Remote booting of x86 machines with PXE and Intel AMT
Two weeks ago, I ordered a refurbished Thinkpad T420, which is compatible with many OS'es, including Genode/Sculpt. It is listed in cnuke's Genode unofficial HCL. So, I, from the beginning, wanted to set up the usual workflow of Genode developers.
This workflow uses standard run tool with special backends, related to remote PXE boot and Intel AMT for remotely controlling the machine.
In case of x86 machines with Intel AMT (usually, all newer Thinkpads), it is possible to remotely control the machine in different ways. Genode run tool uses Intel AMT for:
Remotely powering on/off the machine
getting the log output via AMT SOL (Serial over LAN) COM port.
Intel AMT is a remote management technology. Most newer machines (IIRC, all Intel Core i5/i7/Xeon machines having a "vPRO" label in their specs) have Intel Management Engine built-in. Intel AMT is just one module of Intel ME. Intel ME has a chip with its own firmware (based on Minix3). The firmware has its own IP stack with a HTTP, VNC servers, and a SOL (Serial over LAN) service.
After you go to BIOS setup and enable Intel AMT, and enable required services and configure IP stack, you'll have a second IP enabled on your LAN interface. This IP accepts connections even if your machine is switched off.
AMT SOL is seen on a PCI bus as a standard "Communication" device, so you can access it as a standard PCI COM port, via it's I/O address. You can write to it like to any hardware COM port. The debug output can be got via network with a special utility ("amtterm" package in Debian/Ubuntu).
I'll describe BIOS setup settings for usual Thinkpad's UEFI BIOS. First, you should go to "Config->Intel(R) AMT" and set "Intel (R) AMT Control" to "Enabled". Here, you can also set a timeout (0) and "Console Type" ("PC ANSI" in my case).
Then reboot the system, saving changes and press the blue "ThinkVantage" button, and then Ctrl-P. Press "1" to go to Intel ME setup.
Then you need to set up a valid Intel ME password. The default user is "admin" and password is "admin". But until you change the password to some "secure" password, the setup will refuse to do anything on.
The valid password is "P@ssw0rd", but "Passw0rd" is insufficiently secure. So, you should use a password, using letters of a mixed case, plus some numbers, plus some symbols like "@#$%". Otherwise the setup will consider it being insufficiently secure.
After you set up your ME password, you could go to "Intel(R) AMT Configuration":
At "SOL/IDER/KVM" section, you should enable SOL and "Legacy Redirection Mode".
At "User Consent", you should select:
User Opt-in: None
Opt-in Configurable from Remote IT -> Enable Remote Control of Opt-In Policy
– this is to avoid the user at the controlled machine to acknowledge each your remote action (which is annoying).
At "Password policy" section: "Default Password Only".
At Network Setup section: Here you should enter all usual IP stack settings like
Personally, I did not set up DHCP/DDNS, and assigned all addresses static here.
This should be sufficient for remote powering on/off, accessing SOL and HTTP servers via network.
Web interface is accessible as http://<your-IP>:16992/
AMT SOL service is listening at port 16994, you can attach to it via a command:
amtterm -u admin -p <your-amt-password> <your-IP>
Also, you could try to enable a VNC KVM, so you'll be able to take screenshots, but it's beyond the scope of this article (AFAIK, it requires a windows utility from Intel, didn't tried to use it so far).
The usual Genode development workflow is using a feature of a Genode's run tool, allowing to easily switch the run tool backends. There are backends for:
creating the boot directory for different kernels
creating the boot image (disk, uboot, iso image, etc.)
for loading the image (via tftp, ipxe, jtag debugger etc)
for powering the machine on and off
for getting the LOG output from the system under test (via AMT, serial, qemu etc.)
So, for using PXE and Intel AMT, you should comment the "power_on/qemu", "log/qemu" and "image/iso" backends in your "etc/build.conf" in your build directory and add the following section instead:
# QEMU_RUN_OPT := --include power_on/qemu \ --include log/qemu --include image/iso RUN_OPT := --include power_on/amt \ --power-on-amt-host 172.30.0.140 \ --power-on-amt-password 'P@ssw0rd' \ --power-on-amt-timeout 20 \ --include load/tftp \ --load-tftp-absolute \ --load-tftp-base-dir /var/lib/tftpboot \ --load-tftp-offset-dir /x86 \ --amt-tool wsman \ --include log/amt \ --log-amt-host 172.30.0.140 \ --log-amt-password 'P@ssw0rd'
Here, IP address, and password for "power_on/amt" and "log/amt" are self-explaining. Also, it is useful to increase the power on timeout to 20 seconds, because by default the timeout is as small as 5 seconds, which could be insufficient.
