Re: [PATCH] lib/nmi_backtrace: Serialize even messages about idle CPUs
From: Petr Mladek
Date: Fri Jul 30 2021 - 07:08:44 EST
On Tue 2021-07-27 17:53:04, John Ogness wrote:
> On 2021-07-27, Petr Mladek <pmladek@xxxxxxxx> wrote:
> > The commit 55d6af1d66885059ffc2a ("lib/nmi_backtrace: explicitly serialize
> > banner and regs") serialized backtraces from more CPUs using the re-entrant
> > printk_printk_cpu lock. It was a preparation step for removing the obsolete
> > nmi_safe buffers.
> > The single-line messages about idle CPUs were not serialized against other
> > CPUs and might appear in the middle of backtrace from another CPU,
> > for example:
> > [56394.590068] NMI backtrace for cpu 2
> > [56394.590069] CPU: 2 PID: 444 Comm: systemd-journal Not tainted 5.14.0-rc1-default+ #268
> > [56394.590071] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.12.0-59-gc9ba527-rebuilt.opensuse.org 04/01/2014
> > [56394.590072] RIP: 0010:lock_is_held_type+0x0/0x120
> > [56394.590071] NMI backtrace for cpu 0 skipped: idling at native_safe_halt+0xb/0x10
> > [56394.590076] Code: a2 38 ff 0f 0b 8b 44 24 04 eb bd 48 8d ...
> > [56394.590077] RSP: 0018:ffffab02c07c7e68 EFLAGS: 00000246
> > [56394.590079] RAX: 0000000000000000 RBX: ffff9a7bc0ec8a40 RCX: ffffffffaab8eb40
> > It might cause confusion what CPU the following lines belongs to and
> > whether the backtraces are really serialized.
> I originally implemented this, but later decided against it because it
> causes idle CPUs to begin busy-waiting in NMI context in order to log a
> single line saying they are idle. If the user is aware that there is
> only 1 line for the idle message, then the user knows that it isn't
> causing a problem for reading the stack trace.
> When triggering many such dumps on systems with many CPUs where this
> patch is applied, it seemed like I was making the whole system work
> awfully hard for something that should be trivial.
I understand the concern. I am not super happy with it as well.
> Considering that dump_stack() and show_regs() should be fast and we are
> only dumping to the lockless buffer, it is probably OK to be doing all
> the busy-waiting. Once atomic consoles are introduced, it will have
> quite an impact here, but atomic consoles are mostly only active on
> system crash, so I think that would be OK as well.
Also the single line is printed only on idle CPUs. They have nothing
else to do anyway ;-)
I have pushed the patch into printk/linux.git, branch
We could always revert it when there are some problems with it.