Re: PCI trouble on mvebu (Turris Omnia)

["Toke Høiland-Jørgensen" <toke@redhat.com>]

Pali Rohár <pali@kernel.org> writes:

> On Wednesday 28 October 2020 18:16:26 Bjorn Helgaas wrote:
>> [+cc Pali, Marek, Thomas, Jason]
>> 
>> On Wed, Oct 28, 2020 at 04:40:00PM +0000, ™֟☻̭҇ Ѽ ҉ ® wrote:
>> > On 28/10/2020 16:08, Toke Høiland-Jørgensen wrote:
>> > > Bjorn Helgaas <helgaas@kernel.org> writes:
>> > > > On Wed, Oct 28, 2020 at 02:36:13PM +0100, Toke Høiland-Jørgensen wrote:
>> > > > > Toke Høiland-Jørgensen <toke@redhat.com> writes:
>> > > > > > Bjorn Helgaas <helgaas@kernel.org> writes:
>> > > > > > 
>> > > > > > > [+cc vtolkm]
>> > > > > > > 
>> > > > > > > On Tue, Oct 27, 2020 at 04:43:20PM +0100, Toke Høiland-Jørgensen wrote:
>> > > > > > > > Hi everyone
>> > > > > > > > 
>> > > > > > > > I'm trying to get a mainline kernel to run on my Turris Omnia, and am
>> > > > > > > > having some trouble getting the PCI bus to work correctly. Specifically,
>> > > > > > > > I'm running a 5.10-rc1 kernel (torvalds/master as of this moment), with
>> > > > > > > > the resource request fix[0] applied on top.
>> > > > > > > > 
>> > > > > > > > The kernel boots fine, and the patch in [0] makes the PCI devices show
>> > > > > > > > up. But I'm still getting initialisation errors like these:
>> > > > > > > > 
>> > > > > > > > [    1.632709] pci 0000:01:00.0: BAR 0: error updating (0xe0000004 != 0xffffffff)
>> > > > > > > > [    1.632714] pci 0000:01:00.0: BAR 0: error updating (high 0x000000 != 0xffffffff)
>> > > > > > > > [    1.632745] pci 0000:02:00.0: BAR 0: error updating (0xe0200004 != 0xffffffff)
>> > > > > > > > [    1.632750] pci 0000:02:00.0: BAR 0: error updating (high 0x000000 != 0xffffffff)
>> > > > > > > > 
>> > > > > > > > and the WiFi drivers fail to initialise with what appears to me to be
>> > > > > > > > errors related to the bus rather than to the drivers themselves:
>> > > > > > > > 
>> > > > > > > > [    3.509878] ath: phy0: Mac Chip Rev 0xfffc0.f is not supported by this driver
>> > > > > > > > [    3.517049] ath: phy0: Unable to initialize hardware; initialization status: -95
>> > > > > > > > [    3.524473] ath9k 0000:01:00.0: Failed to initialize device
>> > > > > > > > [    3.530081] ath9k: probe of 0000:01:00.0 failed with error -95
>> > > > > > > > [    3.536012] ath10k_pci 0000:02:00.0: of_irq_parse_pci: failed with rc=134
>> > > > > > > > [    3.543049] pci 0000:00:02.0: enabling device (0140 -> 0142)
>> > > > > > > > [    3.548735] ath10k_pci 0000:02:00.0: can't change power state from D3hot to D0 (config space inaccessible)
>> > > > > > > > [    3.588592] ath10k_pci 0000:02:00.0: failed to wake up device : -110
>> > > > > > > > [    3.595098] ath10k_pci: probe of 0000:02:00.0 failed with error -110
>> > > > > > > > 
>> > > > > > > > lspci looks OK, though:
>> > > > > > > > 
>> > > > > > > > # lspci
>> > > > > > > > 00:01.0 PCI bridge: Marvell Technology Group Ltd. Device 6820 (rev 04)
>> > > > > > > > 00:02.0 PCI bridge: Marvell Technology Group Ltd. Device 6820 (rev 04)
>> > > > > > > > 00:03.0 PCI bridge: Marvell Technology Group Ltd. Device 6820 (rev 04)
>> > > > > > > > 01:00.0 Network controller: Qualcomm Atheros AR9287 Wireless Network Adapter (PCI-Express) (rev 01)
>> > > > > > > > 02:00.0 Network controller: Qualcomm Atheros QCA986x/988x 802.11ac Wireless Network Adapter (rev ff)
>> > > > > > > > 
>> > > > > > > > Does anyone have any clue what could be going on here? Is this a bug, or
>> > > > > > > > did I miss something in my config or other initialisation? I've tried
>> > > > > > > > with both the stock u-boot distributed with the board, and with an
>> > > > > > > > upstream u-boot from latest master; doesn't seem to make any different.
>> > > > > > > Can you try turning off CONFIG_PCIEASPM?  We had a similar recent
>> > > > > > > report at https://bugzilla.kernel.org/show_bug.cgi?id=209833 but I
>> > > > > > > don't think we have a fix yet.
>> > > > > > Yes! Turning that off does indeed help! Thanks a bunch :)
>> > > > > > 
