]> git.ipfire.org Git - thirdparty/squid.git/blame - src/comm/ModEpoll.cc
SourceFormat Enforcement
[thirdparty/squid.git] / src / comm / ModEpoll.cc
CommitLineData
6039b729 1/*
4ac4a490 2 * Copyright (C) 1996-2017 The Squid Software Foundation and contributors
6039b729 3 *
bbc27441
AJ
4 * Squid software is distributed under GPLv2+ license and includes
5 * contributions from numerous individuals and organizations.
6 * Please see the COPYING and CONTRIBUTORS files for details.
6039b729 7 */
8
bbc27441
AJ
9/* DEBUG: section 05 Socket Functions */
10
6039b729 11/*
12 * The idea for this came from these two websites:
13 * http://www.xmailserver.org/linux-patches/nio-improve.html
14 * http://www.kegel.com/c10k.html
15 *
16 * This is to support the epoll sysctl being added to the linux 2.5
17 * kernel tree. The new sys_epoll is an event based poller without
18 * most of the fuss of rtsignals.
19 *
20 * -- David Nicklay <dnicklay@web.turner.com>
21 */
22
23/*
24 * XXX Currently not implemented / supported by this module XXX
25 *
26 * - delay pools
27 * - deferred reads
28 *
29 */
30
f7f3304a 31#include "squid.h"
d841c88d
AJ
32
33#if USE_EPOLL
34
d841c88d 35#include "comm/Loops.h"
1287b9e1 36#include "fde.h"
582c2af2 37#include "globals.h"
d841c88d 38#include "mgr/Registration.h"
582c2af2 39#include "profiler/Profiler.h"
348697ca 40#include "SquidTime.h"
e1656dc4 41#include "StatCounters.h"
00a7574e 42#include "StatHist.h"
d841c88d 43#include "Store.h"
8a02a7f8 44
6039b729 45#define DEBUG_EPOLL 0
46
1a30fdf5 47#include <cerrno>
ad32c661 48#if HAVE_SYS_EPOLL_H
6039b729 49#include <sys/epoll.h>
ad32c661 50#endif
6039b729 51
612fc474 52static int kdpfd = -1;
6039b729 53static int max_poll_time = 1000;
54
55static struct epoll_event *pevents;
56
5acc9f37 57static void commEPollRegisterWithCacheManager(void);
6039b729 58
6039b729 59/* XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX */
60/* Public functions */
61
6039b729 62/*
6039b729 63 * This is a needed exported function which will be called to initialise
64 * the network loop code.
65 */
66void
d841c88d 67Comm::SelectLoopInit(void)
6039b729 68{
6039b729 69 pevents = (struct epoll_event *) xmalloc(SQUID_MAXFD * sizeof(struct epoll_event));
70
71 if (!pevents) {
b69e9ffa
AJ
72 int xerrno = errno;
73 fatalf("comm_select_init: xmalloc() failed: %s\n", xstrerr(xerrno));
6039b729 74 }
75
76 kdpfd = epoll_create(SQUID_MAXFD);
77
78 if (kdpfd < 0) {
b69e9ffa
AJ
79 int xerrno = errno;
80 fatalf("comm_select_init: epoll_create(): %s\n", xstrerr(xerrno));
6039b729 81 }
8cb183ec
FC
82
83 commEPollRegisterWithCacheManager();
6039b729 84}
85
751406fe 86static const char* epolltype_atoi(int x)
87{
26ac0430 88 switch (x) {
751406fe 89
90 case EPOLL_CTL_ADD:
91 return "EPOLL_CTL_ADD";
92
93 case EPOLL_CTL_DEL:
94 return "EPOLL_CTL_DEL";
95
96 case EPOLL_CTL_MOD:
97 return "EPOLL_CTL_MOD";
98
99 default:
100 return "UNKNOWN_EPOLLCTL_OP";
101 }
102}
103
d841c88d 104/**
6039b729 105 * This is a needed exported function which will be called to register
106 * and deregister interest in a pending IO state for a given FD.
