Could ASLR be used to prevent the Spectre attack?

The way Spectre mitigations are shaping up, it's going to require modification of every program that deals with sensitive data, inserting serialization instructions in the right places. Or programs can be compiled with all branch prediction disabled, with more of a speed hit.

Either way, that's going to be piecemeal and error-prone. We'll be stuck with a new class of vulnerabilities for a long time. Perhaps good news for the security industry, but it's going to become as tediously bad as buffer overflows for the rest of us.

Also, so far the mitigations being developed for Spectre only cover branching, but the Spectre paper also suggests the attack can be used in the absence of branches to eg determine the contents of registers, as long as the attacker knows the address of suitable instructions to leverage.

So I wonder if a broader mitigation can be developed, if only to provide some defense in depth from Spectre.

The paper on Spectre has this to say about ASLR:

The algorithm that tracks and matches jump histories appears to use only the low bits of the virtual address (which are further reduced by simple hash function). As a result, an adversary does not need to be able to even execute code at any of the memory addresses containing the victim’s branch instruction. ASLR can also be compensated, since upper bits are ignored and bits 15..0 do not appear to be randomized with ASLR in Win32 or Win64.

I think ASLR on Linux also doesn't currently randomize those bits of the address; cat /proc/self/maps shows the lower 3 bytes always 0. But, could it be made to randomize more of the address, and so guard against Spectre?

(Then we'd only have to get all binaries built PIE, which is easily audited for and distributions are well on their way toward already. Indeed, most executables on my laptop are already built PIE, with the notable exception of /bin/bash and all haskell programs.)

I don't have an answer to this question, but am very curious about thoughts from anyone who knows about ASLR and low level CPU details. I have not been able to find any discussion of it aside from the above part of the Spectre paper.

I wanted to use 1-wire temperature sensors (DS18B20) with my Cubietruck board, running Debian. The only page I could find documenting this is for the sunxi kernel, not the mainline kernel Debian uses. After a couple of hours of research I got it working, so here goes.

wiring

First you need to pick a GPIO pin to use for the 1-wire signal. The Cubietruck's GPIO pins are documented here, and I chose to use pin PG8. Other pins should work as well, although I originally tried to use PB17 and could not get it to work for an unknown reason. I also tried to use PB18 but there was a conflict with something else trying to use that same pin. To find a free pin, cat /sys/kernel/debug/pinctrl/1c20800.pinctrl/pinmux-pins and look for a line like: "pin 200 (PG8): (MUX UNCLAIMED) (GPIO UNCLAIMED)"

Now wire the DS18B20 sensor up. With its flat side facing you, the left pin goes to ground, the center pin to PG8 (or whatever GPIO pin you selected), and the right pin goes to 3.3V. Don't forget to connect the necessary 4.7K ohm resistor between the center and right pins.

You can find plenty of videos showing how to wire up the DS18B20 on youtube, which typically also involve a quick config change to a Raspberry Pi running Raspbian to get it to see the sensor. With Debian it's unfortunately quite a lot more complicated, and so this blog post got kind of long.

configuration

We need to get the kernel to enable the GPIO pin. This seems like a really easy thing, but this is where it gets really annoying and painful.

You have to edit the Cubietruck's device tree. So apt-get source linux and in there edit arch/arm/boot/dts/sun7i-a20-cubietruck.dts

In the root section ('/'), near the top, add this:

    onewire_device {
       compatible = "w1-gpio";
       gpios = <&pio 6 8 GPIO_ACTIVE_HIGH>; /* PG8 */
       pinctrl-names = "default";
       pinctrl-0 = <&my_w1_pin>;
    };

In the '&pio` section, add this:

    my_w1_pin: my_w1_pin@0 {
         allwinner,pins = "PG8";
         allwinner,function = "gpio_in";
    };

Note that if you used a different pin than PG8 you'll need to change that. The "pio 6 8" means letter G, pin 8. The 6 is because G is the 7th letter of the alphabet. I don't know where this is documented; I reverse engineered it from another example. Why this can't be hex, or octal, or symbolic names or anything sane, I don't know.

Now you'll need to compile the dts file into a dtb file. One way is to configure the kernel and use its Makefile; I avoided that by first sudo apt-get install device-tree-compiler and then running, in the top of the linux source tree:

cpp -nostdinc -I include -undef -x assembler-with-cpp \
    ./arch/arm/boot/dts/sun7i-a20-cubietruck.dts | \
    dtc -O dtb -b 0 -o sun7i-a20-cubietruck.dtb -

You'll need to install that into /etc/flash-kernel/dtbs/sun7i-a20-cubietruck.dtb on the cubietruck. Then run flash-kernel to finish installing it.

use

Now reboot, and if all went well, it'll come up and the GPIO pin will finally be turned on:

# grep PG8 /sys/kernel/debug/pinctrl/1c20800.pinctrl/pinmux-pins
pin 200 (PG8): onewire_device 1c20800.pinctrl:200 function gpio_in group PG8

And if you picked a GPIO pin that works and got the sensor wired up correctly, in /sys/bus/w1/devices/ there should be a subdirectory for the sensor, using its unique ID. Here I have two sensors connected, which 1-wire makes easy to do, just hang them all off the same wire.. er wires.

root@honeybee:/sys/bus/w1/devices> ls
28-000008290227@  28-000008645973@  w1_bus_master1@
root@honeybee:/sys/bus/w1/devices> cat *-*/w1_slave
f6 00 4b 46 7f ff 0a 10 d6 : crc=d6 YES
f6 00 4b 46 7f ff 0a 10 d6 t=15375
f6 00 4b 46 7f ff 0a 10 d6 : crc=d6 YES
f6 00 4b 46 7f ff 0a 10 d6 t=15375

So, it's 15.37 Celsius in my house. I need to go feed the fire, this took too long to get set up.

future work

Are you done at this point? I fear not entirely, because what happens when there's a kernel upgrade? If the device tree has changed in some way in the new kernel, you might need to update the modified device tree file. Or it might not boot properly or not work in some way.

With Raspbian, you don't need to modify the device tree. Instead it has support for device tree overlay files, which add some entries to the main device tree. The distribution includes a bunch of useful overlays, including one that enables GPIO pins. The Raspberry Pi's bootloader takes care of merging the main device tree and the selected overlays.

There are u-boot patches to do such merging, or the merging could be done before reboot (by flash-kernel perhaps), but apparently Debian's device tree files are built without phandle based referencing needed for that to work. (See http://elektranox.org/2017/05/0020-dt-overlays/)

There's also a kernel patch to let overlays be loaded on the fly using configfs. It seems to have been around for several years without being merged, for whatever reason, but would avoid this problem nicely if it ever did get merged.