Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 1 | /* |
Willy Tarreau | b277d6e | 2012-05-11 16:59:14 +0200 | [diff] [blame] | 2 | * Functions used to send/receive data using SOCK_STREAM sockets. |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 3 | * |
Willy Tarreau | b277d6e | 2012-05-11 16:59:14 +0200 | [diff] [blame] | 4 | * Copyright 2000-2012 Willy Tarreau <w@1wt.eu> |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 5 | * |
| 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 Tarreau | 6b4aad4 | 2009-01-18 21:59:13 +0100 | [diff] [blame] | 13 | #define _GNU_SOURCE |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 14 | #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 Sivachenko | caf5898 | 2009-08-24 15:11:06 +0400 | [diff] [blame] | 23 | #include <netinet/tcp.h> |
| 24 | |
Willy Tarreau | 2dd0d47 | 2006-06-29 17:53:05 +0200 | [diff] [blame] | 25 | #include <common/compat.h> |
Willy Tarreau | e3ba5f0 | 2006-06-29 18:54:54 +0200 | [diff] [blame] | 26 | #include <common/config.h> |
Willy Tarreau | d6f087e | 2008-01-18 17:20:13 +0100 | [diff] [blame] | 27 | #include <common/debug.h> |
Willy Tarreau | 8374918 | 2007-04-15 20:56:27 +0200 | [diff] [blame] | 28 | #include <common/standard.h> |
Willy Tarreau | 0c303ee | 2008-07-07 00:09:58 +0200 | [diff] [blame] | 29 | #include <common/ticks.h> |
Willy Tarreau | 2dd0d47 | 2006-06-29 17:53:05 +0200 | [diff] [blame] | 30 | #include <common/time.h> |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 31 | |
Willy Tarreau | 2d21279 | 2008-08-27 21:41:35 +0200 | [diff] [blame] | 32 | #include <proto/buffers.h> |
Willy Tarreau | 8b11708 | 2012-08-06 15:06:49 +0200 | [diff] [blame] | 33 | #include <proto/connection.h> |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 34 | #include <proto/fd.h> |
Willy Tarreau | eb47268 | 2010-05-28 18:46:57 +0200 | [diff] [blame] | 35 | #include <proto/freq_ctr.h> |
| 36 | #include <proto/log.h> |
Willy Tarreau | 3eba98a | 2009-01-25 13:56:13 +0100 | [diff] [blame] | 37 | #include <proto/pipe.h> |
Willy Tarreau | fe598a7 | 2010-09-21 21:48:23 +0200 | [diff] [blame] | 38 | #include <proto/protocols.h> |
Willy Tarreau | 75bf2c9 | 2012-08-20 17:01:35 +0200 | [diff] [blame] | 39 | #include <proto/raw_sock.h> |
Willy Tarreau | 73b013b | 2012-05-21 16:31:45 +0200 | [diff] [blame] | 40 | #include <proto/stream_interface.h> |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 41 | #include <proto/task.h> |
| 42 | |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 43 | #include <types/global.h> |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 44 | |
Willy Tarreau | b277d6e | 2012-05-11 16:59:14 +0200 | [diff] [blame] | 45 | /* main event functions used to move data between sockets and buffers */ |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 46 | static void sock_raw_read(struct connection *conn); |
Willy Tarreau | b277d6e | 2012-05-11 16:59:14 +0200 | [diff] [blame] | 47 | |
| 48 | |
Willy Tarreau | af978c4 | 2012-08-20 20:21:01 +0200 | [diff] [blame] | 49 | #if 0 && defined(CONFIG_HAP_LINUX_SPLICE) |
Willy Tarreau | 43d8fb2 | 2011-08-22 17:12:02 +0200 | [diff] [blame] | 50 | #include <common/splice.h> |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 51 | |
| 52 | /* A pipe contains 16 segments max, and it's common to see segments of 1448 bytes |
| 53 | * because of timestamps. Use this as a hint for not looping on splice(). |
| 54 | */ |
| 55 | #define SPLICE_FULL_HINT 16*1448 |
| 56 | |
Willy Tarreau | a9de333 | 2009-11-28 07:47:10 +0100 | [diff] [blame] | 57 | /* how many data we attempt to splice at once when the buffer is configured for |
| 58 | * infinite forwarding */ |
| 59 | #define MAX_SPLICE_AT_ONCE (1<<30) |
| 60 | |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 61 | /* Returns : |
| 62 | * -1 if splice is not possible or not possible anymore and we must switch to |
| 63 | * user-land copy (eg: to_forward reached) |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 64 | * 0 otherwise, including errors and close. |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 65 | * Sets : |
| 66 | * BF_READ_NULL |
| 67 | * BF_READ_PARTIAL |
| 68 | * BF_WRITE_PARTIAL (during copy) |
Willy Tarreau | ba0b63d | 2009-09-20 08:09:44 +0200 | [diff] [blame] | 69 | * BF_OUT_EMPTY (during copy) |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 70 | * SI_FL_ERR |
| 71 | * SI_FL_WAIT_ROOM |
| 72 | * (SI_FL_WAIT_RECV) |
Willy Tarreau | 3eba98a | 2009-01-25 13:56:13 +0100 | [diff] [blame] | 73 | * |
| 74 | * This function automatically allocates a pipe from the pipe pool. It also |
| 75 | * carefully ensures to clear b->pipe whenever it leaves the pipe empty. |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 76 | */ |
Willy Tarreau | 7421efb | 2012-07-02 15:11:27 +0200 | [diff] [blame] | 77 | static int sock_raw_splice_in(struct channel *b, struct stream_interface *si) |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 78 | { |
Willy Tarreau | 82a0456 | 2011-12-11 22:37:06 +0100 | [diff] [blame] | 79 | static int splice_detects_close; |