6039b729 107 */
108void
d841c88d 109Comm::SetSelect(int fd, unsigned int type, PF * handler, void *client_data, time_t timeout)
6039b729 110{
111 fde *F = &fd_table[fd];
751406fe 112 int epoll_ctl_type = 0;
6039b729 113
6039b729 114 assert(fd >= 0);
48e7baac 115 debugs(5, 5, HERE << "FD " << fd << ", type=" << type <<
d841c88d
AJ
116 ", handler=" << handler << ", client_data=" << client_data <<
117 ", timeout=" << timeout);
6039b729 118
612fc474
AJ
119 struct epoll_event ev;
120 memset(&ev, 0, sizeof(ev));
751406fe 121 ev.data.fd = fd;
6039b729 122
a1727eb7 123 if (!F->flags.open) {
124 epoll_ctl(kdpfd, EPOLL_CTL_DEL, fd, &ev);
125 return;
126 }
127
751406fe 128 // If read is an interest
6039b729 129
130 if (type & COMM_SELECT_READ) {
0a080ce3 131 if (handler) {
26ac0430
AJ
132 // Hack to keep the events flowing if there is data immediately ready
133 if (F->flags.read_pending)
134 ev.events |= EPOLLOUT;
751406fe 135 ev.events |= EPOLLIN;
26ac0430 136 }
6039b729 137
138 F->read_handler = handler;
139
140 F->read_data = client_data;
751406fe 141
142 // Otherwise, use previously stored value
143 } else if (F->epoll_state & EPOLLIN) {
144 ev.events |= EPOLLIN;
6039b729 145 }
146
751406fe 147 // If write is an interest
6039b729 148 if (type & COMM_SELECT_WRITE) {
751406fe 149 if (handler)
150 ev.events |= EPOLLOUT;
6039b729 151
152 F->write_handler = handler;
153
154 F->write_data = client_data;
6039b729 155
751406fe 156 // Otherwise, use previously stored value
157 } else if (F->epoll_state & EPOLLOUT) {
158 ev.events |= EPOLLOUT;
159 }
6039b729 160
751406fe 161 if (ev.events)
162 ev.events |= EPOLLHUP | EPOLLERR;
6039b729 163
751406fe 164 if (ev.events != F->epoll_state) {
165 if (F->epoll_state) // already monitoring something.
166 epoll_ctl_type = ev.events ? EPOLL_CTL_MOD : EPOLL_CTL_DEL;
167 else
168 epoll_ctl_type = EPOLL_CTL_ADD;
6039b729 169
751406fe 170 F->epoll_state = ev.events;
6039b729 171
751406fe 172 if (epoll_ctl(kdpfd, epoll_ctl_type, fd, &ev) < 0) {
b69e9ffa
AJ
173 int xerrno = errno;
174 debugs(5, DEBUG_EPOLL ? 0 : 8, "epoll_ctl(," << epolltype_atoi(epoll_ctl_type) <<
175 ",,): failed on FD " << fd << ": " << xstrerr(xerrno));
6039b729 176 }
177 }
178
179 if (timeout)
180 F->timeout = squid_curtime + timeout;
181}
182
3a5a4930 183void
d841c88d 184Comm::ResetSelect(int fd)
3a5a4930 185{
186 fde *F = &fd_table[fd];
187 F->epoll_state = 0;
d841c88d 188 SetSelect(fd, 0, NULL, NULL, 0);
3a5a4930 189}
190
610ee341 191static void commIncomingStats(StoreEntry * sentry);
192
5acc9f37 193static void
dbeed9ea 194commEPollRegisterWithCacheManager(void)
610ee341 195{
8822ebee 196 Mgr::RegisterAction("comm_epoll_incoming",
d9fc6862
A
197 "comm_incoming() stats",
198 commIncomingStats, 0, 1);
610ee341 199}
200
201static void
202commIncomingStats(StoreEntry * sentry)
203{
204 StatCounters *f = &statCounter;
0a515876 205 storeAppendPrintf(sentry, "Total number of epoll(2) loops: %ld\n", statCounter.select_loops);
610ee341 206 storeAppendPrintf(sentry, "Histogram of returned filedescriptors\n");
96886986 207 f->select_fds_hist.dump(sentry, statHistIntDumper);
610ee341 208}
209
d841c88d 210/**
6039b729 211 * Check all connections for new connections and input data that is to be
212 * processed. Also check for connections with data queued and whether we can
213 * write it out.
6039b729 214 *
215 * Called to do the new-style IO, courtesy of of squid (like most of this
216 * new IO code). This routine handles the stuff we've hidden in
217 * comm_setselect and fd_table[] and calls callbacks for IO ready
218 * events.
