Re: kernel 6.2 stuck at boot (efi_call_rts) on arm64

[Andrea Righi]

On Thu, Mar 16, 2023 at 08:58:20AM +0100, Ard Biesheuvel wrote:
> Hello Andrea,
> 
> On Thu, 16 Mar 2023 at 08:54, Andrea Righi <andrea.righi@xxxxxxxxxxxxx> wrote:
> >
> > Hello,
> >
> > the latest v6.2.6 kernel fails to boot on some arm64 systems, the kernel
> > gets stuck and never completes the boot. On the console I see this:
> >
> > [   72.043484] rcu: INFO: rcu_preempt detected stalls on CPUs/tasks:
> > [   72.049571] rcu:     22-...0: (30 GPs behind) idle=b10c/1/0x4000000000000000 softirq=164/164 fqs=6443
> > [   72.058520]     (detected by 28, t=15005 jiffies, g=449, q=174 ncpus=32)
> > [   72.064949] Task dump for CPU 22:
> > [   72.068251] task:kworker/u64:5   state:R  running task     stack:0     pid:447   ppid:2      flags:0x0000000a
> > [   72.078156] Workqueue: efi_rts_wq efi_call_rts
> > [   72.082595] Call trace:
> > [   72.085029]  __switch_to+0xbc/0x100
> > [   72.088508]  0xffff80000fe83d4c
> >
> > After that, as a consequence, I start to get a lot of hung task timeout traces.
> >
> > I tried to bisect the problem and I found that the offending commit is
> > this one:
> >
> >  e7b813b32a42 ("efi: random: refresh non-volatile random seed when RNG is initialized")
> >
> > I've reverted this commit for now and everything works just fine, but I
> > was wondering if the problem could be caused by a lack of entropy on
> > these arm64 boxes or something else.
> >
> > Any suggestion? Let me know if you want me to do any specific test.
> >
> 
> Thanks for the report.
> 
> This is most likely the EFI SetVariable() call going off into the
> weeds and never returning.
> 
> Is this an Ampere Altra system by any chance? Do you see it on
> different types of hardware?

This is: Ampere eMAG / Lenovo ThinkSystem HR330a.

> 
> Could you check whether SetVariable works on this system? E.g. by
> updating the EFI boot timeout (sudo efibootmgr -t <n>)?

ubuntu@kuzzle:~$ sudo efibootmgr -t 10
^C^C^C^C

^ Stuck there, so it really looks like SetVariable is the problem.

Thanks,
-Andrea