Work around race condition in libnss
This change prevents a segfault in ceph::crypto::init when using NSS and
calling rados_connect from multiple threads simultaneously on different
rados_t objects (and updates the documentation for rados_connect to
reflect the fix).
It's pretty simple, just one static mutex wrapping the
NSS definition of ceph::crypto::init. More details regarding the race
condition are in this[0] commit (and pull request #1424).
To reproduce the race condition in the existing codebase, the below[1]
C program will work (depending on number of cores and probably other
things, the number of threads needed to reliably reproduce varies, but
the more the better - in my environment five is sufficient, with four
cores.
[0]:
377c919088423bafcae92759f8d578438717705c
[1]:
```c
void *init(void *p) {
int err;
rados_t cluster;
err = rados_create(&cluster, NULL);
if (err < 0) {
fprintf(stderr, "cannot create cluster handle: %s\n", strerror(-err));
return NULL;
}
err = rados_conf_read_file(cluster, "./ceph.conf");
if (err < 0) {
fprintf(stderr, "error reading config file: %s\n", strerror(-err));
return NULL;
}
rados_connect(cluster);
return NULL;
}
int main() {
pthread_t ts[NTHREAD];
int i;
for (i = 0; i < NTHREAD; i++) {
pthread_create(&ts[i], NULL, init, NULL);
}
for (i = 0; i < NTHREAD; i++) {
int k;
void *p = (void*)&k;
pthread_join(ts[i], p);
}
return 0;
}
```
Signed-off-by: Sharif Olorin <sio@tesser.org>