Willy Tarreau | fb7508a | 2012-05-21 16:47:54 +0200 | [diff] [blame] | 80 | int fd = si_fd(si); |
Willy Tarreau | 31971e5 | 2009-09-20 12:07:52 +0200 | [diff] [blame] | 81 | int ret; |
| 82 | unsigned long max; |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 83 | int retval = 0; |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 84 | |
| 85 | if (!b->to_forward) |
| 86 | return -1; |
| 87 | |
| 88 | if (!(b->flags & BF_KERN_SPLICING)) |
| 89 | return -1; |
| 90 | |
Willy Tarreau | 572bf90 | 2012-07-02 17:01:20 +0200 | [diff] [blame] | 91 | if (buffer_not_empty(&b->buf)) { |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 92 | /* We're embarrassed, there are already data pending in |
| 93 | * the buffer and we don't want to have them at two |
| 94 | * locations at a time. Let's indicate we need some |
| 95 | * place and ask the consumer to hurry. |
| 96 | */ |
| 97 | si->flags |= SI_FL_WAIT_ROOM; |
Willy Tarreau | f9dabec | 2012-08-17 17:33:53 +0200 | [diff] [blame] | 98 | conn_data_stop_recv(&si->conn); |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 99 | b->rex = TICK_ETERNITY; |
Willy Tarreau | 73b013b | 2012-05-21 16:31:45 +0200 | [diff] [blame] | 100 | si_chk_snd(b->cons); |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 101 | return 0; |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 102 | } |
| 103 | |
Willy Tarreau | 3eba98a | 2009-01-25 13:56:13 +0100 | [diff] [blame] | 104 | if (unlikely(b->pipe == NULL)) { |
| 105 | if (pipes_used >= global.maxpipes || !(b->pipe = get_pipe())) { |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 106 | b->flags &= ~BF_KERN_SPLICING; |
| 107 | return -1; |
| 108 | } |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 109 | } |
| 110 | |
Willy Tarreau | 3eba98a | 2009-01-25 13:56:13 +0100 | [diff] [blame] | 111 | /* At this point, b->pipe is valid */ |
| 112 | |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 113 | while (1) { |
Willy Tarreau | a9de333 | 2009-11-28 07:47:10 +0100 | [diff] [blame] | 114 | if (b->to_forward == BUF_INFINITE_FORWARD) |
| 115 | max = MAX_SPLICE_AT_ONCE; |
| 116 | else |
| 117 | max = b->to_forward; |
| 118 | |
Willy Tarreau | 31971e5 | 2009-09-20 12:07:52 +0200 | [diff] [blame] | 119 | if (!max) { |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 120 | /* It looks like the buffer + the pipe already contain |
| 121 | * the maximum amount of data to be transferred. Try to |
| 122 | * send those data immediately on the other side if it |
| 123 | * is currently waiting. |
| 124 | */ |
| 125 | retval = -1; /* end of forwarding */ |
| 126 | break; |
| 127 | } |
| 128 | |
Willy Tarreau | 3eba98a | 2009-01-25 13:56:13 +0100 | [diff] [blame] | 129 | ret = splice(fd, NULL, b->pipe->prod, NULL, max, |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 130 | SPLICE_F_MOVE|SPLICE_F_NONBLOCK); |
| 131 | |
| 132 | if (ret <= 0) { |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 133 | if (ret == 0) { |
Willy Tarreau | 98b306b | 2009-01-25 11:11:32 +0100 | [diff] [blame] | 134 | /* connection closed. This is only detected by |
Willy Tarreau | 82a0456 | 2011-12-11 22:37:06 +0100 | [diff] [blame] | 135 | * recent kernels (>= 2.6.27.13). If we notice |
| 136 | * it works, we store the info for later use. |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 137 | */ |
Willy Tarreau | 82a0456 | 2011-12-11 22:37:06 +0100 | [diff] [blame] | 138 | splice_detects_close = 1; |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 139 | b->flags |= BF_READ_NULL; |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 140 | break; |
| 141 | } |
| 142 | |
| 143 | if (errno == EAGAIN) { |
| 144 | /* there are two reasons for EAGAIN : |
| 145 | * - nothing in the socket buffer (standard) |
| 146 | * - pipe is full |
Willy Tarreau | 98b306b | 2009-01-25 11:11:32 +0100 | [diff] [blame] | 147 | * - the connection is closed (kernel < 2.6.27.13) |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 148 | * Since we don't know if pipe is full, we'll |
| 149 | * stop if the pipe is not empty. Anyway, we |
| 150 | * will almost always fill/empty the pipe. |
| 151 | */ |
| 152 | |
Willy Tarreau | 3eba98a | 2009-01-25 13:56:13 +0100 | [diff] [blame] | 153 | if (b->pipe->data) { |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 154 | si->flags |= SI_FL_WAIT_ROOM; |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 155 | break; |
| 156 | } |
| 157 | |
Willy Tarreau | 82a0456 | 2011-12-11 22:37:06 +0100 | [diff] [blame] | 158 | /* We don't know if the connection was closed, |
| 159 | * but if we know splice detects close, then we |
| 160 | * know it for sure. |
Willy Tarreau | 98b306b | 2009-01-25 11:11:32 +0100 | [diff] [blame] | 161 | * But if we're called upon POLLIN with an empty |
Willy Tarreau | 82a0456 | 2011-12-11 22:37:06 +0100 | [diff] [blame] | 162 | * pipe and get EAGAIN, it is suspect enough to |
Willy Tarreau | 98b306b | 2009-01-25 11:11:32 +0100 | [diff] [blame] | 163 | * try to fall back to the normal recv scheme |
