Stop interrupt storm for fasteoi.
PATCH 1/1 Stop interrupt storm for fasteoi.
Ingo,
I've spent several days banging my head on this bug, and I finally found it. I originally thought we had a bug with the latency tracer, since it seemed to only occur when I turned on latency tracing. But I guess it just changed the timings to cause the bug to happen. Now that I found where the bug is, I don't know how it ever worked, even without the tracing.
When taking an Ethernet interrupt, it was handled by handle_fasteoi_irq.
handle_fasteoi_irq, when the irq is handled by a thread, sets the irq INPROGRESS and masks the irq.
Before leaving handle_fasteoi_irq, a call to desc->chip->eoi is called.
For the apic, this will call move_native_irq.
In the -rt kernel, move_native_irq masks the irq, moves it, and then blindly unmasks it. So when interrupts are turned on next, we take an interrupt storm.
The other handlers besides handle_fasteoi_irq mask the irq regardless of whether the irq is INPROGRESS. But handle_fasteoi_irq will not mask it, if the irq is already INPROGRESS, and just returns. So we keep taking the same interrupt.
So, I change move_native_irq to not mask and unmask if the irq is currently INPROGRESS. But... I'm not sure if this is ok or not, since I don't know all the uses to this.
My original patch was just to do a
+@@ -68,6 +68,9 @@ void move_native_irq(int irq)
if (unlikely(desc->status & IRQ_DISABLED))
return;
if (unlikely(desc->status & IRQ_INPROGRESS))
+ return;
+
desc->chip->mask(irq);
move_masked_irq(irq);
desc->chip->unmask(irq);But I thought that this might be too big of a hammer to this nail. So I changed it to the patch below.
kernel/irq/migration.c +12 -2
@@ -61,6 +61,7 @@ void move_masked_irq(int irq)void move_native_irq (int irq) {struct irq_desc *desc = irq_desc + irq; +int mask =1 ;if (likely (!(desc->status &IRQ_MOVE_PENDING )))return ;
@@ -68,8 +69,17 @@ void move_native_irq(int irq)if (unlikely (desc->status &IRQ_DISABLED ))return ; - desc->chip->mask (irq); +/* +* If the irq is already in progress, it should be masked. +* If we unmask it, we might cause an interrupt storm on RT. +*/ +if (unlikely (desc->status &IRQ_INPROGRESS )) + mask =0 ; + +if (mask) + desc->chip->mask (irq);move_masked_irq (irq); - desc->chip->unmask (irq); +if (mask) + desc->chip->unmask (irq); }