lkml.org 
[lkml]   [2014]   [Aug]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v8 4/4] ARM: Add KGDB/KDB FIQ debugger generic code
On Thu, Jul 10, 2014 at 09:03:47AM +0100, Daniel Thompson wrote:
> From: Anton Vorontsov <anton.vorontsov@linaro.org>
>
> The FIQ debugger may be used to debug situations when the kernel stuck
> in uninterruptable sections, e.g. the kernel infinitely loops or
> deadlocked in an interrupt or with interrupts disabled.
>
> By default KGDB FIQ is disabled in runtime, but can be enabled with
> kgdb_fiq.enable=1 kernel command line option.

I know you've been around the loop on this patch set quite a number of
times. However, there are two issues. The first is a simple concern,
the second is more a design decision...

I've recently been hitting a problem on iMX6Q with a irqs-off deadlock
on CPU0 (somehow, it always hit CPU0 every time I tested.) This wasn't
particularly good as it prevented much in the way of diagnosis.

Of course, things like the spinlock lockup fired... but nothing could
give me a trace from CPU0.

On x86, they have this fixed by using the NMI to trigger a backtrace
on all CPUs when a RCU lockup or spinlock lockup occurs. There's a
generic hook for this called arch_trigger_all_cpu_backtrace().

So, I set about using the contents of some of your patches to implement
this for ARM, and I came out with something which works. In doing this,
I started wondering whether the default FIQ handler should not be just
"subs pc, lr, #4" but mostly your FIQ assembly code you have below.
This, along with your GIC patches to move all IRQs to group 1, then
gives us a way to send a FIQ IPI to CPUs in the system - and the FIQ
IPI could be caught and used to dump a backtrace.

Here's the changes I did for that, which are a tad hacky:

irq-gic.c - SGI 8 gets used to trigger a backtrace. Note it must be
high priority too.

gic_cpu_init()
+ /*
+ * Set all PPI and SGI interrupts to be group 1.
+ *
+ * If grouping is not available (not implemented or prohibited by
+ * security mode) these registers are read-as-zero/write-ignored.
+ */
+ writel_relaxed(0xfffffeff, dist_base + GIC_DIST_IGROUP + 0);
+ writel_relaxed(0xa0a0a000, dist_base + GIC_DIST_PRI + 8);

gic_raise_softirq()
+ softirq = map << 16 | irq;
+ if (irq != 8)
+ softirq |= 0x8000;
+

arch/arm/kernel/smp.c:
+/* For reliability, we're prepared to waste bits here. */
+static DECLARE_BITMAP(backtrace_mask, NR_CPUS) __read_mostly;
+
...
+static void ipi_cpu_backtrace(struct pt_regs *regs)
+{
+ int cpu = smp_processor_id();
+
+ if (cpumask_test_cpu(cpu, to_cpumask(backtrace_mask))) {
+ static arch_spinlock_t lock = __ARCH_SPIN_LOCK_UNLOCKED;
+
+ arch_spin_lock(&lock);
+ printk(KERN_WARNING "FIQ backtrace for cpu %d\n", cpu);
+ show_regs(regs);
+ arch_spin_unlock(&lock);
+ cpumask_clear_cpu(cpu, to_cpumask(backtrace_mask));
+ }
+}
+
...
+void arch_trigger_all_cpu_backtrace(bool include_self)
+{
+ static unsigned long backtrace_flag;
+ int i, cpu = get_cpu();
+
+ if (test_and_set_bit(0, &backtrace_flag)) {
+ /*
+ * If there is already a trigger_all_cpu_backtrace() in progress
+ * (backtrace_flag == 1), don't output double cpu dump infos.
+ */
+ put_cpu();
+ return;
+ }
+
+ cpumask_copy(to_cpumask(backtrace_mask), cpu_online_mask);
+ if (!include_self)
+ cpumask_clear_cpu(cpu, to_cpumask(backtrace_mask));
+
+ if (!cpumask_empty(to_cpumask(backtrace_mask))) {
+ pr_info("Sending FIQ to %s CPUs:\n",
+ (include_self ? "all" : "other"));
+ smp_cross_call(to_cpumask(backtrace_mask), IPI_CPU_BACKTRACE);
+ }
+
+ /* Wait for up to 10 seconds for all CPUs to do the backtrace */
+ for (i = 0; i < 10 * 1000; i++) {
+ if (cpumask_empty(to_cpumask(backtrace_mask)))
+ break;
+
+ mdelay(1);
+ }
+
+ clear_bit(0, &backtrace_flag);
+ smp_mb__after_atomic();
+ put_cpu();
+}
+
+void __fiq_handle(struct pt_regs *regs)
+{
+ ipi_cpu_backtrace(regs);
+}

arch/arm/kernel/setup.c:
+static unsigned int fiq_stack[4][1024];
+
...
cpu_init()
- "msr cpsr_c, %7"
+ "msr cpsr_c, %7\n\t"
+ "mov sp, %8\n\t"
+ "msr cpsr_c, %9"
...
+ PLC (PSR_F_BIT | PSR_I_BIT | FIQ_MODE),
+ "r" (&fiq_stack[cpu][1024]),

The FIQ assembly code is basically the same as yours, but with:

+ .macro fiq_handler
+ bl __fiq_handle
+ .endm

and the code in svc_exit_via_fiq testing the PSR I flag and calling
trace_hardirqs_on removed.

This does have one deficiency, and that is it doesn't EOI the FIQ
interrupt - that's something which should be fixed, but for my
purposes to track down where the locked CPU was, it wasn't strictly
necessary that the system continued to work after this point.

This brings me to my second concern, and is the reason I decided not
to ask for it for this merge window.

Calling the trace_* functions is a no-no from FIQ code.
trace_hardirqs_on() can itself take locks, which can result in a
deadlock.

I thought I'd made it clear that FIQ code can't take locks because
there's no way of knowing what state they're in at the point that the
FIQ fires - _irq() variants won't save you - and that's kind of the
point of FIQ. It's almost never masked by the kernel.

Now, You'll be forgiven if you now point out that in the code above,
I'm taking a spinlock. That's absolutely true. Analyse the code
a little closer and you'll notice it's done in a safe way. There's
only one time where that spinlock is taken and that's from FIQ code,
never from any other code, and only once per CPU - notice how
arch_trigger_all_cpu_backtrace() protects itself against multiple
callers, and how ipi_cpu_backtrace() is careful to check that its
CPU bit is set. This is exactly the same method which x86 code uses
(in fact, much of the above code was stolen from x86!)

So, how about moving the FIQ assembly code to entry-armv.S and making
it less kgdb specific? (Though... we do want to keep a /very/ close
eye on users to ensure that they don't do silly stuff with locking.)

--
FTTC broadband for 0.8mile line: currently at 9.5Mbps down 400kbps up
according to speedtest.net.


\
 
 \ /
  Last update: 2014-08-14 00:21    [W:0.463 / U:3.588 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site