[PATCH v7 4/4] nmi_backtrace: generate one-line reports for idle cpus

On Tue 2016-08-09 12:43:58, Chris Metcalf wrote:

On 8/9/2016 8:43 AM, Petr Mladek wrote:

quoted

On Mon 2016-08-08 12:03:38, Chris Metcalf wrote:

quoted

When doing an nmi backtrace of many cores, most of which are idle,
the output is a little overwhelming and very uninformative.  Suppress
messages for cpus that are idling when they are interrupted and just
emit one line, "NMI backtrace for N skipped: idling at pc 0xNNN".

Hmm, I do not see this message even though the CPU is in the idle state:

[ 7918.884535] CPU: 3 PID: 0 Comm: swapper/3 Not tainted 4.8.0-rc1-4-default+ #3088
[ 7918.884538] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
[ 7918.884539] task: ffff88013a594380 task.stack: ffff88013a598000
[ 7918.884541] RIP: 0010:[<ffffffff81050bc6>]  [<ffffffff81050bc6>] native_safe_halt+0x6/0x10
[ 7918.884543] RSP: 0018:ffff88013a59bea8  EFLAGS: 00000206
[ 7918.884544] RAX: ffff88013a594380 RBX: 0000000000000003 RCX: 0000000000000000
[ 7918.884546] RDX: ffff88013a594380 RSI: 0000000000000001 RDI: ffff88013a594380
[ 7918.884548] RBP: ffff88013a59bea8 R08: 0000000000000000 R09: 0000000000000000
[ 7918.884550] R10: 0000000000000001 R11: 0000000000000001 R12: 0000000000000003
[ 7918.884551] R13: 0000000000000000 R14: ffff88013a598000 R15: ffff88013a598000
[ 7918.884553] FS:  0000000000000000(0000) GS:ffff88013fd80000(0000) knlGS:0000000000000000
[ 7918.884554] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 7918.884556] CR2: 00007f8afc65e000 CR3: 00000001383b8000 CR4: 00000000000006e0
[ 7918.884557] Stack:
[ 7918.884559]  ffff88013a59bec8 ffffffff819573d3 0000000000000003 0000000000000000
[ 7918.884561]  ffff88013a59bed8 ffffffff8102628f ffff88013a59bee8 ffffffff819579ea
[ 7918.884562]  ffff88013a59bf30 ffffffff810bfe1a ffff88013a598000 ffff88013a598000
[ 7918.884563] Call Trace:
[ 7918.884565]  [<ffffffff819573d3>] default_idle+0x23/0x170
[ 7918.884566]  [<ffffffff8102628f>] arch_cpu_idle+0xf/0x20
[ 7918.884568]  [<ffffffff819579ea>] default_idle_call+0x2a/0x50
[ 7918.884570]  [<ffffffff810bfe1a>] cpu_startup_entry+0x16a/0x260
[ 7918.884571]  [<ffffffff8103faf6>] start_secondary+0xf6/0x100
[ 7918.884573] Code: 00 00 00 00 00 55 48 89 e5 fa 5d c3 66 0f 1f 84 00 00 00 00 00 55 48 89 e5 fb 5d c3 66 0f 1f 84 00 00 00 00 00 55 48 89 e5 fb f4 <5d> c3 0f 1f 84 00 00 00 00 00 55
 48 89 e5 f4 5d c3 66 0f 1f 84

Note that I test it in a virtual machine using qemu.

The strange thing is that I do not see .cpuidle.text section in
the vmlinux binary. But it is possible that I have misunderstood
the concept.

The .cpuidle.text sections are merged into the final kernel's overall
text segment.  What you should see is something like this in the "nm -n"
output from the built vmlinux:

[...]
ffffffff81922aa8 T __cpuidle_text_start
ffffffff81922ab0 T default_idle
ffffffff81922b90 t mwait_idle
ffffffff81922d20 T acpi_processor_ffh_cstate_enter
ffffffff81922df0 T default_idle_call
ffffffff81922e30 t cpu_idle_poll
ffffffff81922f50 t intel_idle
ffffffff81923085 t acpi_idle_do_entry
ffffffff819230d0 t poll_idle
ffffffff81923143 T __cpuidle_text_end
[...]

In other words, all the cpuidle functions grouped together and bracketed by
the __cpuidle_text_{start,end} symbols.

Perhaps you were running with a kernel that didn't have the actual patch 4/4
applied, but just the earlier patches?  Or perhaps your host Linux has been
patched, but not the guest Linux running in qemu?  Or perhaps you are
ending up doing an NMI backtrace on the host kernel, not the guest?

Hmm, the problem is that native_safe_halt() is called from default_idle()
here. The function is marked as inline but the compiler did not inline
it.

It helped me to put native_safe_halt() into the __cpuidle_text section:

diff --git a/arch/x86/include/asm/irqflags.h b/arch/x86/include/asm/irqflags.h
index b77f5edb03b0..e31d50acd491 100644
--- a/arch/x86/include/asm/irqflags.h
+++ b/arch/x86/include/asm/irqflags.h

@@ -44,7 +44,7 @@ static inline void native_irq_enable(void)
 	asm volatile("sti": : :"memory");
 }
 
-static inline void native_safe_halt(void)
+static inline __attribute__((__section__(".cpuidle.text"))) void native_safe_halt(void)
 {
 	asm volatile("sti; hlt": : :"memory");
 }

I did not use __cpuidle macro because I was not able to include
linux/cpu.h into that header.

I wonder if it would be possible to detect the idle thread an other
way. For example, I wonder if it would be enough to check for the
PID 0.


Best Regards,
Petr