blob: ad4a0aafa49eec54cd8cdb2b5f56e4b7aa9e955a [file] [log] [blame]
Willy Tarreaubaaee002006-06-26 02:48:02 +02001/*
Willy Tarreauf7bc57c2012-10-03 00:19:48 +02002 * RAW transport layer over SOCK_STREAM sockets.
Willy Tarreaubaaee002006-06-26 02:48:02 +02003 *
Willy Tarreaub277d6e2012-05-11 16:59:14 +02004 * Copyright 2000-2012 Willy Tarreau <w@1wt.eu>
Willy Tarreaubaaee002006-06-26 02:48:02 +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 *
11 */
12
Willy Tarreau6b4aad42009-01-18 21:59:13 +010013#define _GNU_SOURCE
Willy Tarreaubaaee002006-06-26 02:48:02 +020014#include <errno.h>
15#include <fcntl.h>
16#include <stdio.h>
17#include <stdlib.h>
18
19#include <sys/socket.h>
20#include <sys/stat.h>
21#include <sys/types.h>
22
Dmitry Sivachenkocaf58982009-08-24 15:11:06 +040023#include <netinet/tcp.h>
24
Willy Tarreauc7e42382012-08-24 19:22:53 +020025#include <common/buffer.h>
Willy Tarreau2dd0d472006-06-29 17:53:05 +020026#include <common/compat.h>
Willy Tarreaue3ba5f02006-06-29 18:54:54 +020027#include <common/config.h>
Willy Tarreaud6f087e2008-01-18 17:20:13 +010028#include <common/debug.h>
Willy Tarreau83749182007-04-15 20:56:27 +020029#include <common/standard.h>
Willy Tarreau0c303ee2008-07-07 00:09:58 +020030#include <common/ticks.h>
Willy Tarreau2dd0d472006-06-29 17:53:05 +020031#include <common/time.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020032
Willy Tarreau8b117082012-08-06 15:06:49 +020033#include <proto/connection.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020034#include <proto/fd.h>
Willy Tarreaueb472682010-05-28 18:46:57 +020035#include <proto/freq_ctr.h>
36#include <proto/log.h>
Willy Tarreau3eba98a2009-01-25 13:56:13 +010037#include <proto/pipe.h>
Willy Tarreau75bf2c92012-08-20 17:01:35 +020038#include <proto/raw_sock.h>
Willy Tarreau73b013b2012-05-21 16:31:45 +020039#include <proto/stream_interface.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020040#include <proto/task.h>
41
Willy Tarreau5bd8c372009-01-19 00:32:22 +010042#include <types/global.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020043
Willy Tarreaub277d6e2012-05-11 16:59:14 +020044
Willy Tarreau96199b12012-08-24 00:46:52 +020045#if defined(CONFIG_HAP_LINUX_SPLICE)
Willy Tarreau43d8fb22011-08-22 17:12:02 +020046#include <common/splice.h>
Willy Tarreau5bd8c372009-01-19 00:32:22 +010047
48/* A pipe contains 16 segments max, and it's common to see segments of 1448 bytes
49 * because of timestamps. Use this as a hint for not looping on splice().
50 */
51#define SPLICE_FULL_HINT 16*1448
52
Willy Tarreaua9de3332009-11-28 07:47:10 +010053/* how many data we attempt to splice at once when the buffer is configured for
54 * infinite forwarding */
55#define MAX_SPLICE_AT_ONCE (1<<30)
56
Willy Tarreaub6daedd2013-01-07 16:57:09 +010057/* Versions of splice between 2.6.25 and 2.6.27.12 were bogus and would return EAGAIN
58 * on incoming shutdowns. On these versions, we have to call recv() after such a return
59 * in order to find whether splice is OK or not. Since 2.6.27.13 we don't need to do
60 * this anymore, and we can avoid this logic by defining ASSUME_SPLICE_WORKS.
61 */
62
Willy Tarreau5bd8c372009-01-19 00:32:22 +010063/* Returns :
Willy Tarreau96199b12012-08-24 00:46:52 +020064 * -1 if splice() is not supported
65 * >= 0 to report the amount of spliced bytes.
