blob: 437ee39290fcd87684933831210ad2e592a1aff6 [file] [log] [blame]
Willy Tarreau172945f2019-08-08 15:28:52 +02001/*
2 * Ring buffer management
3 *
4 * Copyright (C) 2000-2019 Willy Tarreau - w@1wt.eu
5 *
6 * This library is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation, version 2.1
9 * exclusively.
10 *
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
15 *
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with this library; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19 */
20
21#include <stdlib.h>
Willy Tarreau3f0f82e2020-06-04 19:42:41 +020022#include <haproxy/applet.h>
Willy Tarreau4c7e4b72020-05-27 12:58:42 +020023#include <haproxy/api.h>
Willy Tarreau8dabda72020-05-27 17:22:10 +020024#include <haproxy/buf.h>
Willy Tarreau83487a82020-06-04 20:19:54 +020025#include <haproxy/cli.h>
Willy Tarreau3f567e42020-05-28 15:29:19 +020026#include <haproxy/thread.h>
Willy Tarreaud2ad57c2020-06-03 19:43:35 +020027#include <haproxy/ring.h>
Willy Tarreau5e539c92020-06-04 20:45:39 +020028#include <haproxy/stream_interface.h>
Willy Tarreau172945f2019-08-08 15:28:52 +020029
30/* Creates and returns a ring buffer of size <size> bytes. Returns NULL on
31 * allocation failure.
32 */
33struct ring *ring_new(size_t size)
34{
35 struct ring *ring = NULL;
36 void *area = NULL;
37
38 if (size < 2)
39 goto fail;
40
41 ring = malloc(sizeof(*ring));
42 if (!ring)
43 goto fail;
44
45 area = malloc(size);
46 if (!area)
47 goto fail;
48
49 HA_RWLOCK_INIT(&ring->lock);
Willy Tarreau1d181e42019-08-30 11:17:01 +020050 LIST_INIT(&ring->waiters);
Willy Tarreau172945f2019-08-08 15:28:52 +020051 ring->readers_count = 0;
52 ring->ofs = 0;
53 ring->buf = b_make(area, size, 0, 0);
54 /* write the initial RC byte */
55 b_putchr(&ring->buf, 0);
56 return ring;
57 fail:
58 free(area);
59 free(ring);
60 return NULL;
61}
62
63/* Resizes existing ring <ring> to <size> which must be larger, without losing
64 * its contents. The new size must be at least as large as the previous one or
65 * no change will be performed. The pointer to the ring is returned on success,
66 * or NULL on allocation failure. This will lock the ring for writes.
67 */
68struct ring *ring_resize(struct ring *ring, size_t size)
69{
70 void *area;
71
72 if (b_size(&ring->buf) >= size)
73 return ring;
74
75 area = malloc(size);
76 if (!area)
77 return NULL;
78
79 HA_RWLOCK_WRLOCK(LOGSRV_LOCK, &ring->lock);
80
81 /* recheck the buffer's size, it may have changed during the malloc */
82 if (b_size(&ring->buf) < size) {
83 /* copy old contents */
84 b_getblk(&ring->buf, area, ring->buf.data, 0);
85 area = HA_ATOMIC_XCHG(&ring->buf.area, area);
86 ring->buf.size = size;
87 ring->buf.head = 0;
88 }
89
90 HA_RWLOCK_WRUNLOCK(LOGSRV_LOCK, &ring->lock);
91
92 free(area);
93 return ring;
94}
95
96/* destroys and frees ring <ring> */
97void ring_free(struct ring *ring)
98{
99 if (!ring)
100 return;
101 free(ring->buf.area);
102 free(ring);
103}
104
Willy Tarreaube978532019-08-27 11:44:13 +0200105/* Tries to send <npfx> parts from <prefix> followed by <nmsg> parts from <msg>
106 * to ring <ring>. The message is sent atomically. It may be truncated to
107 * <maxlen> bytes if <maxlen> is non-null. There is no distinction between the
108 * two lists, it's just a convenience to help the caller prepend some prefixes
109 * when necessary. It takes the ring's write lock to make sure no other thread
110 * will touch the buffer during the update. Returns the number of bytes sent,
111 * or <=0 on failure.
