This repository contains an (almost) stand-alone environment for running qemu as well as some scripts for creating and managing the image and a script to run qemu. This environemnt is targetted at emulated NVM Express (NVMe), Persistent Memory (PMEM) and LightNVM (aka OpenChannel) SSD testing but can be used for many other things.
There are some minimalist kernel config in this repo as well that can be used as a starting point for building a suitable kernel.
To run qemu using this image from the command should be:
./runqemu <path_to_bzImage>
You can run
./runqemu -h
to get the command line arguments supported.
This script will automatically create a snapshot image so you can revert to the original image by deleting images/jessie.qcow2.
There are a few methods for VM creation in this repo. They are all in various stages of repair.
A QEMU-only (no libvirt) way to generate new VMs. Uses the Ubuntu cloud image ISO and a cloud-init script. Note there is an issue that the files for this cannot reside on a guest's VirtFS (i.e. running this script inside a VM to generate a nested VM won't work due to file permission issues). This may be better with virtio-fs.
A libvirt-based script that uses cloud images and cloud-init and virt-install. This seems to work quite well on bare-metal.
The original and somewhat outdated method from Logan that starts with a base backing issue (based on Debian) and then uses deboostrap to setup the user and install some packages.
This repo does not include the QEMU executable. You can specific a path to the exe you want to run using the -q option. This is useful if you are using a fork (or your own branch) of QEMU that has support for specific hardware (e.g. LightNVM SSDs). Some useful links include the upstream QEMU repo, Keith Busch's NVMe repo and Matias Bjorling's LightNVM fork.
Note you might want to track all three of these and install them all since certain things are only supported in certain forks. With luck, over time, support for all things will end up upstream ;-). A useful tool for multiple installs is the --prefix= option in the ./configure step of the QEMU install.
Note that by default KVM support is turned on. Use the -k switch to turn this off (and suffer the wrath of slowness or if you are already running inside some form of a VM).
The runqemu script will automatically boot queitly and login as root giving a shell over stdio (which is managable but can have some rough edges). When the user logs out it will automatically powerdown the machine and exit qemu. An SSH login is also available, while running, forwarded to localhost port 3324. The root password is 'awhisten'.
[Top Tip: When on the login console use "^c a" to switch to the qemu command line (and vice-versa).]
When the NVMe option is chosen there is an nvme drive mounted on /mnt/nvme with the corresponding image in images/nvme.qcow2. This image just contains an ext4 partition with a single 4MB random test file. It is snapshotted so changes do not get saved run to run.
A second NVMe drive exists at /dev/nvme1 but this is not mounted. This second drive has a Controller Memory Buffer (CMB) advertised on it.
Note that since upstream has a different level of support for NVMe drive options than the other forks we use a -u option to handle that differently.
The host's /home file is also passthrough mounted to the guests /home directory so test scripts, etc can be stored and run directly from the users home directory on the host. In order for this to work you will use to make sure that QEMU is configured with VirtFS enabled and that the kernel you are running has the relevant Plan9 support.
QEMU's gdb feature is turned on so you may debug the kernel using gdb. Note that you run the first command below from a shell prompt and the second from within gdb. Note you should run these two command before or after invoking QEMU.
(shell) gdb vmlinux
(gdb) target remote :1234
If you are trying to debug a kernel module you need to use the gdb add-symbol-file command. This command needs to point to the .ko and provide memory offsets for .bbs, .data and .text sections. For example something like this:
(gdb) add-symbol-file <path to module> <text_addr> -s .data
<data_addr> -s .bss <bss_addr>
Note you can get the addresses from the running kernel in /sys/modules//sections directory and that the kernel has to be running before these addresses can be determined.
There are a few scripts in the scripts sub-folder that can be used to re-generate the jessie-clean.qcow2 image. To regenerate this file or create a new base image use the following steps:
-
sudo modprobe nbd max_part=8 - this makes sure that Network Block Device kernel module is loaded. We need this for step 2, 3 and 4.
-
./create
- this creates the bare qcow2 image, partitions the image and then uses deboostrap to setup the image as a chroot with a basic Debian Jessie install.
-
./setup
- this configures the Debian Jessie install. Setups up networking, machine name and some other key attributes. Note you can change the machine name and root password by editing this script.
