Skip to contentSkip to conversation

how to perform a safe NMI stack trace on all CPUs on x86?

RFC 王龙 4 participants 5 messages May 13, 2015 – May 14, 2015

lore
Share & tools
lore

王龙

author
#0 lore raw

Hi all,

In kernel before 3.19, when trigger_all_cpu_backtrace() is called on x86, it will trigger an NMI on each CPU and call show_regs(). But this can lead to a hard lock up if the NMI comes in on another printk().

The commit a9edc88093287183ac934be44f295f183b2c62dd (x86/nmi: Perform a safe NMI stack trace on all CPUs) fix this problem on kernel mainline. when the NMI triggers, it switches the printk routine for that CPU to call a NMI safe printk function that records the printk in a per_cpu seq_buf descriptor. After all NMIs have finished recording its data, the seq_bufs are printed in a safe context. But how do we fix this problem in older version of kernel(eg, 3.10 stable)? The 3.10 stable has no "switch printk routine" and "seq_buf" infrastructures.

Could anyone give me some ideas?

Best Regards
Wang Long���{.n�+�������+%�����ݶ��w��{.n�+����{��G�����{ay�ʇڙ�,j��f���h��������z_�(�階�ݢj"���m������G������?����&���~��iO��z��v�^�m���� ������?�I�

#1 lore raw
Quoted 5 lines · On Wed, 13 May 2015 22:14:54 +0800 "王龙" <[email protected]> wrote:

context. But how do we fix this problem in older version of kernel(eg, 3.10 stable)? The 3.10 stable has no "switch printk routine" and "seq_buf" infrastructures.

Could anyone give me some ideas?

Backport the necessary patches.

-- Steve

#2 lore raw
Quoted 7 lines · Steven Rostedt
Quoted 3 lines · On Wed, 13 May 2015 22:14:54 +0800 "王龙" <[email protected]> wrote:

context. But how do we fix this problem in older version of kernel(eg, 3.10 stable)? The 3.10 stable has no "switch printk routine" and "seq_buf" infrastructures.

Quoted 7 lines
Quoted 2 lines

Could anyone give me some ideas?

Backport the necessary patches.

-- Steve

Hi Steve,

Thank you for your reply, I will backport necessary patches to 3.10 stable. Welcome you to review my backport patches.

Best Regards
Wang Long

Quoted 2 lines

.

#3 lore raw
In reply to 王龙
Quoted 14 lines · 王龙

Hi all,

In kernel before 3.19, when trigger_all_cpu_backtrace() is called on x86, it will trigger an NMI on each CPU and call show_regs(). But this can lead to a hard lock up if the NMI comes in on another printk().

The commit a9edc88093287183ac934be44f295f183b2c62dd (x86/nmi: Perform a safe NMI stack trace on all CPUs) fix this problem on kernel mainline. when the NMI triggers, it switches the printk routine for that CPU to call a NMI safe printk function that records the printk in a per_cpu seq_buf descriptor. After all NMIs have finished recording its data, the seq_bufs are printed in a safe context. But how do we fix this problem in older version of kernel(eg, 3.10 stable)? The 3.10 stable has no "switch printk routine" and "seq_buf" infrastructures.

Quoted 1 line

Could anyone give me some ideas?

Either you backport seq_buf-based aproach to the older kernel, or, if you are working on 3.4 kernel or earlier (basically any kernel preceeding the printk() revamp that happened in 7ff9554bb57 and after), you can use slightly simpler aproach.

It's an aproach we used initially when finding out the issue for the first time, and it is proven to work as well (but it's not applicable after Kay added all the complexity to printk()).

You can see it in our SLE11 kernel tree, available on

	http://kernel.suse.com/cgit/kernel/commit/?h=SLE11-SP4&id=8d62ae68ff61d77ae3c4899f05dbd9c9742b14c9

for example.

It's up to you to judget which is the least painful way :)

Signature · 2 lines
Jiri Kosina
SUSE Labs
#4 lore raw
Quoted 27 lines · Jiri Kosina
Quoted 14 lines · 王龙

Hi all,

In kernel before 3.19, when trigger_all_cpu_backtrace() is called on x86, it will trigger an NMI on each CPU and call show_regs(). But this can lead to a hard lock up if the NMI comes in on another printk().

The commit a9edc88093287183ac934be44f295f183b2c62dd (x86/nmi: Perform a safe NMI stack trace on all CPUs) fix this problem on kernel mainline. when the NMI triggers, it switches the printk routine for that CPU to call a NMI safe printk function that records the printk in a per_cpu seq_buf descriptor. After all NMIs have finished recording its data, the seq_bufs are printed in a safe context. But how do we fix this problem in older version of kernel(eg, 3.10 stable)? The 3.10 stable has no "switch printk routine" and "seq_buf" infrastructures.

Quoted 1 line

Could anyone give me some ideas?

Either you backport seq_buf-based aproach to the older kernel, or, if you are working on 3.4 kernel or earlier (basically any kernel preceeding the printk() revamp that happened in 7ff9554bb57 and after), you can use slightly simpler aproach.

It's an aproach we used initially when finding out the issue for the first time, and it is proven to work as well (but it's not applicable after Kay added all the complexity to printk()).

Quoted 8 lines

You can see it in our SLE11 kernel tree, available on

	http://kernel.suse.com/cgit/kernel/commit/?h=SLE11-SP4&id=8d62ae68ff61d77ae3c4899f05dbd9c9742b14c9

for example.

It's up to you to judget which is the least painful way :)

Hi Jiri Kosina,

For 3.10 stable, the only way to solve this problem is backport seq_buf-based aproach.

I will backport necessary patches to 3.10 stable. Welcome you to review my backport patches.

Best Regards
Wang Long