112 */
113ssize_t ring_write(struct ring *ring, size_t maxlen, const struct ist pfx[], size_t npfx, const struct ist msg[], size_t nmsg)
114{
115 struct buffer *buf = &ring->buf;
Willy Tarreau1d181e42019-08-30 11:17:01 +0200116 struct appctx *appctx;
Willy Tarreaube978532019-08-27 11:44:13 +0200117 size_t totlen = 0;
118 size_t lenlen;
Willy Tarreau30362902019-08-30 15:06:10 +0200119 uint64_t dellen;
Willy Tarreaube978532019-08-27 11:44:13 +0200120 int dellenlen;
121 ssize_t sent = 0;
122 int i;
123
124 /* we have to find some room to add our message (the buffer is
125 * never empty and at least contains the previous counter) and
126 * to update both the buffer contents and heads at the same
127 * time (it's doable using atomic ops but not worth the
128 * trouble, let's just lock). For this we first need to know
129 * the total message's length. We cannot measure it while
130 * copying due to the varint encoding of the length.
131 */
132 for (i = 0; i < npfx; i++)
133 totlen += pfx[i].len;
134 for (i = 0; i < nmsg; i++)
135 totlen += msg[i].len;
136
137 if (totlen > maxlen)
138 totlen = maxlen;
139
140 lenlen = varint_bytes(totlen);
141
142 HA_RWLOCK_WRLOCK(LOGSRV_LOCK, &ring->lock);
143 if (lenlen + totlen + 1 + 1 > b_size(buf))
144 goto done_buf;
145
146 while (b_room(buf) < lenlen + totlen + 1) {
147 /* we need to delete the oldest message (from the end),
148 * and we have to stop if there's a reader stuck there.
149 * Unless there's corruption in the buffer it's guaranteed
150 * that we have enough data to find 1 counter byte, a
151 * varint-encoded length (1 byte min) and the message
152 * payload (0 bytes min).
153 */
154 if (*b_head(buf))
155 goto done_buf;
156 dellenlen = b_peek_varint(buf, 1, &dellen);
157 if (!dellenlen)
158 goto done_buf;
159 BUG_ON(b_data(buf) < 1 + dellenlen + dellen);
160
161 b_del(buf, 1 + dellenlen + dellen);
162 ring->ofs += 1 + dellenlen + dellen;
163 }
164
165 /* OK now we do have room */
166 __b_put_varint(buf, totlen);
167
168 totlen = 0;
169 for (i = 0; i < npfx; i++) {
170 size_t len = pfx[i].len;
171
172 if (len + totlen > maxlen)
173 len = maxlen - totlen;
174 if (len)
175 __b_putblk(buf, pfx[i].ptr, len);
176 totlen += len;
177 }
178
179 for (i = 0; i < nmsg; i++) {
180 size_t len = msg[i].len;
181
182 if (len + totlen > maxlen)
183 len = maxlen - totlen;
184 if (len)
185 __b_putblk(buf, msg[i].ptr, len);
186 totlen += len;
187 }
188
189 *b_tail(buf) = 0; buf->data++;; // new read counter
190 sent = lenlen + totlen + 1;
Willy Tarreau1d181e42019-08-30 11:17:01 +0200191
192 /* notify potential readers */
Willy Tarreau9597cbd2020-05-19 17:07:30 +0200193 list_for_each_entry(appctx, &ring->waiters, wait_entry)
Willy Tarreau1d181e42019-08-30 11:17:01 +0200194 appctx_wakeup(appctx);
195
Willy Tarreaube978532019-08-27 11:44:13 +0200196 done_buf:
197 HA_RWLOCK_WRUNLOCK(LOGSRV_LOCK, &ring->lock);
198 return sent;
199}
Willy Tarreau172945f2019-08-08 15:28:52 +0200200
Willy Tarreau928068a2020-05-19 19:14:42 +0200201/* Tries to attach appctx <appctx> as a new reader on ring <ring>. This is
202 * meant to be used by low level appctx code such as CLI or ring forwarding.
203 * For higher level functions, please see the relevant parts in appctx or CLI.
204 * It returns non-zero on success or zero on failure if too many users are
205 * already attached. On success, the caller MUST call ring_detach_appctx()
206 * to detach itself, even if it was never woken up.
207 */
Emeric Brundcd58af2020-05-28 14:39:30 +0200208int ring_attach(struct ring *ring)
Willy Tarreau928068a2020-05-19 19:14:42 +0200209{
210 int users = ring->readers_count;
211
212 do {
213 if (users >= 255)
214 return 0;
215 } while (!_HA_ATOMIC_CAS(&ring->readers_count, &users, users + 1));
216 return 1;
217}
218
219/* detach an appctx from a ring. The appctx is expected to be waiting at
220 * offset <ofs>. Nothing is done if <ring> is NULL.
