blob: ba1da1b168688f0a0bc9514cecef12ac02cb9ba6 [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 Tarreau4c7e4b72020-05-27 12:58:42 +020022#include <haproxy/api.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020023#include <haproxy/applet.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 Tarreaud2ad57c2020-06-03 19:43:35 +020026#include <haproxy/ring.h>
Willy Tarreau5e539c92020-06-04 20:45:39 +020027#include <haproxy/stream_interface.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020028#include <haproxy/thread.h>
Willy Tarreau172945f2019-08-08 15:28:52 +020029
Emeric Brune14b98c2021-01-12 14:21:00 +010030/* Initialize a pre-allocated ring with the buffer area
31 * of size */
32void ring_init(struct ring *ring, void *area, size_t size)
33{
34 HA_RWLOCK_INIT(&ring->lock);
35 LIST_INIT(&ring->waiters);
36 ring->readers_count = 0;
37 ring->ofs = 0;
38 ring->buf = b_make(area, size, 0, 0);
39 /* write the initial RC byte */
40 b_putchr(&ring->buf, 0);
41}
42
Willy Tarreau172945f2019-08-08 15:28:52 +020043/* Creates and returns a ring buffer of size <size> bytes. Returns NULL on
44 * allocation failure.
45 */
46struct ring *ring_new(size_t size)
47{
48 struct ring *ring = NULL;
49 void *area = NULL;
50
51 if (size < 2)
52 goto fail;
53
54 ring = malloc(sizeof(*ring));
55 if (!ring)
56 goto fail;
57
58 area = malloc(size);
59 if (!area)
60 goto fail;
61
Emeric Brune14b98c2021-01-12 14:21:00 +010062 ring_init(ring, area, size);
Willy Tarreau172945f2019-08-08 15:28:52 +020063 return ring;
64 fail:
65 free(area);
66 free(ring);
67 return NULL;
68}
69
70/* Resizes existing ring <ring> to <size> which must be larger, without losing
71 * its contents. The new size must be at least as large as the previous one or
72 * no change will be performed. The pointer to the ring is returned on success,
73 * or NULL on allocation failure. This will lock the ring for writes.
74 */
75struct ring *ring_resize(struct ring *ring, size_t size)
76{
77 void *area;
78
79 if (b_size(&ring->buf) >= size)
80 return ring;
81
82 area = malloc(size);
83 if (!area)
84 return NULL;
85
86 HA_RWLOCK_WRLOCK(LOGSRV_LOCK, &ring->lock);
87
88 /* recheck the buffer's size, it may have changed during the malloc */
89 if (b_size(&ring->buf) < size) {
90 /* copy old contents */
91 b_getblk(&ring->buf, area, ring->buf.data, 0);
92 area = HA_ATOMIC_XCHG(&ring->buf.area, area);
93 ring->buf.size = size;
94 ring->buf.head = 0;
95 }
96
97 HA_RWLOCK_WRUNLOCK(LOGSRV_LOCK, &ring->lock);
98
99 free(area);
100 return ring;
101}
102
103/* destroys and frees ring <ring> */
104void ring_free(struct ring *ring)
105{
106 if (!ring)
107 return;
108 free(ring->buf.area);
109 free(ring);
110}
111
Willy Tarreaube978532019-08-27 11:44:13 +0200112/* Tries to send <npfx> parts from <prefix> followed by <nmsg> parts from <msg>
113 * to ring <ring>. The message is sent atomically. It may be truncated to
114 * <maxlen> bytes if <maxlen> is non-null. There is no distinction between the
115 * two lists, it's just a convenience to help the caller prepend some prefixes
116 * when necessary. It takes the ring's write lock to make sure no other thread
117 * will touch the buffer during the update. Returns the number of bytes sent,
118 * or <=0 on failure.
