Blogging about open source virtualization

SOCAT is a CLI utility which enables the concatenation of two sockets together. It establishes two bidirectional byte streams and transfers data between them.

socat supports several address types (e.g. TCP, UDP, UNIX domain sockets, etc.) to construct the streams. The latest version 1.7.4, released earlier this year [2021-01-04], supports also AF_VSOCK addresses:

• VSOCK-LISTEN:<port>

  • Listen on port and accepts a VSOCK connection.

• VSOCK-CONNECT:<cid>:<port>

  • Establishes a VSOCK stream connection to the specified cid and port.

FOSDEM 2021

If you are interested on VSOCK, I’ll talk witn Andra Paraschiv (AWS) about it at FOSDEM 2021. The talk is titled Leveraging virtio-vsock in the cloud and containers and it’s scheduled for Saturday, February 6th 2021 at 11:30 AM (CET).

We will show cool VSOCK use cases and some demos about developing, debugging, and measuring the VSOCK performance, including socat demos.

Examples

socat could be very useful for concatenating and redirecting sockets. In this section we will see some examples.

All examples below refer to a guest with CID 42 that we created using virt-builder and virt-install .

VM setup

virt-builder is able to download the installer and create the disk image with Fedora 33 or other distros. It is also able to set the root password and inject the ssh public key, simplifying the creation of guest disk image:

VM_IMAGE="vsockguest_f33.qcow2"

host$ virt-builder --root-password=password:mypassword \
        --ssh-inject root:file:/home/user/.ssh/id_rsa.pub \
        --output=${VM_IMAGE} \
        --format=qcow2 --size 10G --selinux-relabel \
        --update fedora-33

Once the disk image is ready, we create our VM with virt-install. We can specify the VM settings like the number of vCPUs, the amount of RAM, and the CID assigned to the VM [42]:

host$ virt-install --name vsockguest \
        --ram 2048 --vcpus 2 --os-variant fedora33 \
        --import --disk path=${VM_IMAGE},bus=virtio \
        --graphics none --vsock cid.address=42

After the creation of the VM, we will remain attached to the console and we can detach from it by pressing ctrl-].

We can reattach to the console in this way:

host$ virsh console vsockguest

If the VM is turned off, we can boot it and attach directly to the console in this way:

host$ virsh start --console vsockguest

ncat like

It’s possible to use socat like ncat, transferring stdin and stdout via VSOCK.

Guest listening

In this example we start socat in the guest listening on port 1234:

guest$ socat - VSOCK-LISTEN:1234

Then we connect from the host using the CID 42 assigned to the VM:

host$ socat - VSOCK-CONNECT:42:1234

At this point we can exchange characters between guest and host, since stdin and stdout are linked through the VSOCK socket.

Host listening

In this example we do the opposite, starting socat in the host listening on port 1234:

host$ socat - VSOCK-LISTEN:1234

Then, in the guest, we connect to the host using the well defined CID 2. It’s always used to reach the host:

guest$ socat - VSOCK-CONNECT:2:1234

ssh over VSOCK

The coolest feature of socat is to concatenate sockets of different address families, so in this example we redirect ssh traffic through VSOCK socket exposed by the VM.

This example could be useful if the VM doesn’t have any NIC attached and we want to provide some network connectivity, like the ssh access.

First of all, in the guest we start socat linking the VSOCK socket listening on port 22, to a TCP socket which will connect to the local TCP port 22 where the ssh server is listening:

guest$ socat VSOCK-LISTEN:22,reuseaddr,fork TCP:localhost:22

On the host we link a TCP socket listening on a port of our choice (e.g. 4321) to the guest port 22 just opened using VSOCK:

host$ socat TCP4-LISTEN:4321,reuseaddr,fork VSOCK-CONNECT:42:22

Finally from the host we can connect to the guest using ssh on the local port 4321, where socat is listening:

host$ ssh -p 4321 root@localhost

socat redirects all the traffic between the sockets and allow us to use ssh over VSOCK to reach the guest.

The previous article in this series introduced QEMU storage concepts. Now we move on to look at the two most popular emulated storage controllers for virtualization: virtio-blk and virtio-scsi.

This post provides recommendations for configuring virtio-blk and virtio-scsi and how to choose between the two devices. The recommendations provide good performance in a wide range of use cases and are suitable as default settings in tools that use QEMU.

Virtio storage devices

Key points

• Prefer virtio storage devices over other emulated storage controllers.
• Use the latest virtio drivers.

Virtio devices are recommended over other emulated storage controllers as they are generally the most performant and fully-featured storage controllers in QEMU.

Unlike emulations of hardware storage controllers, virtio-blk and virtio-scsi are specifically designed and optimized for virtualization. The details of how they work are published for driver and device implementors in the VIRTIO specification.

Virtio drivers are available for Linux, Windows, and other operating systems. Installing the latest version is recommended for the latest bug fixes and performance enhancements.

If virtio drivers are not available, the AHCI (SATA) device is widely supported by modern x86 operating systems and can be used as a fallback. On non-x86 guests the default storage controller can be used as a fallback.

Comparing virtio-blk and virtio-scsi

Key points

• Prefer virtio-blk in performance-critical use cases.
• Prefer virtio-scsi for attaching more than 28 disks or for full SCSI support.
• With virtio-scsi, use scsi-block for SCSI passthrough and otherwise use scsi-hd.

