Hi,
we are using inject_irq() from Martin Stein to inject interrupts into the normal world VM running Linux.
When measuring the time between issuing inject_irq() and actually arriving at the ISR in normal world (using CPU performance counters in CPU cycles granularity) we can see that this always requires at least 10 million cycles, which is 10ms at 1Ghz and exactly what is defined as cpu_fill_ms in base-hw/src/core/include/kernel/configuration.h.
If we set cpu_fill_ms to 1, we can measure that it only requires 1 million cycles. Apparently it is not possible to set values below 1 (some assertion fails on compile).
Is there another option, probably a more elegant one, that allows the acceleration of interrupt injection?
Best, Stefan.
Hi Stefan,
On 07/15/2016 04:43 PM, Stefan Brenner wrote:
Hi,
we are using inject_irq() from Martin Stein to inject interrupts into the normal world VM running Linux.
When measuring the time between issuing inject_irq() and actually arriving at the ISR in normal world (using CPU performance counters in CPU cycles granularity) we can see that this always requires at least 10 million cycles, which is 10ms at 1Ghz and exactly what is defined as cpu_fill_ms in base-hw/src/core/include/kernel/configuration.h.
If we set cpu_fill_ms to 1, we can measure that it only requires 1 million cycles. Apparently it is not possible to set values below 1 (some assertion fails on compile).
Is there another option, probably a more elegant one, that allows the acceleration of interrupt injection?
Normally, as long as no other component is runable and scheduled, you should immediately receive the interrupt after calling inject_irq(...). Are you sure that no other component is actively polling or otherwise consuming a lot of cpu time?
Regards Stefan
Best, Stefan.
As an addition to Stefans answer:
Be aware that calling vm->inject_irq(X) does not explicitely request the kernel to do an injection but merely marks the VM context as "Injection of interrupt X requested". The caller of vm->inject_irq(X) must not enter the kernel for that and so, the operation does not necessarily trigger any scheduling. The injection is not executed before you enter the kernel again and the VM is the next job in your schedule. Both depends on your individual Genode setup.
Cheers, Martin
Am 25.07.2016 um 16:13 schrieb Stefan Kalkowski:
Hi Stefan,
On 07/15/2016 04:43 PM, Stefan Brenner wrote:
Hi,
we are using inject_irq() from Martin Stein to inject interrupts into the normal world VM running Linux.
When measuring the time between issuing inject_irq() and actually arriving at the ISR in normal world (using CPU performance counters in CPU cycles granularity) we can see that this always requires at least 10 million cycles, which is 10ms at 1Ghz and exactly what is defined as cpu_fill_ms in base-hw/src/core/include/kernel/configuration.h.
If we set cpu_fill_ms to 1, we can measure that it only requires 1 million cycles. Apparently it is not possible to set values below 1 (some assertion fails on compile).
Is there another option, probably a more elegant one, that allows the acceleration of interrupt injection?
Normally, as long as no other component is runable and scheduled, you should immediately receive the interrupt after calling inject_irq(...). Are you sure that no other component is actively polling or otherwise consuming a lot of cpu time?
Regards Stefan
Best, Stefan.
-
Martin Stein -
Stefan Brenner -
Stefan Kalkowski