Limit irq nesting
PATCH 1/1 Limit irq nesting
Hi
During examination of stack-overflows on sparc64, it was found that there is no limit for a number of nested IRQ handlers. Sparc64 is especially stack-hungry architecture, minimum frame size is 192 bytes and after 75 frames it overflows.
If someone provides a legitimate reason for more than 2 nested handlers, you could increase the constant in the patch --- but there really should be some limit, so that many simultaneous interrupts can't blow the stack.
Mikulas
---
IRQs without IRQF_DISABLED could nest to arbitrary level.
At worst this would mean having as many IRQ handlers stack frames, as there are interrupts registered --- enough to cause a stack overflow.
This patch makes a limit to have at most two handlers on the stack.
--- include/linux/interrupt.h | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-)
include/linux/interrupt.h +6 -1
@@ -16,6 +16,11 @@#include <asm/system.h> /* +* Max number of interrupt handlers on a stack. To prevent stack overflow. +*/ +#define MAX_NESTED_INTERRUPTS 2 + +/* * These correspond to the IORESOURCE_IRQ_* defines in * linux/ioport.h to select the interrupt line behaviour. When * requesting an interrupt without specifying a IRQF_TRIGGER, the
@@ -95,7 +100,7 @@ extern void devm_free_irq(struct device#ifdef CONFIG_LOCKDEP # define local_irq_enable_in_hardirq () do { } while (0)#else -# define local_irq_enable_in_hardirq () local_irq_enable() +# define local_irq_enable_in_hardirq () do { if (hardirq_count() < (MAX_NESTED_INTERRUPTS << HARDIRQ_SHIFT)) local_irq_enable(); } while (0)#endif extern void disable_irq_nosync (unsigned int irq);
Hi Mikulas,
Quoted 10 lines
--- include/linux/interrupt.h | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-)Index:
linux-2.6.26-rc8/include/linux/interrupt.h
===================================================================include/linux/interrupt.h +0
-0@@ -16,6 +16,11 @@#include <asm/system.h>
include/linux/interrupt.h line 26
Quoted 10 lines
include/linux/interrupt.h +5
-0/* +* Max number of interrupt handlers on a stack. To prevent stack overflow. +*/ +#define MAX_NESTED_INTERRUPTS 2 + +/* * These correspond to the IORESOURCE_IRQ_* defines in * linux/ioport.h to select the interrupt line behaviour. When * requesting an interrupt without specifying a IRQF_TRIGGER, the
I have no objection to your approach.
but I don't know MAX_NESTED_INTERRUPTS is properly or not.
Could you explain why MAX_NESTED_INTERRUPTS is two? Is it architecture independent properly number?
Quoted 9 lines
include/linux/interrupt.h +1
-1@@ -95,7 +100,7 @@ extern void devm_free_irq(struct device#ifdef CONFIG_LOCKDEP # define local_irq_enable_in_hardirq () do { } while (0)#else -# define local_irq_enable_in_hardirq () local_irq_enable() +# define local_irq_enable_in_hardirq () do { if (hardirq_count() < (MAX_NESTED_INTERRUPTS << HARDIRQ_SHIFT)) local_irq_enable(); } while (0)#endif extern void disable_irq_nosync (unsigned int irq);
Quoted 26 lines · KOSAKI Motohiro
Hi Mikulas,
Quoted 10 lines
--- include/linux/interrupt.h | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-)Index:
linux-2.6.26-rc8/include/linux/interrupt.h
===================================================================include/linux/interrupt.h +0
-0@@ -16,6 +16,11 @@#include <asm/system.h> include/linux/interrupt.h line 26
Quoted 10 lines
include/linux/interrupt.h +5
-0/* +* Max number of interrupt handlers on a stack. To prevent stack overflow. +*/ +#define MAX_NESTED_INTERRUPTS 2 + +/* * These correspond to the IORESOURCE_IRQ_* defines in * linux/ioport.h to select the interrupt line behaviour. When * requesting an interrupt without specifying a IRQF_TRIGGER, the I have no objection to your approach.
but I don't knowMAX_NESTED_INTERRUPTSis properly or not.
Quoted 2 lines
Could you explain why
MAX_NESTED_INTERRUPTSis two? Is it architecture independent properly number?
I assume that there is at most one active bad-behaving device with long interrupt processing time. For example IDE disk in PIO mode.
If someone starts whining that he is simultaneously using two disks in PIO mode and is losing characters on serial line, we can make an architecture override for this (after we check that the architecture has enough stack space to hold more interrupts). But PIO mode is dead anyway, so there's no need to preemptively care about it.
Mikulas