-
./shrink
- This script shrinks the image file as much as possible by doing zerofree and then running qemu-img convert.
-
./cross [NEW] - This script uses multistrap to build a rootfs image for an ARCH that is different to the host machine (e.g. building an ARM64 rootfs on x86_64). Refer to the header of the script for more information. You will need some packages installed in order to run this script and it has to be run as root for the chroot command to work.
We are happy to consider PRs for other -clean images as long as they utilize the Large File Storage (LFS) feature or are pulled in via curl or wget.
This repo utilizes git Large File Storage (lfs) in order to avoid having to host the large jessie-clean.qcow2 image inside the repo. See here for more information.
The simple script in the scripts folder generates a really simple initramfs with a statically linked rootfs. Useful for really simple sanity testing. To build and run this simple initramfs do the following:
- cd scripts
- ./simple
- cd ..
- qemu-system-x86_64 -m 512 -kernel kernels/bzImage-4.8 -initrd scripts/simple.cpio.gz -append "console=ttyS0" -serial mon:stdio -nographic
For a more intresting initramfs example you can run busybox. To do this perform the following steps.
- Download the busybox source and build it to include all the tools that you want. Make sure this is a statically linked executable.
- cd scripts
- ./busybox
- cd ..
- emu-system-x86_64 -m 512 -kernel kernels/bzImage-4.8 -initrd scripts/initramfs.cpio.gz -append "console=ttyS0" -serial mon:stdio -nographic
This should boot into busybox shell and you can execute your installed command from there. Enjoy! Note this will not then pass on to a subsequent root filesystem (yet).
By default we map the /home folder on the host to the /home folder on the guest using Plan 9 folder sharing over VirtFS. However this assumes the host has the kernel support to do this. To disable this option use the -f switch (if you do this you might want to attach a image to replace the /home folder). Note you also need to make sure the QEMU executable you are using was compiled with VirtFS support enabled.
There are some good websites on how to do OS debug via QEMU. See for example:
http://stackoverflow.com/questions/11408041/how-to-debug-the-linux-kernel-with-gdb-and-qemu
A good command to start gdb would be:
gdb -ex 'target remote localhost:1234' -ex 'set architecture i386:x86-64:intel'
-ex 'break ' -ex c ./vmlinux
Make sure you have CONFIG_DEBUG_INFO set in the .config when you build the quest kernel.
Phew. After quite a bit of fun and games I can get a NVMf setup running using two QEMU guests. The kernel config files are checked in (v4.8-rc5 for now). There are two helper scripts in the scripts folders and the jessie image needs updating with some of the libs (including the librxe).
For a few reasons it is easier to keep the rdma_rxe.ko module seperate to the kernel. Also you need the linux headers inside the VM root filesystem to compile librxe. You currently need to copy (or symlink) /usr/lib64/* to /usr/lib. You also have to setup VM to VM networking for which I (for now) used -netdev socket. QEMU perfers ntap but that's alot more work. The IPv4 addresses on the eth0 interfaces of the two VMs are statically configured using techniques discussed in http://csortu.blogspot.ca/2009/12/building-virtual-network-with-qemu.html.
I started the nvmef target VM using the following command:
./runqemu -v -m 2048 -t -i images/nvmf-target.qcow2
./kernels/bzImage-4.8-nvmf-soft-roce
and the nvmef host VM using the following command:
./runqemu -v -m 2048 -s 3235 -n -i images/nvmf-host.qcow2
./kernels/bzImage-4.8-nvmf-soft-roce
I then logged into the target system and executed the nvmf-target script in the scripts folder using the target VMs assigned IP address as the input argument. On the host side I executed the nvmf-host script with the IP address of the target as the first argument.
nvme discover worked. nvme connect worked.
IO on the system caused some panics so need to investigate that. Note that the NVMf connection should be established between the eth1 interfaces and not the eth0 interfaces.
Note you need the rdma_rxe module installed on both target and host right now as there is an issue with monolithic kernels and rxe we need to root-cause.
Once you get a VM up and running you can use Ansible to install a bunch more things into it. I have a rather nice project called batesste-ansible that you can use for this.