219 */
c8407295 220Comm::Flag
d841c88d 221Comm::DoSelect(int msec)
6039b729 222{
223 int num, i,fd;
224 fde *F;
225 PF *hdl;
226
227 struct epoll_event *cevents;
6039b729 228
a1727eb7 229 PROF_start(comm_check_incoming);
230
6039b729 231 if (msec > max_poll_time)
232 msec = max_poll_time;
233
234 for (;;) {
235 num = epoll_wait(kdpfd, pevents, SQUID_MAXFD, msec);
098346fd 236 ++ statCounter.select_loops;
6039b729 237
238 if (num >= 0)
239 break;
240
241 if (ignoreErrno(errno))
242 break;
243
244 getCurrentTime();
245
a1727eb7 246 PROF_stop(comm_check_incoming);
247
4ee57cbe 248 return Comm::COMM_ERROR;
6039b729 249 }
250
a1727eb7 251 PROF_stop(comm_check_incoming);
6039b729 252 getCurrentTime();
253
f30f7998 254 statCounter.select_fds_hist.count(num);
751406fe 255
6039b729 256 if (num == 0)
f53969cc 257 return Comm::TIMEOUT; /* No error.. */
751406fe 258
259 PROF_start(comm_handle_ready_fd);
6039b729 260
cbebe602 261 for (i = 0, cevents = pevents; i < num; ++i, ++cevents) {
6039b729 262 fd = cevents->data.fd;
263 F = &fd_table[fd];
d841c88d 264 debugs(5, DEBUG_EPOLL ? 0 : 8, HERE << "got FD " << fd << " events=" <<
26ac0430
AJ
265 std::hex << cevents->events << " monitoring=" << F->epoll_state <<
266 " F->read_handler=" << F->read_handler << " F->write_handler=" << F->write_handler);
751406fe 267
268 // TODO: add EPOLLPRI??
6039b729 269
0a080ce3 270 if (cevents->events & (EPOLLIN|EPOLLHUP|EPOLLERR) || F->flags.read_pending) {
a1727eb7 271 if ((hdl = F->read_handler) != NULL) {
d841c88d 272 debugs(5, DEBUG_EPOLL ? 0 : 8, HERE << "Calling read handler on FD " << fd);
751406fe 273 PROF_start(comm_write_handler);
0a080ce3 274 F->flags.read_pending = 0;
6039b729 275 F->read_handler = NULL;
276 hdl(fd, F->read_data);
751406fe 277 PROF_stop(comm_write_handler);
098346fd 278 ++ statCounter.select_fds;
751406fe 279 } else {
d841c88d 280 debugs(5, DEBUG_EPOLL ? 0 : 8, HERE << "no read handler for FD " << fd);
751406fe 281 // remove interest since no handler exist for this event.
d841c88d 282 SetSelect(fd, COMM_SELECT_READ, NULL, NULL, 0);
6039b729 283 }
284 }
285
751406fe 286 if (cevents->events & (EPOLLOUT|EPOLLHUP|EPOLLERR)) {
a1727eb7 287 if ((hdl = F->write_handler) != NULL) {
d841c88d 288 debugs(5, DEBUG_EPOLL ? 0 : 8, HERE << "Calling write handler on FD " << fd);
751406fe 289 PROF_start(comm_read_handler);
6039b729 290 F->write_handler = NULL;
291 hdl(fd, F->write_data);
751406fe 292 PROF_stop(comm_read_handler);
098346fd 293 ++ statCounter.select_fds;
751406fe 294 } else {
d841c88d 295 debugs(5, DEBUG_EPOLL ? 0 : 8, HERE << "no write handler for FD " << fd);
751406fe 296 // remove interest since no handler exist for this event.
d841c88d 297 SetSelect(fd, COMM_SELECT_WRITE, NULL, NULL, 0);
6039b729 298 }
299 }
300 }
301
751406fe 302 PROF_stop(comm_handle_ready_fd);
303
c8407295 304 return Comm::OK;
6039b729 305}
306
307void
d841c88d 308Comm::QuickPollRequired(void)
6039b729 309{
768d3d2f 310 max_poll_time = 10;
6039b729 311}
312
8a02a7f8 313#endif /* USE_EPOLL */
f53969cc 314