Re: [BUG: 6.6.y] bpf fentry hook on _raw_spin_lock_irqsave triggers hard lockup

[Greg KH <gregkh@linuxfoundation.org>]

On Wed, Jul 29, 2026 at 03:53:22PM +0800, Feng Yang wrote:
> When attaching fentry programs to _raw_spin_lock_irqsave/_raw_spin_unlock_irqrestore on
> the v6.6 kernel and using the BPF ring buffer within BPF programs, we encounter a hard LOCKUP.
> 
> Reproducible BPF code is shown below:
> 
> static inline int test(void *ctx)
> {
>         struct event *event;
> 
>         event = reserve_buf(sizeof(*event));
>         if (!event)
>                 return 0;
> 
>         submit_buf(ctx, event, sizeof(*event));
> 
>         return 0;
> }
> 
> SEC("fentry/_raw_spin_lock_irqsave")
> int BPF_PROG(aa)
> {
>         return test(ctx);
> }
> 
> SEC("fentry/_raw_spin_unlock_irqrestore")
> int BPF_PROG(bb)
> {
>         return test(ctx);
> }
> 
> We suspect the hard LOCKUP is caused by the call chain:
> aa->test->reserve_buf->_raw_spin_lock_irqsave->_raw_spin_unlock_irqrestore->bb->test->reserve_buf->_raw_spin_lock_irqsave.
> 
> Mainline has merged the rqspinlock patchset
> (https://lore.kernel.org/all/20250411101759.4061366-1-memxor@gmail.com/)
> which resolves this issue. Should we backport the fix?
> 
> Alternatively, would a fix modeled after this patch
> https://lore.kernel.org/bpf/20201029071925.3103400-3-songliubraving@fb.com/
> as outlined below be viable?

Do not do stable-only patches, that's not how the stable trees work.
Fix the issue in Linus's tree first, and then backport the needed
changes to the older trees.

thanks,

greg k-h