221 */
222void ring_detach_appctx(struct ring *ring, struct appctx *appctx, size_t ofs)
223{
224 if (!ring)
225 return;
226
227 HA_RWLOCK_WRLOCK(LOGSRV_LOCK, &ring->lock);
228 if (ofs != ~0) {
229 /* reader was still attached */
230 ofs -= ring->ofs;
231 BUG_ON(ofs >= b_size(&ring->buf));
232 LIST_DEL_INIT(&appctx->wait_entry);
233 HA_ATOMIC_SUB(b_peek(&ring->buf, ofs), 1);
234 }
235 HA_ATOMIC_SUB(&ring->readers_count, 1);
236 HA_RWLOCK_WRUNLOCK(LOGSRV_LOCK, &ring->lock);
237}
238
Willy Tarreau072931c2019-08-27 11:55:39 +0200239/* Tries to attach CLI handler <appctx> as a new reader on ring <ring>. This is
240 * meant to be used when registering a CLI function to dump a buffer, so it
241 * returns zero on success, or non-zero on failure with a message in the appctx
Willy Tarreaufcf94982019-11-15 15:07:21 +0100242 * CLI context. It automatically sets the io_handler and io_release callbacks if
243 * they were not set.
Willy Tarreau072931c2019-08-27 11:55:39 +0200244 */
245int ring_attach_cli(struct ring *ring, struct appctx *appctx)
246{
Emeric Brundcd58af2020-05-28 14:39:30 +0200247 if (!ring_attach(ring))
Willy Tarreau928068a2020-05-19 19:14:42 +0200248 return cli_err(appctx,
249 "Sorry, too many watchers (255) on this ring buffer. "
250 "What could it have so interesting to attract so many watchers ?");
Willy Tarreau072931c2019-08-27 11:55:39 +0200251
Willy Tarreaufcf94982019-11-15 15:07:21 +0100252 if (!appctx->io_handler)
253 appctx->io_handler = cli_io_handler_show_ring;
254 if (!appctx->io_release)
255 appctx->io_release = cli_io_release_show_ring;
Willy Tarreau072931c2019-08-27 11:55:39 +0200256 appctx->ctx.cli.p0 = ring;
Willy Tarreau13696ff2019-08-30 10:16:14 +0200257 appctx->ctx.cli.o0 = ~0; // start from the oldest event
Willy Tarreau072931c2019-08-27 11:55:39 +0200258 return 0;
259}
260
261/* This function dumps all events from the ring whose pointer is in <p0> into
Willy Tarreau13696ff2019-08-30 10:16:14 +0200262 * the appctx's output buffer, and takes from <o0> the seek offset into the
Willy Tarreau1d181e42019-08-30 11:17:01 +0200263 * buffer's history (0 for oldest known event). It looks at <i0> for boolean
264 * options: bit0 means it must wait for new data or any key to be pressed. Bit1
265 * means it must seek directly to the end to wait for new contents. It returns
266 * 0 if the output buffer or events are missing is full and it needs to be
267 * called again, otherwise non-zero. It is meant to be used with
268 * cli_release_show_ring() to clean up.
Willy Tarreau072931c2019-08-27 11:55:39 +0200269 */
270int cli_io_handler_show_ring(struct appctx *appctx)
271{
272 struct stream_interface *si = appctx->owner;
273 struct ring *ring = appctx->ctx.cli.p0;
274 struct buffer *buf = &ring->buf;
Willy Tarreau13696ff2019-08-30 10:16:14 +0200275 size_t ofs = appctx->ctx.cli.o0;
Willy Tarreau072931c2019-08-27 11:55:39 +0200276 uint64_t msg_len;
277 size_t len, cnt;
278 int ret;
279
280 if (unlikely(si_ic(si)->flags & (CF_WRITE_ERROR|CF_SHUTW)))
281 return 1;
282
Willy Tarreau223dded2020-05-19 19:21:45 +0200283 HA_RWLOCK_WRLOCK(LOGSRV_LOCK, &ring->lock);
Willy Tarreau9597cbd2020-05-19 17:07:30 +0200284 LIST_DEL_INIT(&appctx->wait_entry);
Willy Tarreau223dded2020-05-19 19:21:45 +0200285 HA_RWLOCK_WRUNLOCK(LOGSRV_LOCK, &ring->lock);
286
287 HA_RWLOCK_RDLOCK(LOGSRV_LOCK, &ring->lock);
Willy Tarreau1d181e42019-08-30 11:17:01 +0200288
Willy Tarreau072931c2019-08-27 11:55:39 +0200289 /* explanation for the initialization below: it would be better to do
290 * this in the parsing function but this would occasionally result in
291 * dropped events because we'd take a reference on the oldest message
292 * and keep it while being scheduled. Thus instead let's take it the
293 * first time we enter here so that we have a chance to pass many
Willy Tarreau13696ff2019-08-30 10:16:14 +0200294 * existing messages before grabbing a reference to a location. This
295 * value cannot be produced after initialization.