| 164 | * which will be able to deal with the situation. |
| 165 | */ |
Willy Tarreau | 82a0456 | 2011-12-11 22:37:06 +0100 | [diff] [blame] | 166 | if (splice_detects_close) |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 167 | conn_data_poll_recv(&si->conn); /* we know for sure that it's EAGAIN */ |
Willy Tarreau | 82a0456 | 2011-12-11 22:37:06 +0100 | [diff] [blame] | 168 | else |
| 169 | retval = -1; |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 170 | break; |
| 171 | } |
Willy Tarreau | dc340a9 | 2009-06-28 23:10:19 +0200 | [diff] [blame] | 172 | |
Willy Tarreau | a9de333 | 2009-11-28 07:47:10 +0100 | [diff] [blame] | 173 | if (errno == ENOSYS || errno == EINVAL) { |
Willy Tarreau | dc340a9 | 2009-06-28 23:10:19 +0200 | [diff] [blame] | 174 | /* splice not supported on this end, disable it */ |
| 175 | b->flags &= ~BF_KERN_SPLICING; |
| 176 | si->flags &= ~SI_FL_CAP_SPLICE; |
| 177 | put_pipe(b->pipe); |
| 178 | b->pipe = NULL; |
| 179 | return -1; |
| 180 | } |
| 181 | |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 182 | /* here we have another error */ |
| 183 | si->flags |= SI_FL_ERR; |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 184 | break; |
| 185 | } /* ret <= 0 */ |
| 186 | |
Willy Tarreau | 31971e5 | 2009-09-20 12:07:52 +0200 | [diff] [blame] | 187 | if (b->to_forward != BUF_INFINITE_FORWARD) |
| 188 | b->to_forward -= ret; |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 189 | b->total += ret; |
Willy Tarreau | 3eba98a | 2009-01-25 13:56:13 +0100 | [diff] [blame] | 190 | b->pipe->data += ret; |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 191 | b->flags |= BF_READ_PARTIAL; |
Willy Tarreau | ba0b63d | 2009-09-20 08:09:44 +0200 | [diff] [blame] | 192 | b->flags &= ~BF_OUT_EMPTY; |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 193 | |
Willy Tarreau | 6f4a82c | 2009-03-21 20:43:57 +0100 | [diff] [blame] | 194 | if (b->pipe->data >= SPLICE_FULL_HINT || |
| 195 | ret >= global.tune.recv_enough) { |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 196 | /* We've read enough of it for this time. */ |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 197 | break; |
| 198 | } |
| 199 | } /* while */ |
| 200 | |
Willy Tarreau | 3eba98a | 2009-01-25 13:56:13 +0100 | [diff] [blame] | 201 | if (unlikely(!b->pipe->data)) { |
| 202 | put_pipe(b->pipe); |
| 203 | b->pipe = NULL; |
| 204 | } |
| 205 | |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 206 | return retval; |
| 207 | } |
| 208 | |
Willy Tarreau | 6b4aad4 | 2009-01-18 21:59:13 +0100 | [diff] [blame] | 209 | #endif /* CONFIG_HAP_LINUX_SPLICE */ |
| 210 | |
| 211 | |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 212 | /* Receive up to <count> bytes from connection <conn>'s socket and store them |
| 213 | * into buffer <buf>. The caller must ensure that <count> is always smaller |
| 214 | * than the buffer's size. Only one call to recv() is performed, unless the |
| 215 | * buffer wraps, in which case a second call may be performed. The connection's |
| 216 | * flags are updated with whatever special event is detected (error, read0, |
| 217 | * empty). The caller is responsible for taking care of those events and |
| 218 | * avoiding the call if inappropriate. The function does not call the |
| 219 | * connection's polling update function, so the caller is responsible for this. |
| 220 | */ |
| 221 | static int raw_sock_to_buf(struct connection *conn, struct buffer *buf, int count) |
| 222 | { |
| 223 | int ret, done = 0; |
| 224 | int try = count; |
| 225 | |
| 226 | /* stop here if we reached the end of data */ |
| 227 | if ((fdtab[conn->t.sock.fd].ev & (FD_POLL_IN|FD_POLL_HUP)) == FD_POLL_HUP) |
| 228 | goto read0; |
| 229 | |
| 230 | /* compute the maximum block size we can read at once. */ |
| 231 | if (buffer_empty(buf)) { |
| 232 | /* let's realign the buffer to optimize I/O */ |
| 233 | buf->p = buf->data; |
| 234 | } |
| 235 | else if (buf->data + buf->o < buf->p && |
| 236 | buf->p + buf->i < buf->data + buf->size) { |
| 237 | /* remaining space wraps at the end, with a moving limit */ |
| 238 | if (try > buf->data + buf->size - (buf->p + buf->i)) |
| 239 | try = buf->data + buf->size - (buf->p + buf->i); |
| 240 | } |
| 241 | |
| 242 | /* read the largest possible block. For this, we perform only one call |
| 243 | * to recv() unless the buffer wraps and we exactly fill the first hunk, |
| 244 | * in which case we accept to do it once again. A new attempt is made on |
| 245 | * EINTR too. |
| 246 | */ |
| 247 | while (try) { |
| 248 | ret = recv(conn->t.sock.fd, bi_end(buf), try, 0); |
| 249 | |
| 250 | if (ret > 0) { |
| 251 | buf->i += ret; |
| 252 | done += ret; |
| 253 | if (ret < try) { |
| 254 | /* unfortunately, on level-triggered events, POLL_HUP |
| 255 | * is generally delivered AFTER the system buffer is |
| 256 | * empty, so this one might never match. |
| 257 | */ |
| 258 | if (fdtab[conn->t.sock.fd].ev & FD_POLL_HUP) |
| 259 | goto read0; |
| 260 | break; |
| 261 | } |
| 262 | count -= ret; |
| 263 | try = count; |
| 264 | } |
