2 static int __fd_clr(const int fd, int dir)
259{
260 __label__ switch_state;
261 unsigned int i;
262
Willy Tarreau7a52a5c2008-08-16 16:06:02 +0200263#if DEBUG_DEV
264 if (fdtab[fd].state == FD_STCLOSE) {
265 fprintf(stderr, "sepoll.fd_clr called on closed fd #%d.\n", fd);
266 ABORT_NOW();
267 }
268#endif
Willy Tarreaude99e992007-04-16 00:53:59 +0200269 i = ((unsigned)fd_list[fd].e >> dir) & FD_EV_MASK_DIR;
270
271 if (i == FD_EV_SPEC) {
272 // switch to IDLE state
273 goto switch_state;
274 }
275 else if (likely(i == FD_EV_WAIT)) {
276 // switch to STOP state
277 /* We will create a queue entry for this one because we want to
278 * process it later in order to merge it with other events on
279 * the same FD.
280 */
281 alloc_spec_entry(fd);
282 switch_state:
283 fd_list[fd].e ^= (unsigned int)(FD_EV_IN_SL << dir);
284 return 1;
285 }
286 return 0;
287}
288
Willy Tarreau6653d172007-05-13 01:52:05 +0200289/* normally unused */
Willy Tarreaude99e992007-04-16 00:53:59 +0200290REGPRM1 static void __fd_rem(int fd)
291{
292 __fd_clr(fd, DIR_RD);
293 __fd_clr(fd, DIR_WR);
294}
295
296/*
297 * On valid epoll() implementations, a call to close() automatically removes
298 * the fds. This means that the FD will appear as previously unset.
299 */
300REGPRM1 static void __fd_clo(int fd)
301{
Willy Tarreau7a52a5c2008-08-16 16:06:02 +0200302 release_spec_entry(fd);
Willy Tarreaude99e992007-04-16 00:53:59 +0200303 fd_list[fd].e &= ~(FD_EV_MASK);
304}
305
Willy Tarreaudc246a72007-05-09 21:57:51 +0200306/*
Willy Tarreaude99e992007-04-16 00:53:59 +0200307 * speculative epoll() poller
308 */
Willy Tarreau0c303ee2008-07-07 00:09:58 +0200309REGPRM2 static void _do_poll(struct poller *p, int exp)
Willy Tarreaude99e992007-04-16 00:53:59 +0200310{
311 static unsigned int last_skipped;
Willy Tarreauf2e8ee22008-05-25 10:39:02 +0200312 static unsigned int spec_processed;
Willy Tarreau6653d172007-05-13 01:52:05 +0200313 int status, eo;
Willy Tarreaude99e992007-04-16 00:53:59 +0200314 int fd, opcode;
315 int count;
316 int spec_idx;
Willy Tarreaud825eef2007-05-12 22:35:00 +0200317 int wait_time;
Willy Tarreaude99e992007-04-16 00:53:59 +0200318
319
320 /* Here we have two options :
Willy Tarreau6653d172007-05-13 01:52:05 +0200321 * - either walk the list forwards and hope to match more events
Willy Tarreaude99e992007-04-16 00:53:59 +0200322 * - or walk it backwards to minimize the number of changes and
323 * to make better use of the cache.
324 * Tests have shown that walking backwards improves perf by 0.2%.
325 */
326
Willy Tarreau6653d172007-05-13 01:52:05 +0200327 status = 0;
Willy Tarreaude99e992007-04-16 00:53:59 +0200328 spec_idx = nbspec;
329 while (likely(spec_idx > 0)) {
330 spec_idx--;
331 fd = spec_list[spec_idx];
Willy Tarreau6653d172007-05-13 01:52:05 +0200332 eo = fd_list[fd].e; /* save old events */
Willy Tarreaude99e992007-04-16 00:53:59 +0200333
Willy Tarreau6653d172007-05-13 01:52:05 +0200334 /*
335 * Process the speculative events.
336 *
337 * Principle: events which are marked FD_EV_SPEC are processed
338 * with their assigned function. If the function returns 0, it
339 * means there is nothing doable without polling first. We will
340 * then convert the event to a pollable one by assigning them
341 * the WAIT status.