Two virtio storage controllers are available: virtio-blk and virtio-scsi.

virtio-blk

The virtio-blk device presents a block device to the virtual machine. Each virtio-blk device appears as a disk inside the guest. virtio-blk was available before virtio-scsi and is the most widely deployed virtio storage controller.

The virtio-blk device offers high performance thanks to a thin software stack and is therefore a good choice when performance is a priority. It does not support non-disk devices such as CD-ROM drives.

CD-ROMs and in general any application that sends SCSI commands are better served by the virtio-scsi device, which has full SCSI support. SCSI passthrough was removed from the Linux virtio-blk driver in v5.6 in favor of using virtio-scsi.

Virtual machines that require access to many disks can hit limits based on availability of PCI slots, which are under contention with other devices exposed to the guest, such as NICs. For example a typical i440fx machine type default configuration allows for about 28 disks. It is possible to use multi-function devices to pack multiple virtio-blk devices into a single PCI slot at the cost of losing hotplug support, or additional PCI busses can be defined. Generally though it is simpler to use a single virtio-scsi PCI adapter instead.

virtio-scsi

The virtio-scsi device presents a SCSI Host Bus Adapter to the virtual machine. SCSI offers a richer command set than virtio-blk and supports more use cases.

Each device supports up to 16,383 LUNs (disks) per target and up to 255 targets. This allows a single virtio-scsi device to handle all disks in a virtual machine, although defining more virtio-scsi devices makes it possible to tune for NUMA topology as we will see in a later blog post.

Emulated LUNs can be exposed as hard disk drives or CD-ROMs. Physical SCSI devices can be passed through into the virtual machine, including CD-ROM drives, tapes, and other devices besides hard disk drives.

Clustering software that uses SCSI Persistent Reservations is supported by virtio-scsi, but not by virtio-blk.

Performance of virtio-scsi may be lower than virtio-blk due to a thicker software stack, but in many use cases, this is not a significant factor. The following graph compares 4KB random read performance at various queue depths:

virtio-scsi configuration

The following SCSI devices are available with virtio-scsi:

The scsi-hd device is suitable for disk image files and host block devices when SCSI passthrough is not required.

The scsi-block device offers SCSI passthrough and is preferred over scsi-generic due to higher performance.

The following graph compares the sequential I/O performance of these devices using virtio-scsi with an iothread:

Conclusion

The virtio-blk and virtio-scsi offer a choice between a single block device and a full-fledged SCSI Host Bus Adapter. Virtualized guests typically use one or both of them depending on functional and performance requirements. This post compared the two and offered recommendations on how to choose between them.

The next post in this series will discuss the iothreads feature that both virtio-blk and virtio-scsi support for increased performance.

QEMU v5.2 is out. A highlight from a KVM on Z perspective:

• PCI passthrough support now includes any PCI devices other than RoCE Express cards, including ISM devices as needed for SMC-D.
  

For further details, see the Release Notes.

The following videos and publications are now available on the IBM Knowledge Center:

• IBM Secure Execution for Linux
  
• Video: Securing your Workload in the Cloud An Introduction
  Explains the risk factors against which IBM Secure Execution for Linux helps to protect for your workload.
• Video: Securing your Workload in the Cloud
  Shows how IBM Secure Execution for Linux protects your workload in the cloud.
• These videos are also reachable in the IBM Knowledge Center.
• Publication: Introducing IBM Secure Execution for Linux
  The new version of the guide includes new troubleshooting information and additional image hardening tips.
  Also available as PDF.
• Publication: KVM Virtual Server Management
  New information includes configuring KVM virtual servers for IBM Secure Execution for Linux, backing virtual block devices with PCIe-attached NVMe devices, and descriptions of new tools.
  Also available as PDF.

The two previous blog posts about why git forges are von Neumann machines and the Radicle peer-to-peer git forge explored models for git forges. In this final post I want to cover yet another model that draws from the previous ones but has its own unique twist.

Peer-to-peer git apps

I previously showed how applications can be built on centralized git forges using CI/CD functionality for executing code, webhooks for interacting with the outside world, and disjoint branches for storing data.

A more elegant architecture is a peer-to-peer one where instead of many clients and one server there are just peers. Each peer has full access to the data. There is no client/server application code split, instead each peer runs an application for itself.

First, this makes it easier to move the data to new hosting infrastructure or fork a project since all data resides in the git repository. Merge requests, issues, wikis, and even the app settings are all stored in the git repo itself.

Second, this gives more power to the users who can process data however they want without being limited by the server's API. All peers are on equal footing and users don't need permission to alter applications, because they run locally.

Finally, it is easier to develop a local application than a client/server application. Being able to open a file and tweak the code is immediate and less hassle than testing and deploying a server-side application.

Internet peer-to-peer systems typically still require some central point for bootstrapping and this is no exception. A publicly-accessible git repository is still needed so that peers can fetch and push changes. However, in this model the git server does not run application code but "git apps" like merge requests, issue trackers, wikis, etc can still be implemented. Here is how it works...

The anti-application server

The git server is not allowed to run application code in our model, so apps like merge requests won't be processing data on the server side. However, the repository does need some primitives to make peer-to-peer git apps possible. These primitives are access control policies for refs and directories/files.

Peers run applications locally and the git server is "dumb" with the sole job of enforcing access control. You can imagine this like a multi-user UNIX machine where users have access to a shared directory. UNIX file permissions determine how processes can access the data. By choosing permissions carefully, multiple users can collaborate in the shared directory in a safe and controlled manner.

This is an anti-application server because no application code runs on the server side. The server is just a git repository that stores data and enforces access control on git push.