| 265 | else if (ret == 0) { |
| 266 | goto read0; |
| 267 | } |
| 268 | else if (errno == EAGAIN) { |
| 269 | conn->flags |= CO_FL_WAIT_DATA; |
| 270 | break; |
| 271 | } |
| 272 | else if (errno != EINTR) { |
| 273 | conn->flags |= CO_FL_ERROR; |
| 274 | break; |
| 275 | } |
| 276 | } |
| 277 | return done; |
| 278 | |
| 279 | read0: |
| 280 | conn_sock_read0(conn); |
| 281 | return done; |
| 282 | } |
| 283 | |
| 284 | |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 285 | /* |
Willy Tarreau | d797128 | 2006-07-29 18:36:34 +0200 | [diff] [blame] | 286 | * this function is called on a read event from a stream socket. |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 287 | */ |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 288 | static void sock_raw_read(struct connection *conn) |
Willy Tarreau | b277d6e | 2012-05-11 16:59:14 +0200 | [diff] [blame] | 289 | { |
Willy Tarreau | 8018471 | 2012-07-06 14:54:49 +0200 | [diff] [blame] | 290 | struct stream_interface *si = container_of(conn, struct stream_interface, conn); |
Willy Tarreau | 7421efb | 2012-07-02 15:11:27 +0200 | [diff] [blame] | 291 | struct channel *b = si->ib; |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 292 | int ret, max, cur_read; |
Willy Tarreau | b8949f1 | 2007-03-23 22:39:59 +0100 | [diff] [blame] | 293 | int read_poll = MAX_READ_POLL_LOOPS; |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 294 | |
| 295 | #ifdef DEBUG_FULL |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 296 | fprintf(stderr,"sock_raw_read : fd=%d, ev=0x%02x, owner=%p\n", conn->t.sock.fd, fdtab[conn->t.sock.fd].ev, fdtab[conn->t.sock.fd].owner); |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 297 | #endif |
Willy Tarreau | 7154365 | 2009-07-14 19:55:05 +0200 | [diff] [blame] | 298 | /* stop immediately on errors. Note that we DON'T want to stop on |
| 299 | * POLL_ERR, as the poller might report a write error while there |
| 300 | * are still data available in the recv buffer. This typically |
| 301 | * happens when we send too large a request to a backend server |
| 302 | * which rejects it before reading it all. |
| 303 | */ |
Willy Tarreau | 8018471 | 2012-07-06 14:54:49 +0200 | [diff] [blame] | 304 | if (conn->flags & CO_FL_ERROR) |
Willy Tarreau | 6996e15 | 2007-04-30 14:37:43 +0200 | [diff] [blame] | 305 | goto out_error; |
Willy Tarreau | d6f087e | 2008-01-18 17:20:13 +0100 | [diff] [blame] | 306 | |
| 307 | /* stop here if we reached the end of data */ |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 308 | if (conn_data_read0_pending(conn)) |
Willy Tarreau | d6f087e | 2008-01-18 17:20:13 +0100 | [diff] [blame] | 309 | goto out_shutdown_r; |
Willy Tarreau | 8374918 | 2007-04-15 20:56:27 +0200 | [diff] [blame] | 310 | |
Willy Tarreau | d06e711 | 2009-03-29 10:18:41 +0200 | [diff] [blame] | 311 | /* maybe we were called immediately after an asynchronous shutr */ |
| 312 | if (b->flags & BF_SHUTR) |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 313 | return; |
Willy Tarreau | d06e711 | 2009-03-29 10:18:41 +0200 | [diff] [blame] | 314 | |
Willy Tarreau | af978c4 | 2012-08-20 20:21:01 +0200 | [diff] [blame] | 315 | #if 0 && defined(CONFIG_HAP_LINUX_SPLICE) |
Willy Tarreau | 14acc70 | 2011-05-11 20:47:24 +0200 | [diff] [blame] | 316 | if (b->to_forward >= MIN_SPLICE_FORWARD && b->flags & BF_KERN_SPLICING) { |
Willy Tarreau | 98b306b | 2009-01-25 11:11:32 +0100 | [diff] [blame] | 317 | |
| 318 | /* Under Linux, if FD_POLL_HUP is set, we have reached the end. |
| 319 | * Since older splice() implementations were buggy and returned |
| 320 | * EAGAIN on end of read, let's bypass the call to splice() now. |
| 321 | */ |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 322 | if (fdtab[conn->t.sock.fd].ev & FD_POLL_HUP) |
Willy Tarreau | 98b306b | 2009-01-25 11:11:32 +0100 | [diff] [blame] | 323 | goto out_shutdown_r; |
| 324 | |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 325 | if (sock_raw_splice_in(b, si) >= 0) { |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 326 | if (si->flags & SI_FL_ERR) |
| 327 | goto out_error; |
| 328 | if (b->flags & BF_READ_NULL) |
| 329 | goto out_shutdown_r; |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 330 | return; |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 331 | } |
| 332 | /* splice not possible (anymore), let's go on on standard copy */ |
| 333 | } |
| 334 | #endif |
Willy Tarreau | 8a7af60 | 2008-05-03 23:07:14 +0200 | [diff] [blame] | 335 | cur_read = 0; |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 336 | conn->flags &= ~(CO_FL_WAIT_DATA | CO_FL_WAIT_ROOM); |
| 337 | while (!(conn->flags & (CO_FL_ERROR | CO_FL_SOCK_RD_SH | CO_FL_DATA_RD_SH | CO_FL_WAIT_DATA | CO_FL_WAIT_ROOM | CO_FL_HANDSHAKE))) { |
Willy Tarreau | 9dab5fc | 2012-05-07 11:56:55 +0200 | [diff] [blame] | 338 | max = bi_avail(b); |
Willy Tarreau | 864e825 | 2009-12-28 17:36:37 +0100 | [diff] [blame] | 339 | |
Willy Tarreau | 9dab5fc | 2012-05-07 11:56:55 +0200 | [diff] [blame] | 340 | if (!max) { |