Willy Tarreaude99e992007-04-16 00:53:59 +0200342 */
343
Willy Tarreau7a52a5c2008-08-16 16:06:02 +0200344#ifdef DEBUG_DEV
345 if (fdtab[fd].state == FD_STCLOSE) {
346 fprintf(stderr,"fd=%d, fdtab[].ev=%x, fd_list[].e=%x, .s=%d, idx=%d\n",
347 fd, fdtab[fd].ev, fd_list[fd].e, fd_list[fd].s1, spec_idx);
348 }
349#endif
Willy Tarreaud6f087e2008-01-18 17:20:13 +0100350 fdtab[fd].ev &= FD_POLL_STICKY;
Willy Tarreau6653d172007-05-13 01:52:05 +0200351 if ((eo & FD_EV_MASK_R) == FD_EV_SPEC_R) {
352 /* The owner is interested in reading from this FD */
Willy Tarreau7a52a5c2008-08-16 16:06:02 +0200353 if (fdtab[fd].state != FD_STERROR) {
Willy Tarreau6653d172007-05-13 01:52:05 +0200354 /* Pretend there is something to read */
Willy Tarreaude99e992007-04-16 00:53:59 +0200355 fdtab[fd].ev |= FD_POLL_IN;
Willy Tarreau6653d172007-05-13 01:52:05 +0200356 if (!fdtab[fd].cb[DIR_RD].f(fd))
357 fd_list[fd].e ^= (FD_EV_WAIT_R ^ FD_EV_SPEC_R);
358 else
359 status++;
Willy Tarreaude99e992007-04-16 00:53:59 +0200360 }
361 }
Willy Tarreau6653d172007-05-13 01:52:05 +0200362 else if ((eo & FD_EV_MASK_R) == FD_EV_STOP_R) {
363 /* This FD was being polled and is now being removed. */
364 fd_list[fd].e &= ~FD_EV_MASK_R;
365 }
Willy Tarreaude99e992007-04-16 00:53:59 +0200366
Willy Tarreau6653d172007-05-13 01:52:05 +0200367 if ((eo & FD_EV_MASK_W) == FD_EV_SPEC_W) {
368 /* The owner is interested in writing to this FD */
Willy Tarreau7a52a5c2008-08-16 16:06:02 +0200369 if (fdtab[fd].state != FD_STERROR) {
Willy Tarreau6653d172007-05-13 01:52:05 +0200370 /* Pretend there is something to write */
Willy Tarreaude99e992007-04-16 00:53:59 +0200371 fdtab[fd].ev |= FD_POLL_OUT;
Willy Tarreau6653d172007-05-13 01:52:05 +0200372 if (!fdtab[fd].cb[DIR_WR].f(fd))
373 fd_list[fd].e ^= (FD_EV_WAIT_W ^ FD_EV_SPEC_W);
374 else
375 status++;
Willy Tarreaude99e992007-04-16 00:53:59 +0200376 }
377 }
Willy Tarreau6653d172007-05-13 01:52:05 +0200378 else if ((eo & FD_EV_MASK_W) == FD_EV_STOP_W) {
379 /* This FD was being polled and is now being removed. */
380 fd_list[fd].e &= ~FD_EV_MASK_W;
381 }
Willy Tarreaude99e992007-04-16 00:53:59 +0200382
Willy Tarreau6653d172007-05-13 01:52:05 +0200383 /* Now, we will adjust the event in the poll list. Indeed, it
384 * is possible that an event which was previously in the poll
385 * list now goes out, and the opposite is possible too. We can
386 * have opposite changes for READ and WRITE too.
