On Wed, 28 Oct 2020 13:17:58 +0100 Heiner Kallweit wrote:
> On 28.10.2020 12:43, Serge Belyshev wrote:
> >> For several network drivers it was reported that using
> >> __napi_schedule_irqoff() is unsafe with forced threading. One way to
> >> fix this is switching back to __napi_schedule, but then we lose the
> >> benefit of the irqoff version in general. As stated by Eric it doesn't
> >> make sense to make the minimal hard irq handlers in drivers using NAPI
> >> a thread. Therefore ensure that the hard irq handler is never
> >> thread-ified.
> >>
> >> Fixes: 9a899a35b0d6 ("r8169: switch to napi_schedule_irqoff")
> >> Link: https://lkml.org/lkml/2020/10/18/19
> >> Signed-off-by: Heiner Kallweit <hkallwe...@gmail.com>
> >> ---
> >> drivers/net/ethernet/realtek/r8169_main.c | 8 ++++----
> >> 1 file changed, 4 insertions(+), 4 deletions(-)
> >>
> >> diff --git a/drivers/net/ethernet/realtek/r8169_main.c
> >> b/drivers/net/ethernet/realtek/r8169_main.c
> >> index 7d366b036..3b6ddc706 100644
> >> --- a/drivers/net/ethernet/realtek/r8169_main.c
> >> +++ b/drivers/net/ethernet/realtek/r8169_main.c
> > ...
> >
> > Hi! This patch actually breaks r8169 with threadirqs on an old box
> > where it was working before:
> >
> > [ 0.000000] DMI: Gigabyte Technology Co., Ltd.
> > GA-MA790FX-DQ6/GA-MA790FX-DQ6, BIOS F7g 07/19/2010
> > ...
> > [ 1.072676] r8169 0000:02:00.0 eth0: RTL8168b/8111b, 00:1a:4d:5d:6b:c3,
> > XID 380, IRQ 18
> > ...
> > [ 8.850099] genirq: Flags mismatch irq 18. 00010080 (eth0) vs. 00002080
> > (ahci[0000:05:00.0])
> >
> > (error is reported to userspace, interface failed to bring up).
> > Reverting the patch fixes the problem.
> >
> Thanks for the report. On this old chip version MSI is unreliable,
> therefore r8169 falls back to a PCI legacy interrupt. On your system
> this PCI legacy interrupt seems to be shared between network and
> disk. Then the IRQ core tries to threadify the disk interrupt
> (setting IRQF_ONESHOT), whilst the network interrupt doesn't have
> this flag set. This results in the flag mismatch error.
>
> Maybe, if one source of a shared interrupt doesn't allow forced
> threading, this should be applied to the other sources too.
> But this would require a change in the IRQ core, therefore
> +Thomas to get his opinion on the issue.
Other handles may take spin_locks, which will sleep on RT.
I guess we may need to switch away from the _irqoff() variant for
drivers with IRQF_SHARED after all :(
