Re: [PATCH v2 2/2] mwifiex: pcie: add reset_d3cold quirk for Surface gen4+ devices

From: Maximilian Luz
Date: Fri Jul 09 2021 - 17:00:22 EST


On 7/9/21 9:44 PM, Pali Rohár wrote:
On Friday 09 July 2021 21:27:51 Maximilian Luz wrote:
On 7/9/21 8:44 PM, Pali Rohár wrote:

[...]

My (very) quick attempt ('echo 1 > /sys/bus/pci/.../reset) at
reproducing this didn't work, so I think at very least a network
connection needs to be active.

This is doing PCIe function level reset. Maybe you can get more luck
with PCIe Hot Reset. See following link how to trigger PCIe Hot Reset
from userspace: https://alexforencich.com/wiki/en/pcie/hot-reset-linux

Thanks for that link! That does indeed do something which breaks the
adapter. Running the script produces

[ 178.388414] mwifiex_pcie 0000:01:00.0: PREP_CMD: card is removed
[ 178.389128] mwifiex_pcie 0000:01:00.0: PREP_CMD: card is removed
[ 178.461365] mwifiex_pcie 0000:01:00.0: performing cancel_work_sync()...
[ 178.461373] mwifiex_pcie 0000:01:00.0: cancel_work_sync() done
[ 178.984106] pci 0000:01:00.0: [11ab:2b38] type 00 class 0x020000
[ 178.984161] pci 0000:01:00.0: reg 0x10: [mem 0x00000000-0x000fffff 64bit pref]
[ 178.984193] pci 0000:01:00.0: reg 0x18: [mem 0x00000000-0x000fffff 64bit pref]
[ 178.984430] pci 0000:01:00.0: supports D1 D2
[ 178.984434] pci 0000:01:00.0: PME# supported from D0 D1 D3hot D3cold
[ 178.984871] pcieport 0000:00:1c.0: ASPM: current common clock configuration is inconsistent, reconfiguring
[ 179.297919] pci 0000:01:00.0: BAR 0: assigned [mem 0xd4400000-0xd44fffff 64bit pref]
[ 179.297961] pci 0000:01:00.0: BAR 2: assigned [mem 0xd4500000-0xd45fffff 64bit pref]
[ 179.298316] mwifiex_pcie 0000:01:00.0: enabling device (0000 -> 0002)
[ 179.298752] mwifiex_pcie: PCI memory map Virt0: 00000000c4593df1 PCI memory map Virt2: 0000000039d67daf
[ 179.300522] mwifiex_pcie 0000:01:00.0: WLAN read winner status failed!
[ 179.300552] mwifiex_pcie 0000:01:00.0: info: _mwifiex_fw_dpc: unregister device
[ 179.300622] mwifiex_pcie 0000:01:00.0: Read register failed
[ 179.300912] mwifiex_pcie 0000:01:00.0: performing cancel_work_sync()...
[ 179.300928] mwifiex_pcie 0000:01:00.0: cancel_work_sync() done

after which the card is unusable (there is no WiFi interface availabel
any more). Reloading the driver module doesn't help and produces

[ 376.906833] mwifiex_pcie: PCI memory map Virt0: 0000000025149d28 PCI memory map Virt2: 00000000c4593df1
[ 376.907278] mwifiex_pcie 0000:01:00.0: WLAN read winner status failed!
[ 376.907281] mwifiex_pcie 0000:01:00.0: info: _mwifiex_fw_dpc: unregister device
[ 376.907293] mwifiex_pcie 0000:01:00.0: Read register failed
[ 376.907404] mwifiex_pcie 0000:01:00.0: performing cancel_work_sync()...
[ 376.907406] mwifiex_pcie 0000:01:00.0: cancel_work_sync() done

again. Performing a function level reset produces

[ 402.489572] mwifiex_pcie 0000:01:00.0: mwifiex_pcie_reset_prepare: adapter structure is not valid
[ 403.514219] mwifiex_pcie 0000:01:00.0: mwifiex_pcie_reset_done: adapter structure is not valid

and doesn't help either.

More Qualcomm/Atheros wifi cards are broken in a way that they stop
responding after PCIe Hot Reset and completely disappear from the PCIe
bus. It is possible that similar issue have also these Marvell cards?

As now we know that bride does not support hotplug it cannot inform
system when card disconnect from the bus. The one possible way how to
detect if PCIe card is available at specific address is trying to read
its device and vendor id. Kernel caches device/vendor id, so from
userspace is needed to call lspci with -b argument (to ignore kernel's
reported cached value). Could you provide output of following command
after you do PCIe Hot Reset?

lspci -s 01:00.0 -b -vv

(and compare with output which you have already provided if there are
any differences)

There do seem to be some differences, specifically regarding memory.
See https://paste.ubuntu.com/p/Rz2CDjwkCv/ for the full output.

If PCIe Hot Reset is breaking the card then the only option how to
"reset" card into working state again is PCIe Warm Reset. Unfortunately
there is no common mechanism how to do it from system. PCIe Warm Reset
is done by asserting PERST# signal on card itself, in mPCIe form factor
it is pin 22. In most cases pin 22 is connected to some GPIO so via GPIO
subsystem it could be controlled.