387 */
388
389 if ((eo ^ fd_list[fd].e) & FD_EV_RW_PL) {
390 /* poll status changed*/
391 if ((fd_list[fd].e & FD_EV_RW_PL) == 0) {
392 /* fd removed from poll list */
393 opcode = EPOLL_CTL_DEL;
394 }
395 else if ((eo & FD_EV_RW_PL) == 0) {
396 /* new fd in the poll list */
Willy Tarreaude99e992007-04-16 00:53:59 +0200397 opcode = EPOLL_CTL_ADD;
398 }
Willy Tarreau6653d172007-05-13 01:52:05 +0200399 else {
400 /* fd status changed */
401 opcode = EPOLL_CTL_MOD;
402 }
403
404 /* construct the epoll events based on new state */
405 ev.events = 0;
406 if (fd_list[fd].e & FD_EV_WAIT_R)
407 ev.events |= EPOLLIN;
408
409 if (fd_list[fd].e & FD_EV_WAIT_W)
410 ev.events |= EPOLLOUT;
411
412 ev.data.fd = fd;
Willy Tarreaude99e992007-04-16 00:53:59 +0200413 epoll_ctl(epoll_fd, opcode, fd, &ev);
Willy Tarreau6653d172007-05-13 01:52:05 +0200414 }
Willy Tarreaude99e992007-04-16 00:53:59 +0200415
Willy Tarreaude99e992007-04-16 00:53:59 +0200416
Willy Tarreau6653d172007-05-13 01:52:05 +0200417 if (!(fd_list[fd].e & FD_EV_RW_SL)) {
418 /* This fd switched to combinations of either WAIT or
419 * IDLE. It must be removed from the spec list.
420 */
Willy Tarreau4eac2092007-08-31 17:01:18 +0200421 release_spec_entry(fd);
Willy Tarreau6653d172007-05-13 01:52:05 +0200422 continue;
Willy Tarreaude99e992007-04-16 00:53:59 +0200423 }
424 }
425
Willy Tarreau6653d172007-05-13 01:52:05 +0200426 /* It may make sense to immediately return here if there are enough
427 * processed events, without passing through epoll_wait() because we
428 * have exactly done a poll.
429 * Measures have shown a great performance increase if we call the
430 * epoll_wait() only the second time after speculative accesses have
431 * succeeded. This reduces the number of unsucessful calls to
432 * epoll_wait() by a factor of about 3, and the total number of calls
433 * by about 2.
Willy Tarreauf2e8ee22008-05-25 10:39:02 +0200434 * However, when we do that after having processed too many events,
435 * events waiting in epoll() starve for too long a time and tend to
436 * become themselves eligible for speculative polling. So we try to
437 * limit this practise to reasonable situations.
Willy Tarreaude99e992007-04-16 00:53:59 +0200438 */
Willy Tarreau6653d172007-05-13 01:52:05 +0200439
Willy Tarreauf2e8ee22008-05-25 10:39:02 +0200440 spec_processed += status;
441 if (status >= MIN_RETURN_EVENTS && spec_processed < absmaxevents) {
Willy Tarreau6653d172007-05-13 01:52:05 +0200442 /* We have processed at least MIN_RETURN_EVENTS, it's worth
443 * returning now without checking epoll_wait().
444 */
445 if (++last_skipped <= 1) {
Willy Tarreaub0b37bc2008-06-23 14:00:57 +0200446 tv_update_date(0, 1);
Willy Tarreaude99e992007-04-16 00:53:59 +0200447 return;
448 }
Willy Tarreaude99e992007-04-16 00:53:59 +0200449 }
450 last_skipped = 0;
451
Willy Tarreau3a628112008-06-13 21:06:56 +0200452 if (nbspec || status || run_queue) {
Willy Tarreau6653d172007-05-13 01:52:05 +0200453 /* Maybe we have processed some events that we must report, or
Willy Tarreau3a628112008-06-13 21:06:56 +0200454 * maybe we still have events in the spec list, or there are
455 * some tasks left pending in the run_queue, so we must not
Willy Tarreau6653d172007-05-13 01:52:05 +0200456 * wait in epoll() otherwise we will delay their delivery by
457 * the next timeout.