119 */
120ssize_t ring_write(struct ring *ring, size_t maxlen, const struct ist pfx[], size_t npfx, const struct ist msg[], size_t nmsg)
121{
122 struct buffer *buf = &ring->buf;
Willy Tarreau1d181e42019-08-30 11:17:01 +0200123 struct appctx *appctx;
Willy Tarreaube978532019-08-27 11:44:13 +0200124 size_t totlen = 0;
125 size_t lenlen;
Willy Tarreau30362902019-08-30 15:06:10 +0200126 uint64_t dellen;
Willy Tarreaube978532019-08-27 11:44:13 +0200127 int dellenlen;
128 ssize_t sent = 0;
129 int i;
130
131 /* we have to find some room to add our message (the buffer is
132 * never empty and at least contains the previous counter) and
133 * to update both the buffer contents and heads at the same
134 * time (it's doable using atomic ops but not worth the
135 * trouble, let's just lock). For this we first need to know
136 * the total message's length. We cannot measure it while
137 * copying due to the varint encoding of the length.
138 */
139 for (i = 0; i < npfx; i++)
140 totlen += pfx[i].len;
141 for (i = 0; i < nmsg; i++)
142 totlen += msg[i].len;
143
144 if (totlen > maxlen)
145 totlen = maxlen;
146
147 lenlen = varint_bytes(totlen);
148
149 HA_RWLOCK_WRLOCK(LOGSRV_LOCK, &ring->lock);
150 if (lenlen + totlen + 1 + 1 > b_size(buf))
151 goto done_buf;
152
153 while (b_room(buf) < lenlen + totlen + 1) {
154 /* we need to delete the oldest message (from the end),
155 * and we have to stop if there's a reader stuck there.
156 * Unless there's corruption in the buffer it's guaranteed
157 * that we have enough data to find 1 counter byte, a
158 * varint-encoded length (1 byte min) and the message
159 * payload (0 bytes min).
160 */
161 if (*b_head(buf))
162 goto done_buf;
163 dellenlen = b_peek_varint(buf, 1, &dellen);
164 if (!dellenlen)
165 goto done_buf;
166 BUG_ON(b_data(buf) < 1 + dellenlen + dellen);
167
168 b_del(buf, 1 + dellenlen + dellen);
169 ring->ofs += 1 + dellenlen + dellen;
170 }
171
172 /* OK now we do have room */
173 __b_put_varint(buf, totlen);
174
175 totlen = 0;
176 for (i = 0; i < npfx; i++) {
177 size_t len = pfx[i].len;
178
179 if (len + totlen > maxlen)
180 len = maxlen - totlen;
181 if (len)
182 __b_putblk(buf, pfx[i].ptr, len);
183 totlen += len;
184 }
185
186 for (i = 0; i < nmsg; i++) {
187 size_t len = msg[i].len;
188
189 if (len + totlen > maxlen)
190 len = maxlen - totlen;
191 if (len)
192 __b_putblk(buf, msg[i].ptr, len);
193 totlen += len;
194 }
195
William Dauchy477757c2020-08-07 22:19:23 +0200196 *b_tail(buf) = 0; buf->data++; // new read counter
Willy Tarreaube978532019-08-27 11:44:13 +0200197 sent = lenlen + totlen + 1;
Willy Tarreau1d181e42019-08-30 11:17:01 +0200198
199 /* notify potential readers */
Willy Tarreau9597cbd2020-05-19 17:07:30 +0200200 list_for_each_entry(appctx, &ring->waiters, wait_entry)
Willy Tarreau1d181e42019-08-30 11:17:01 +0200201 appctx_wakeup(appctx);
202
Willy Tarreaube978532019-08-27 11:44:13 +0200203 done_buf:
204 HA_RWLOCK_WRUNLOCK(LOGSRV_LOCK, &ring->lock);
205 return sent;
206}
Willy Tarreau172945f2019-08-08 15:28:52 +0200207
Willy Tarreau928068a2020-05-19 19:14:42 +0200208/* Tries to attach appctx <appctx> as a new reader on ring <ring>. This is
209 * meant to be used by low level appctx code such as CLI or ring forwarding.