66 * connection flags are updated (error, read0, wait_room, wait_data).
67 * The caller must have previously allocated the pipe.
Willy Tarreau5bd8c372009-01-19 00:32:22 +010068 */
Willy Tarreau96199b12012-08-24 00:46:52 +020069int raw_sock_to_pipe(struct connection *conn, struct pipe *pipe, unsigned int count)
Willy Tarreau5bd8c372009-01-19 00:32:22 +010070{
Willy Tarreaub6daedd2013-01-07 16:57:09 +010071#ifndef ASSUME_SPLICE_WORKS
Willy Tarreau82a04562011-12-11 22:37:06 +010072 static int splice_detects_close;
Willy Tarreaub6daedd2013-01-07 16:57:09 +010073#endif
Willy Tarreau31971e52009-09-20 12:07:52 +020074 int ret;
Willy Tarreauafad0e02012-08-09 14:45:22 +020075 int retval = 0;
Willy Tarreau5bd8c372009-01-19 00:32:22 +010076
Willy Tarreau96199b12012-08-24 00:46:52 +020077 /* Under Linux, if FD_POLL_HUP is set, we have reached the end.
78 * Since older splice() implementations were buggy and returned
79 * EAGAIN on end of read, let's bypass the call to splice() now.
80 */
Willy Tarreau6f5d1412012-10-04 20:38:49 +020081 if (unlikely(!(fdtab[conn->t.sock.fd].ev & FD_POLL_IN))) {
82 /* stop here if we reached the end of data */
83 if ((fdtab[conn->t.sock.fd].ev & (FD_POLL_ERR|FD_POLL_HUP)) == FD_POLL_HUP)
84 goto out_read0;
85
86 /* report error on POLL_ERR before connection establishment */
87 if ((fdtab[conn->t.sock.fd].ev & FD_POLL_ERR) && (conn->flags & CO_FL_WAIT_L4_CONN)) {
88 conn->flags |= CO_FL_ERROR;
89 return retval;
90 }
91 }
Willy Tarreaua9de3332009-11-28 07:47:10 +010092
Willy Tarreau96199b12012-08-24 00:46:52 +020093 while (count) {
94 if (count > MAX_SPLICE_AT_ONCE)
95 count = MAX_SPLICE_AT_ONCE;
Willy Tarreau5bd8c372009-01-19 00:32:22 +010096
Willy Tarreau96199b12012-08-24 00:46:52 +020097 ret = splice(conn->t.sock.fd, NULL, pipe->prod, NULL, count,
Willy Tarreau5bd8c372009-01-19 00:32:22 +010098 SPLICE_F_MOVE|SPLICE_F_NONBLOCK);
99
100 if (ret <= 0) {
Willy Tarreau5bd8c372009-01-19 00:32:22 +0100101 if (ret == 0) {
Willy Tarreau98b306b2009-01-25 11:11:32 +0100102 /* connection closed. This is only detected by
Willy Tarreau82a04562011-12-11 22:37:06 +0100103 * recent kernels (>= 2.6.27.13). If we notice
104 * it works, we store the info for later use.
Willy Tarreau5bd8c372009-01-19 00:32:22 +0100105 */
Willy Tarreaub6daedd2013-01-07 16:57:09 +0100106#ifndef ASSUME_SPLICE_WORKS
Willy Tarreau82a04562011-12-11 22:37:06 +0100107 splice_detects_close = 1;
Willy Tarreaub6daedd2013-01-07 16:57:09 +0100108#endif
Willy Tarreau96199b12012-08-24 00:46:52 +0200109 goto out_read0;
Willy Tarreau5bd8c372009-01-19 00:32:22 +0100110 }
111
112 if (errno == EAGAIN) {
113 /* there are two reasons for EAGAIN :
114 * - nothing in the socket buffer (standard)
115 * - pipe is full
Willy Tarreau98b306b2009-01-25 11:11:32 +0100116 * - the connection is closed (kernel < 2.6.27.13)
Willy Tarreau96199b12012-08-24 00:46:52 +0200117 * The last case is annoying but know if we can detect it
118 * and if we can't then we rely on the call to recv() to
119 * get a valid verdict. The difference between the first
120 * two situations is problematic. Since we don't know if
121 * the pipe is full, we'll stop if the pipe is not empty.