458 */
Willy Tarreaude99e992007-04-16 00:53:59 +0200459 wait_time = 0;
460 }
Willy Tarreaud825eef2007-05-12 22:35:00 +0200461 else {
Willy Tarreau0c303ee2008-07-07 00:09:58 +0200462 if (!exp)
Willy Tarreaub0b37bc2008-06-23 14:00:57 +0200463 wait_time = MAX_DELAY_MS;
Willy Tarreau0c303ee2008-07-07 00:09:58 +0200464 else if (tick_is_expired(exp, now_ms))
Willy Tarreaubdefc512007-05-14 02:02:04 +0200465 wait_time = 0;
Willy Tarreaub0b37bc2008-06-23 14:00:57 +0200466 else {
Willy Tarreau0c303ee2008-07-07 00:09:58 +0200467 wait_time = TICKS_TO_MS(tick_remain(now_ms, exp)) + 1;
Willy Tarreaub0b37bc2008-06-23 14:00:57 +0200468 if (wait_time > MAX_DELAY_MS)
469 wait_time = MAX_DELAY_MS;
470 }
Willy Tarreaud825eef2007-05-12 22:35:00 +0200471 }
Willy Tarreaude99e992007-04-16 00:53:59 +0200472
Willy Tarreauf2e8ee22008-05-25 10:39:02 +0200473 /* now let's wait for real events. We normally use maxpollevents as a
474 * high limit, unless <nbspec> is already big, in which case we need
475 * to compensate for the high number of events processed there.
476 */
477 fd = MIN(absmaxevents, spec_processed);
478 fd = MAX(global.tune.maxpollevents, fd);
479 fd = MIN(maxfd, fd);
480 spec_processed = 0;
Willy Tarreau1db37712007-06-03 17:16:49 +0200481 status = epoll_wait(epoll_fd, epoll_events, fd, wait_time);
Willy Tarreaub0b37bc2008-06-23 14:00:57 +0200482 tv_update_date(wait_time, status);
Willy Tarreaude99e992007-04-16 00:53:59 +0200483
484 for (count = 0; count < status; count++) {
485 int e = epoll_events[count].events;
486 fd = epoll_events[count].data.fd;
487
488 /* it looks complicated but gcc can optimize it away when constants
489 * have same values.
490 */
Willy Tarreaud6f087e2008-01-18 17:20:13 +0100491 DPRINTF(stderr, "%s:%d: fd=%d, ev=0x%08x, e=0x%08x\n",
492 __FUNCTION__, __LINE__,
493 fd, fdtab[fd].ev, e);
494
495 fdtab[fd].ev &= FD_POLL_STICKY;
496 fdtab[fd].ev |=
Willy Tarreaude99e992007-04-16 00:53:59 +0200497 ((e & EPOLLIN ) ? FD_POLL_IN : 0) |
498 ((e & EPOLLPRI) ? FD_POLL_PRI : 0) |
499 ((e & EPOLLOUT) ? FD_POLL_OUT : 0) |
500 ((e & EPOLLERR) ? FD_POLL_ERR : 0) |
501 ((e & EPOLLHUP) ? FD_POLL_HUP : 0);
502
503 if ((fd_list[fd].e & FD_EV_MASK_R) == FD_EV_WAIT_R) {
Willy Tarreau8bb46f42007-04-30 12:56:21 +0200504 if (fdtab[fd].state == FD_STCLOSE || fdtab[fd].state == FD_STERROR)
Willy Tarreaude99e992007-04-16 00:53:59 +0200505 continue;
Willy Tarreaud6f087e2008-01-18 17:20:13 +0100506 if (fdtab[fd].ev & (FD_POLL_IN|FD_POLL_HUP|FD_POLL_ERR))
Willy Tarreaude99e992007-04-16 00:53:59 +0200507 fdtab[fd].cb[DIR_RD].f(fd);
508 }
509
510 if ((fd_list[fd].e & FD_EV_MASK_W) == FD_EV_WAIT_W) {
Willy Tarreau8bb46f42007-04-30 12:56:21 +0200511 if (fdtab[fd].state == FD_STCLOSE || fdtab[fd].state == FD_STERROR)
Willy Tarreaude99e992007-04-16 00:53:59 +0200512 continue;
Willy Tarreaud6f087e2008-01-18 17:20:13 +0100513 if (fdtab[fd].ev & (FD_POLL_OUT|FD_POLL_ERR))
Willy Tarreaude99e992007-04-16 00:53:59 +0200514 fdtab[fd].cb[DIR_WR].f(fd);
515 }
516 }
517}
518
519/*
520 * Initialization of the speculative epoll() poller.
521 * Returns 0 in case of failure, non-zero in case of success. If it fails, it
522 * disables the poller by setting its pref to 0.