Willy Tarreau | 864e825 | 2009-12-28 17:36:37 +0100 | [diff] [blame] | 341 | b->flags |= BF_FULL; |
| 342 | si->flags |= SI_FL_WAIT_ROOM; |
| 343 | break; |
| 344 | } |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 345 | |
Willy Tarreau | 1fe6bc3 | 2012-08-20 20:27:59 +0200 | [diff] [blame^] | 346 | ret = conn->data->rcv_buf(conn, &b->buf, max); |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 347 | if (ret <= 0) |
| 348 | break; |
Willy Tarreau | d6d0690 | 2009-08-19 11:22:33 +0200 | [diff] [blame] | 349 | |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 350 | cur_read += ret; |
Willy Tarreau | b38903c | 2008-11-23 21:33:29 +0100 | [diff] [blame] | 351 | |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 352 | /* if we're allowed to directly forward data, we must update ->o */ |
| 353 | if (b->to_forward && !(b->flags & (BF_SHUTW|BF_SHUTW_NOW))) { |
| 354 | unsigned long fwd = ret; |
| 355 | if (b->to_forward != BUF_INFINITE_FORWARD) { |
| 356 | if (fwd > b->to_forward) |
| 357 | fwd = b->to_forward; |
| 358 | b->to_forward -= fwd; |
Willy Tarreau | 0abebcc | 2009-01-08 00:09:41 +0100 | [diff] [blame] | 359 | } |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 360 | b_adv(b, fwd); |
| 361 | } |
Willy Tarreau | f890dc9 | 2008-12-13 21:12:26 +0100 | [diff] [blame] | 362 | |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 363 | if (conn->flags & CO_FL_WAIT_L4_CONN) { |
| 364 | conn->flags &= ~CO_FL_WAIT_L4_CONN; |
| 365 | si->exp = TICK_ETERNITY; |
| 366 | } |
Willy Tarreau | b38903c | 2008-11-23 21:33:29 +0100 | [diff] [blame] | 367 | |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 368 | b->flags |= BF_READ_PARTIAL; |
| 369 | b->total += ret; |
Willy Tarreau | 9641e8f | 2007-03-23 23:02:09 +0100 | [diff] [blame] | 370 | |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 371 | if (bi_full(b)) { |
| 372 | /* The buffer is now full, there's no point in going through |
| 373 | * the loop again. |
| 374 | */ |
| 375 | if (!(b->flags & BF_STREAMER_FAST) && (cur_read == buffer_len(&b->buf))) { |
| 376 | b->xfer_small = 0; |
| 377 | b->xfer_large++; |
| 378 | if (b->xfer_large >= 3) { |
| 379 | /* we call this buffer a fast streamer if it manages |
| 380 | * to be filled in one call 3 consecutive times. |
| 381 | */ |
| 382 | b->flags |= (BF_STREAMER | BF_STREAMER_FAST); |
| 383 | //fputc('+', stderr); |
Willy Tarreau | 8a7af60 | 2008-05-03 23:07:14 +0200 | [diff] [blame] | 384 | } |
Willy Tarreau | 6996e15 | 2007-04-30 14:37:43 +0200 | [diff] [blame] | 385 | } |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 386 | else if ((b->flags & (BF_STREAMER | BF_STREAMER_FAST)) && |
| 387 | (cur_read <= b->buf.size / 2)) { |
| 388 | b->xfer_large = 0; |
| 389 | b->xfer_small++; |
| 390 | if (b->xfer_small >= 2) { |
| 391 | /* if the buffer has been at least half full twice, |
| 392 | * we receive faster than we send, so at least it |
| 393 | * is not a "fast streamer". |
| 394 | */ |
| 395 | b->flags &= ~BF_STREAMER_FAST; |
| 396 | //fputc('-', stderr); |
Willy Tarreau | 8a7af60 | 2008-05-03 23:07:14 +0200 | [diff] [blame] | 397 | } |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 398 | } |
| 399 | else { |
| 400 | b->xfer_small = 0; |
| 401 | b->xfer_large = 0; |
| 402 | } |
Willy Tarreau | 2bea3a1 | 2008-08-28 09:47:43 +0200 | [diff] [blame] | 403 | |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 404 | b->flags |= BF_FULL; |
| 405 | si->flags |= SI_FL_WAIT_ROOM; |
| 406 | break; |
| 407 | } |
Willy Tarreau | ab3e1d3 | 2007-06-03 14:10:36 +0200 | [diff] [blame] | 408 | |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 409 | if ((b->flags & BF_READ_DONTWAIT) || --read_poll <= 0) |
| 410 | break; |
Willy Tarreau | 6f4a82c | 2009-03-21 20:43:57 +0100 | [diff] [blame] | 411 | |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 412 | /* if too many bytes were missing from last read, it means that |
| 413 | * it's pointless trying to read again because the system does |
| 414 | * not have them in buffers. |
| 415 | */ |
| 416 | if (ret < max) { |
| 417 | if ((b->flags & (BF_STREAMER | BF_STREAMER_FAST)) && |
| 418 | (cur_read <= b->buf.size / 2)) { |
| 419 | b->xfer_large = 0; |
| 420 | b->xfer_small++; |
| 421 | if (b->xfer_small >= 3) { |
| 422 | /* we have read less than half of the buffer in |
| 423 | * one pass, and this happened at least 3 times. |
| 424 | * This is definitely not a streamer. |
| 425 | */ |
| 426 | b->flags &= ~(BF_STREAMER | BF_STREAMER_FAST); |
| 427 | //fputc('!', stderr); |
| 428 | } |
Willy Tarreau | 6f4a82c | 2009-03-21 20:43:57 +0100 | [diff] [blame] | 429 | } |
Willy Tarreau | 8374918 | 2007-04-15 20:56:27 +0200 | [diff] [blame] | 430 | |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 431 | /* if a streamer has read few data, it may be because we |
| 432 | * have exhausted system buffers. It's not worth trying |
| 433 | * again. |
| 434 | */ |
| 435 | if (b->flags & BF_STREAMER) |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 436 | break; |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 437 | |
