active swapping using rumpdisk

[Michael Kelly <mike@weatherwax.co.uk>]

Swap activity in a rumpdisk based system soon results in a locked 
system. I think I've identified 1 part of the problem occurring within 
ipc_kmsg_copyin_body(). Using rumpdisk, vm_map_copyin() is being used to 
copy in data for messages rather than vm_map_copyin_page_list(). The 
system lock occurs when a page that is referenced in the message is 
swapped out before that message is processed by the receiving task. This 
wouldn't happen using page_list copy as those pages are kept in a busy 
state that prevents them from being paged out.

The reason that page lists are not used is that the rump disk port is 
not configured as an IKOT_DEVICE. The equivalent non-rumpdisk device 
port is set using:

(gnumach/linux/dev/glue/block.c)

  ipc_kobject_set (bd->port, (ipc_kobject_t) &bd->device, IKOT_DEVICE);

I cannot see a simple method of achieving the same in rumpdisk but to 
test a similar effect I 'hacked' ipc_kmsg_copyin_body() to use page 
lists for many of the messages intended for the rumpdisk task to 
pseudo-test the alteration:

        if (kmsg->ikm_header.msgh_id >= 2800 &&
            kmsg->ikm_header.msgh_id < 2900)
          use_page_lists = TRUE;

This results in many successful iterations of my stress test case 
whereas normally it locks very swiftly.

Is there a mechanism that can be used outside of the kernel to configure 
the rumpdisk ports as IKOT_DEVICE ports so that it can be tested properly?

Regards,

Mike.