Hi Robert,
On Tue, Jun 06, 2023 at 07:40:15PM +0000, Robert Hancock wrote:
> On Tue, 2023-06-06 at 21:24 +0200, Andi Shyti wrote:
> > Hi Robert,
> >
> > On Tue, Jun 06, 2023 at 12:25:58PM -0600, Robert Hancock wrote:
> > > In xiic_process, it is possible that error events such as
> > > arbitration
> > > lost or TX error can be raised in conjunction with other interrupt
> > > flags
> > > such as TX FIFO empty or bus not busy. Error events result in the
> > > controller being reset and the error returned to the calling
> > > request,
> > > but the function could potentially try to keep handling the other
> > > events, such as by writing more messages into the TX FIFO. Since
> > > the
> > > transaction has already failed, this is not helpful and will just
> > > cause
> > > issues.
> >
> > what kind of issues?
> >
>
> The one I ran into is what I alluded to further down, where that
> WARN_ON was triggering repeatedly, which ended up flooding the kernel
> log and causing the device's watchdog timer to timeout. I'm not sure
> what other forms of havoc might ensue, other than "nothing good"..
>
> > > This problem has been present ever since:
> > >
> > > commit 7f9906bd7f72 ("i2c: xiic: Service all interrupts in isr")
> > >
> > > which allowed non-error events to be handled after errors, but
> > > became
> > > more obvious after:
> > >
> > > commit 743e227a8959 ("i2c: xiic: Defer xiic_wakeup() and
> > > __xiic_start_xfer() in xiic_process()")
> > >
> > > which reworked the code to add a WARN_ON which triggers if both the
> > > xfer_more and wakeup_req flags were set, since this combination is
> > > not supposed to happen, but was occurring in this scenario.
> > >
> > > Skip further interrupt handling after error flags are detected to
> > > avoid
> > > this problem.
> > >
> > > Fixes: 7f9906bd7f72 ("i2c: xiic: Service all interrupts in isr")
> > > Signed-off-by: Robert Hancock <robert.hancock@xxxxxxxxxx>
> >
> > please add:
> >
> > Cc: <stable@xxxxxxxxxxxxxxx> # v4.3+
> >
>
> Can add for a v2 - although with the Fixes tag it would most likely be
> picked up for stable anyway..
It's just a courtesy to stable maintainers :)
> > > ---
> > > drivers/i2c/busses/i2c-xiic.c | 2 ++
> > > 1 file changed, 2 insertions(+)
> > >
> > > diff --git a/drivers/i2c/busses/i2c-xiic.c
> > > b/drivers/i2c/busses/i2c-xiic.c
> > > index 8a3d9817cb41..ee6edc963dea 100644
> > > --- a/drivers/i2c/busses/i2c-xiic.c
> > > +++ b/drivers/i2c/busses/i2c-xiic.c
> > > @@ -721,6 +721,8 @@ static irqreturn_t xiic_process(int irq, void
> > > *dev_id)
> > > wakeup_req = 1;
> > > wakeup_code = STATE_ERROR;
> > > }
> > > + /* don't try to handle other events */
> > > + goto out;
> >
> > why don't we have goto's after every irq evaluation but only
> > here? Do the issues you mentioned happen olny in this particular
> > error case?
> >
>
> As far as I can tell, yes. For example you could legitimately have
> XIIC_INTR_TX_EMPTY_MASK and XIIC_INTR_BNB_MASK both being handled. The
> error case is special as we end up resetting the controller, so it
> doesn't make sense to try to continue with the rest of the transaction
> while also completing it with an error.
I think the patch is correct and I will ack it:
Acked-by: Andi Shyti <andi.shyti@xxxxxxxxxx>
I think, though, that this needs a proper fix and testing, in
order to cover all the possible combinations. The scenario you
highlighted is indeed one, but not only, potential situation that
could arise.
Can I just ask you to write a bit more in the comment to
highlight the possible failure?
Thanks a lot,
Andi
> > Thanks,
> > Andi
> >
> > > }
> > > if (pend & XIIC_INTR_RX_FULL_MASK) {
> > > /* Receive register/FIFO is full */
> > > --
> > > 2.40.1
> > >
>
> --
> Robert Hancock <robert.hancock@xxxxxxxxxx>
