[PATCH] arm64: defconfig: disable CONFIG_DEVMEM

Leif Lindholm leif.lindholm at linaro.org
Sun Jan 15 04:42:54 PST 2017

On Fri, Jan 13, 2017 at 11:48:01AM +0000, Will Deacon wrote:
> On Fri, Jan 13, 2017 at 11:37:34AM +0000, Leif Lindholm wrote:
> > /dev/mem is the opposite of what an operating system is for.
> > Additionally, on arm* it opens up for denial-of-service attacks from
> > userspace. So leave it disabled by default, requiring people who need
> > it to enable it explicitly.
> I really like the idea, but are we sure that nothing common breaks without
> this? For example, does Debian still boot nicely with this patch applied?

Getting distros to not have to shop around for config fragments in
order to be able to a stable system is one of my main purposes of this

Since Debian just published a 4.9 kernel, I gave that a spin (both DT
and ACPI). No issues.

> Just trying to get a feel for how much userspace this has seen (particularly
> on ACPI-based systems, which I seem to remember like poking about in here).

There are various tools that let you figure out various things about
the system (a bit like running dtc on a dump of the active
device-tree), but nothing actually used for booting a system (much
like dtc).

/dev/mem has been used for things like dmidecode and acpidump in the
past, but we fixed those tools back in 2015 to use /sys interfaces
instead of blindly groping around and hoping for the best.

Stretch version of dmidecode operates as expected on both DT and ACPI
boot, and acpidump does in the ACPI case (no tables in /sys otherwise).

There may be other tools that will also break if not implementing
support for access via /sys, but none critical for system operation
(and currently a denial-of-service waiting to happen anyway).

Systemd will try to access /dev/mem to extract boot timestamp stuff,
but handles a failure gracefully (i.e. not even a warning message).

On a side note, comparing the resulting configs, there is a
semi-broken config dependency in lib/Kconfig.debug, meaning
CONFIG_*STRICT_DEVMEM get set even if CONFIG_DEVMEM is not. 
But I'll send that out as a separate patch.


More information about the linux-arm-kernel mailing list