| 438 | /* if we read a large block smaller than what we requested, |
| 439 | * it's almost certain we'll never get anything more. |
Willy Tarreau | 9f19529 | 2007-04-15 21:26:58 +0200 | [diff] [blame] | 440 | */ |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 441 | if (ret >= global.tune.recv_enough) |
| 442 | break; |
Willy Tarreau | 8374918 | 2007-04-15 20:56:27 +0200 | [diff] [blame] | 443 | } |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 444 | } /* while !flags */ |
| 445 | |
| 446 | if (conn->flags & CO_FL_ERROR) |
| 447 | goto out_error; |
| 448 | |
| 449 | if (conn->flags & CO_FL_WAIT_DATA) { |
| 450 | /* we don't automatically ask for polling if we have |
| 451 | * read enough data, as it saves some syscalls with |
| 452 | * speculative pollers. |
| 453 | */ |
| 454 | if (cur_read < MIN_RET_FOR_READ_LOOP) |
| 455 | __conn_data_poll_recv(conn); |
| 456 | else |
| 457 | __conn_data_want_recv(conn); |
| 458 | } |
| 459 | |
| 460 | if (conn_data_read0_pending(conn)) |
| 461 | /* connection closed */ |
| 462 | goto out_shutdown_r; |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 463 | |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 464 | return; |
Willy Tarreau | 6996e15 | 2007-04-30 14:37:43 +0200 | [diff] [blame] | 465 | |
Willy Tarreau | d6f087e | 2008-01-18 17:20:13 +0100 | [diff] [blame] | 466 | out_shutdown_r: |
Willy Tarreau | e5ed406 | 2008-08-30 03:17:31 +0200 | [diff] [blame] | 467 | /* we received a shutdown */ |
Willy Tarreau | d6f087e | 2008-01-18 17:20:13 +0100 | [diff] [blame] | 468 | b->flags |= BF_READ_NULL; |
Willy Tarreau | 520d95e | 2009-09-19 21:04:57 +0200 | [diff] [blame] | 469 | if (b->flags & BF_AUTO_CLOSE) |
Willy Tarreau | 418fd47 | 2009-09-06 21:37:23 +0200 | [diff] [blame] | 470 | buffer_shutw_now(b); |
Willy Tarreau | 9bf9c14 | 2012-08-20 15:38:41 +0200 | [diff] [blame] | 471 | stream_sock_read0(si); |
Willy Tarreau | 2ba4465 | 2012-08-20 17:30:32 +0200 | [diff] [blame] | 472 | conn_data_read0(conn); |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 473 | return; |
Willy Tarreau | d6f087e | 2008-01-18 17:20:13 +0100 | [diff] [blame] | 474 | |
Willy Tarreau | 6996e15 | 2007-04-30 14:37:43 +0200 | [diff] [blame] | 475 | out_error: |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 476 | /* Read error on the connection, report the error and stop I/O */ |
Willy Tarreau | 8018471 | 2012-07-06 14:54:49 +0200 | [diff] [blame] | 477 | conn->flags |= CO_FL_ERROR; |
Willy Tarreau | f9dabec | 2012-08-17 17:33:53 +0200 | [diff] [blame] | 478 | conn_data_stop_both(conn); |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 479 | } |
| 480 | |
| 481 | |
| 482 | /* |
Willy Tarreau | 0c2fc1f | 2009-01-18 15:30:37 +0100 | [diff] [blame] | 483 | * This function is called to send buffer data to a stream socket. |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 484 | * It returns -1 in case of unrecoverable error, otherwise zero. |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 485 | */ |
Willy Tarreau | fae4499 | 2012-08-20 14:02:10 +0200 | [diff] [blame] | 486 | static int sock_raw_write_loop(struct connection *conn) |
Willy Tarreau | 0c2fc1f | 2009-01-18 15:30:37 +0100 | [diff] [blame] | 487 | { |
Willy Tarreau | fae4499 | 2012-08-20 14:02:10 +0200 | [diff] [blame] | 488 | struct stream_interface *si = container_of(conn, struct stream_interface, conn); |
Willy Tarreau | 7421efb | 2012-07-02 15:11:27 +0200 | [diff] [blame] | 489 | struct channel *b = si->ob; |
Willy Tarreau | 8374918 | 2007-04-15 20:56:27 +0200 | [diff] [blame] | 490 | int write_poll = MAX_WRITE_POLL_LOOPS; |
Willy Tarreau | 0c2fc1f | 2009-01-18 15:30:37 +0100 | [diff] [blame] | 491 | int ret, max; |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 492 | |
Willy Tarreau | af978c4 | 2012-08-20 20:21:01 +0200 | [diff] [blame] | 493 | #if 0 && defined(CONFIG_HAP_LINUX_SPLICE) |
Willy Tarreau | 3eba98a | 2009-01-25 13:56:13 +0100 | [diff] [blame] | 494 | while (b->pipe) { |
Willy Tarreau | fb7508a | 2012-05-21 16:47:54 +0200 | [diff] [blame] | 495 | ret = splice(b->pipe->cons, NULL, si_fd(si), NULL, b->pipe->data, |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 496 | SPLICE_F_MOVE|SPLICE_F_NONBLOCK); |
| 497 | if (ret <= 0) { |
| 498 | if (ret == 0 || errno == EAGAIN) { |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 499 | conn_data_poll_send(&si->conn); |
| 500 | return 0; |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 501 | } |
| 502 | /* here we have another error */ |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 503 | return -1; |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 504 | } |
| 505 | |
| 506 | b->flags |= BF_WRITE_PARTIAL; |
Willy Tarreau | 3eba98a | 2009-01-25 13:56:13 +0100 | [diff] [blame] | 507 | b->pipe->data -= ret; |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 508 | |
Willy Tarreau | 3eba98a | 2009-01-25 13:56:13 +0100 | [diff] [blame] | 509 | if (!b->pipe->data) { |
| 510 | put_pipe(b->pipe); |
| 511 | b->pipe = NULL; |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 512 | break; |
