Re: [PATCH] coresight: etm3x: convert struct etm_drvdata's spinlock to raw_spinlock

[Suzuki K Poulose]

On 11/07/2023 16:45, James Clark wrote:
> On 11/07/2023 15:05, Greg Kroah-Hartman wrote:
> > On Tue, Jul 11, 2023 at 03:05:36PM +0800, quanyang.wang@xxxxxxxxxxxxx wrote:
> > > From: Quanyang Wang <quanyang.wang@xxxxxxxxxxxxx>
> > >
> > > For PREEMPT_RT kernel, spinlock_t locks become sleepable. The functions
> > > etm_dying_cpu and etm_starting_cpu which call spin_lock/unlock run in
> > > an irq-disabled context, this will trigger the following calltrace:
> > >
> > >      BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:46
> > >      in_atomic(): 1, irqs_disabled(): 128, non_block: 0, pid: 25, name: migration/1
> > >      preempt_count: 1, expected: 0
> > >      RCU nest depth: 0, expected: 0
> > >      1 lock held by migration/1/25:
> > >       #0: 82a7587c (&drvdata->spinlock){....}-{2:2}, at: etm_dying_cpu+0x28/0x54
> > >      Preemption disabled at:
> > >      [<801ec760>] cpu_stopper_thread+0x94/0x120
> > >      CPU: 1 PID: 25 Comm: migration/1 Not tainted 6.1.35-rt10-yocto-preempt-rt #30
> > >      Hardware name: Xilinx Zynq Platform
> > >      Stopper: multi_cpu_stop+0x0/0x174 <- __stop_cpus.constprop.0+0x48/0x88
> > >       unwind_backtrace from show_stack+0x18/0x1c
> > >       show_stack from dump_stack_lvl+0x58/0x70
> > >       dump_stack_lvl from __might_resched+0x14c/0x1c0
> > >       __might_resched from rt_spin_lock+0x4c/0x84
> > >       rt_spin_lock from etm_dying_cpu+0x28/0x54
> > >       etm_dying_cpu from cpuhp_invoke_callback+0x140/0x33c
> > >       cpuhp_invoke_callback from __cpuhp_invoke_callback_range+0xa4/0x104
> > >       __cpuhp_invoke_callback_range from take_cpu_down+0x7c/0xa8
> > >       take_cpu_down from multi_cpu_stop+0x15c/0x174
> > >       multi_cpu_stop from cpu_stopper_thread+0x9c/0x120
> > >       cpu_stopper_thread from smpboot_thread_fn+0x31c/0x360
> > >       smpboot_thread_fn from kthread+0x100/0x124
> > >       kthread from ret_from_fork+0x14/0x2c
> > >
> > > Convert struct etm_drvdata's spinlock to raw_spinlock to fix it.
> >
> > wait, why will a raw_spinlock fix this?  Why not fix the root problem
> > here, that of calling these locks inproperly in irq context?
> >
> > How is changing to a raw_spinlock going to fix the above splat?
> >
> > thanks,
> >
> > greg k-h
>
> If it's just etm_starting_cpu() and etm_dying_cpu() as mentioned in the
> commit message then can those spinlocks be removed?
>
> Surely there can't be any concurrent access to the per-cpu data when the
> hotplug callbacks are called?

Accessing the per-cpu data is not a problem. The spinlocks are there to
protect the accesses to drvdata->mode. etm_starting_cpu() would try to
enable the etm (i.e., start the tracing) if the mode is not DISABLED.
Especially for SYSFS mode, this could be controlled from a different
CPU, affecting the mode. I think we may still be able to avoid this
lock, by allowing the modifications to the mode performed via
enable_hw/disable_hw on the CPU. That way, there cannot be concurrent
modifications to the mode for a given ETM bound to the CPU.

Suzuki