210 * For higher level functions, please see the relevant parts in appctx or CLI.
211 * It returns non-zero on success or zero on failure if too many users are
212 * already attached. On success, the caller MUST call ring_detach_appctx()
213 * to detach itself, even if it was never woken up.
214 */
Emeric Brundcd58af2020-05-28 14:39:30 +0200215int ring_attach(struct ring *ring)
Willy Tarreau928068a2020-05-19 19:14:42 +0200216{
217 int users = ring->readers_count;
218
219 do {
220 if (users >= 255)
221 return 0;
222 } while (!_HA_ATOMIC_CAS(&ring->readers_count, &users, users + 1));
223 return 1;
224}
225
226/* detach an appctx from a ring. The appctx is expected to be waiting at
227 * offset <ofs>. Nothing is done if <ring> is NULL.
228 */
229void ring_detach_appctx(struct ring *ring, struct appctx *appctx, size_t ofs)
230{
231 if (!ring)
232 return;
233
234 HA_RWLOCK_WRLOCK(LOGSRV_LOCK, &ring->lock);
235 if (ofs != ~0) {
236 /* reader was still attached */
237 ofs -= ring->ofs;
238 BUG_ON(ofs >= b_size(&ring->buf));
239 LIST_DEL_INIT(&appctx->wait_entry);
Willy Tarreau4781b152021-04-06 13:53:36 +0200240 HA_ATOMIC_DEC(b_peek(&ring->buf, ofs));
Willy Tarreau928068a2020-05-19 19:14:42 +0200241 }
Willy Tarreau4781b152021-04-06 13:53:36 +0200242 HA_ATOMIC_DEC(&ring->readers_count);
Willy Tarreau928068a2020-05-19 19:14:42 +0200243 HA_RWLOCK_WRUNLOCK(LOGSRV_LOCK, &ring->lock);
244}
245
Willy Tarreau072931c2019-08-27 11:55:39 +0200246/* Tries to attach CLI handler <appctx> as a new reader on ring <ring>. This is
247 * meant to be used when registering a CLI function to dump a buffer, so it
248 * returns zero on success, or non-zero on failure with a message in the appctx
Willy Tarreaufcf94982019-11-15 15:07:21 +0100249 * CLI context. It automatically sets the io_handler and io_release callbacks if
250 * they were not set.
Willy Tarreau072931c2019-08-27 11:55:39 +0200251 */
252int ring_attach_cli(struct ring *ring, struct appctx *appctx)
253{
Emeric Brundcd58af2020-05-28 14:39:30 +0200254 if (!ring_attach(ring))
Willy Tarreau928068a2020-05-19 19:14:42 +0200255 return cli_err(appctx,
256 "Sorry, too many watchers (255) on this ring buffer. "
257 "What could it have so interesting to attract so many watchers ?");
Willy Tarreau072931c2019-08-27 11:55:39 +0200258
Willy Tarreaufcf94982019-11-15 15:07:21 +0100259 if (!appctx->io_handler)
260 appctx->io_handler = cli_io_handler_show_ring;
261 if (!appctx->io_release)
262 appctx->io_release = cli_io_release_show_ring;
Willy Tarreau072931c2019-08-27 11:55:39 +0200263 appctx->ctx.cli.p0 = ring;
Willy Tarreau13696ff2019-08-30 10:16:14 +0200264 appctx->ctx.cli.o0 = ~0; // start from the oldest event
Willy Tarreau072931c2019-08-27 11:55:39 +0200265 return 0;
266}
267
268/* This function dumps all events from the ring whose pointer is in <p0> into
Willy Tarreau13696ff2019-08-30 10:16:14 +0200269 * the appctx's output buffer, and takes from <o0> the seek offset into the
Willy Tarreau1d181e42019-08-30 11:17:01 +0200270 * buffer's history (0 for oldest known event). It looks at <i0> for boolean
271 * options: bit0 means it must wait for new data or any key to be pressed. Bit1
272 * means it must seek directly to the end to wait for new contents. It returns
273 * 0 if the output buffer or events are missing is full and it needs to be
274 * called again, otherwise non-zero. It is meant to be used with
275 * cli_release_show_ring() to clean up.
