rcu_sched stall OR kernel panic on PowerEdge R640

[djgalloway]

Here’s the latest output using the debug kernel:

console [ttyS1] enabled
bootconsole [earlyvga0] disabled
ACPI: Core revision 20180810
clocksource: hpet: mask: 0xffffffff max_cycles: 0xffffffff, max_idle_ns: 79635855245 ns
APIC: Switch to symmetric I/O mode setup
x2apic: IRQ remapping doesn't support X2APIC mode
x2apic disabled
Switched APIC routing to flat.
..TIMER: vector=0x30 apic1=0 pin1=2 apic2=-1 pin2=-1
clocksource: tsc-early: mask: 0xffffffffffffffff max_cycles: 0x1fb633008a4, max_idle_ns: 440795292230 ns
Calibrating delay loop (skipped), value calculated using timer frequency.. 4400.00 BogoMIPS (lpj=2200000)
pid_max: default: 32768 minimum: 301
Mount-cache hash table entries: 131072 (order: 8, 1048576 bytes)
Mountpoint-cache hash table entries: 131072 (order: 8, 1048576 bytes)
ENERGY_PERF_BIAS: Set to 'normal', was 'performance'
ENERGY_PERF_BIAS: View and update with x86_energy_perf_policy(8)
process: using mwait in idle threads
Last level iTLB entries: 4KB 64, 2MB 8, 4MB 8
Last level dTLB entries: 4KB 64, 2MB 0, 4MB 0, 1GB 4
Spectre V1 : Mitigation: usercopy/swapgs barriers and __user pointer sanitization
Spectre V2 : Mitigation: Full generic retpoline
Spectre V2 : Spectre v2 / SpectreRSB mitigation: Filling RSB on context switch
Spectre V2 : Enabling Restricted Speculation for firmware calls
Spectre V2 : mitigation: Enabling conditional Indirect Branch Prediction Barrier
Spectre V2 : User space: Mitigation: STIBP via seccomp and prctl
Speculative Store Bypass: Mitigation: Speculative Store Bypass disabled via prctl and seccomp
MDS: Mitigation: Clear CPU buffers
Freeing SMP alternatives memory: 52K
smpboot: CPU0: Intel(R) Xeon(R) Silver 4114 CPU @ 2.20GHz (family: 0x6, model: 0x55, stepping: 0x4)
Performance Events: PEBS fmt3+, Skylake events, 32-deep LBR, full-width counters, Intel PMU driver.
... version:                4
... bit width:              48
... generic registers:      4
... value mask:             0000ffffffffffff
... max period:             00007fffffffffff
... fixed-purpose events:   3
... event mask:             000000070000000f
rcu: Hierarchical SRCU implementation.
smp: Bringing up secondary CPUs ...
x86: Booting SMP configuration:
.... node  #0, CPUs:      #1 #2 #3 #4 #5 #6 #7
smp: Brought up 1 node, 8 CPUs
smpboot: Max logical packages: 10
smpboot: Total of 8 processors activated (35220.85 BogoMIPS)
devtmpfs: initialized
clocksource: jiffies: mask: 0xffffffff max_cycles: 0xffffffff, max_idle_ns: 1911260446275000 ns
futex hash table entries: 2048 (order: 5, 131072 bytes)
xor: automatically using best checksumming function   avx       
pinctrl core: initialized pinctrl subsystem
rcu: INFO: rcu_sched self-detected stall on CPU
rcu:    0-....: (20999 ticks this GP) idle=04a/1/0x4000000000000002 softirq=10/10 fqs=5241 
rcu:     (t=21000 jiffies g=-1175 q=19)
NMI backtrace for cpu 0
CPU: 0 PID: 1 Comm: swapper/0 Not tainted 4.19.65 #12
Hardware name: Dell Inc. PowerEdge R640/08HT8T, BIOS 2.2.11 06/13/2019
Call Trace:
 <IRQ>
 0xffffffff81d6ecad
 0xffffffff81d7222f
 ? 0xffffffff8102b073
 0xffffffff81d7228a
 0xffffffff8107ce90
 0xffffffff8107c41d
 0xffffffff810806b4
 0xffffffff8108a34e
 0xffffffff81e017d5
 0xffffffff81e013af
 </IRQ>
RIP: 0010:0xffffffff8108fa1d
Code: 36 48 89 de 89 c7 e8 ca ef cd 00 3b 05 c0 13 86 01 73 24 48 63 f0 49 8b 16 48 03 14 f5 30 83 61 82 8b 72 18 40 80 e6 01 74 04 <f3> 90 eb f3 eb d1 0f 0b e9 72 fe ff ff 48 83 c4 10 5b 5d 41 5c 41
RSP: 0000:ffffc9000007fae0 EFLAGS: 00000202 ORIG_RAX: ffffffffffffff13
RAX: 0000000000000001 RBX: ffff8897e101fac8 RCX: 0000000000000001
RDX: ffff8897e10621c0 RSI: 0000000000000001 RDI: ffff8897e101fac8
RBP: 000000000001fa80 R08: 0000000000000000 R09: 00000000016daed4
R10: ffffc9000007fb58 R11: 000fffffffe00000 R12: 0000000000000001
R13: 0000000000000008 R14: ffff8897e101fac0 R15: 0000000000000000
 ? 0xffffffff81039a

[george1421]

Just for grins I had the OP boot a 486 kernel I built for another poster for a specific dedicated machine to image with FOG. That kernel gave a bit more details than the full system kernel .