Willy Tarreau | 3eba98a | 2009-01-25 13:56:13 +0100 | [diff] [blame] | 513 | } |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 514 | |
| 515 | if (--write_poll <= 0) |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 516 | return 0; |
Willy Tarreau | eb9fd51 | 2011-12-11 22:11:47 +0100 | [diff] [blame] | 517 | |
| 518 | /* The only reason we did not empty the pipe is that the output |
| 519 | * buffer is full. |
| 520 | */ |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 521 | conn_data_poll_send(&si->conn); |
Willy Tarreau | eb9fd51 | 2011-12-11 22:11:47 +0100 | [diff] [blame] | 522 | return 0; |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 523 | } |
| 524 | |
| 525 | /* At this point, the pipe is empty, but we may still have data pending |
| 526 | * in the normal buffer. |
| 527 | */ |
Willy Tarreau | 5bd8c37 | 2009-01-19 00:32:22 +0100 | [diff] [blame] | 528 | #endif |
Willy Tarreau | 572bf90 | 2012-07-02 17:01:20 +0200 | [diff] [blame] | 529 | if (!b->buf.o) { |
Willy Tarreau | ba0b63d | 2009-09-20 08:09:44 +0200 | [diff] [blame] | 530 | b->flags |= BF_OUT_EMPTY; |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 531 | return 0; |
Willy Tarreau | ba0b63d | 2009-09-20 08:09:44 +0200 | [diff] [blame] | 532 | } |
Willy Tarreau | 8374918 | 2007-04-15 20:56:27 +0200 | [diff] [blame] | 533 | |
Willy Tarreau | 0c2fc1f | 2009-01-18 15:30:37 +0100 | [diff] [blame] | 534 | /* when we're in this loop, we already know that there is no spliced |
| 535 | * data left, and that there are sendable buffered data. |
| 536 | */ |
Willy Tarreau | 6996e15 | 2007-04-30 14:37:43 +0200 | [diff] [blame] | 537 | while (1) { |
Willy Tarreau | 572bf90 | 2012-07-02 17:01:20 +0200 | [diff] [blame] | 538 | max = b->buf.o; |
Willy Tarreau | 8374918 | 2007-04-15 20:56:27 +0200 | [diff] [blame] | 539 | |
Willy Tarreau | 89fa706 | 2012-03-02 16:13:16 +0100 | [diff] [blame] | 540 | /* outgoing data may wrap at the end */ |
Willy Tarreau | 572bf90 | 2012-07-02 17:01:20 +0200 | [diff] [blame] | 541 | if (b->buf.data + max > b->buf.p) |
| 542 | max = b->buf.data + max - b->buf.p; |
Willy Tarreau | f890dc9 | 2008-12-13 21:12:26 +0100 | [diff] [blame] | 543 | |
Willy Tarreau | 6db06d3 | 2009-08-19 11:14:11 +0200 | [diff] [blame] | 544 | /* check if we want to inform the kernel that we're interested in |
| 545 | * sending more data after this call. We want this if : |
| 546 | * - we're about to close after this last send and want to merge |
| 547 | * the ongoing FIN with the last segment. |
| 548 | * - we know we can't send everything at once and must get back |
| 549 | * here because of unaligned data |
Willy Tarreau | d38b53b | 2010-01-03 11:18:34 +0100 | [diff] [blame] | 550 | * - there is still a finite amount of data to forward |
Willy Tarreau | 6db06d3 | 2009-08-19 11:14:11 +0200 | [diff] [blame] | 551 | * The test is arranged so that the most common case does only 2 |
| 552 | * tests. |
Willy Tarreau | fb14edc | 2009-06-14 15:24:37 +0200 | [diff] [blame] | 553 | */ |
Willy Tarreau | fb14edc | 2009-06-14 15:24:37 +0200 | [diff] [blame] | 554 | |
Willy Tarreau | face839 | 2010-01-03 11:37:54 +0100 | [diff] [blame] | 555 | if (MSG_NOSIGNAL && MSG_MORE) { |
Willy Tarreau | 6db06d3 | 2009-08-19 11:14:11 +0200 | [diff] [blame] | 556 | unsigned int send_flag = MSG_DONTWAIT | MSG_NOSIGNAL; |
| 557 | |
Willy Tarreau | 96e3121 | 2011-05-30 18:10:30 +0200 | [diff] [blame] | 558 | if ((!(b->flags & BF_NEVER_WAIT) && |
| 559 | ((b->to_forward && b->to_forward != BUF_INFINITE_FORWARD) || |
| 560 | (b->flags & BF_EXPECT_MORE))) || |
Willy Tarreau | 572bf90 | 2012-07-02 17:01:20 +0200 | [diff] [blame] | 561 | ((b->flags & (BF_SHUTW|BF_SHUTW_NOW|BF_HIJACK)) == BF_SHUTW_NOW && (max == b->buf.o)) || |
| 562 | (max != b->buf.o)) { |
Willy Tarreau | face839 | 2010-01-03 11:37:54 +0100 | [diff] [blame] | 563 | send_flag |= MSG_MORE; |
| 564 | } |
Willy Tarreau | 6db06d3 | 2009-08-19 11:14:11 +0200 | [diff] [blame] | 565 | |
Willy Tarreau | 2be3939 | 2010-01-03 17:24:51 +0100 | [diff] [blame] | 566 | /* this flag has precedence over the rest */ |
| 567 | if (b->flags & BF_SEND_DONTWAIT) |
| 568 | send_flag &= ~MSG_MORE; |
| 569 | |
Willy Tarreau | 572bf90 | 2012-07-02 17:01:20 +0200 | [diff] [blame] | 570 | ret = send(si_fd(si), bo_ptr(&b->buf), max, send_flag); |
Willy Tarreau | d6d0690 | 2009-08-19 11:22:33 +0200 | [diff] [blame] | 571 | } else { |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 572 | int skerr; |
| 573 | socklen_t lskerr = sizeof(skerr); |
| 574 | |
Willy Tarreau | fb7508a | 2012-05-21 16:47:54 +0200 | [diff] [blame] | 575 | ret = getsockopt(si_fd(si), SOL_SOCKET, SO_ERROR, &skerr, &lskerr); |
Willy Tarreau | c642348 | 2006-10-15 14:59:03 +0200 | [diff] [blame] | 576 | if (ret == -1 || skerr) |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 577 | ret = -1; |
| 578 | else |
Willy Tarreau | 572bf90 | 2012-07-02 17:01:20 +0200 | [diff] [blame] | 579 | ret = send(si_fd(si), bo_ptr(&b->buf), max, MSG_DONTWAIT); |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 580 | } |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 581 | |
