[Barrelfish-users] Fwd: Interrupt Handlers

[Timothy Roscoe]

Piyus,

It sounds like you're working within a Unix model of per-process kernel 
stacks, call-and-return patterns for exceptions, and kernel preemption.

The Barrelfish kernel is different: it is more of a set of short event 
handlers which run with interrupts disabled, do not generate exceptions, 
do not return (the stack is discarded, not necessarily unwound) and 
instead jump at a continuation when they're done.   They execute 
serially.  The kernel cannot take a page fault, as it never accesses 
virtual memory that might be paged out, and cannot be handed a user 
reference to non-existent memory.

When the kernel is entered from user mode (whether due to an interrupt 
or a user trap), the userspace execution context (including the PC) is 
saved in the dispatcher (think: process) control block.  This PC value 
*is* (in some versions of Barrelfish, I think) inspected by the kernel 
to determine the dispatcher's state - within a particular range of 
addresses.

In the common case, this PC address is not returned to when the process 
is rescheduled; instead the kernel initiates an upcall to a different 
continuation in the dispatcher, and the thread can then be resumed in 
user space.   Consequently, it's pretty important that this value isn't 
changed by the kernel, but it is always the address of a userspace 
instruction.

It might help us to answer your question if we understood the 
assumptions (independent of OS design choices) that you want to verify 
hold in Barrelfish.

Of course, the source (and documentation) is always available from the 
web site...

Best regards,

  -- Mothy

On 13-08-01 01:46 PM, Piyus Kedia wrote:
> Hi,
>
> I am working on Kernel Binary Translator which pushes the translated PC
> instead of the original PC on the stack when an interrupt or exception
> occurs while executing in kernel mode. There could be some problem if
> the interrupt or exception handler uses the PC value pushed on the
> stack. Since interrupts are disabled in kernel mode this problem won't
> exist. But still some exceptions like page fault can occur in kernel
> mode. I just want to make sure that these handlers don't use the PC
> value from the stack or don't overwrite the return address on the stack.
>
> Thanks,
> Piyus
>
>
> On Fri, Aug 2, 2013 at 4:20 AM, Timothy Roscoe <troscoe at inf.ethz.ch
> <mailto:troscoe at inf.ethz.ch>> wrote:
>
>
>     As Andrew says, the Barrelfish CPU driver processes all exceptions
>     (whether hardware interrupts or traps) serially with interrupts
>     disabled, unless it's idling.
>
>     Note that (like K42 and Psyche) Barrelfish uses upcall dispatch most
>     of the time (think Scheduler Activations), so exit from an exception
>     handler is typically a jump to a completely different IP/PC value
>     from when that (or any other) exception occurred.
>
>     Note also that Barrelfish doesn't have any kernel threads - each CPU
>     driver has a single, static, dedicated kernel stack.
>
>     Could you say a little more about what you mean?
>
>       -- Mothy
>
>
>     On 13-08-01 05:32 AM, Piyus Kedia wrote:
>
>         Hi All,
>
>         I am working on dynamic binary translator. Please tell me if anybody
>         know if in Barrelfish interrupted kernel PC is ever used in
>         interrupt
>         handlers?
>
>         Thanks and regards,
>         Piyus
>
>
>
>
>         _________________________________________________
>         Barrelfish-users mailing list
>         Barrelfish-users at lists.inf.__ethz.ch
>         <mailto:Barrelfish-users at lists.inf.ethz.ch>
>         https://lists.inf.ethz.ch/__mailman/listinfo/barrelfish-__users
>         <https://lists.inf.ethz.ch/mailman/listinfo/barrelfish-users>
>
>
>