>> > > > > > You mention that bisecting this would be helpful - I can try that
>> > > > > > tomorrow; any idea when this was last working?
>> > > > > OK, so I tried to bisect this, but, erm, I couldn't find a working
>> > > > > revision to start from? I went all the way back to 4.10 (which is the
>> > > > > first version to include the device tree file for the Omnia), and even
>> > > > > on that, the wireless cards were failing to initialise with ASPM
>> > > > > enabled...
>> > > > I have no personal experience with this device; all I know is that the
>> > > > bugzilla suggests that it worked in v5.4, which isn't much help.
>> > > > 
>> > > > Possibly the apparent regression was really a .config change, i.e.,
>> > > > CONFIG_PCIEASPM was disabled in the v5.4 kernel vtolkm@ tested and it
>> > > > "worked" but got enabled later and it started failing?
>> > > Yeah, I suspect so. The OpenWrt config disables CONFIG_PCIEASPM by
>> > > default and only turns it on for specific targets. So I guess that it's
>> > > most likely that this has never worked...
>> > > 
>> > > > Maybe the debug patch below would be worth trying to see if it makes
>> > > > any difference?  If it *does* help, try omitting the first hunk to see
>> > > > if we just need to apply the quirk_enable_clear_retrain_link() quirk.
>> > > Tried, doesn't help...
>> > > 
>> > > -Toke
>> > 
>> > Found this patch
>> > 
>> > https://github.com/openwrt/openwrt/blob/7c0496f29bed87326f1bf591ca25ace82373cfc7/target/linux/mvebu/patches-5.4/405-PCI-aardvark-Improve-link-training.patch
>> > 
>> > that mentions the Compex WLE900VX card, which reading the lspci verbose
>> > output from the bugtracker seems to the device being troubled.
>> 
>> Interesting.  Indeed, the Compex WLE900VX card seems to have the
>> Qualcomm Atheros QCA9880 on it, and it looks like Toke's system has
>> the same device in it.
>> 
>> The patch you mention (https://git.kernel.org/linus/43fc679ced18) is
>> for aardvark, so of course doesn't help mvebu.
>> 
>> PCIe hardware is supposed to automatically negotiate the highest link
>> speed supported by both ends.  But software *is* allowed to set an
>> upper limit (the Target Link Speed in Link Control 2).  If we initiate
>> a retrain and the link doesn't come back up, I wonder if we should try
>> to help the hardware out by using Target Link Speed to limit to a
>> lower speed and attempting another retrain, something like this hacky
>> patch: (please collect the dmesg log if you try this)
>
> My experience with that WLE900VX card, aardvark driver and aspm code:
>
> Link training in GEN2 mode for this card succeed only once after reset.
> Repeated link retraining fails and it fails even when aardvark is
> reconfigured to GEN1 mode. Reset via PERST# signal is required to have
> working link training.
>
> What I did in aardvark driver: Set mode to GEN2, do link training. If
> success read "negotiated link speed" from "Link Control Status Register"
> (for WLE900VX it is 0x1 - GEN1) and set it into aardvark. And then
> retrain link again (for WLE900VX now it would be at GEN1). After that
> card is stable and all future retraining (e.g. from aspm.c) also passes.
>
> If I do not change aardvark mode from GEN2 to GEN1 the second link
> training fails. And if I change mode to GEN1 after this failed link
> training then nothing happen, link training do not success.
>
> So just speculation now... In current setup initialization of card does
> one link training at GEN2. Then aspm.c is called which is doing second
> link retraining at GEN2. And if it fails then below patch issue third
> link retraining at GEN1. If A38x/pci-mvebu has same problem as aardvark
> then second link retraining must be at GEN1 (not GEN2) to workaround
> this issue.
>
> Bjorn, Toke: what about trying to hack aspm.c code to never do link
> retraining at GEN2 speed? And always force GEN1 speed prior link
> training?

Sounds like a plan. I poked around in aspm.c and must confess to being a
bit lost in the soup of registers ;)

So if one of you can cook up a patch, that would be most helpful!

-Toke