Message-ID: <CAOuPNLg4tM6Cw8Yf9rdpkBOU=+Fh654yFQeAhJf1y88Gb3MAew@mail.gmail.com>
Date:   Fri, 16 Nov 2018 20:10:28 +0530
From:   Pintu Agarwal <pintu.ping@...il.com>
To:     Valdis Kletnieks <valdis.kletnieks@...edu>
Cc:     open list <linux-kernel@...r.kernel.org>,
        linux-arm-kernel@...ts.infradead.org,
        Russell King - ARM Linux <linux@...linux.org.uk>,
        kernelnewbies@...nelnewbies.org,
        Jungseok Lee <jungseoklee85@...il.com>,
        catalin.marinas@....com, will.deacon@....com,
        Takahiro Akashi <takahiro.akashi@...aro.org>,
        mark.rutland@....com, Sungjinn Chung <barami97@...il.com>
Subject: Re: [ARM64] Printing IRQ stack usage information

On Fri, Nov 16, 2018 at 5:03 PM <valdis.kletnieks@...edu> wrote:
>
> On Fri, 16 Nov 2018 11:44:36 +0530, Pintu Agarwal said:
>
> > > If your question is "Did one
> > > of the CPUs blow out its IRQ stack (or come close to doing so)?" there's better
> > > approaches.
> > >
> > Yes, exactly, this is what the main intention.
> > If you have any better idea about this approach, please refer me.
> > It will be of great help.
>
> Look at the code controlled by '#ifdef CONFIG_DEBUG_STACK_USAGE'
> which does the same thing for process stacks, or CONFIG_SCHED_STACK_END_CHECK
> or the use of guard pages for detecting stack overrun....

Hi,

Thank you so much for your reference.
Yes, I have already gone through the process stack usage, which I
found slightly different.
However, I will go through it in more detail, and see if I can gain
some ideas from there.

I found a similar irq_stack_usage implementation in parisc architecture:
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/arch/parisc/kernel/irq.c?h=v4.19.1

I have also gone through the unwind_frame() part in arch/arm64/stacktrace.c:
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/arch/arm64/kernel/stacktrace.c?h=v4.9.137

By referring to these, I tried to make a similar approach for arm64:
I created a new function: dump_irq_stack_info()
[arch/arm64/kernel/traps.c], and called it as part of show_stack().

This is the experimental patch I created.
Note: This is just for my experiment purpose. I know this is ugly and
in very bad shape right now.
It is only to get some idea about irq stack usage.

diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c
index 11e5eae..6ac855d 100644
--- a/arch/arm64/kernel/traps.c
+++ b/arch/arm64/kernel/traps.c
@@ -214,9 +214,39 @@ static void dump_backtrace(struct pt_regs *regs,
struct task_struct *tsk)
        }
 }

+void dump_irq_stack_info(void)
+{
+        int cpu, actual;
+        unsigned long irq_stack_ptr;
+        unsigned long stack_start;
+        unsigned long free_stack;
+
+        actual = IRQ_STACK_SIZE;
+        free_stack = 0;
+        pr_info("CPU UNUSED-STACK ACTUAL-STACK\n");
+
+       for_each_present_cpu(cpu) {
+               unsigned long sp;
+               irq_stack_ptr = IRQ_STACK_PTR(cpu);
+               sp = current_stack_pointer;
+               //sp = IRQ_STACK_TO_TASK_STACK(irq_stack_ptr);
+               stack_start = (unsigned long)per_cpu(irq_stack, cpu);
+               if (on_irq_stack(sp, cpu)) {
+                       pr_info("cpu:%d : sp: on irq_stack\n", cpu);
+                       free_stack = sp - stack_start;
+               } else {
+                       free_stack = irq_stack_ptr - stack_start;
+               }
+               pr_info("%2d %10lu %10d\n", cpu, free_stack, actual);
+       }
+}
+
 void show_stack(struct task_struct *tsk, unsigned long *sp)
 {
        dump_backtrace(NULL, tsk);
+       dump_irq_stack_info();
        barrier();
 }

Then, I developed a sample kernel module for timer handler
(timerirq.c) and called the dump_stack() function from inside my timer
interrupt handler.
The dump_stack() will internally call show_stack(), which will then
call our function: dump_irq_stack_info().

/* From interrupt context */
static void my_timer_irq_handler(unsigned long ptr)
{
        int i;
        unsigned long flags;

        if (in_interrupt()) {
                pr_info("[timerirq]: %s: in interrupt context, count: %d\n",
                                        __func__, count);
                spin_lock_irqsave(&mylock, flags);
+               dump_stack();
                spin_unlock_irqrestore(&mylock, flags);
        } else {
                /* This is not needed here*/
        }
        tasklet_schedule(&my_tasklet);
}

OUTPUT:
------------
With this, I got the below output as part of dump_stack() and backtrace:
<snip>
[   43.267923] CPU UNUSED-STACK ACTUAL-STACK
[   43.271925]  0      16368      16384
[   43.275493]  1      16368      16384
[   43.279061]  2      16368      16384
[   43.282628] cpu:3 : sp: on irq_stack
[   43.286195]  3      15616      16384
[   43.289762]  4      16368      16384
[   43.293330]  5      16368      16384
[   43.296898]  6      16368      16384
[   43.300465]  7      16368      16384
<snip>

So, I observed that my interrupt handler was executed by cpu3, and
it's irq_stack usage is shown:
3      15616      16384

With this information, I can know that which interrupt handler is
using how much irq_stack ?

Is this approach valid ?
Or still there is much better way to dump the information ?

For example: is it possible to keep storing the irq_stack_usage (for
each cpu in a variable) information from boot time, and then use this
variable to dump the irq_stack information, after the system booted,
may be from proc entry ?


Thanks,
Pintu

Powered by blists - more mailing lists