Re: holding locks while kalloc?

[Roland McGrath]

> I noticed that in gnumach/i386/i386/iopb.c, i386_io_port_add is careful
> to release all locks while running kalloc() to allocate some memory for
> the I/O bitmap, and goes back to retry via goto.
> 
> Is this in general a good strategy?  What is the danger that purposedly
> is avoided here?  Do I need to do a similar thing when locking the
> task I/O bitmap in the new code?

You never want to hold a spin lock (simple_lock) for any length of time
that is not strictly bounded and small.  The easiest way to be sure of the
bounds is never to hold a spin lock while calling other code.  The danger
is that one processor will spin for a long time while the other processor
does this more complex processing (e.g. calling kalloc).