Willy Tarreau072931c2019-08-27 11:55:39 +0200296 */
Willy Tarreau13696ff2019-08-30 10:16:14 +0200297 if (unlikely(ofs == ~0)) {
Willy Tarreau1d181e42019-08-30 11:17:01 +0200298 ofs = 0;
299
300 /* going to the end means looking at tail-1 */
301 if (appctx->ctx.cli.i0 & 2)
302 ofs += b_data(buf) - 1;
303
304 HA_ATOMIC_ADD(b_peek(buf, ofs), 1);
305 ofs += ring->ofs;
Willy Tarreau072931c2019-08-27 11:55:39 +0200306 }
307
308 /* we were already there, adjust the offset to be relative to
309 * the buffer's head and remove us from the counter.
310 */
311 ofs -= ring->ofs;
312 BUG_ON(ofs >= buf->size);
313 HA_ATOMIC_SUB(b_peek(buf, ofs), 1);
314
315 /* in this loop, ofs always points to the counter byte that precedes
316 * the message so that we can take our reference there if we have to
317 * stop before the end (ret=0).
318 */
319 ret = 1;
320 while (ofs + 1 < b_data(buf)) {
321 cnt = 1;
322 len = b_peek_varint(buf, ofs + cnt, &msg_len);
323 if (!len)
324 break;
325 cnt += len;
326 BUG_ON(msg_len + ofs + cnt + 1 > b_data(buf));
327
328 if (unlikely(msg_len + 1 > b_size(&trash))) {
329 /* too large a message to ever fit, let's skip it */
330 ofs += cnt + msg_len;
331 continue;
332 }
333
334 chunk_reset(&trash);
335 len = b_getblk(buf, trash.area, msg_len, ofs + cnt);
336 trash.data += len;
337 trash.area[trash.data++] = '\n';
338
339 if (ci_putchk(si_ic(si), &trash) == -1) {
340 si_rx_room_blk(si);
341 ret = 0;
342 break;
343 }
344 ofs += cnt + msg_len;
345 }
346
347 HA_ATOMIC_ADD(b_peek(buf, ofs), 1);
348 ofs += ring->ofs;
Willy Tarreau13696ff2019-08-30 10:16:14 +0200349 appctx->ctx.cli.o0 = ofs;
Willy Tarreau072931c2019-08-27 11:55:39 +0200350 HA_RWLOCK_RDUNLOCK(LOGSRV_LOCK, &ring->lock);
Willy Tarreau1d181e42019-08-30 11:17:01 +0200351
352 if (ret && (appctx->ctx.cli.i0 & 1)) {
353 /* we've drained everything and are configured to wait for more
354 * data or an event (keypress, close)
355 */
356 if (!si_oc(si)->output && !(si_oc(si)->flags & CF_SHUTW)) {
357 /* let's be woken up once new data arrive */
Willy Tarreau223dded2020-05-19 19:21:45 +0200358 HA_RWLOCK_WRLOCK(LOGSRV_LOCK, &ring->lock);
Willy Tarreau9597cbd2020-05-19 17:07:30 +0200359 LIST_ADDQ(&ring->waiters, &appctx->wait_entry);
Willy Tarreau223dded2020-05-19 19:21:45 +0200360 HA_RWLOCK_WRUNLOCK(LOGSRV_LOCK, &ring->lock);
Willy Tarreau1d181e42019-08-30 11:17:01 +0200361 si_rx_endp_done(si);
362 ret = 0;
363 }
364 /* always drain all the request */
365 co_skip(si_oc(si), si_oc(si)->output);
366 }
Willy Tarreau072931c2019-08-27 11:55:39 +0200367 return ret;
368}
369
370/* must be called after cli_io_handler_show_ring() above */
371void cli_io_release_show_ring(struct appctx *appctx)
372{
373 struct ring *ring = appctx->ctx.cli.p0;
Willy Tarreau13696ff2019-08-30 10:16:14 +0200374 size_t ofs = appctx->ctx.cli.o0;
Willy Tarreau072931c2019-08-27 11:55:39 +0200375
Willy Tarreau928068a2020-05-19 19:14:42 +0200376 ring_detach_appctx(ring, appctx, ofs);
Willy Tarreau072931c2019-08-27 11:55:39 +0200377}
378
379
Willy Tarreau172945f2019-08-08 15:28:52 +0200380/*
381 * Local variables:
382 * c-indent-level: 8
383 * c-basic-offset: 8
384 * End:
385 */