Willy Tarreau072931c2019-08-27 11:55:39 +0200276 */
277int cli_io_handler_show_ring(struct appctx *appctx)
278{
279 struct stream_interface *si = appctx->owner;
280 struct ring *ring = appctx->ctx.cli.p0;
281 struct buffer *buf = &ring->buf;
Willy Tarreau13696ff2019-08-30 10:16:14 +0200282 size_t ofs = appctx->ctx.cli.o0;
Willy Tarreau6b267992022-08-04 17:00:21 +0200283 size_t last_ofs;
Willy Tarreau072931c2019-08-27 11:55:39 +0200284 uint64_t msg_len;
285 size_t len, cnt;
286 int ret;
287
288 if (unlikely(si_ic(si)->flags & (CF_WRITE_ERROR|CF_SHUTW)))
289 return 1;
290
Willy Tarreau223dded2020-05-19 19:21:45 +0200291 HA_RWLOCK_WRLOCK(LOGSRV_LOCK, &ring->lock);
Willy Tarreau9597cbd2020-05-19 17:07:30 +0200292 LIST_DEL_INIT(&appctx->wait_entry);
Willy Tarreau223dded2020-05-19 19:21:45 +0200293 HA_RWLOCK_WRUNLOCK(LOGSRV_LOCK, &ring->lock);
294
295 HA_RWLOCK_RDLOCK(LOGSRV_LOCK, &ring->lock);
Willy Tarreau1d181e42019-08-30 11:17:01 +0200296
Willy Tarreau072931c2019-08-27 11:55:39 +0200297 /* explanation for the initialization below: it would be better to do
298 * this in the parsing function but this would occasionally result in
299 * dropped events because we'd take a reference on the oldest message
300 * and keep it while being scheduled. Thus instead let's take it the
301 * first time we enter here so that we have a chance to pass many
Willy Tarreau13696ff2019-08-30 10:16:14 +0200302 * existing messages before grabbing a reference to a location. This
303 * value cannot be produced after initialization.
Willy Tarreau072931c2019-08-27 11:55:39 +0200304 */
Willy Tarreau13696ff2019-08-30 10:16:14 +0200305 if (unlikely(ofs == ~0)) {
Willy Tarreau1d181e42019-08-30 11:17:01 +0200306 ofs = 0;
307
308 /* going to the end means looking at tail-1 */
309 if (appctx->ctx.cli.i0 & 2)
310 ofs += b_data(buf) - 1;
311
Willy Tarreau4781b152021-04-06 13:53:36 +0200312 HA_ATOMIC_INC(b_peek(buf, ofs));
Willy Tarreau1d181e42019-08-30 11:17:01 +0200313 ofs += ring->ofs;
Willy Tarreau072931c2019-08-27 11:55:39 +0200314 }
315
316 /* we were already there, adjust the offset to be relative to
317 * the buffer's head and remove us from the counter.
318 */
319 ofs -= ring->ofs;
320 BUG_ON(ofs >= buf->size);
Willy Tarreau4781b152021-04-06 13:53:36 +0200321 HA_ATOMIC_DEC(b_peek(buf, ofs));
Willy Tarreau072931c2019-08-27 11:55:39 +0200322
323 /* in this loop, ofs always points to the counter byte that precedes
324 * the message so that we can take our reference there if we have to
325 * stop before the end (ret=0).
