RE: client_socket bogosity...
The create_connection hook has a fatal design flaw. create_conn is run before ap_update_vhost_given_ip(), which means that it is impossible to install input and output filters based on vhost info. More concisely, it is now impossible to install filters in place of CORE_IN and CORE_OUT based on vhost config. It was always impossible to do that. The patches you are talking about actually make it quite a bit cleaner for you to insert your filters above the CORE filters however, because now there is no way for somebody else to use the socket that was a part conn_rec. The other advantage to that patch is that if you want to insert filters based on where the connection came from, you can. Those patches were meant to allow different bottom-level filters based on the type of socket used, not based on vhost config. Ryan
RE: client_socket bogosity...
One Nov. 12, Ryan committed a patch creating the create_conn hook. The idea was to move the client_socket out of the conn_rec presumably to make available only to the core_in and core_out filters. However, I just found a backdoor... In core_create_conn() the socket is saved away thusly: ap_set_module_config(net-c-conn_config, core_module, csd); And whoever needs to access the socket does this: apr_socket_t *csd = ap_get_module_config(f-c-conn_config, core_module); That hack was added because the proxy does the completely wrong thing with regard to handing sockets. In order to finish the Nov. 12 patch, I need to rip a lot of logic out of the proxy and re-implement, which I haven't had time to do recently. The only other module that should use the get_module_config hack is the perchild module, which is also doing to completely wrong thing with regard to sockets, but I haven't had time to fix that one either. So the goal of hiding the socket is completely blown. The Nov. 11 change added a lot of complexity to the server (hard to read/understand code) in pursuit of a goal that is then immediately circumvented by the ap_get|set_module_config. So we made the server more complex for no reason. It actually isn't blown. Try writing a module that implements a non TCP socket, and it will work as long as you don't use the proxy or the perchild module. As proof, look at the fact that the Unix MPMs have been using that mechanism to handle the pipe_of_death. This allowed me to remove the ugly hacks at the beginning of the accept loop, which checked for the POD. Also, a big portion of the Nov 12 patch was to consolidate the accept functions for Unix and BeOS, which has meant far less duplicated code in the server. I am on the verge of vetoing the Nov. 12 patch in favor of moving the client_socket back into the con_rec. Comments? Please don't let two mis-behaved modules color your judgment on this. Both proxy and perchild must be re-written if they are going to be clean, and once that is done the stupid set_module_config can be removed. In fact, the server ran for over a day without the set_module_config, but that broke the proxy, so I added the hack to allow the proxy to continue to work, while I worked to solve the underlying problem. Unfortunately, work and some extracurricular activities have stopped me from contributing much code recently. Hopefully, I will have time to code again soon. Ryan
RE: client_socket bogosity...
Please don't let two mis-behaved modules color your judgment on this. Both proxy and perchild must be re-written if they are going to be clean, and once that is done the stupid set_module_config can be removed. In fact, the server ran for over a day without the set_module_config, but that broke the proxy, so I added the hack to allow the proxy to continue to work, while I worked to solve the underlying problem. mmm... I'd be interested to know what the solution is for net_time_filter since it is using the ap_get_module_config hack also. Allan
RE: client_socket bogosity...
Please don't let two mis-behaved modules color your judgment on this. Both proxy and perchild must be re-written if they are going to be clean, and once that is done the stupid set_module_config can be removed. In fact, the server ran for over a day without the set_module_config, but that broke the proxy, so I added the hack to allow the proxy to continue to work, while I worked to solve the underlying problem. mmm... I'd be interested to know what the solution is for net_time_filter since it is using the ap_get_module_config hack also. The real solution is to pass the core_net_rec structure to the NET_TIME filter as user-data for the filter, the same way we do for CORE_IN and CORE_OUT. Ryan
RE: client_socket bogosity...
The real solution is to pass the core_net_rec structure to the NET_TIME filter as user-data for the filter, the same way we do for CORE_IN and CORE_OUT. But you'll have to play some sort of trick to do that since the NET_TIME filter is added from the create_request hook which does not know about core_net_rec, only core_create_conn knows about that. I suppose some core specific magic could be played but that would prevent any other module from running it's own net_time hook. Allan
RE: client_socket bogosity...
prevent any other module from running it's own net_time hook. I meant net_time filter...
RE: client_socket bogosity...
The real solution is to pass the core_net_rec structure to the NET_TIME filter as user-data for the filter, the same way we do for CORE_IN and CORE_OUT. But you'll have to play some sort of trick to do that since the NET_TIME filter is added from the create_request hook which does not know about core_net_rec, only core_create_conn knows about that. I suppose some core specific magic could be played but that would prevent any other module from running it's own net_time hook. This is why I didn't do this originally, but it wouldn't be hard to add. As long as it's just the core that knows about the core_net_rec, we are okay. I just ran out of time while doing the implementation. Ryan
RE: client_socket bogosity...
On Fri, Jan 25, 2002 at 02:01:26PM -0800, Ryan Bloom wrote: Please don't let two mis-behaved modules color your judgment on this. Both proxy and perchild must be re-written if they are going to be clean, and once that is done the stupid set_module_config can be removed. In fact, the server ran for over a day without the set_module_config, but that broke the proxy, so I added the hack to allow the proxy to continue to work, while I worked to solve the underlying problem. mmm... I'd be interested to know what the solution is for net_time_filter since it is using the ap_get_module_config hack also. The real solution is to pass the core_net_rec structure to the NET_TIME filter as user-data for the filter, the same way we do for CORE_IN and CORE_OUT. You're kidding, right? The core_net_rec is private to the core. You can't just start passing that around the server. That just breaks the abstraction that you've worked so hard to implement. No... something else needs to be figured out [for the net_time_filter]. But the net_time_filter is implemented by the core. If we really want to abstract this more, we can create a function that, given a conn_rec can set a timeout on the socket. I think I see how that can be done. Or, we can actually create a net_module, that has a spot in the c-conf_vector that is the net_rec. That module can implement the logic to set a timeout on a connection. Which is really what we want, to be able to set a timeout on a connection, not on a socket. Ryan
Re: client_socket bogosity...
