[haiku-development] Re: graphics server
- From: Joseph Groover <looncraz@xxxxxxxxxxx>
- To: haiku-development@xxxxxxxxxxxxx
- Date: Thu, 08 Dec 2011 22:44:08 -0600
In any case, I am with Clemens on this one. There shouldn't be a
separate process. At some point, you have to deal with a crashing
component and have to reconnect the clients to the server. It makes no
difference what you reconnect to what, so you may as well implement
reconnecting applications to the app_server (and registrar while
you're at it). I don't even think it would be very hard.
My idea was to implement a kernel-served server application type with
its own resource management policies. The idea would be that critical
servers would have a dedicated kernel connection with which critical
skeleton data can be stored by the server so critical servers could be
restarted at will in addition to tracking any kernel-side aborts that
may need to be made (such as interrupts, memory allocations, etc...).
Critical system server failures (app_server and input_server mostly)
would be caught and the server given a re-launch to attempt rescue, then
be launched in a fail-safe mode to attempt rescue. A server could also
implement a special-purpose recovery mode, so that the, possibly
corrupt, data structures and connections could be manually re-built -
but that would be limited to the app_server, really - in the future...
When a new graphics driver is installed, the app_server could quickly
suspend connections, the kernel would suspend all connected threads
correctly, and the app_server would shut itself down. The kernel would
then re-launch the app_server in a controlled-manner, keeping the data
structures as read-only until the app_server gives a server resume
message to signify successful launch. If it fails, the graphics driver
could then possibly be rolled-back to the previous version with an extra
chance to resurrect the server state before falling back to fail-safe.
I'd rather the kernel take up the task rather than having a watcher
daemon, just seems cleaner...
Time to resume the old "uptime wars!"
--The loon
Other related posts:
