On Mon, Sep 28 2020 at 21:05, Edward Cree wrote:

efx_ef10_try_update_nic_stats_vf() used in_interrupt() to figure out
 whether it is safe to sleep (for MCDI) or not.
The only caller from which it was not is efx_net_stats(), which can be
 invoked under dev_base_lock from net-sysfs::netstat_show().
So add a new update_stats_atomic() method to struct efx_nic_type, and
 call it from efx_net_stats(), removing the need for
 efx_ef10_try_update_nic_stats_vf() to behave differently for this case
 (which it wasn't doing correctly anyway).
For all nic_types other than EF10 VF, this method is NULL and so we
 call the regular update_stats() methods, which are happy with being
 called from atomic contexts.

Fixes: f00bf2305cab ("sfc: don't update stats on VF when called in atomic context")
Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Edward Cree <redacted>

That's much nicer.

---
Only compile-tested so far, because I'm waiting for my kernel to
 finish rebuilding with CONFIG_DEBUG_ATOMIC_SLEEP which I'm hoping
 is the right thing to detect the bug in the existing code.
I also wasn't quite sure how to give credit to the thorough analysis
 in the commit message of Sebastian's patch.  I don't think we have
 a Whatever-by: tag to cover that, do we?

Sebastian did the analysis and I did some word polishing, but the credit
surely goes to him.

Thanks,

        tglx