Fix critical bug in default TCP keepalive user timeout
It turns out that our calculation of the TCP_USER_TIMEOUT value was
flawed in several ways:
* there should have been parenthesis around the + operator
(line->keepalive_probe_interval + line->keepalive_idle_timeout) as the
keepalive_idle_timeout is in seconds, not milli-seconds.
* in the default case, all those values are configured to -1
(E1INP_USE_DEFAULT). This means we're using
1000 * -1 * -1 + -1 = 999
i.e. just below a second which clearly is not enough for a lossy
satellite or wifi back-haul.
This fixes a regression introduced in Ia7659c209aea0d26eb37d31e771adc91b17ae668
(libosmo-abis >= 1.4.0) when TCP keepalive user timeouts became enabled
by default.
The initial support for TCP_USER_TIMEOUT was merged in
I5e7425958472aa5d758e09bfbefc7d7d37bf6f5f (libosmo-abis >= 0.7.0) but
since TCP keepalives were not yet enabled by default, only users with
explicit TCP keepalive configuration in their config files would be
affected - and then only of the second part of the bug (operator
precedence).
In addition, let's print the actually-used values to the log, helping to
spot unintended values.
Change-Id: Idca24d3e676a45d860d9eec60dc2097d8d87f3bf
Closes: OS#6375
Related: OS#5785, SYS#6801
Fixes: Ia7659c209aea0d26eb37d31e771adc91b17ae668
diff --git a/src/input/ipaccess.c b/src/input/ipaccess.c
index 75d9693..22ac0ef 100644
--- a/src/input/ipaccess.c
+++ b/src/input/ipaccess.c
@@ -588,7 +588,7 @@
static void update_fd_settings(struct e1inp_line *line, int fd)
{
int ret;
- int val;
+ int val, idle_val, interval_val, retry_count_val, user_timeout_val;
if (line->keepalive_num_probes) {
/* Enable TCP keepalive to find out if the connection is gone */
@@ -599,33 +599,33 @@
else
LOGPIL(line, DLINP, LOGL_NOTICE, "TCP Keepalive is enabled\n");
+ idle_val = line->keepalive_idle_timeout > 0 ?
+ line->keepalive_idle_timeout :
+ DEFAULT_TCP_KEEPALIVE_IDLE_TIMEOUT;
+ interval_val = line->keepalive_probe_interval > -1 ?
+ line->keepalive_probe_interval :
+ DEFAULT_TCP_KEEPALIVE_INTERVAL;
+ retry_count_val = line->keepalive_num_probes > 0 ?
+ line->keepalive_num_probes :
+ DEFAULT_TCP_KEEPALIVE_RETRY_COUNT;
+ user_timeout_val = 1000 * retry_count_val * (interval_val + idle_val);
+ LOGPIL(line, DLINP, LOGL_NOTICE, "TCP keepalive idle_timeout=%us, interval=%us, retry_count=%u "
+ "user_timeout=%ums\n", idle_val, interval_val, retry_count_val, user_timeout_val);
/* The following options are not portable! */
- val = line->keepalive_idle_timeout > 0 ?
- line->keepalive_idle_timeout :
- DEFAULT_TCP_KEEPALIVE_IDLE_TIMEOUT;
- ret = setsockopt(fd, IPPROTO_TCP, TCP_KEEPIDLE, &val, sizeof(val));
+ ret = setsockopt(fd, IPPROTO_TCP, TCP_KEEPIDLE, &idle_val, sizeof(idle_val));
if (ret < 0) {
LOGPIL(line, DLINP, LOGL_ERROR, "Failed to set TCP keepalive idle time: %s\n",
strerror(errno));
}
- val = line->keepalive_probe_interval > -1 ?
- line->keepalive_probe_interval :
- DEFAULT_TCP_KEEPALIVE_INTERVAL;
- ret = setsockopt(fd, IPPROTO_TCP, TCP_KEEPINTVL, &val, sizeof(val));
+ ret = setsockopt(fd, IPPROTO_TCP, TCP_KEEPINTVL, &interval_val, sizeof(interval_val));
if (ret < 0) {
LOGPIL(line, DLINP, LOGL_ERROR, "Failed to set TCP keepalive interval: %s\n",
strerror(errno));
}
- val = line->keepalive_num_probes > 0 ?
- line->keepalive_num_probes :
- DEFAULT_TCP_KEEPALIVE_RETRY_COUNT;
- ret = setsockopt(fd, IPPROTO_TCP, TCP_KEEPCNT, &val, sizeof(val));
+ ret = setsockopt(fd, IPPROTO_TCP, TCP_KEEPCNT, &retry_count_val, sizeof(retry_count_val));
if (ret < 0)
LOGPIL(line, DLINP, LOGL_ERROR, "Failed to set TCP keepalive count: %s\n", strerror(errno));
- val = 1000 * line->keepalive_num_probes *
- line->keepalive_probe_interval +
- line->keepalive_idle_timeout;
- ret = setsockopt(fd, IPPROTO_TCP, TCP_USER_TIMEOUT, &val, sizeof(val));
+ ret = setsockopt(fd, IPPROTO_TCP, TCP_USER_TIMEOUT, &user_timeout_val, sizeof(user_timeout_val));
if (ret < 0)
LOGPIL(line, DLINP, LOGL_ERROR, "Failed to set TCP user timeout: %s\n", strerror(errno));
}