326 */
327 ret = 1;
328 while (ofs + 1 < b_data(buf)) {
329 cnt = 1;
330 len = b_peek_varint(buf, ofs + cnt, &msg_len);
331 if (!len)
332 break;
333 cnt += len;
334 BUG_ON(msg_len + ofs + cnt + 1 > b_data(buf));
335
336 if (unlikely(msg_len + 1 > b_size(&trash))) {
337 /* too large a message to ever fit, let's skip it */
338 ofs += cnt + msg_len;
339 continue;
340 }
341
342 chunk_reset(&trash);
343 len = b_getblk(buf, trash.area, msg_len, ofs + cnt);
344 trash.data += len;
345 trash.area[trash.data++] = '\n';
346
347 if (ci_putchk(si_ic(si), &trash) == -1) {
348 si_rx_room_blk(si);
349 ret = 0;
350 break;
351 }
352 ofs += cnt + msg_len;
353 }
354
Willy Tarreau4781b152021-04-06 13:53:36 +0200355 HA_ATOMIC_INC(b_peek(buf, ofs));
Willy Tarreau072931c2019-08-27 11:55:39 +0200356 ofs += ring->ofs;
Willy Tarreau6b267992022-08-04 17:00:21 +0200357 last_ofs = ring->ofs;
Willy Tarreau13696ff2019-08-30 10:16:14 +0200358 appctx->ctx.cli.o0 = ofs;
Willy Tarreau072931c2019-08-27 11:55:39 +0200359 HA_RWLOCK_RDUNLOCK(LOGSRV_LOCK, &ring->lock);
Willy Tarreau1d181e42019-08-30 11:17:01 +0200360
361 if (ret && (appctx->ctx.cli.i0 & 1)) {
362 /* we've drained everything and are configured to wait for more
363 * data or an event (keypress, close)
364 */
365 if (!si_oc(si)->output && !(si_oc(si)->flags & CF_SHUTW)) {
366 /* let's be woken up once new data arrive */
Willy Tarreau223dded2020-05-19 19:21:45 +0200367 HA_RWLOCK_WRLOCK(LOGSRV_LOCK, &ring->lock);
Willy Tarreau2b718102021-04-21 07:32:39 +0200368 LIST_APPEND(&ring->waiters, &appctx->wait_entry);
Willy Tarreau6b267992022-08-04 17:00:21 +0200369 ofs = ring->ofs;
Willy Tarreau223dded2020-05-19 19:21:45 +0200370 HA_RWLOCK_WRUNLOCK(LOGSRV_LOCK, &ring->lock);
Willy Tarreau6b267992022-08-04 17:00:21 +0200371 if (ofs != last_ofs) {
372 /* more data was added into the ring between the
373 * unlock and the lock, and the writer might not
374 * have seen us. We need to reschedule a read.
375 */
376 si_rx_endp_more(si);
377 } else
378 si_rx_endp_done(si);
Willy Tarreau1d181e42019-08-30 11:17:01 +0200379 ret = 0;
380 }
381 /* always drain all the request */
382 co_skip(si_oc(si), si_oc(si)->output);
383 }
Willy Tarreau072931c2019-08-27 11:55:39 +0200384 return ret;
385}
386
387/* must be called after cli_io_handler_show_ring() above */
388void cli_io_release_show_ring(struct appctx *appctx)
389{
390 struct ring *ring = appctx->ctx.cli.p0;
Willy Tarreau13696ff2019-08-30 10:16:14 +0200391 size_t ofs = appctx->ctx.cli.o0;
Willy Tarreau072931c2019-08-27 11:55:39 +0200392
Willy Tarreau928068a2020-05-19 19:14:42 +0200393 ring_detach_appctx(ring, appctx, ofs);
Willy Tarreau072931c2019-08-27 11:55:39 +0200394}
395
396
Willy Tarreau172945f2019-08-08 15:28:52 +0200397/*
398 * Local variables:
399 * c-indent-level: 8
400 * c-basic-offset: 8
401 * End:
402 */