| 582 | if (ret > 0) { |
Willy Tarreau | 505e34a | 2012-07-06 10:17:53 +0200 | [diff] [blame] | 583 | if (si->conn.flags & CO_FL_WAIT_L4_CONN) { |
| 584 | si->conn.flags &= ~CO_FL_WAIT_L4_CONN; |
Willy Tarreau | 8ae52cb | 2012-05-20 10:38:46 +0200 | [diff] [blame] | 585 | si->exp = TICK_ETERNITY; |
| 586 | } |
Willy Tarreau | b38903c | 2008-11-23 21:33:29 +0100 | [diff] [blame] | 587 | |
Willy Tarreau | 3da77c5 | 2008-08-29 09:58:42 +0200 | [diff] [blame] | 588 | b->flags |= BF_WRITE_PARTIAL; |
Willy Tarreau | e393fe2 | 2008-08-16 22:18:07 +0200 | [diff] [blame] | 589 | |
Willy Tarreau | 572bf90 | 2012-07-02 17:01:20 +0200 | [diff] [blame] | 590 | b->buf.o -= ret; |
| 591 | if (likely(!buffer_len(&b->buf))) |
Willy Tarreau | 0c2fc1f | 2009-01-18 15:30:37 +0100 | [diff] [blame] | 592 | /* optimize data alignment in the buffer */ |
Willy Tarreau | 572bf90 | 2012-07-02 17:01:20 +0200 | [diff] [blame] | 593 | b->buf.p = b->buf.data; |
Willy Tarreau | 8374918 | 2007-04-15 20:56:27 +0200 | [diff] [blame] | 594 | |
Willy Tarreau | 9dab5fc | 2012-05-07 11:56:55 +0200 | [diff] [blame] | 595 | if (likely(!bi_full(b))) |
Willy Tarreau | 02d6cfc | 2012-03-01 18:19:58 +0100 | [diff] [blame] | 596 | b->flags &= ~BF_FULL; |
| 597 | |
Willy Tarreau | 572bf90 | 2012-07-02 17:01:20 +0200 | [diff] [blame] | 598 | if (!b->buf.o) { |
Willy Tarreau | f17810e | 2012-03-09 18:10:44 +0100 | [diff] [blame] | 599 | /* Always clear both flags once everything has been sent, they're one-shot */ |
| 600 | b->flags &= ~(BF_EXPECT_MORE | BF_SEND_DONTWAIT); |
Willy Tarreau | ba0b63d | 2009-09-20 08:09:44 +0200 | [diff] [blame] | 601 | if (likely(!b->pipe)) |
| 602 | b->flags |= BF_OUT_EMPTY; |
Willy Tarreau | 0c2fc1f | 2009-01-18 15:30:37 +0100 | [diff] [blame] | 603 | break; |
Willy Tarreau | ba0b63d | 2009-09-20 08:09:44 +0200 | [diff] [blame] | 604 | } |
Willy Tarreau | 8374918 | 2007-04-15 20:56:27 +0200 | [diff] [blame] | 605 | |
Willy Tarreau | ab3e1d3 | 2007-06-03 14:10:36 +0200 | [diff] [blame] | 606 | /* if the system buffer is full, don't insist */ |
| 607 | if (ret < max) |
| 608 | break; |
| 609 | |
Willy Tarreau | 6996e15 | 2007-04-30 14:37:43 +0200 | [diff] [blame] | 610 | if (--write_poll <= 0) |
| 611 | break; |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 612 | } |
Willy Tarreau | 6996e15 | 2007-04-30 14:37:43 +0200 | [diff] [blame] | 613 | else if (ret == 0 || errno == EAGAIN) { |
Willy Tarreau | 0c2fc1f | 2009-01-18 15:30:37 +0100 | [diff] [blame] | 614 | /* nothing written, we need to poll for write first */ |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 615 | conn_data_poll_send(&si->conn); |
| 616 | return 0; |
Willy Tarreau | 8374918 | 2007-04-15 20:56:27 +0200 | [diff] [blame] | 617 | } |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 618 | else { |
Willy Tarreau | 0c2fc1f | 2009-01-18 15:30:37 +0100 | [diff] [blame] | 619 | /* bad, we got an error */ |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 620 | return -1; |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 621 | } |
Willy Tarreau | 6996e15 | 2007-04-30 14:37:43 +0200 | [diff] [blame] | 622 | } /* while (1) */ |
Willy Tarreau | afad0e0 | 2012-08-09 14:45:22 +0200 | [diff] [blame] | 623 | return 0; |
Willy Tarreau | 0c2fc1f | 2009-01-18 15:30:37 +0100 | [diff] [blame] | 624 | } |
Willy Tarreau | 6996e15 | 2007-04-30 14:37:43 +0200 | [diff] [blame] | 625 | |
Willy Tarreau | 0c2fc1f | 2009-01-18 15:30:37 +0100 | [diff] [blame] | 626 | |
Willy Tarreau | 5c979a9 | 2012-05-07 17:15:39 +0200 | [diff] [blame] | 627 | /* stream sock operations */ |
Willy Tarreau | 75bf2c9 | 2012-08-20 17:01:35 +0200 | [diff] [blame] | 628 | struct sock_ops raw_sock = { |
Willy Tarreau | 100c467 | 2012-08-20 12:06:26 +0200 | [diff] [blame] | 629 | .update = stream_int_update_conn, |
Willy Tarreau | 4a36b56 | 2012-08-06 19:31:45 +0200 | [diff] [blame] | 630 | .shutr = NULL, |
| 631 | .shutw = NULL, |
Willy Tarreau | 46a8d92 | 2012-08-20 12:38:36 +0200 | [diff] [blame] | 632 | .chk_rcv = stream_int_chk_rcv_conn, |
Willy Tarreau | de5722c | 2012-08-20 15:01:10 +0200 | [diff] [blame] | 633 | .chk_snd = stream_int_chk_snd_conn, |
Willy Tarreau | b277d6e | 2012-05-11 16:59:14 +0200 | [diff] [blame] | 634 | .read = sock_raw_read, |
Willy Tarreau | eecf6ca | 2012-08-20 15:09:53 +0200 | [diff] [blame] | 635 | .write = si_conn_send_cb, |
Willy Tarreau | fae4499 | 2012-08-20 14:02:10 +0200 | [diff] [blame] | 636 | .snd_buf = sock_raw_write_loop, |
Willy Tarreau | 1fe6bc3 | 2012-08-20 20:27:59 +0200 | [diff] [blame^] | 637 | .rcv_buf = raw_sock_to_buf, |
Willy Tarreau | 2420827 | 2012-05-21 17:28:50 +0200 | [diff] [blame] | 638 | .close = NULL, |
Willy Tarreau | 5c979a9 | 2012-05-07 17:15:39 +0200 | [diff] [blame] | 639 | }; |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 640 | |
| 641 | /* |
| 642 | * Local variables: |
| 643 | * c-indent-level: 8 |
| 644 | * c-basic-offset: 8 |
| 645 | * End: |
| 646 | */ |