122 * Anyway, we will almost always fill/empty the pipe.
Willy Tarreau5bd8c372009-01-19 00:32:22 +0100123 */
Willy Tarreau96199b12012-08-24 00:46:52 +0200124 if (pipe->data) {
125 /* alway stop reading until the pipe is flushed */
126 conn->flags |= CO_FL_WAIT_ROOM;
Willy Tarreau5bd8c372009-01-19 00:32:22 +0100127 break;
128 }
129
Willy Tarreau82a04562011-12-11 22:37:06 +0100130 /* We don't know if the connection was closed,
131 * but if we know splice detects close, then we
132 * know it for sure.
Willy Tarreau98b306b2009-01-25 11:11:32 +0100133 * But if we're called upon POLLIN with an empty
Willy Tarreau82a04562011-12-11 22:37:06 +0100134 * pipe and get EAGAIN, it is suspect enough to
Willy Tarreau98b306b2009-01-25 11:11:32 +0100135 * try to fall back to the normal recv scheme
136 * which will be able to deal with the situation.
137 */
Willy Tarreaub6daedd2013-01-07 16:57:09 +0100138#ifndef ASSUME_SPLICE_WORKS
Willy Tarreau82a04562011-12-11 22:37:06 +0100139 if (splice_detects_close)
Willy Tarreaub6daedd2013-01-07 16:57:09 +0100140#endif
Willy Tarreau56a77e52012-09-02 18:34:44 +0200141 __conn_data_poll_recv(conn); /* we know for sure that it's EAGAIN */
Willy Tarreau5bd8c372009-01-19 00:32:22 +0100142 break;
143 }
Willy Tarreau45b88932012-11-12 12:00:09 +0100144 else if (errno == ENOSYS || errno == EINVAL || errno == EBADF) {
Willy Tarreau96199b12012-08-24 00:46:52 +0200145 /* splice not supported on this end, disable it.
146 * We can safely return -1 since there is no
147 * chance that any data has been piped yet.
148 */
Willy Tarreaudc340a92009-06-28 23:10:19 +0200149 return -1;
150 }
Willy Tarreau96199b12012-08-24 00:46:52 +0200151 else if (errno == EINTR) {
152 /* try again */
153 continue;
154 }
Willy Tarreau5bd8c372009-01-19 00:32:22 +0100155 /* here we have another error */
Willy Tarreau96199b12012-08-24 00:46:52 +0200156 conn->flags |= CO_FL_ERROR;
Willy Tarreau5bd8c372009-01-19 00:32:22 +0100157 break;
158 } /* ret <= 0 */
159
Willy Tarreau96199b12012-08-24 00:46:52 +0200160 retval += ret;
161 pipe->data += ret;
Willy Tarreau5bd8c372009-01-19 00:32:22 +0100162
Willy Tarreaubaf2a502013-01-07 16:38:26 +0100163 /* if a POLL_HUP was present, we've read the last segment */
164 if ((fdtab[conn->t.sock.fd].ev & (FD_POLL_ERR|FD_POLL_HUP)) == FD_POLL_HUP)
165 goto out_read0;
166
Willy Tarreau96199b12012-08-24 00:46:52 +0200167 if (pipe->data >= SPLICE_FULL_HINT || ret >= global.tune.recv_enough) {
168 /* We've read enough of it for this time, let's stop before
169 * being asked to poll.