523 */
524REGPRM1 static int _do_init(struct poller *p)
525{
526 __label__ fail_fd_list, fail_spec, fail_ee, fail_fd;
527
528 p->private = NULL;
529
530 epoll_fd = epoll_create(global.maxsock + 1);
531 if (epoll_fd < 0)
532 goto fail_fd;
533
Willy Tarreauf2e8ee22008-05-25 10:39:02 +0200534 /* See comments at the top of the file about this formula. */
535 absmaxevents = MAX(global.tune.maxpollevents, global.maxsock/4);
Willy Tarreaude99e992007-04-16 00:53:59 +0200536 epoll_events = (struct epoll_event*)
Willy Tarreauf2e8ee22008-05-25 10:39:02 +0200537 calloc(1, sizeof(struct epoll_event) * absmaxevents);
Willy Tarreaude99e992007-04-16 00:53:59 +0200538
539 if (epoll_events == NULL)
540 goto fail_ee;
541
542 if ((spec_list = (uint32_t *)calloc(1, sizeof(uint32_t) * global.maxsock)) == NULL)
543 goto fail_spec;
544
545 fd_list = (struct fd_status *)calloc(1, sizeof(struct fd_status) * global.maxsock);
546 if (fd_list == NULL)
547 goto fail_fd_list;
548
549 return 1;
550
551 fail_fd_list:
552 free(spec_list);
553 fail_spec:
554 free(epoll_events);
555 fail_ee:
556 close(epoll_fd);
557 epoll_fd = 0;
558 fail_fd:
559 p->pref = 0;
560 return 0;
561}
562
563/*
564 * Termination of the speculative epoll() poller.
565 * Memory is released and the poller is marked as unselectable.
566 */
567REGPRM1 static void _do_term(struct poller *p)
568{
Willy Tarreaua534fea2008-08-03 12:19:50 +0200569 free(fd_list);
570 free(spec_list);
571 free(epoll_events);
Willy Tarreaude99e992007-04-16 00:53:59 +0200572
573 close(epoll_fd);
574 epoll_fd = 0;
575
576 fd_list = NULL;
577 spec_list = NULL;
578 epoll_events = NULL;
579
580 p->private = NULL;
581 p->pref = 0;
582}
583
584/*
585 * Check that the poller works.
586 * Returns 1 if OK, otherwise 0.
587 */
588REGPRM1 static int _do_test(struct poller *p)
589{
590 int fd;
591
592 fd = epoll_create(global.maxsock + 1);
593 if (fd < 0)
594 return 0;
595 close(fd);
596 return 1;
597}
598
599/*
Willy Tarreaufb8983f2007-06-03 16:40:44 +0200600 * Recreate the epoll file descriptor after a fork(). Returns 1 if OK,
601 * otherwise 0. It will ensure that all processes will not share their
602 * epoll_fd. Some side effects were encountered because of this, such
603 * as epoll_wait() returning an FD which was previously deleted.
604 */
605REGPRM1 static int _do_fork(struct poller *p)
606{
607 close(epoll_fd);
608 epoll_fd = epoll_create(global.maxsock + 1);
609 if (epoll_fd < 0)
610 return 0;
611 return 1;
612}
613
614/*
Willy Tarreaude99e992007-04-16 00:53:59 +0200615 * It is a constructor, which means that it will automatically be called before
616 * main(). This is GCC-specific but it works at least since 2.95.
617 * Special care must be taken so that it does not need any uninitialized data.
618 */
619__attribute__((constructor))
620static void _do_register(void)
621{
622 struct poller *p;
623
624 if (nbpollers >= MAX_POLLERS)
625 return;
626 p = &pollers[nbpollers++];
627
628 p->name = "sepoll";
629 p->pref = 400;
630 p->private = NULL;
631
632 p->test = _do_test;
633 p->init = _do_init;
634 p->term = _do_term;
635 p->poll = _do_poll;
Willy Tarreaufb8983f2007-06-03 16:40:44 +0200636 p->fork = _do_fork;
Willy Tarreaude99e992007-04-16 00:53:59 +0200637
638 p->is_set = __fd_is_set;
639 p->cond_s = p->set = __fd_set;
640 p->cond_c = p->clr = __fd_clr;
641 p->rem = __fd_rem;
642 p->clo = __fd_clo;
643}
644
645
646/*
647 * Local variables:
648 * c-indent-level: 8
649 * c-basic-offset: 8
650 * End:
651 */