Access control

Repositories that accept push requests need a pre-receive hook (see githooks(5)) that checks incoming requests against the access control policy. If the request complies with the access control policy then the git push is accepted. Otherwise the git push is rejected and changes are not made to the git repository.

The first type of access control is on git refs. Git refs are the namespace where branches and tags are stored in a git repository. If a regular expression matches the ref and the operation type (create, fast-forward, force, delete) then it is allowed. For example, this policy rule allows any user to push to refs/heads/foo but force pushes and deletion are not allowed:

anyone create,fast-forward ^heads/foo$

The operations available on refs include:

What's more interesting is that $user_id is expanded to the git push user's identifier so we can write rules to limit access to per-user ref namespaces:

anyone create-branch,fast-forward,force,delete ^heads/$user_id/.*$

This would allow Alice to push her own branches but Alice could not push to Bob's branches.

We have covered how to define access control policies on refs. Access control policies are also needed on branches so that multiple users can modify the same branch in a controlled and safe manner. The syntax is similar but the policy applies to changes made by commits to directories/files (what git calls a tree). The following allows users to create files in a directory but not delete or modify them (somewhat similar to the UNIX restricted deletion or "sticky" bit on world-writable directories):

anyone create-file ^shared-dir/.*$

The operations available on branches include:

$user_id expansion is also available for branch access control. Here the user can create, modify, and delete files in a per-user directory:

anyone create-file,modify,delete-file ^$user_id/.*$

User IDs

You might be wondering how user identifiers work. Git supports GPG-signed push requests with git push --signed. We can use the GPG key ID as the user identifier, eliminating the need for centralized user accounts. Remember that the GPG key ID is based on the public key. Key pairs are randomly generated and it is improbable that the same key will be generated by two different users. That said, GPG key ID uniqueness has been weak in the past when the default size was 32 bits. Git explicitly enables long 64-bit GPG key IDs but I wonder if collisions could be a problem. Maybe an ID with more bits based on the public key should be used instead, but for now let's assume the GPG key ID is unique.

The downside of this approach is that user IDs are not human-friendly. Git apps can allow the user to assign aliases to avoid displaying raw user IDs. Doing this automatically either requires an external ID issuer like confirming email address ownership, which is tedious for new users, or by storing a registry of usernames in the git repo, which means a first-come-first-server policy for username allocation and possible conflicts when merging from two repositories that don't share history. Due to these challenges I think it makes sense to use raw GPG key IDs at the data storage level and make them prettier at the user interface level.

The GPG key ID approach works well for desktop clients but not for web clients. The web application (even if implemently on the client side) would need access to the private key so it can push to the git repository. Users should not trust remotely hosted web applications with their private keys. Maybe there is a standard Web API that can help but I'm not aware one. More thought is needed here.

The pre-receive git hook checks that signature verification passed and has access to the GPG key ID in the GIT_PUSH_CERT_KEY environment variable. Then the access control policy can be checked.

Access control is a git app

Access control is the first and most fundamental git app. The access control policies that were described above are stored as files in the apps/access-control branch in the repository. Pushes to that branch are also subject to access control checks. Here is the branch's initial layout:

branches/ - access control policies for branches
  owner.conf
groups/ - group definitions (see below)
  ...
refs/ - access control policies for refs
  owner.conf

The default branches/owner.conf access control policy is as follows:

owner create-file,create-directory,modify,delete ^.*$

The default refs/owner.conf access control policy is as follows:

owner create-branch,create-tag,fast-foward,force,delete ^.*$

This gives the owner the ability to push refs and modify branches as they wish. The owner can grant other users access by pushing additional access control policy files or changing exsting files on the apps/access-control branch.

Each access control policy file in refs/ or branches/ is processed in turn. If no access control rule matches the operation then the entire git push is rejected.

Groups can be defined to alias one or more user identifiers. This avoids duplicating access control rules when more than one user should have the same access. There are two automatic groups: owner contains just the user who owns the git repository and anyone is the group of all users.

This completes the description of the access control app. Now let's look at how other functionality is built on top of this.

The merge requests app

A merge requests app can be built on top of this model. The refs access control policy is as follows:

# The data branch contains the titles, comments, etc
anyone modify ^apps/merge-reqs/data$

# Each merge request revision is pushed as a tag in a per-user namespace
anyone create-tag ^apps/merge-reqs/$user_id/[0-9]+-v[0-9]+$

The branch access control policy is:

# Merge requests are per-user and numbered
anyone create-directory ^merge-reqs/$user_id/[0-9]+$

# Title string
anyone create-file,modify ^merge-reqs/$user_id/[0-9]+/title$

# Labels (open, needs-review, etc) work like this:
#
#   merge-reqs/<user-id>/<merge-req-num>/labels/
#     needs-review -> /labels/needs-review
#     ...
#   labels/
#     needs-review/
#       <user-id>/
#         <merge-req-num> -> /merge-reqs/<user-id>/<merge-req-num>
#         ...
#       ...
#     ...
#
# This directory and symlink layout makes it possible to enumerate labels for a
# given merge request and to enumerate merge requests for a given label.
#
# Both the merge request author and maintainers can add/remove labels to/from a
# merge request.
anyone create-directory ^merge-reqs/[^/]+/[0-9]+/labels$
anyone create-symlink,delete ^merge-reqs/$user_id/[0-9]+/labels/.*$
maintainers create-symlink,delete ^merge-reqs/[^/]+/[0-9]+/labels/.*$
maintainers create-directory ^labels/[^/]+$
anyone create-symlink,delete ^labels/[^/]+/$user_id/[0-9]+$
maintainers create-symlink,delete ^labels/[^/]+/[^/]+/[0-9]+$

# Comments are stored as individual files in per-user directories. Each file
# contains a timestamp and the contents of the comment. The timestamp can be
# used to sort comments chronologically.
anyone create-directory ^merge-reqs/[^/]+/[0-9]+/comments$
anyone create-directory ^merge-reqs/[^/]+/[0-9]+/comments/$user_id$
anyone create-file,modify ^merge-reqs/[^/]+/[0-9]+/comments/$user_id/[0-9]+$

When a user creates a merge request they provide a title, an initial comment, apply labels, and push a v1 tag for review and merging. Other users can comment by adding files into the merge request's per-user comments directory. Labels can be added and removed by changing symlinks in the labels directories.

The user can publish a new revision of the merge request by pushing a v2 tag and adding a comment describing the changes. Once the maintainers are satisfied they merge the final revision tag into the relevant branch (e.g. "main") and relabel the merge request from open/needs-review to closed/merged.

This workflow can be implemented by a tool that performs the necessary git operations so users do not need to understand the git app's internal data layout. Users just need to interact with the tool that displays merge requests, allows commenting, provides searches, etc. A natural way to implement this tool is as a git alias so it integrates alongside git's built-in commands.

One issue with this approach is that it uses the file system as a database. Performance and scalability are likely to be worst than using a database or application-specific file format. However, the reason for this approach is that it allows the access control app to enforce a policy that ensures users cannot modify or delete other user's data without running application-specific code on the server and while keeping everything stored in a git repository.

An example where this approach performs poorly is for full-text search. The application would need to search all title and comment files for a string. There is no index for efficient lookups. However, if applications find that git-grep(1) does not perform well they can maintain their own index and cache files locally.

I hope that this has shown how git apps can be built without application code running on the server.

Continuous integration bots

Now that we have the merge requests app it's time to think how a continuous integration service could interface with it. The goal is to run tests on each revision of a merge request and report failures so the author of the merge request can rectify the situation.

A CI bot watches the repository for changes. In particular, it needs to watch for tags created with the ref name apps/merge-reqs/[^/]+/[0-9]+-v[0-9]+.

When a new tag is found the CI bot checks it out and runs tests. The results of the tests are posted as a comment by creating a file in merge-regs/<user-id>/>merge-req-num>/comments/<ci-bot-user-id>/0 on the apps/merge-reqs/data branch. A ci-pass or ci-fail label can also be applied to the merge request so that the CI status can be easily queried by users and tools.

Going further

There are many loose ends. How can non-git users participate on issue trackers and wikis? It might be possible to implement a full peer as a client-side web application using isomorphic-git, a JavaScript git implementation. As mentioned above, the GPG key ID approach is not very browser-friendly because it requires revealing the private key to the web page and since keys are user identifiers using temporary keys does not work well.

The data model does not allow efficient queries. A full copy of the data is necessary in order to query it. That's acceptable for local applications because they can maintain their own indexes and are expected to keep the data for a long period of time. It works less well for short-lived web page sessions like a casual user filing a new bug on the issue tracker.

The git push --signed technique is not the only option. Git also supports signed commits and signed tags. The difference between signed pushes and signed tags/commits is significant. The signed push approach only validates the access control policy when the repository is changed and leaves no audit log for future reference. The signed commit/tag approach keeps the signatures in the git history. Signed commits/tags can be propagated in a peer-to-peer network and each peer can validate the access control policy itself. While signed commits/tags apply the access control policy to each object in the repository, signed pushes apply the access control policy to each change made to the repository. The difference is that it's easy to rebase and include work from different authors with signed pushes. Signed commits/tags require re-signing for rebasing and each commit is validated against its signature, which may be different from the user who is making the push request.

There are a lot of interesting approaches and trade-offs to explore here. This model we've discussed fits closely with how I've seen developers use git in open source projects. It is designed around a "main" repository/server that contributors push their code to. But each clone of the repository has all the data and can be published as a new "main" repository, if necessary.

Although these ideas are unfinished I decided to write them up with the knowledge that I probably won't implement them myself. QEMU is moving to GitLab with a traditional centralized git forge. I don't think this is the right time to develop this idea and try to convince the QEMU community to use it. For projects that have fewer infrastructure requirements it would give their contributors more power than being confined to a centralized git forge.

I hope this was an interesting read for anyone thinking about git forges and building git apps.

We’d like to announce the availability of the QEMU 5.2.0 release. This release contains 3200+ commits from 216 authors.

You can grab the tarball from our download page. The full list of changes are available in the Wiki.

Note that QEMU has switched build systems so you will need to install ninja to compile it. See the “Build Information” section of the Changelog for more information about this change.

Highlights include:

• block: support for using qemu-storage-daemon as a vhost-user-blk device backend, and new ‘block-export-add’ QMP command which replaces now-deprecated ‘nbd-server-add’ with support for qemu-storage-daemon
• block: qcow2 support for subcluster-based allocation (via extended_l2=on qemu-img option), improvements to NBD client network stalls, and qemu-nbd support for exposing multiple dirty bitmaps at once
• migration: higher-bandwidth encrypted migration via TLS+multifd, new ‘block-bitmap-mapping’ option for finer-grained control over which bitmaps to migrate, and support for migration over a ‘vsock’ device (for nested environments and certain hardware classes)
• qemu-ga: support for guest-get-devices, guest-get-disks, and guest-ssh-{get,add-remove}-authorized-keys commands.
• virtiofs: virtiofsd support for new options to control how xattr names are seen by the guest, specify sandboxing alternative to pivot_root, and allowing different host mounts to be seen as separate submounts in the guest to avoid inode clashes
• ARM: new board support for mp2-an386 (Cortex-M4 based), mp2-an500 (Cortex-M7 based), raspi3ap (Raspberry Pi 3 model A+), raspi0 (Raspberry Pi Zero), raspi1ap (Raspberry Pi A+), and npcm750-evb/quanta-gsj (Nuvoton iBMC)
• ARM: ARMv8.2 FEAT_FP16 (half-precision flaoting point) support for AArch32 (already supported for AArch64)
• ARM: virt: support for kvm-steal-time accounting
• HPPA: support for booting NetBSD and older Linux distros like debian-0.5 and debian-0.6.1
• PowerPC: pseries: improved support for user-specified NUMA distance topologies
• RISC-V: live migration support
• RISC-V: experimental hypervisor support updated to v0.6.1 and other improvements
• RISC-V: support for NUMA sockets on virt/Spike machine types
• s390: KVM support for diagnose 0x318 instruction, TCG support for additional z14 instructions
• s390: vfio-pci devices now report real hardware features for functions instead of just emulated values
• Xtensa: DFPU co-processor with single/double-precision FP opcodes is now supported
• x86: improved support for asynchronous page faults via new kvm-async-pf-int -cpu option
• and lots more…

Thank you to everyone involved!

In this post I want to explore an idea for a new type of application made possible by the power of git forges like GitLab and GitHub. I don't have a proof-of-concept and in fact we'll discuss hurdles that could make the idea infeasible. But it's an interesting way of thinking that is worth sharing as it offers a new way of looking at and using git forges.

Git forges offer git repository hosting at their core and then add on related features like pull requests, wikis, issue trackers, static website hosting, continuous integration, and more. They are now a long way away from the initial idea of a hosted git repository where you can publish code. They do so much more. They have a von Neumann architecture and are Turing complete. It's possible to do interesting things with that.

A git forge app (GFA) is an application that runs on a git forge. A hosted git repository is not just a place to publish and back up the source code. It's where the app runs. You can view the git forge as a Platform-as-a-Service (PaaS) or serverless computing provider. The app doesn't need to be deployed elsewhere, because the git forge itself is the execution environment.

A GFA must be able to:

1. Process data.
2. Store data.
3. Interact with the outside world.

This is basically what computer applications do. Git forges have become powerful enough to do this.

Imagine the following: you fork a repository and this instantiates the GFA. The GFA is now running under your git forge account. The web interface is available at https://<user>.gitlab.com/<app>/ and HTTP POST requests can also be used to interact with the application. The GFA could be a todo list, an RSS reader, a blog, etc. To the website visitor it appears like any other web application but everything happens within the git forge and no other hosting provider is necessary.

Let's look at how it's possible to use a git forge as the execution environment for an app.

Processing Data

The first order of business is executing code so that the application can process data. A configuration file is placed into a git repository to define runnable jobs. The job execution systems offered by git forges are GitLab CI and GitHub Actions, respectively. When the job is triggered it executes in a temporary environment, for example a Linux container image. It allows code hosted in the git repository to run on a server somewhere for a little while. There is a free allowance that grants a certain number of hours of unpaid execution time. This is how GFAs process data.

For example, say we are building an RSS reader. Our repo looks like this:

rss-reader/
    .gitlab-ci.yml - the job configuration file
    fetch-new-feeds.sh - download RSS feeds and check for new posts

The .gitlab-ci.yml file will contain a scheduled pipeline that runs fetch-new-feeds.sh every 15 minutes.

Storing Data

Applications need to store data persistently. The primary way of doing that in a git forge is by storing data in the git repository itself. Mutable data can be kept on a separate branch called data and can be force-pushed to avoid forever increasing the repository size when we don't need to store previous revisions of the data.

This works because each git branch has an independent commit history. The file namespace it stores is completely separate from other branches. This means it's possible to have the GFA's code on the main branch and to store data on the data branch without upsetting the commit history or files on the main branch.

Data storage is available to jobs because they are allowed to manipulate their own git repository thanks to an authentication token. On GitHub GITHUB_TOKEN allows jobs to push to their git repository. This gives jobs a way of storing data.

There are other forms of data storage besides the git repository. Artifactsare files produced by a job that can be downloaded via a URL. Artifacts are subject to expiry time and file size limitations. Caching is available to temporarily store files between job runs, although this data can be lost at any time.

At this point you may be thinking that this is nice but there is no way to store private data if the git repository is public. Git forges offer a secrets mechanism where variables can be stored privately and only made available during job execution. This can be used to stash an encryption key so that the data is stored in public but is encrypted. Anyone can download the encrypted data but they will not have the key needed to decrypt it. Some applications can also take advantage of client-side encryption and homomorphic encryption.

Interacting with the Outside World

Git forges offer a wealth of ways to interact with the outside world, called triggers. Triggers can run a job when an HTTP POST request is made. POST requests typically need a trigger token for authentication, but that token can be published if the triggered job is safe to be invoked by anyone. Both browser clients and Webhooks can invoke triggers through HTTP POST requests.

For example, imagine our RSS reader app needs an API to mark feeds as read. We define an HTTP POST trigger that runs a mark-feed-read.sh script that updates the feeds stored in the data branch. Since only the git repository owner should be able to invoke this trigger we do not publish the trigger token. Instead it is stored encrypted in the repository and the user provides a passphrase for decrypting this secret.

At this point it is also useful to offer a web interface. This is possible thanks to the static pages hosting feature called GitLab Pages and GitHub Pages. Pushing HTML, JavaScript, CSS, and images to a special branch publishes the static website at https://<user>.gitlab.com/<app>/.

For the most part GFAs are asynchronous, they cannot handle HTTP requests synchronously like a web application that outputs an HTTP response. Incoming HTTP requests simply schedule GFA jobs that run sometime later. There are a few ways around this. The browser can poll for results using XMLHttpRequest or similar techniques. Alternatively, a GFA can fire up a job that runs for several minutes although I haven't investigated if there is a practical way to communicate with a web application running in a job (I guess incoming HTTPS is tricky to achieve).

Triggers offer a lot more fun than what I've already covered. They can respond to the creation and modification of wiki pages, issues, and pull request comments. This means it's possible to use those entities for interacting with the outside world. The GFA could act as a chat bot on its issue tracker, for example.

Limitations

Git forges weren't designed for GFAs. But then they weren't initially designed to be Turing complete and offering ways to interact with the outside world either. That was added incrementally as demand for that functionality grew. Maybe git forges will evolve into full-blown serverless hosting competitors to today's cloud hosting providers.

GFAs are a hack that uses features like static pages, CI jobs, and webhooks in a creative way to run applications on a git forge. Building GFAs that are actually useful is likely to hit some challenges:

• No synchronous requests - it is hard to implement search queries and other dynamic behavior in GFAs because they are primarily asynchronous (and slow!). This limitation matters for certain classes of applications. A lot can be done client-side instead to make up for this deficiency. But maybe someone can figure out how to do synchronous requests in GFAs.
• Security - I have outlined how to make data publicly available and also how to encrypt it so only the git repository owner can view the plaintext. This is enough for personal web applications, but it's not sufficient for multi-user applications. Maybe git forges offer a form of authentication that works with GFAs so multiple users can store private data in a single GFA instance (the git repository owner could view all users' data but other users could not).
• Free usage tiers - job execution time, storage capacity, and request throttling will limit how resource-intensive a GFA can be before it outgrows the free tier and eventually even the paid tier.

The first generation of GFAs could be written from scratch with each job carefully designed. Then a second generation of GFAs could be built on top of frameworks that abstract the tedious git forge integration with standard APIs and data models. For example, a Vue.js frontend could use a key/value store API and the whole thing can be hosted as a GFA.

Even if GFAs don't become a thing because git forges decide there is not enough demand to make them work really well, changing your perspective to think of git forges in this way is valuable. For example, I have a git repo for building a container image that I deploy on my server and that pushes output files to a web server host. All of that can be replaced with a git repository that runs a job and publishes to GitLab Pages. This simplifies things and frees me from maintaining infrastructure.

Conclusion

Git forges offer jobs for processing data, git repositories and artifacts for storing data, and triggers for interacting with the outside world. It is possible to build applications that exist solely as a git repository on a git forge. There is no longer a need to deploy code because the git forge itself is powerful enough to run non-trivial applications. I look forward to how this evolves and whether git forges eventually become full-blown cloud service providers.

Git is a distributed version control system and does not require a central server. Although repositories are usually published at a well-known location for convenient cloning and fetching of the latest changes, this is actually not necessary. Each clone can have the full commit history and evolve independently. Furthermore, code changes can be exchanged via email or other means. Finally, even the clone itself does not need to be made from a well-known domain that hosts a git repository (see git-bundle(1)).

Given that git itself is already fully decentralized one would think there is no further work to do. I came across the Radicle project and its somewhat psychedelic website. Besides having a website with a wild color scheme, the project aims to offer a social coding experiment or git forge functionality using a peer-to-peer network architecture. According to the documentation the motivation seems to be that git's built-in functionality works but is not user-friendly enough to make it accessible. In particular, it lacks social coding features.

The goal is to add git forge features like project and developer discovery, issue trackers, wikis, etc. Additional, distinctly decentralized functionality, is also touched on involving Ethereum as a way to anchor project metadata, pay contributors, etc. Radicle is still in early development so these features are not yet implemented. Here is my take on the How it Works documentation, which is a little confusing due to its early stage and some incomplete sentences or typos. I don't know whether my understanding actually corresponds to the Radicle implementation that exists today or its eventual vision, because I haven't studied the code or tried running the software. However, the ideas that the documentation has brought up are interesting and fruitful in their own right, so I wanted to share them and explain them in my own words in case you also find them worth exploring.

The git data model

Let's quickly review the git data model because it is important for understanding peer-to-peer git forges. A git repository contains a refs/ subdirectory that provides a namespace for local branch heads (refs/heads/), local and remotely fetched tags (refs/tags/), and remotely fetched branches (refs/remotes/<remote>/). Actually this namespace layout is just a convention for everyday git usage and it's possible to use the refs/ namespace differently as we will see. The git client fetches refs from a remote according to a refspec rule that maps remote refs to local refs. This gives the client the power to fetch only certain refs from the server. The client can also put them in a different location in its local refs/ directory than the server. For details, see the git-fetch(1) man page.

Refs files contain the commit hash of an object stored in the repository's object database. An object can be a commit, tree (directory), tag, or a blob (file). Branch refs point to the latest commit object. A commit object refers to a tree object that may refer to further tree objects for sub-directories and finally the blob objects that make up the files being stored. Note that a git repository supports disjoint branches that share no history. Perhaps the most well-known example of disjoint branches are the GitHub Pages and GitLab Pages features where these git forges publish static websites from the HTML/CSS/JavaScript/image files on a specific branch in the repository. That branch shares no version history with other branches and the directories/files typically have no similarity to the repository's main branch.

Now we have covered enough git specifics to talk about peer-to-peer git forges. If you want to learn more about how git objects are actually stored, check out my article on the repository layout and pack files.

Identity and authority

Normally a git repository has one or more owners who are allowed to push refs. No one else has permission to modify the refs namespace. What if we tried to share a single refs namespace with the whole world and everyone could push? There would be chaos due to naming conflicts and malicious users would delete or change other users' refs. So it seems like an unworkable idea unless there is some way to enforce structure on the global refs namespace.

Peer-to-peer systems have solutions to these problems. First, a unique identity can be created by picking a random number with a sufficient number of bits so that the chance of collision is improbable. That unique identity can be used as a prefix in the global ref namespace to avoid accidental collisions. Second, there needs to be a way to prevent unauthorized users from modifying the part of the global namespace that is owned by other users.

Public-key cryptography provides the primitive for achieving both these things. A public key or its hash can serve as the unique identifier that provides identity and prevents accidental collisions. Ownership can be enforced by verifying that changes to the global namespace are signed with the private key corresponding to the unique identity.

For example, we fetch the following refs from a peer:

<identity>/
  heads/
    main
  metadata/
    signed_refs

This is a simplified example based on the Radicle documentation. Here identity is the unique identity based on a public key. Remember no one else in the world has the same identity because the chance of generating the same public key is improbable. The heads/ refs are normal git refs to commit objects - these are published branches. The signed_refs ref points to an git object that contains a list of commit hashes and a signature generated using the public key. The signature can be verified using the public key.

Next we need to verify these changes to check that they were created with the private key that is only known to the identity's owner. First, we check the signature on the object pointed to by the signed_refs ref. If the signature is not valid we reject these changes and do not store them in our local repository. Next, we look up each ref in heads/ against the list in signed_refs. If a ref is missing from the list then we reject these refs and do not allow them into our local repository.

This scheme lends itself to peer-to-peer systems because the refs can be propagated (copied) between peers and verified at each step. The identity owner does not need to be present at each copy step since their cryptographic signature is all we need to be certain that they authorized these refs. So I can receive refs originally created by identity A from peer B and still be sure that peer B did not modify them since identity A's signature is intact.

Now we have a global refs namespace that is partitioned so that each identity is able to publish refs and peers can verify that these changes are authorized.

Gossip

It may not be clear yet that it's not necessary to clone the entire global namespace. In fact, it's possible that no single peer will ever have a full copy of the entire global namespace! That's because this is a distributed system. Peers only fetch refs that they care about from their peers. Peers fetch from each other and this forms a network. The network does not need to be fully connected and it's possible to have multiple clusters of peers running without full global connectivity.

To bootstrap the global namespace there are seed repositories. Seeds are a common concept in peer-to-peer systems. They provide an entry point for new peers to learn about and start participating with other peers. In BitTorrent this is called a "tracker" rather than a "seed".

According to the Radicle documentation it is possible to directly fetch from peers. This probably means a git-daemon(1) or git-http-backend(1) needs to be accessible on the public internet. Many peers will not have sufficient network connectivity due to NAT limitations. I guess Radicle does not expect every user to participate as a repository.

Interestingly, there is a gossip system for propagating refs through the network. Let's revisit the refs for an identity in the global namespace:

<identity>/
  heads/
    main
  metadata/
    signed_refs
  remotes/
    <another-identity>/
      heads/
        main
        foo
      metadata/
        signed_refs
      remotes/
        ...

We can publish identities that we track in remotes/. It's a recursive refs layout. This is how someone tracking our refs can find out about related identities and their refs.

Thanks to git's data model the commit, tree, and blob objects can be shared even though we duplicate refs published by another identity. Since git is a content-addressable object database the data is stored once even though multiple refs point to it.

Now we not only have a global namespace where anyone can publish git refs, but also ways to build a peer-to-peer network and propagate data throughout the network. It's important to note that data is only propagated if peers are interested in fetching it. Peers are not forced to store data that they are not interested in.

How data is stored locally

Let's bring the pieces together and show how the system stores data. The peer creates a local git repository called the monorepo for the purpose of storing portions of the global namespace. It fetches refs from seeds or direct peers to get started. Thanks to the remotes/ refs it also learns about other refs on the network that it did not request directly.

This git repository is just a data store, it is not usable for normal git workflows. The conventional git branch and git tag commands would not work well with the global namespace layout and verification requirements. Instead we can clone a local file:/// repository from the monorepo that fetches a subset of the refs into the conventional git refs layout. The files can be shared because git-clone(1) supports hard links to local repositories. Thanks to githooks(5) and/or extensible git-push(1) remote helper support it's possible to generate the necessary global namespace metadata (e.g. signatures) when we push from the local clone to the local monorepo. The monorepo can then publish the final refs to other peers.

