Re: False positive "do_IRQ: #.55 No irq handler for vector" messages on AMD ryzen based laptops

From: Hans de Goede
Date: Tue Mar 05 2019 - 11:02:09 EST


Hi,

On 05-03-19 15:06, Lendacky, Thomas wrote:
On 3/3/19 4:57 AM, Hans de Goede wrote:
Hi,

On 21-02-19 13:30, Hans de Goede wrote:
Hi,

On 19-02-19 22:47, Lendacky, Thomas wrote:
On 2/19/19 3:01 PM, Thomas Gleixner wrote:
Hans,

On Tue, 19 Feb 2019, Hans de Goede wrote:

Cc+: ACPI/AMD folks

Various people are reporting false positive "do_IRQ: #.55 No irq
handler for
vector"
messages on AMD ryzen based laptops, see e.g.:

https://bugzilla.redhat.com/show_bug.cgi?id=1551605

Which contains this dmesg snippet:

Feb 07 20:14:29 localhost.localdomain kernel: smp: Bringing up
secondary CPUs
...
Feb 07 20:14:29 localhost.localdomain kernel: x86: Booting SMP
configuration:
Feb 07 20:14:29 localhost.localdomain kernel: .... node #0,
CPUs:ÂÂÂÂÂ #1
Feb 07 20:14:29 localhost.localdomain kernel: do_IRQ: 1.55 No irq
handler for
vector
Feb 07 20:14:29 localhost.localdomain kernel:Â #2
Feb 07 20:14:29 localhost.localdomain kernel: do_IRQ: 2.55 No irq
handler for
vector
Feb 07 20:14:29 localhost.localdomain kernel:Â #3
Feb 07 20:14:29 localhost.localdomain kernel: do_IRQ: 3.55 No irq
handler for
vector
Feb 07 20:14:29 localhost.localdomain kernel: smp: Brought up 1 node,
4 CPUs
Feb 07 20:14:29 localhost.localdomain kernel: smpboot: Max logical
packages: 1
Feb 07 20:14:29 localhost.localdomain kernel: smpboot: Total of 4
processors
activated (15968.49 BogoMIPS)

It seems that we get an IRQ for each CPU as we bring it online,
which feels to me like it is some sorta false-positive.

Sigh, that looks like BIOS value add again.

It's not a false positive. Something _IS_ sending a vector 55 to these
CPUs
for whatever reason.


I remember seeing something like this in the past and it turned out to be
a BIOS issue. BIOS was enabling the APs to interact with the legacy 8259
interrupt controller when only the BSP should. During POST the APs were
exposed to ExtINT/INTR events as a result of the mis-configuration
(probably due to a UEFI timer-tick using the 8259) and this left a pending
ExtINT/INTR interrupt latched on the APs.

When the APs were started by the OS, the latched ExtINT/INTR interrupt is
processed shortly after the OS enables interrupts. The AP then queries the
8259 to identify the vector number (which is the value of the 8259's ICW2
register + the IRQ level). The master 8259's ICW2 was set to 0x30 and,
since no interrupts are actually pending, the 8259 will respond with IRQ7
(spurious interrupt) yielding a vector of 0x37 or 55.

The OS was not expecting vector 55 and printed the message.

ÂFrom the Intel Developer's Manual: Vol 3a, Section 10.5.1:
"Only one processor in the system should have an LVT entry configured to
use the ExtINT delivery mode."

Not saying this is the problem, but very well could be.

That sounds like a likely candidate, esp. also since this only happens
once per CPU when we first only the CPU.

Can you provide me with a patch with some printk-s / pr_debugs to
test for this, then I can build a kernel with that patch added and
we can see if your hypothesis is right.

Ping? I like your theory, can you provide some help with debugging this
further (to prove that your theory is correct ) ?

It's been a very long time since I dealt with this and I was only on the
periphery. You might be able to print the LVT entries from the APIC and
see if any of them have an un-masked ExtINT delivery mode. You would need
to do this very early before Linux modifies any values.

I'm afraid I'm not familiar enough with the interrupt / APIC parts of
the kernel to do something like this myself.

Or you can report the issue to the OEM and have them check their BIOS
code to see if they are doing this.

I will try to go this route, but I'm not really hopeful that will
lead to a solution.

Regards,

Hans