The "–load-tftp-absolute" enables absolute paths passed by the run tool to a tftp server. "–load-tftp-base-dir" sets the tftp server root. The build system creates a "x86" symlink, pointing to your build directory, in the TFTP root. The name of this symlink is set in the "–load-tftp-offset-dir" parameter.
Also, my config contains the "–amt-tool wsman" parameter, which seems to be not mandatory, but I have it. In addition to "amtterm", the run tool uses a "wsman" utility, to remotely power-on the machine. Wsman is a pair of an "openwsman" library and "wsmancli" command line utility, which uses SOAP RPC to an AMT HTTP server.
Unfortunately, "wsman" is missing in Debian packages (so, I needed to build it from sources), but there is "wsmancli" package in Ubuntu. "Amtterm" is present in both Debian and Ubuntu.
So, you should do
apt-get install wsmancli amtterm
to have the packages installed.
To boot Genode remotely, you need to have a usual PXE boot setup. First, you should change the boot device priorities in BIOS setup, so that, your test machine will boot from network by default. Most probably, you should use a wired LAN card, as wifi cards with boot ROM's are rare. So, when the system tries to boot from network, it sends a broadcast DHCP "discover" request, to find a DHCP server. When DHCP server is found, the boot ROM gets IP settings from it, and different DHCP options, which set up the bootable image name at the tftp server, tftp server IP address, gateway IP address etc.
As I had a DNSMasq DHCP server installed on my OpenWRT router, I reused it as a DHCP server for remote boot. I added the following lines to "/etc/dnsmasq.conf" file at the DHCP server:
dhcp-option=option:router,172.30.0.1 dhcp-option=6,172.30.0.1 dhcp-option=15,lan dhcp-option=17,/var/lib/tftpboot tftp-root=/var/lib/tftpboot dhcp-boot=/var/lib/tftpboot/x86/pulsar,tpx61t,172.30.0.101
Here "tftp-root" sets up the tftp server root (but it seems to be for a "dnsmasq" built-in tftp server, which I don't use in my setup. Instead, I set up a tftp server on my development machine).
"dhcp-boot" sets up the path to a boot loader. Genode currently uses "Pulsar" as a PXE loader. It is very small and supports the multiboot protocol, like GRUB does. Here we can see the path to the bootloader on the tftp server, followed by a tftp server hostname and IP address, separated by commas.
Also, here are the following DHCP options used:
"router": the IP gateway address
6: DNS server address
15: local domain
17: root-path: TFTP server root
As a tftp server, I use tftpd-hpa. The Debian package is named the same way. My settings in "/etc/default/tftpd-hpa" are as follows:
TFTP_USERNAME="tftp" TFTP_DIRECTORY="/var/lib/tftpboot" TFTP_ADDRESS=":69" # TFTP_OPTIONS="--secure --listen" TFTP_OPTIONS="--verbosity 4"
Here we have "–secure" option disabled, so tftp server does not use chroot, but uses absolute pathnames instead. Set TFTP_USERNAME to your tftp server's user. This user should be able to access your build directory, so you should set up permissions accordingly.
Now, you should put "pulsar" binary and its config file in the form: "config-<your-mac-address-bytes-delimited-by-dashes>" to your tftp server root:
# tftp 172.30.0.101 root /var/lib/tftpboot/x86 config config-00-00-00-00-00-00
The "root" command changes the root directory for "pulsar" to your build directory (pointed to by the /var/lib/tftpboot/x86 symlink). Then it switches to the new config file "config-00-00-00-00-00-00" located there. The build system automatically creates the "config-00-00-00-00-00-00" file in the build directory root, which, in turn, points to a config-52-54-00-12-34-56 file in /var/lib/tftpboot/x86/var/run/sculpt_test directory, in case you issued a
make KERNEL=nova run/sculpt_test
command in your build directory. Now if you typed the above command, the Sculpt image should be built from sources, then the build system creates a bootable image, all required symlinks and configs. Then the test machine is powered on, the image is loaded to it via network, it starts up, and you'll see a Genode log on your development machine.
During my attempts to boot Sculpt remotely, I encountered a problem: I had a NX bit disabled in my BIOS setup. This caused most 64-bit kernels to reboot or trap. 32-bit kernels booted successfully, though, because NX bit (aka Execute Prevention) is usually enabled on 32-bit systems only in conjunction with PAE feature, which was disabled in most used kernels, and these kernels don't expect it to be enabled, unlike 64-bit ones. So, the advice is to not disable the "Execute Prevention" in BIOS.
Alexander Boettcher for pointing me at my fault with the NX bit disabled
Christian Helmuth and Josef Soentgen on #genode irc channel, for helping me to get a working setup with AMT SOL and PXE boot.