On Tue, Jan 17, 2017 at 11:48:22AM -0800, Brian Norris wrote:
> On Sun, Jan 15, 2017 at 04:54:52PM -0800, Dmitry Torokhov wrote:
> > On Fri, Jan 13, 2017 at 03:35:37PM -0800, Brian Norris wrote:
> > > The following sequence occurs when using IEEE power-save on 8997:
> > > (a) driver sees SLEEP event
> > > (b) driver issues SLEEP CONFIRM
> > > (c) driver recevies CMD interrupt; within the interrupt processing loop,
> > > we do (d) and (e):
> > > (d) wait for FW sleep cookie (and often time out; it takes a while), FW
> > > is putting card into low power mode
> > > (e) re-check PCIE_HOST_INT_STATUS register; quit loop with 0 value
> > >
> > > But at (e), no one actually signaled an interrupt (i.e., we didn't check
> > > adapter->int_status). And what's more, because the card is going to
> > > sleep, this register read appears to take a very long time in some cases
> > > -- 3 milliseconds in my case!
> > >
> > > Now, I propose that (e) is completely unnecessary. If there were any
> > > additional interrupts signaled after the start of this loop, then the
> > > interrupt handler would have set adapter->int_status to non-zero and
> > > queued more work for the main loop -- and we'd catch it on the next
> > > iteration of the main loop.
> > >
> > > So this patch drops all the looping/re-reading of PCIE_HOST_INT_STATUS,
> > > which avoids the problematic (and slow) register read in step (e).
> > >
> > > Incidentally, this is a very similar issue to the one fixed in commit
> > > ec815dd2a5f1 ("mwifiex: prevent register accesses after host is
> > > sleeping"), except that the register read is just very slow instead of
> > > fatal in this case.
> > >
> > > Tested on 8997 in both MSI and (though not technically supported at the
> > > moment) MSI-X mode.
> >
> > Well, that kills interrupt mitigation and with PCIE that might be
> > somewhat important (SDIO is too slow to be important I think) and might
> > cost you throughput.
>
> Hmm, well I don't see us disabling interrupts in here, at least for MSI
> mode, so it doesn't actually look like a mitigation mechanism. More like
> a redundancy. But I'm not an expert on MSI, and definitely not on
> network performance.
Well, right, maybe not mitigation, but at least you have a chance to
avoid scheduling latency at times.
>
> Also, FWIW, I did some fairly non-scientific tests of this on my
> systems, and I didn't see much difference. I can run better tests, and
> even collect data on how often we loop here vs. see new interrupts.
That would be great. Maybe packet aggregation takes care of interrupts
arriving "too closely" together most of the time, I dunno.
>
> > OTOH maybe Marvell should convert PICE to NAPI to make this more
> > obvious and probably more correct.
>
> From my brief reading, that sounds like a better way to make this
> configurable.
>
> So I'm not sure which way you'd suggest then; take a patch like this,
> which makes the driver more clear and less buggy? Or write some
> different patch that isolates just the power-save related condition, so
> we break out of this look [1]?
I think it really depends on the test results. If we do not see
degradation in both throughput and latency then I think we can take this
patch and then see if switching to NAPI would be the ultimate solution.
>
> I'm also interested in any opinions from the Marvell side -- potentially
> testing results, rationale behind this code structure, or even a better
> alternative patch.
>
> Brian
>
> [1] i.e., along the lines of commit ec815dd2a5f1 ("mwifiex: prevent
> register accesses after host is sleeping").
Thanks.
--
Dmitry
