Message-ID: <CAH3Oq6Q-XgNvTB9CFdD2bj+rbUBUsUV2C6drx-JsRYziNAJ=vQ@mail.gmail.com>
Date:	Mon, 4 Jun 2012 13:17:11 +0800
From:	Ning Jiang <ning.n.jiang@...il.com>
To:	Thomas Gleixner <tglx@...utronix.de>
Cc:	linux-kernel@...r.kernel.org
Subject: Re: Make kstat_incr_irqs_this_cpu() called when an irq is actually
 being handled

2012/6/1 Ning Jiang <ning.n.jiang@...il.com>:
> 2012/6/1 Thomas Gleixner <tglx@...utronix.de>:
>> On Thu, 31 May 2012, Ning Jiang wrote:
>>> kstat_incr_irqs_this_cpu() is supposed to record the irq number
>>> that is actually taken by a certain cpu. But it failed to do so
>>> in its current form.
>>
>> It exactly counts the number of interrupts which are delivered to and
>> handled by a CPU. That includes interrupts which are delivered in a
>> state where the device handler cannot be called. And this is on
>> purpose. We want to see the real number of interruptions not the
>> number of calls to a device handler.
>>
>>> For level irq, if its disabled or no action available, we'll skip
>>> irq handling. When it is enabled later, the irq will be retriggered
>>> and kstat_incr_irqs_this_cpu() will be counted twice for this
>>> single irq.
>>
>> This counts the number of interrupts arrived at the CPU and therefor
>> it counts TWO in this situation. The CPU is interrupted twice, not
>> once.
>>
>
> I've got your point. For level irq, no problem. But how do you account
> for edge irq in my comments?
>
> For edge irq, we'll not have the same problem as level irq because
> kstat_incr_irqs_this_cpu() is placed after status checking for
> disable or no action. But it has its own problem. Suppose a second
> irq happens when the first is being handled, it will mark itself
> pending and quit, then it will be handled in the same loop after the
> first one has been handled. Thus kstat_incr_irqs_this_cpu() will
> only be counted once for two irqs.


I've made a another patch based on your conception of irq accounting.
It fixes the
corner case in which there comes up a second irq when the first is
still in progress,
we'll increment the irq count before we quit.

>From 93960c69e588989346589fac78b311ac62fb7552 Mon Sep 17 00:00:00 2001
From: Ning Jiang <ning.n.jiang@...il.com>
Date: Mon, 4 Jun 2012 09:54:08 +0800
Subject: [PATCH] genirq: Increment irq count for second edge irq when
the first is in progress

kstat_incr_irqs_this_cpu() records the real number of interruptions
to a certain CPU, regardless of the current irq state. That includes
interrupts which are delivered in a state where the device handler
cannot be called.

For level irq, if its disabled or no action available, we'll skip
irq handling. When it is enabled later, the irq will be retriggered
and kstat_incr_irqs_this_cpu() will be incremented again. This counts
the number of interrupts arrived at the CPU and therefor it counts
TWO in this situation. The CPU is interrupted twice, not once. The
accounting works correctly as expected.

But for edge irq, we have a subtle problem. Suppose a second irq
happens when the first is still in progress, it will mark itself
pending and quit, then it will be handled in the same loop after
the first one has been handled. Thus kstat_incr_irqs_this_cpu()
will only be counted once for two irqs.

Add a check in edge handler if there is a irq already in progress,
we'll first kstat_incr_irqs_this_cpu() before we mask and quit, so
that we can get correct irq count in this case.

[ Thanks Thomas for his excellent comments ]

Signed-off-by: Ning Jiang <ning.n.jiang@...il.com>
---
 kernel/irq/chip.c |    4 ++++
 1 files changed, 4 insertions(+), 0 deletions(-)

diff --git a/kernel/irq/chip.c b/kernel/irq/chip.c
index eebd6d5..254b8aa 100644
--- a/kernel/irq/chip.c
+++ b/kernel/irq/chip.c
@@ -489,6 +489,8 @@ handle_edge_irq(unsigned int irq, struct irq_desc *desc)
 	if (unlikely(irqd_irq_disabled(&desc->irq_data) ||
 		     irqd_irq_inprogress(&desc->irq_data) || !desc->action)) {
 		if (!irq_check_poll(desc)) {
+			if (irqd_irq_inprogress(&desc->irq_data))
+				kstat_incr_irqs_this_cpu(irq, desc);
 			desc->istate |= IRQS_PENDING;
 			mask_ack_irq(desc);
 			goto out_unlock;
@@ -550,6 +552,8 @@ void handle_edge_eoi_irq(unsigned int irq, struct
irq_desc *desc)
 	if (unlikely(irqd_irq_disabled(&desc->irq_data) ||
 		     irqd_irq_inprogress(&desc->irq_data) || !desc->action)) {
 		if (!irq_check_poll(desc)) {
+			if (irqd_irq_inprogress(&desc->irq_data))
+				kstat_incr_irqs_this_cpu(irq, desc);
 			desc->istate |= IRQS_PENDING;
 			goto out_eoi;
 		}
-- 
1.7.1

Download attachment "0001-genirq-Increment-irq-count-for-second-edge-irq-when-.patch" of type "application/octet-stream" (2431 bytes)

Powered by blists - more mailing lists