Re: [PATCH v2] net: fec: Use unlocked timecounter reads for saving state

[Andrew Lunn <andrew@lunn.ch>]

On Tue, Aug 30, 2022 at 05:05:24PM +0200, Csókás Bence wrote:
> 
> 
> On 2022. 08. 30. 14:29, Andrew Lunn wrote:
> > > -	fec_ptp_gettime(&fep->ptp_caps, &fep->ptp_saved_state.ts_phc);
> > > +	if (preempt_count_equals(0)) {
> > 
> > ~/linux/drivers$ grep -r preempt_count_equals *
> > ~/linux/drivers$
> > 
> > No other driver plays games like this.
> > 
> > Why not unconditionally take the lock?
> 
> Because then we would be back at the original problem (see Marc's message):
> 
> | [   14.001542] BUG: sleeping function called from invalid context at
> kernel/locking/mutex.c:283
> 
> | [   14.010604] in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 13,
> name: kworker/0:1
> 
> | [   14.018737] preempt_count: 201, expected: 0
> 
> We cannot take a mutex in atomic context. However, we also don't *need to*
> take a mutex in atomic context.

You are not taking a mutex, you are taking a spinlock. You can do that
in atomic context. Can you protect everything which needs protecting
with a spinlock? And avoid sleeping...

     Andrew