Checking if this processor honours the WP bit even in supervisor mode...Ok.
SLUB: HWalign=64, Order=0-3, MinObjects=0, CPUs=8, Nodes=1
rcu: Hierarchical RCU implementation.
NR_IRQS: 2304, nr_irqs: 1848, preallocated irqs: 16
Console: colour VGA+ 80x25
console [tty0] enabled
console [ttyS1] enabled
ACPI: Core revision 20180810
clocksource: hpet: mask: 0xffffffff max_cycles: 0xffffffff, max_idle_ns: 79635855245 ns
APIC: Switch to symmetric I/O mode setup
Enabling APIC mode:  Flat.  Using 9 I/O APICs
------------[ cut here ]------------
Kernel BUG at 0xc1028128 [verbose debug info unavailable]
invalid opcode: 0000 [#1] SMP
CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.19.65 #2
Hardware name: Dell Inc. PowerEdge R640/08HT8T, BIOS 2.2.11 06/13/2019
EIP: 0xc1028128

It looks like the kernel is crashing at enabling apic mode or with the apic IO. The clock source hpet also is memorable for some reason.

So the kernel is crashing at the same point. For reference the 486 compatible kernel is also “Linux version 4.19.65”

acpi=ht acpi=oldboot acpi_osi=Linux

noapic

[Junkhacker]

i was googling the problem a bit and i was curious, will it boot if you remove the raid card?
just trying to understand the source of the panic.

[george1421]

@Junkhacker @Sebastian-Roth

I was able to get the OP going by doing this and that.

We are not sure if it was this or that that got the kernel to boot. What I did was unlocked the max CPUs (that was capped at in the kernel and I also enabled almost all of the ACPI modules in the kernel. We also tried the acpi_osi=Linux kernel parameter.

We ruled out the acpi_osi=Linux kernel parameter fixing the issue so it must be something I enabled in the kernel. Tomorrow AM I’m going to reset the kernel environment and only unlock the max CPUs. The OP is going to test that new kernel to see if it was unlocking the max cpu or it was the acpi modules I enabled.

Either way I’ll report where we ended up and which kernel change fixed the issue. I have also seen other recent CPU stalls like this that was fixed by setting acpi=off so we may need to move what ever fixed the issue into the main kernel build because new hardware/cpus may require it.

[george1421]

@developers Here’s the final update on this issue.

I reset my kernel build environment and then created 2 new kernel builds. The first was to remove the imposed CPU limit on the linux kernel this kernel was called bzImageMaxCPU. I reset the kernel build environment and then went through the ACPI settings turning on what I turned on in the debug kernel. This kernel was called bzImageACPI.

The OP tested both and the bzImageMaxCPU was the only kernel that booted on those Dell servers. So in the end @Quazz was right about the CPU not liking some of its cores disabled.

So I would recommend that we add the following settings to the official kernel build

CONFIG_INTEL_IDLE
and
Processor type and features —>

Enable Maximum number of SMP Processors and NUMA Nodes

We have seen a recent uptick in reports of rcu_sched stalls with kernel panics Maybe we are running into this issue more often as the core counts go up on these processors.

[Sebastian Roth]

@george1421 @Quazz @djgalloway Great work!!! Thanks to you all. I will add this in the next days!

[Sebastian Roth]

@george1421 @Quazz I found a bit of time to look into this. Adding CONFIG_INTEL_IDLE should be just fine I think. But I am not exactly sure about adding CONFIG_MAXSMP (Enable Maximum number of SMP Processors and NUMA Nodes). Found this topic: https://www.xenomai.org/pipermail/xenomai/2018-July/039297.html

Though I am not convinced this will actually cause trouble it’s still a bit risky. @Testers @Moderators. Would you be able to run a test kernel on several different client machines so we get a feeling of this being troublesome or not?

[george1421]

@Sebastian-Roth I can test it here, but I don’t have a system that is causing this rcu_sched issue. But I can surely test it against our current fleet of Dell systems to see if it does any harm.

We can also hold this “test” kernel in reserve in case this issue comes up again if you don’t want to release it as general availability. What I would not like to see is having a special kernel for this, and a different special kernel for that.