170 */
Willy Tarreau5bd8c372009-01-19 00:32:22 +0100171 break;
172 }
173 } /* while */
174
Willy Tarreau665e6ee2012-10-04 20:20:46 +0200175 if (unlikely(conn->flags & CO_FL_WAIT_L4_CONN) && retval)
176 conn->flags &= ~CO_FL_WAIT_L4_CONN;
Willy Tarreau96199b12012-08-24 00:46:52 +0200177 return retval;
Willy Tarreau3eba98a2009-01-25 13:56:13 +0100178
Willy Tarreau96199b12012-08-24 00:46:52 +0200179 out_read0:
180 conn_sock_read0(conn);
Willy Tarreau665e6ee2012-10-04 20:20:46 +0200181 conn->flags &= ~CO_FL_WAIT_L4_CONN;
Willy Tarreau5bd8c372009-01-19 00:32:22 +0100182 return retval;
183}
184
Willy Tarreau96199b12012-08-24 00:46:52 +0200185/* Send as many bytes as possible from the pipe to the connection's socket.
186 */
187int raw_sock_from_pipe(struct connection *conn, struct pipe *pipe)
188{
189 int ret, done;
190
191 done = 0;
192 while (pipe->data) {
193 ret = splice(pipe->cons, NULL, conn->t.sock.fd, NULL, pipe->data,
194 SPLICE_F_MOVE|SPLICE_F_NONBLOCK);
195
196 if (ret <= 0) {
197 if (ret == 0 || errno == EAGAIN) {
Willy Tarreau56a77e52012-09-02 18:34:44 +0200198 __conn_data_poll_send(conn);
Willy Tarreau96199b12012-08-24 00:46:52 +0200199 break;
200 }
201 else if (errno == EINTR)
202 continue;
203
204 /* here we have another error */
205 conn->flags |= CO_FL_ERROR;
206 break;
207 }
208
209 done += ret;
210 pipe->data -= ret;
211 }
Willy Tarreau665e6ee2012-10-04 20:20:46 +0200212 if (unlikely(conn->flags & CO_FL_WAIT_L4_CONN) && done)
213 conn->flags &= ~CO_FL_WAIT_L4_CONN;
Willy Tarreau96199b12012-08-24 00:46:52 +0200214 return done;
215}
216
Willy Tarreau6b4aad42009-01-18 21:59:13 +0100217#endif /* CONFIG_HAP_LINUX_SPLICE */
218
219
Willy Tarreau2ba44652012-08-20 17:30:32 +0200220/* Receive up to <count> bytes from connection <conn>'s socket and store them
221 * into buffer <buf>. The caller must ensure that <count> is always smaller
222 * than the buffer's size. Only one call to recv() is performed, unless the
223 * buffer wraps, in which case a second call may be performed. The connection's
224 * flags are updated with whatever special event is detected (error, read0,
225 * empty). The caller is responsible for taking care of those events and
226 * avoiding the call if inappropriate. The function does not call the
227 * connection's polling update function, so the caller is responsible for this.
228 */
229static int raw_sock_to_buf(struct connection *conn, struct buffer *buf, int count)
230{
231 int ret, done = 0;
232 int try = count;
233
Willy Tarreau6f5d1412012-10-04 20:38:49 +0200234 if (unlikely(!(fdtab[conn->t.sock.fd].ev & FD_POLL_IN))) {
235 /* stop here if we reached the end of data */
236 if ((fdtab[conn->t.sock.fd].ev & (FD_POLL_ERR|FD_POLL_HUP)) == FD_POLL_HUP)
237 goto read0;
238
239 /* report error on POLL_ERR before connection establishment */
240 if ((fdtab[conn->t.sock.fd].ev & FD_POLL_ERR) && (conn->flags & CO_FL_WAIT_L4_CONN)) {
241 conn->flags |= CO_FL_ERROR;
242 return done;
243 }
244 }
Willy Tarreau2ba44652012-08-20 17:30:32 +0200245
246 /* compute the maximum block size we can read at once. */
247 if (buffer_empty(buf)) {
248 /* let's realign the buffer to optimize I/O */
249 buf->p = buf->data;
250 }
251 else if (buf->data + buf->o < buf->p &&
252 buf->p + buf->i < buf->data + buf->size) {
253 /* remaining space wraps at the end, with a moving limit */
254 if (try > buf->data + buf->size - (buf->p + buf->i))
255 try = buf->data + buf->size - (buf->p + buf->i);
256 }
257
258 /* read the largest possible block. For this, we perform only one call
259 * to recv() unless the buffer wraps and we exactly fill the first hunk,
260 * in which case we accept to do it once again. A new attempt is made on
261 * EINTR too.
