On Wed, Oct 17, 2018 at 11:38:11AM +0800, Hanjun Guo wrote:
> We met NVMe and RAID probe failure on 4.19-rc1+ based kernel on
> Hisilicon D06, then we bisect to this commit:
>
> cdf6b7362108 "PCI: pciehp: Always enable occupied slot on probe"
>
> Reverting this patch makes the system back to functional, I'm not
> sure why this lead some regression on our board, could you share
> some idea on this?
>
> Boot log below, [0] is the failure one, [1] is the good one,
> please let me you if you need something more.
Please cherry-pick the following linux-next commit and test if the
issue goes away:
commit 80696f991424d05a784c0cf9c314ac09ac280406
Author: Lukas Wunner <lukas@xxxxxxxxx>
Date: Sat Sep 8 09:59:01 2018 +0200
PCI: pciehp: Tolerate Presence Detect hardwired to zero
If this does not fix the issue, please provide full dmesg output for
the working and non-working case with "pciehp.pciehp_debug=1" on the
kernel command line, preferrably attached to a bugzilla entry.
Detailed analysis of the dmesg output you've already provided:
There are 9 hotplug ports but devices are only present below 4 of them:
a RAID adapter, a BMC with VGA, a 2x 10G Ethernet adapter and the NVMe drive:
0000:00:00.0 [19e5:a120] bridge to [bus 01]
0000:00:04.0 [19e5:a120] bridge to [bus 02]
0000:00:08.0 [19e5:a120] bridge to [bus 03]: 0000:03:00.0 [1000:0097] mpt3sas *
0000:00:10.0 [19e5:a120] bridge to [bus 04]: 0000:04:00.0 [19e5:1711] hibmc *
0000:00:12.0 [19e5:a120] bridge to [bus 05]: 0000:05:00.0 [8086:10fb] ixgbe
0000:05:00.1 [8086:10fb] ixgbe
0000:80:00.0 [19e5:a120] bridge to [bus 81]: 0000:81:00.0 [19e5:0123] nvme *
0000:80:08.0 [19e5:a120] bridge to [bus 82]
0000:80:0c.0 [19e5:a120] bridge to [bus 83]
0000:80:10.0 [19e5:a120] bridge to [bus 84]
Of the 4 occupied hotplug ports, 3 exhibit a hardware bug wherein the link
is reported to be up but the Presence Detect State bit in the Slot Status
register is zero. Consequently pciehp deems the slot unoccupied.
I've marked the broken ports with an asterisk *.
The above-mentioned commit works around such broken hardware but so far
we knew only of a single affected chip, the Wilocity/QCA wil6210.
It looks like this bug is more common than we thought, so it might be
worth applying it already to v4.19. However v4.19 is expected to be
released in 4 days, so backporting to a 4.19 stable release might be
more practical than squeezing the commit in before the coming Sunday.
For some reason the hotplug port 0000:00:12.0 with the 2x 10G Ethernet
adapter does not exhibit the hardware bug.
HTH,
Lukas