Running the same command on a kernel with (among other) this patch
unfortunately also breaks the adapter in the same way. As far as I can
tell though, it doesn't run through the reset code added by this patch
(as indicated by the log message when performing FLR), which I assume
in a non-forced scenario, e.g. firmware issues (which IIRC is why this
patch exists), it would?

Err... I have caught this part. Is proposed patch able to recover also
from state which happens after PCIe Hot Reset?

I'm not sure at this point if the power-cycle through D3cold would fix
this (I think it might, but have no evidence for that). This patch alone
isn't able to recover the device, as, when triggering the hot-reset via
that script, the code never seems to run mwifiex_pcie_reset_d3cold_quirk().

If I remember correctly, the main issue was that the firmware state
doesn't get reset completely. This can be somewhat observed when doing
'echo 1 > /sys/bus/pci/devices/.../reset' via the difference in log
messages:

For an unpatched kernel:

[ 64.159509] mwifiex_pcie 0000:01:00.0: info: shutdown mwifiex...
[ 64.159546] mwifiex_pcie 0000:01:00.0: PREP_CMD: card is removed
[ 64.159922] mwifiex_pcie 0000:01:00.0: PREP_CMD: card is removed
[ 65.240272] mwifiex_pcie 0000:01:00.0: WLAN FW already running! Skip FW dnld
[ 65.240285] mwifiex_pcie 0000:01:00.0: WLAN FW is active
[ 65.327359] mwifiex_pcie 0000:01:00.0: info: MWIFIEX VERSION: mwifiex 1.0 (15.68.19.p21)
[ 65.327370] mwifiex_pcie 0000:01:00.0: driver_version = mwifiex 1.0 (15.68.19.p21)

For a patched kernel:

[ 41.966094] mwifiex_pcie 0000:01:00.0: info: shutdown mwifiex...
[ 41.966451] mwifiex_pcie 0000:01:00.0: PREP_CMD: card is removed
[ 41.967227] mwifiex_pcie 0000:01:00.0: PREP_CMD: card is removed
[ 42.063543] mwifiex_pcie 0000:01:00.0: Using reset_d3cold quirk to perform FW reset
[ 42.063558] mwifiex_pcie 0000:01:00.0: putting into D3cold...
[ 42.081010] usb 1-6: USB disconnect, device number 9
[ 42.339922] pcieport 0000:00:1c.0: putting into D3cold...
[ 42.425766] pcieport 0000:00:1c.0: putting into D0...
[ 42.695987] mwifiex_pcie 0000:01:00.0: putting into D0...
[ 42.956673] mwifiex_pcie 0000:01:00.0: enabling device (0000 -> 0002)
[ 45.012736] mwifiex_pcie 0000:01:00.0: info: FW download over, size 723540 bytes
[ 45.740882] usb 1-6: new high-speed USB device number 10 using xhci_hcd
[ 45.881294] usb 1-6: New USB device found, idVendor=1286, idProduct=204c, bcdDevice=32.01
[ 45.881308] usb 1-6: New USB device strings: Mfr=1, Product=2, SerialNumber=3
[ 45.881313] usb 1-6: Product: Bluetooth and Wireless LAN Composite Device
[ 45.881318] usb 1-6: Manufacturer: Marvell
[ 45.881322] usb 1-6: SerialNumber: 0000000000000000
[ 45.884882] Bluetooth: hci0: unexpected event for opcode 0x0000
[ 45.885128] Bluetooth: hci0: unexpected event for opcode 0x0000
[ 45.966218] mwifiex_pcie 0000:01:00.0: WLAN FW is active
[ 46.157474] mwifiex_pcie 0000:01:00.0: info: MWIFIEX VERSION: mwifiex 1.0 (15.68.19.p21)
[ 46.157485] mwifiex_pcie 0000:01:00.0: driver_version = mwifiex 1.0 (15.68.19.p21)

Note the absence of "WLAN FW already running! Skip FW dnld" on the
second log. Due to the power cycle we're essentially forcing the device
to re-download and re-initialize its firmware. On an unpatched kernel,
it looks like the firmware itself is kept and state may not be cleared
properly. So in other words it looks like the firmware, while being
prompted to do a reset, doesn't do that properly (at least when it has
crashed before).

IIRC this then allowed us to recover from firmware issues that the
"normal" firmware reset that mwifiex supposedly performs on a FLR
(or reloading driver modules) didn't help with.

I'm not so sure any more if resetting actively caused issues or if it
just showed different symptoms of some firmware issue that prompted us
to do the reset in the first place (again, been quite a while since I
last dealt with this stuff, sorry). All I know is that this patched
reset gets the card going again.

As a side note: There are also more patches by Jonas (and Tsuchiya?)
building on top of the quirk implementation introduced here which
significantly reduce the need for doing resets in the first place
(nevertheless having a reset that actually does properly reset the
device is a good thing IMHO). Those patches can for example be found
here:

https://github.com/linux-surface/kernel/compare/eaaf96ba58a5fe5999b89fe3afaded74caa96480...989c8725a6d4e62db6370dd0fefe45498274d3ce

Unfortunately I can't test that with a
network connection (and without compiling a custom kernel for which I
don't have the time right now) because there's currently another bug
deadlocking on device removal if there's an active connection during
removal (which also seems to trigger on reset). That one ill be fixed
by

https://lore.kernel.org/linux-wireless/20210515024227.2159311-1-briannorris@xxxxxxxxxxxx/

Jonas might know more.

[...]