arm64: 4.14 of_match_node() issues

Hi All,

On 2017-10-09 14:04, Robin Murphy wrote:

On 09/10/17 11:58, Robin Murphy wrote:

quoted

On 09/10/17 10:24, Will Deacon wrote:

quoted

Hi Andreas,

On Sat, Oct 07, 2017 at 02:42:13AM +0200, Andreas F?rber wrote:

quoted

Since 4.14-rc1 I am seeing frequent oopses during module loading (e.g.,
MMC, USB) from initrd on aarch64. Symptoms are similar to this in -rc3:

[  OK  ] Started udev Coldplug all Devices.
[   10.117775] usbcore: registered new interface driver usbfs
[   10.118235] Unable to handle kernel paging request at virtual address
ffff000008e5abc0
[   10.118238] Mem abort info:
[   10.118245]   Exception class = DABT (current EL), IL = 32 bits
[   10.118249]   SET = 0, FnV = 0
[   10.118253]   EA = 0, S1PTW = 0
[   10.118256] Data abort info:
[   10.118261]   ISV = 0, ISS = 0x00000006
[   10.118264]   CM = 0, WnR = 0
[   10.118274] swapper pgtable: 4k pages, 48-bit VAs, pgd = ffff0000094a5000
[   10.118279] [ffff000008e5abc0] *pgd=00000000bfffe003,
*pud=00000000bfffd003, *pmd=0000000000000000
[   10.118299] Internal error: Oops: 96000006 [#1] SMP
[   10.118305] Modules linked in: fixed usbcore(+) sunxi_mmc mmc_core
phy_sun4i_usb sg
[   10.118341] CPU: 3 PID: 49 Comm: kworker/3:1 Not tainted
4.14.0-rc3-2.gf27997b-default #1
[   10.118345] Hardware name: sunxi sunxi/sunxi, BIOS 2017.05-rc1 04/13/2017
[   10.118369] Workqueue: events deferred_probe_work_func
[   10.118378] task: ffff80007c8f4000 task.stack: ffff0000099d8000
[   10.118394] PC is at __of_match_node.part.1+0x48/0x88
[   10.118403] LR is at of_match_node+0x40/0x70
[   10.118411] pc : [<ffff00000879aed0>] lr : [<ffff00000879af50>]

[...]

quoted

This has been observed on Pine64 (>60%; also by Stefan) and Odroid-C2;
my other arm64 boards such as Raspberry Pi 3 have not run into this so
far. No such problems on 32-bit boards.

This is using the openSUSE config:
https://kernel.opensuse.org/cgit/kernel-source/plain/config/arm64/default

Hmm, hard to know what to suggest without a concrete reproducer. Do you know
which driver is being probed in the log above? Also, does this still break
if you pass "keepinitrd" on the cmdline? Finally, can you dump the kernel
virtual memory layout, please?

FWIW, this looks a lot like what happens when a built-in driver's
of_match_table is marked __init, but for whatever reason (deferred probe
etc.) winds up getting poked by a module load after it no longer exists.

Heh, synchronicity...

https://lists.linuxfoundation.org/pipermail/iommu/2017-October/024572.html

Looks like we either have to revert plenty of patches from the const
brigade, avoid freeing init, or come up with some way to make the driver
core cleverer about the whole deal :(

It's not that bad. I did a quick check with

# git grep "of_device_id.*init"

and Exynos SYSMMU driver was the only candidate for a fix. Maybe someone
else should double check that list to make sure that there is no other
platform device/driver related code there.

Best regards
-- 
Marek Szyprowski, PhD
Samsung R&D Institute Poland