[Sebastian Roth]

@george1421 said in rcu_sched stall OR kernel panic on PowerEdge R640:

We can also hold this “test” kernel in reserve in case this issue comes up again if you don’t want to release it as general availability.

Don’t get me wrong on this. I am more than happy to make this the default kernel for everyone. It comes at low cost. But I’d like to see this tested on several different machines (PC as well as notebooks and even servers if possible) before we make it the new default kernel.

[Quazz]

@Sebastian-Roth As far as I understand it, Xenomai implements a patch to the kernel that does all kinds of stuff, potentially it’s not compatible with their patches, but as far as I know CONFIG_MAXSMP is in fact enabled by default on Kernel 4.4+ or so on all major distributions without issues.

That said, I don’t mind testing it.

[Sebastian Roth]

@Quazz said in rcu_sched stall OR kernel panic on PowerEdge R640:

as far as I know CONFIG_MAXSMP is in fact enabled by default on Kernel 4.4+ or so on all major distributions without issues

That’s valuable information! Any reference for this?

[Quazz]

@Sebastian-Roth Hmm, I may have been misremembering, though their CONFIG_NR_CPUS
is going to be much higher than 8 at the very least. (at least 512 afaik)

The only difference I can find is that CONFIG_MAXSMP enables CPUMASK_OFFSTACK
, which it requires to function correctly I believe (or any high CONFIG_NR_CPUS would at least)

[Sebastian Roth]

@Quazz @george1421 Ok, back from travels… what shall we do with this pending topic. I do understand that adding CONFIG_MAXSMP does fix the rcu_sched stall issue on PowerEdge R640. But do we know if this fixes rcu_sched stalls on other platforms as well? Would we get at least two more people to test this before we add it to the official kernel?

@george1421 Did you get to test this kernel on your fleet of Dell hardware to see if it might cause any other harm?

[george1421]

@Sebastian-Roth said in rcu_sched stall OR kernel panic on PowerEdge R640:

Did you get to test this kernel on your fleet of Dell hardware to see if it might cause any other harm?

TBH, no I did not test it. I haven’t found any other system that the max-cpu value fixed either. We had one dual core with the rcu_sched stall, but that was fixed with the current kernel and changing the acpi clock source that Quazz posted. I think the max-cpu will only impact CPUs with more than 8 cores.

[Sebastian Roth]

@george1421 Looking through a stack of other rcu_sched stall topics in the forums I can’t seem to find any thread where I’d think that people had CPUs with more than 8 cores. Sure sooner or later this will be state of the art but I don’t reckon we should step ahead of this. We know the current kernel works pretty good on most CPUs and I’d rather point people to this topic and provide compile instructions than setting CONFIG_MAXSMP as default. Hmm?

[george1421]

@Sebastian-Roth I’m still on the fence about this, I would say turn it on because the core count continues to rise on these processors. What can it hurt? And on the other side we really don’t know what the impact could be.

[Sebastian Roth]

@george1421 said in rcu_sched stall OR kernel panic on PowerEdge R640:

And on the other side we really don’t know what the impact could be.

Yes, because of that I don’t like switching it on.

What can it hurt?

I don’t know. Stalls on older CPUs?

[Quazz]

@Sebastian-Roth

There are kernel flags to disable SMP if necessary, so I think it’s pretty safe to compile with MAXSMP. Just my opinion of course; without a diverse test fleet it’s hard to say for sure since kernels can always have bugs or unforeseen interactions. But that would be true for any change we make.

I can’t find anything googling about stalls/problems with MAXSMP either. Only some people on embedded systems who want to reduce the size of their kernel, but that’s a targetted compile anyway.

There will be more and more systems entering the floor with more than 8 cores (our current NR_CPU value) given the recent CPU releases as well, so at the very least that number could use a bump.

[Sebastian Roth]

@Quazz said in rcu_sched stall OR kernel panic on PowerEdge R640:

I can’t find anything googling about stalls/problems with MAXSMP either. Only some people on embedded systems who want to reduce the size of their kernel, but that’s a targetted compile anyway.

Ok, you and George have convinced me this is most probably not going to cause us much trouble, so I will add the options as mentioned below.

[Sebastian Roth]

@george1421 @Quazz I just added the two kernel options as mentioned below to the x64 kernel config (not pushed the change yet).

While CONFIG_INTEL_IDLE is available in x86 (32 bit) config as well CONFIG_MAXSMP is not (depends on X86_64 [=y]). Should I leave CONFIG_NR_CPUS set to 8 (default I think) or increase it to 16, 32, 64?

For ARM kernel config we don’t have CONFIG_INTEL_IDLE nor CONFIG_MAXSMP but can adjust CONFIG_NR_CPUS too. 16, 32, 64?