Already posted in https://github.com/grpc/grpc/issues/21471 but I will post 
it here as well:
===================================================================

*Note: This is NOT a proper fix but rather a quick&dirty workaround to 
allow process forking after grpc_init() has been called. Hopefully there 
will be a proper fix soon.*


The current problem with GRPC as described in many places (e.g. #17695 
<https://github.com/grpc/grpc/issues/17695>, #13412 
<https://github.com/grpc/grpc/issues/13412>, #11814 
<https://github.com/grpc/grpc/pull/11814>) is that once grpc_init() has 
been called in a process, then this process should not be forked. There is 
a somewhat partial forking support but only if no connection has been 
established yet.


In my case I have an extension for PHP which is running under Apache. When 
Apache runs in prefork mode, it dynamically forks child processes to handle 
the incoming load. This means I need to be able to initialize a new GRPC 
instance in forked processes even when the parent process already has an 
active GRPC connection. Since this use case is not yet supported I made a 
workaround.


Essentially what this workaround does is to reset the internal GRPC state 
just enough so that a new and clean GRPC instance can be created in a 
forked child process.


My code change is based on the 1.24.3 release and can be found here 
<https://github.com/librato/grpc/compare/v1.24.3...librato:v1.24.3-fork-support>
.


There are several functions that are called internally to collect active 
sockets, pthread_once states, mutexes, and conditional variables:

void grpc_add_socket_fd(int fd);
void grpc_add_once_init(pthread_once_t *once);
void grpc_add_mutex(pthread_mutex_t *mutex);
void grpc_remove_mutex(pthread_mutex_t *mutex);
void grpc_add_cond(pthread_cond_t *cond);
void grpc_remove_cond(pthread_cond_t *cond);


Then there are new globals which hold a pointer to critical internal GRPC 
states:

void **fork_g_event_engine;
void **fork_g_resolver_registry_state;
...

------------------------------


>From the user perspective, all that is needed is to call the newly added 
function grpc_clean_after_fork() before initializing GRPC. Be sure to call 
it ONLY ONCE per process.


grpc_clean_after_fork() closes all inherited sockets, resets mutexes, 
conditional variables, pthread_once states, and other internal states that 
are left-overs from the parent process. Afterwards GRPC will initialize 
properly in every forked child.


I'm just sharing my findings here in case somebody else is in a similar 
situation and wants to give it a try. However, while this workaround works 
for me, it might not work in other cases.


-- 
You received this message because you are subscribed to the Google Groups 
"grpc.io" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
To view this discussion on the web visit 
https://groups.google.com/d/msgid/grpc-io/b3630784-8108-4f00-8124-70b1868c9718%40googlegroups.com.

Reply via email to