Re: [PATCH 1/2] mwifiex: Use non-posted PCI register writes

[Jonas Dreßler]

On 9/20/21 7:48 PM, Brian Norris wrote:
> On Sat, Sep 18, 2021 at 12:37 AM Jonas Dreßler <verdre@xxxxxxx> wrote:
> > Thanks for the pointer to that commit Brian, it turns out this is
> > actually the change that causes the "Firmware wakeup failed" issues that
> > I'm trying to fix with the second patch here.
>
> Huh. That's interesting, although I guess it makes some sense given
> your theory of "dropped writes". FWIW, this strategy (post a single
> write, then wait for wakeup) is the same used by some other
> chips/drivers too (e.g., ath10k/pci), although in those cases card
> wakeup is much much faster. But if the bus was dropping writes
> somehow, those strategies would fail too.
>
> > Also my approach is a lot messier than just reverting
> > 062e008a6e83e7c4da7df0a9c6aefdbc849e2bb3 and also appears to be blocking
> > even longer...
>
> For the record, in case you're talking about my data ("blocking even
> longer"): I was only testing patch 1. Patch 2 isn't really relevant to
> my particular systems (Rockchip RK3399 + Marvell 8997/PCIe), because
> (a) I'm pretty sure my system isn't "dropping" any reads or writes
> (b) all my delay is in the read-back; the Rockchip PCIe bus is waiting
> indefinitely for the card to wake up, instead of timing out and
> reporting all-1's like many x86 systems appear to do (I've tested
> this).
>
> So, the 6ms delay is entirely sitting in the ioread32(), not a delay loop.
>
> I haven't yet tried your version 2 (which avoids the blocking read to
> wake up; good!), but it sounds like in theory it could solve your
> problem while avoiding 6ms delays for me. I intend to test your v2
> this week.

With "blocking even longer" I meant that (on my system) the delay-loop 
blocks even longer than waking up the card via mwifiex_read_reg() (both 
are in the orders of milliseconds). And given that in certain cases the 
card wakeup (or a write getting through to the card, I have no idea) can 
take extremely long, I'd feel more confident going with the 
mwifiex_read_reg() method to wake up the card.

Anyway, you know what's even weirder with all this: I've been testing 
the first commit of patch v2 (so just the single read-back instead of 
the big hammer) together with 062e008a6e83e7c4da7df0a9c6aefdbc849e2bb3 
reverted for a good week now and haven't seen any wakeup failure yet. 
Otoh I'm fairly sure the big hammer with reading back every write wasn't 
enough to fix the wakeup failures, otherwise I wouldn't even have 
started working on the second commit.

So that would mean there's a difference between writing and then reading 
back vs only reading to wake up the card: Only the latter fixes the 
wakeup failures.

> > Does anyone have an idea what could be the reason for the posted write
> > not going through, or could that also be a potential firmware bug in the
> > chip?
>
> I have no clue about that. That does sound downright horrible, but so
> are many things when dealing with this family of hardware/firmware.
> I'm not sure how to prove out whether this is a host bus problem, or
> an endpoint/firmware problem, other than perhaps trying the same
> module/firmware on another system, if that's possible.
>
> Anyway, to reiterate: I'm not fundamentally opposed to v2 (pending a
> test run here), even if it is a bit ugly and perhaps not 100%
> understood.

I'm not 100% sure about all this yet, I think I'm gonna try to confirm 
my older findings once again now and then we'll see. FTR, would you be 
fine with using the mwifiex_read_reg() method to wake up the card and 
somehow quirking your system to use write_reg()?

> Brian