KVM: arm64: Regression in at least linux-6.1.y tree with recent FPSIMD/SVE/SME fix

Wed Oct 1 10:47:58 PDT 2025

Hi Will, thanks for the prompt reply and thorough explanation!

On the 6.1 branch, those functions have additional logic around realtime kernels, like so:

static void get_cpu_fpsimd_context(void)
{
        if (!IS_ENABLED(CONFIG_PREEMPT_RT))
                local_bh_disable();
        else
                preempt_disable();
        __get_cpu_fpsimd_context();
}

...

static void put_cpu_fpsimd_context(void)
{
        __put_cpu_fpsimd_context();
        if (!IS_ENABLED(CONFIG_PREEMPT_RT))
                local_bh_enable();
        else
                preempt_enable();
}

Honestly, we don't even set CONFIG_PREEMPT_RT for our ARM64 kernels, but in the interest of being complete, should we also not call preempt_enable/disable if (!irqs_disabled()) ?

I'm assuming this will get queued up for a fix in the stable branches soon in any case.

Thanks,

--
Kenneth Van Alstyne, Jr.
(228) 547-8045

> On Oct 1, 2025, at 05:32, Will Deacon <will at kernel.org> wrote:
> 
> Hi Kenneth,
> 
> Thanks for the report.
> 
> On Tue, Sep 30, 2025 at 05:31:38PM +0000, Kenneth Van Alstyne wrote:
>> Sending via plain text email -- apologies if you receive this twice.
>> 
>> If this isn't the process for reporting a regression in a LTS kernel per
>> https://www.kernel.org/doc/html/latest/admin-guide/reporting-issues.html,
>> I'm happy to follow another process.
>> 
>> Kernel 6.1.149 introduced a regression, at least on our ARM Cortex
>> A57-based platforms, via commit 8f4dc4e54eed4bebb18390305eb1f721c00457e1
>> in arch/arm64/kernel/fpsimd.c where booting KVM VMs eventually leads to a
>> spinlock recursion BUG and crash of the box.
>> 
>> Reverting that commit via the below reverts to the old (working) behavior:
>> 
>> diff --git a/arch/arm64/kernel/fpsimd.c b/arch/arm64/kernel/fpsimd.c
>> index 837d1937300a57..bc42163a7fd1f0 100644
>> --- a/arch/arm64/kernel/fpsimd.c
>> +++ b/arch/arm64/kernel/fpsimd.c
>> @@ -1851,10 +1851,10 @@ void fpsimd_save_and_flush_cpu_state(void)
>>  if (!system_supports_fpsimd())
>>  return;
>>  WARN_ON(preemptible());
>> - get_cpu_fpsimd_context();
>> + __get_cpu_fpsimd_context();
>>  fpsimd_save();
>>  fpsimd_flush_cpu_state();
>> - put_cpu_fpsimd_context();
>> + __put_cpu_fpsimd_context();
>> }
>>  #ifdef CONFIG_KERNEL_MODE_NEON
> 
> Hmm, the problem with doing that is it will reintroduce the bug that
> 8f4dc4e54eed ("KVM: arm64: Fix kernel BUG() due to bad backport of
> FPSIMD/SVE/SME fix") was trying to fix (see the backtrace in the commit
> message). So the old behaviour is still broken, just in a slightly
> different way.
> 
>> It's not entirely clear to me if this is specific to our firmware,
>> specific to ARM Cortex A57, or more systemic as we lack sufficiently
>> differentiated hardware to know.  I've tested on the latest 6.1 kernel in
>> addition to the one in the log below and have also tested a number of
>> firmware versions available for these boxes.
>> 
>> Steps to reproduce:
>> 
>> Boot VM in qemu-system-aarch64 with "-accel kvm" and "-cpu host" flags set -- no other arguments seem to matter
>> Generate CPU load in VM
>> 
>> Kernel log:
>> 
>> [sjc1] root at si-compute-kvm-e0fff70016b4:/# [  805.905413] BUG: spinlock recursion on CPU#7, CPU 3/KVM/57616
>> [  805.905452]  lock: 0xffff3045ef850240, .magic: dead4ead, .owner: CPU 3/KVM/57616, .owner_cpu: 7
>> [  805.905477] CPU: 7 PID: 57616 Comm: CPU 3/KVM Tainted: G           O       6.1.152 #1
>> [  805.905495] Hardware name: SoftIron SoftIron Platform Mainboard/SoftIron Platform Mainboard, BIOS 1.31 May 11 2023
>> [  805.905516] Call trace:
>> [  805.905524]  dump_backtrace+0xe4/0x110
>> [  805.905538]  show_stack+0x20/0x30
>> [  805.905548]  dump_stack_lvl+0x6c/0x88
>> [  805.905561]  dump_stack+0x18/0x34
>> [  805.905571]  spin_dump+0x98/0xac
>> [  805.905583]  do_raw_spin_lock+0x70/0x128
>> [  805.905596]  _raw_spin_lock+0x18/0x28
>> [  805.905607]  raw_spin_rq_lock_nested+0x18/0x28
>> [  805.905620]  update_blocked_averages+0x70/0x550
>> [  805.905634]  run_rebalance_domains+0x50/0x70
>> [  805.905645]  handle_softirqs+0x198/0x328
>> [  805.905659]  __do_softirq+0x1c/0x28
>> [  805.905669]  ____do_softirq+0x18/0x28
>> [  805.905680]  call_on_irq_stack+0x30/0x48
>> [  805.905691]  do_softirq_own_stack+0x24/0x30
>> [  805.905703]  do_softirq+0x74/0x90
>> [  805.905714]  __local_bh_enable_ip+0x64/0x80
> 
> Argh, this is because we can't simply mask/unmask softirqs and so when
> they get re-enabled we process anything pending. I _think_ irqs are
> disabled at this point, so perhaps we should only bother with
> disabling/enabling softirqs if hardirqs are enabled, a bit like the hack
> Ard had in:
> 
> https://lore.kernel.org/all/20250924152651.3328941-13-ardb+git@google.com/
> 
> Hacky diff at the end.
> 
>> [  805.905727]  fpsimd_save_and_flush_cpu_state+0x5c/0x68
>> [  805.905740]  kvm_arch_vcpu_put_fp+0x4c/0x88
>> [  805.905752]  kvm_arch_vcpu_put+0x28/0x88
>> [  805.905764]  kvm_sched_out+0x38/0x58
> 
> (I think we run context_switch() => prepare_task_switch() here, so irqs
> are disabled)
> 
> Will
> 
> --->8
> 
> diff --git a/arch/arm64/kernel/fpsimd.c b/arch/arm64/kernel/fpsimd.c
> index fc51cdd5aaa7..a79df0804d67 100644
> --- a/arch/arm64/kernel/fpsimd.c
> +++ b/arch/arm64/kernel/fpsimd.c
> @@ -184,7 +184,8 @@ static void __get_cpu_fpsimd_context(void)
>  */
> static void get_cpu_fpsimd_context(void)
> {
> -       local_bh_disable();
> +       if (!irqs_disabled())
> +               local_bh_disable();
>        __get_cpu_fpsimd_context();
> }
> 
> @@ -205,7 +206,8 @@ static void __put_cpu_fpsimd_context(void)
> static void put_cpu_fpsimd_context(void)
> {
>        __put_cpu_fpsimd_context();
> -       local_bh_enable();
> +       if (!irqs_disabled())
> +               local_bh_enable();
> }
> 
> static bool have_cpu_fpsimd_context(void)