[uml-devel] irq handling deadlock since 2.4.2

[Lennert B. <bu...@gn...>]

Hi,

as SIGIO doesn't seem to be disabled during the SIGIO handler, it looks like
there is a potential deadlock since 2.4.2 (as the 2.4.2 SIGIO handler keeps
looping until all active descriptors have had all their input handled).

- sigio_handler calls do_IRQ to handle input on a certain fd.
- During the handling of that IRQ, another event on the same file descriptor
  comes in, so do_IRQ or one of its children is preempted by sigio_handler.
- sigio_handler will call do_IRQ on the active fd, but do_IRQ will return
  immediately because the irq is marked INPROGRESS. sigio_handler will notice
  that the fd is still active, so will keep calling do_IRQ ad infinitum.

Included is a patch which makes do_IRQ return 0 if the IRQ wasn't handled
because it was in progress or disabled, and makes sigio_handler respect
do_IRQ's return code. It seems to solve the deadlock for me rather nicely,
as my test uml* is still up after about an hour of ping flooding, whereas it
never took me more than about 10 seconds to hang it with vanilla 2.4.2.


cheers,
Lennert



* Hostname dev0.math.leidenuniv.nl. It has an account 'guest' with password
  'guest'. Feel free to check it out and give it a beating, but please don't
  abuse it.


--- arch/um/kernel/irq_user.c.orig	Sat Mar  3 18:33:57 2001
+++ arch/um/kernel/irq_user.c	Sat Mar  3 18:33:37 2001
@@ -71,9 +71,11 @@
 			next = irq_fd->next;
 			if(irq_fd->active){
 				irq_fd->active = 0;
-				do_IRQ(irq_fd->irq, user_mode);
+				if (!do_IRQ(irq_fd->irq, user_mode))
+					count--;
 			}
 		}
+		if(count == 0) break;
 	}
 	if(user_mode){
 		interrupt_end();
--- arch/um/kernel/irq.c.orig	Sat Mar  3 18:34:04 2001
+++ arch/um/kernel/irq.c	Sat Mar  3 18:34:24 2001
@@ -328,7 +328,11 @@
 
 	if (softirq_active(cpu) & softirq_mask(cpu))
 		do_softirq();
-	return 1;
+
+	if (action != NULL)
+		return 1;
+
+	return 0;
 }
 
 /**