login causing udp sento() calls to block

[jeroen]

hello,

we have an application that MUST send an udp packet every 10ms. so we use a realtime kernel patch and a high prio thread. all well until someone tries to login using either ssh (on a different network adapter) or even the local console. then we experience hick-ups. reason is the sendto() call blocking for more than 10 ms.

does anyone have a clue what can be going on during login? or even entering a user name. or logging out. I'm confused :(

[G309216C]

Hi,

We would love you to help you but first provide the UDP send thread or some of the code you suspect might be the cause of this "bug".

Thanks

[jeroen]

don't think it's a bug - just something going on that I do not yet understand ...


// start thread
pthread_t my_thread;
pthread_create(&my_thread, NULL, &thread_main, NULL);


// thread main sets own prio
void* thread_main(void* ptr)
{
struct sched_param param;
param.sched_priority = 70;
pthread_setschedparam(pthread_self(), SCHED_FIFO, &param);

// open socket
socket_fd = socket(AF_INET, SOCK_DGRAM, 0);
int reuse_addr = 1;
setsockopt(socket_fd, SOL_SOCKET, SO_REUSEADDR, (void*)&reuse_addr, sizeof(reuse_addr));
struct sockaddr_in source;
source.sin_family = AF_INET;
source.sin_port = htons(port);
source.sin_addr.s_addr = inet_addr(source_address);
bind(socket_fd, (struct sockaddr*)&source, sizeof(source));
unsigned char time_to_live = 1;
socket_error = setsockopt(socket_fd, IPPROTO_IP, IP_MULTICAST_TTL, &time_to_live, sizeof(time_to_live));


// send messages
struct timespec t;
clock_gettime(CLOCK_MONOTONIC, &t);
while(true)
{
// send message - this call may block for more than 10ms
int bytes_sent = sendto(socket_fd, buffer, buffer_size, 0, (struct sockaddr*)&destination, sizeof(destination));

// sleep until next message must be sent
t.tv_nsec += INTERVAL; // 10 ms
while (t.tv_nsec >= NSEC_PER_SEC)
{
t.tv_nsec -= NSEC_PER_SEC;
t.tv_sec++;
}
clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME, &t, NULL);
}

[kbw]

Are you running on a VM?

[jeroen]

it seems selinux was my foe ... now it's switched off my application is no longer bothered by logins. even a SCHED_FIFO thread at prio 99 was affected (same calculations took more time without getting nonvoluntary context switches) while it selinux was still on ...

[kbw]

Thanks for the update.