Hi all I'm trying to port the SBCL compiler to the Hurd but I've found a serious port leak which prevents me from completing the port.
So, the first thing I did was to compile CLISP in order to bootstrap SBCL.
CLISP compiled just fine and runs well as far as I know.
Next, I started the bootstrapping process to compile SBCL but it would
always fail with a kernel panic since Mach was unable to allocate more
slabs for the ipc_port data structure. I figured out that the kernel had
thousands of inactive ports (which had a single reference and one send-only
right). I wrote a very small script in lisp to reproduce the bug, which
simply loads a file several times, like this:
(loop for i from 0 to 100
do (load "hello-world"))
and the number of inactive ports in the kernel will increase (never
reclaimed), even when the CLISP process is killed. I've also attempted to
use a separate partition and the ports would still hang around after the
ext2fs translator was terminated.
After looking around and using KDB, I've figured out that the following
loop in kern/exceptions.c/exception_raise_continue_slow was the culprit:
while (mr == MACH_RCV_INTERRUPTED) {
/*
* Somebody is trying to force this thread
* to a clean point. We must cooperate
* and then resume the receive.
*/
while (thread_should_halt(self)) {
/* don't terminate while holding a reference */
if (self->ast & AST_TERMINATE)
ipc_port_release(reply_port);
thread_halt_self();
}
ip_lock(reply_port);
....
}
The reply port of the target thread (from CLISP?) is not being released
since it enters the while(thread_should_halt(self)) loop and the thread
terminates inside thread_halt_self but the previous condition does not
hold, which fails to release the port. I also think that once the code
enters thread_halt_self, it never comes back since the stack is discarded
(for both if cases inside the function).
I've changed the code to make sure the port is released when
thread_should_halt is true. So far, the kernel works great and I was
finally able to complete the second bootstrapping stage for SBCL. Please
see the attached patch and let me know what you think.
Regards
Flavio
port-leak.patch
Description: Binary data