262 */
263 while (try) {
264 ret = recv(conn->t.sock.fd, bi_end(buf), try, 0);
265
266 if (ret > 0) {
267 buf->i += ret;
268 done += ret;
269 if (ret < try) {
270 /* unfortunately, on level-triggered events, POLL_HUP
271 * is generally delivered AFTER the system buffer is
272 * empty, so this one might never match.
273 */
274 if (fdtab[conn->t.sock.fd].ev & FD_POLL_HUP)
275 goto read0;
276 break;
277 }
278 count -= ret;
279 try = count;
280 }
281 else if (ret == 0) {
282 goto read0;
283 }
284 else if (errno == EAGAIN) {
Willy Tarreau56a77e52012-09-02 18:34:44 +0200285 __conn_data_poll_recv(conn);
Willy Tarreau2ba44652012-08-20 17:30:32 +0200286 break;
287 }
288 else if (errno != EINTR) {
289 conn->flags |= CO_FL_ERROR;
290 break;
291 }
292 }
Willy Tarreau665e6ee2012-10-04 20:20:46 +0200293
294 if (unlikely(conn->flags & CO_FL_WAIT_L4_CONN) && done)
295 conn->flags &= ~CO_FL_WAIT_L4_CONN;
Willy Tarreau2ba44652012-08-20 17:30:32 +0200296 return done;
297
298 read0:
299 conn_sock_read0(conn);
Willy Tarreau665e6ee2012-10-04 20:20:46 +0200300 conn->flags &= ~CO_FL_WAIT_L4_CONN;
Willy Tarreaudebdc4b2012-12-07 00:01:33 +0100301
302 /* Now a final check for a possible asynchronous low-level error
303 * report. This can happen when a connection receives a reset
304 * after a shutdown, both POLL_HUP and POLL_ERR are queued, and
305 * we might have come from there by just checking POLL_HUP instead
306 * of recv()'s return value 0, so we have no way to tell there was
307 * an error without checking.
308 */
309 if (unlikely(fdtab[conn->t.sock.fd].ev & FD_POLL_ERR))
310 conn->flags |= CO_FL_ERROR;
Willy Tarreau2ba44652012-08-20 17:30:32 +0200311 return done;
312}
313
314
Willy Tarreau5368d802012-08-21 18:22:06 +0200315/* Send all pending bytes from buffer <buf> to connection <conn>'s socket.
316 * <flags> may contain MSG_MORE to make the system hold on without sending
317 * data too fast.
318 * Only one call to send() is performed, unless the buffer wraps, in which case
319 * a second call may be performed. The connection's flags are updated with
320 * whatever special event is detected (error, empty). The caller is responsible
321 * for taking care of those events and avoiding the call if inappropriate. The
322 * function does not call the connection's polling update function, so the caller
323 * is responsible for this.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200324 */
Willy Tarreau5368d802012-08-21 18:22:06 +0200325static int raw_sock_from_buf(struct connection *conn, struct buffer *buf, int flags)
Willy Tarreau0c2fc1f2009-01-18 15:30:37 +0100326{
Willy Tarreau5368d802012-08-21 18:22:06 +0200327 int ret, try, done, send_flag;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200328
Willy Tarreau5368d802012-08-21 18:22:06 +0200329 done = 0;
330 /* send the largest possible block. For this we perform only one call
331 * to send() unless the buffer wraps and we exactly fill the first hunk,
332 * in which case we accept to do it once again.
Willy Tarreau5bd8c372009-01-19 00:32:22 +0100333 */
Willy Tarreau5368d802012-08-21 18:22:06 +0200334 while (buf->o) {
335 try = buf->o;
Willy Tarreau89fa7062012-03-02 16:13:16 +0100336 /* outgoing data may wrap at the end */
Willy Tarreau5368d802012-08-21 18:22:06 +0200337 if (buf->data + try > buf->p)
338 try = buf->data + try - buf->p;
Willy Tarreauf890dc92008-12-13 21:12:26 +0100339
Willy Tarreau5368d802012-08-21 18:22:06 +0200340 send_flag = MSG_DONTWAIT | MSG_NOSIGNAL;
341 if (try < buf->o)
342 send_flag = MSG_MORE;
Willy Tarreaufb14edc2009-06-14 15:24:37 +0200343
Willy Tarreau5368d802012-08-21 18:22:06 +0200344 ret = send(conn->t.sock.fd, bo_ptr(buf), try, send_flag | flags);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200345
346 if (ret > 0) {
Willy Tarreau5368d802012-08-21 18:22:06 +0200347 buf->o -= ret;
348 done += ret;
Willy Tarreaub38903c2008-11-23 21:33:29 +0100349
Willy Tarreau5fb38032012-12-16 19:39:09 +0100350 if (likely(buffer_empty(buf)))
Willy Tarreau0c2fc1f2009-01-18 15:30:37 +0100351 /* optimize data alignment in the buffer */
Willy Tarreau5368d802012-08-21 18:22:06 +0200352 buf->p = buf->data;
Willy Tarreau83749182007-04-15 20:56:27 +0200353
Willy Tarreauab3e1d32007-06-03 14:10:36 +0200354 /* if the system buffer is full, don't insist */
Willy Tarreau5368d802012-08-21 18:22:06 +0200355 if (ret < try)
Willy Tarreau6996e152007-04-30 14:37:43 +0200356 break;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200357 }
Willy Tarreau0ea0cf62012-11-11 20:38:30 +0100358 else if (ret == 0 || errno == EAGAIN || errno == ENOTCONN) {
Willy Tarreau0c2fc1f2009-01-18 15:30:37 +0100359 /* nothing written, we need to poll for write first */
Willy Tarreau56a77e52012-09-02 18:34:44 +0200360 __conn_data_poll_send(conn);
Willy Tarreau5368d802012-08-21 18:22:06 +0200361 break;
Willy Tarreau83749182007-04-15 20:56:27 +0200362 }
Willy Tarreau5368d802012-08-21 18:22:06 +0200363 else if (errno != EINTR) {
364 conn->flags |= CO_FL_ERROR;
365 break;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200366 }
Willy Tarreau5368d802012-08-21 18:22:06 +0200367 }
Willy Tarreau665e6ee2012-10-04 20:20:46 +0200368 if (unlikely(conn->flags & CO_FL_WAIT_L4_CONN) && done)
369 conn->flags &= ~CO_FL_WAIT_L4_CONN;
Willy Tarreau5368d802012-08-21 18:22:06 +0200370 return done;
Willy Tarreau0c2fc1f2009-01-18 15:30:37 +0100371}
Willy Tarreau6996e152007-04-30 14:37:43 +0200372
Willy Tarreau0c2fc1f2009-01-18 15:30:37 +0100373
Willy Tarreauf7bc57c2012-10-03 00:19:48 +0200374/* transport-layer operations for RAW sockets */
375struct xprt_ops raw_sock = {
Willy Tarreauc5788912012-08-24 18:12:41 +0200376 .snd_buf = raw_sock_from_buf,
377 .rcv_buf = raw_sock_to_buf,
Willy Tarreau96199b12012-08-24 00:46:52 +0200378#if defined(CONFIG_HAP_LINUX_SPLICE)
379 .rcv_pipe = raw_sock_to_pipe,
380 .snd_pipe = raw_sock_from_pipe,
381#endif
Willy Tarreauc5788912012-08-24 18:12:41 +0200382 .shutr = NULL,
383 .shutw = NULL,
384 .close = NULL,
Willy Tarreau5c979a92012-05-07 17:15:39 +0200385};
Willy Tarreaubaaee002006-06-26 02:48:02 +0200386
387/*
388 * Local variables:
389 * c-indent-level: 8
390 * c-basic-offset: 8
391 * End:
392 */