The create_connection hook has a fatal design flaw. create_conn is run before ap_update_vhost_given_ip(), which means that it is impossible to install input and output filters based on vhost info. I want to install SSL_IN and SSL_OUT filters if the request is coming in to a vhost/port enabled for SSL and that can't be done with the create_connection hook. Bill One Nov. 12, Ryan committed a patch creating the create_conn hook. The idea was to move the client_socket out of the conn_rec presumably to make available only to the core_in and core_out filters. However, I just found a backdoor... In core_create_conn() the socket is saved away thusly: ap_set_module_config(net-c-conn_config, core_module, csd); And whoever needs to access the socket does this: apr_socket_t *csd = ap_get_module_config(f-c-conn_config, core_module); That hack was added because the proxy does the completely wrong thing with regard to handing sockets. In order to finish the Nov. 12 patch, I need to rip a lot of logic out of the proxy and re-implement, which I haven't had time to do recently. The only other module that should use the get_module_config hack is the perchild module, which is also doing to completely wrong thing with regard to sockets, but I haven't had time to fix that one either. So the goal of hiding the socket is completely blown. The Nov. 11 change added a lot of complexity to the server (hard to read/understand code) in pursuit of a goal that is then immediately circumvented by the ap_get|set_module_config. So we made the server more complex for no reason. It actually isn't blown. Try writing a module that implements a non TCP socket, and it will work as long as you don't use the proxy or the perchild module. As proof, look at the fact that the Unix MPMs have been using that mechanism to handle the pipe_of_death. This allowed me to remove the ugly hacks at the beginning of the accept loop, which checked for the POD. Also, a big portion of the Nov 12 patch was to consolidate the accept functions for Unix and BeOS, which has meant far less duplicated code in the server. I am on the verge of vetoing the Nov. 12 patch in favor of moving the client_socket back into the con_rec. Comments? Please don't let two mis-behaved modules color your judgment on this. Both proxy and perchild must be re-written if they are going to be clean, and once that is done the stupid set_module_config can be removed. In fact, the server ran for over a day without the set_module_config, but that broke the proxy, so I added the hack to allow the proxy to continue to work, while I worked to solve the underlying problem. Unfortunately, work and some extracurricular activities have stopped me from contributing much code recently. Hopefully, I will have time to code again soon. Ryan
Re: client_socket bogosity...
Bill Stoddard wrote: The create_connection hook has a fatal design flaw. create_conn is run before ap_update_vhost_given_ip(), which means that it is impossible to install input and output filters based on vhost info. I want to install SSL_IN and SSL_OUT filters if the request is coming in to a vhost/port enabled for SSL and that can't be done with the create_connection hook. Bill On that point. I don't think there is any way of inserting a proxy specifc filter either, as their is now way for the hook to know what kind of request is the connection is for. One Nov. 12, Ryan committed a patch creating the create_conn hook. The idea was to move the client_socket out of the conn_rec presumably to make available only to the core_in and core_out filters. However, I just found a backdoor... In core_create_conn() the socket is saved away thusly: ap_set_module_config(net-c-conn_config, core_module, csd); And whoever needs to access the socket does this: apr_socket_t *csd = ap_get_module_config(f-c-conn_config, core_module); That hack was added because the proxy does the completely wrong thing with regard to handing sockets. In order to finish the Nov. 12 patch, I need to rip a lot of logic out of the proxy and re-implement, which I haven't had time to do recently. The only other module that should use the get_module_config hack is the perchild module, which is also doing to completely wrong thing with regard to sockets, but I haven't had time to fix that one either. So the goal of hiding the socket is completely blown. The Nov. 11 change added a lot of complexity to the server (hard to read/understand code) in pursuit of a goal that is then immediately circumvented by the ap_get|set_module_config. So we made the server more complex for no reason. It actually isn't blown. Try writing a module that implements a non TCP socket, and it will work as long as you don't use the proxy or the perchild module. As proof, look at the fact that the Unix MPMs have been using that mechanism to handle the pipe_of_death. This allowed me to remove the ugly hacks at the beginning of the accept loop, which checked for the POD. Also, a big portion of the Nov 12 patch was to consolidate the accept functions for Unix and BeOS, which has meant far less duplicated code in the server. I am on the verge of vetoing the Nov. 12 patch in favor of moving the client_socket back into the con_rec. Comments? Please don't let two mis-behaved modules color your judgment on this. Both proxy and perchild must be re-written if they are going to be clean, and once that is done the stupid set_module_config can be removed. In fact, the server ran for over a day without the set_module_config, but that broke the proxy, so I added the hack to allow the proxy to continue to work, while I worked to solve the underlying problem. Unfortunately, work and some extracurricular activities have stopped me from contributing much code recently. Hopefully, I will have time to code again soon. Ryan
Re: client_socket bogosity...
The create_connection hook has a fatal design flaw. create_conn is run before ap_update_vhost_given_ip(), which means that it is impossible to install input and output filters based on vhost info. More concisely, it is now impossible to install filters in place of CORE_IN and CORE_OUT based on vhost config. Bill