Building a peer-to-peer git forge

There are neat ideas in Radicle and it remains to be seen how well it will grow to support git forge functionality. A number of challenges need to be addressed:

• Usability - Radicle is a middle-ground between centralized git forges and email-based decentralized development. The goal is to be easy to use like git forges. Peer-to-peer systems often have challenges providing a human-friendly interface on top of public key identities (having usernames without centralized user accounts). Users will probably prefer to think in terms of repositories, merge requests, issues, and wikis instead of peers, gossip, identities, etc.
• Security - The global namespace and peer-to-peer model is a target that malicious users will attack by trying to impersonate or steal identities, flood the system with garbage, game reputation systems with sockpuppets, etc.
• Scalability - Peers only care about certain repositories and don't want to be slowed down by all the other refs in the global namespace. The recursive refs layout seems like it could cause performance problems and maybe users will configure clients to limit the depth to a low number like 3. At first glance Radicle should be able to scale well since peers only need to fetch refs they are interested in, but it's a novel way of using git refs, so we can expect scalability problems as the system grows.
• Data model - How will this data model grow to handle wikis, issue trackers, etc? Issue tracker comments are an example of a data structure that requires conflict resolution in a distributed system. If two users post comments on an issue, how will this be resolved without a conflict? Luckily there is quite a lot of research on distributed data structures such as Conflict-free Replicated Data Types (CRDTs) and it may be possible to avoid most conflicts by eliminating concepts like linear comment numbering.
• CI/CD - As mentioned in my blog post about why git forges are von Neumann machines, git forges are more than just data stores. They also have a computing model, initially used for Continuous Integration and Continuous Delivery, but really a general serverless computing platform. This is hard to do securely and without unwanted resource usage in a peer-to-peer system. Maybe Radicle will use Ethereum for compute credits?

Conclusion

Radicle is a cool idea and I look forward to seeing where it goes. It is still at an early stage but already shows interesting approaches with the global refs namespace and monorepo data store.

In previous blog posts I talked about QEMU’s qcow2 file format and how to make it faster. This post gives an overview of how the data is structured inside the image and how that affects performance, and this presentation at KVM Forum 2017 goes further into the topic.

This time I will talk about a new extension to the qcow2 format that seeks to improve its performance and reduce its memory requirements.

Let’s start by describing the problem.

Limitations of qcow2

One of the most important parameters when creating a new qcow2 image is the cluster size. Much like a filesystem’s block size, the qcow2 cluster size indicates the minimum unit of allocation. One difference however is that while filesystems tend to use small blocks (4 KB is a common size in ext4, ntfs or hfs+) the standard qcow2 cluster size is 64 KB. This adds some overhead because QEMU always needs to write complete clusters so it often ends up doing copy-on-write and writing to the qcow2 image more data than what the virtual machine requested. This gets worse if the image has a backing file because then QEMU needs to copy data from there, so a write request not only becomes larger but it also involves additional read requests from the backing file(s).

Because of that qcow2 images with larger cluster sizes tend to:

• grow faster, wasting more disk space and duplicating data.
• increase the amount of necessary I/O during cluster allocation,
  reducing the allocation performance.

Unfortunately, reducing the cluster size is in general not an option because it also has an impact on the amount of metadata used internally by qcow2 (reference counts, guest-to-host cluster mapping). Decreasing the cluster size increases the number of clusters and the amount of necessary metadata. This has direct negative impact on I/O performance, which can be mitigated by caching it in RAM, therefore increasing the memory requirements (the aforementioned post covers this in more detail).

Subcluster allocation

The problems described in the previous section are well-known consequences of the design of the qcow2 format and they have been discussed over the years.

I have been working on a way to improve the situation and the work is now finished and available in QEMU 5.2 as a new extension to the qcow2 format called extended L2 entries.

The so-called L2 tables are used to map guest addresses to data clusters. With extended L2 entries we can store more information about the status of each data cluster, and this allows us to have allocation at the subcluster level.

The basic idea is that data clusters are now divided into 32 subclusters of the same size, and each one of them can be allocated separately. This allows combining the benefits of larger cluster sizes (less metadata and RAM requirements) with the benefits of smaller units of allocation (less copy-on-write, smaller images). If the subcluster size matches the block size of the filesystem used inside the virtual machine then we can eliminate the need for copy-on-write entirely.

So with subcluster allocation we get:

• Sixteen times less metadata per unit of allocation, greatly reducing the amount of necessary L2 cache.
• Much faster I/O during allocating when the image has a backing file, up to 10-15 times more I/O operations per second for the same cluster size in my tests (see chart below).
• Smaller images and less duplication of data.

This figure shows the average number of I/O operations per second that I get with 4KB random write requests to an empty 40GB image with a fully populated backing file.

Things to take into account:

• The performance improvements described earlier happen during allocation. Writing to already allocated (sub)clusters won’t be any faster.
• If the image does not have a backing file chances are that the allocation performance is equally fast, with or without extended L2 entries. This depends on the filesystem, so it should be tested before enabling this feature (but note that the other benefits mentioned above still apply).
• Images with extended L2 entries are sparse, that is, they have holes and because of that their apparent size will be larger than the actual disk usage.
• It is not recommended to enable this feature in compressed images, as compressed clusters cannot take advantage of any of the benefits.
• Images with extended L2 entries cannot be read with older versions of QEMU.

How to use this?

Extended L2 entries are available starting from QEMU 5.2. Due to the nature of the changes it is unlikely that this feature will be backported to an earlier version of QEMU.

In order to test this you simply need to create an image with extended_l2=on, and you also probably want to use a larger cluster size (the default is 64 KB, remember that every cluster has 32 subclusters). Here is an example:

$ qemu-img create -f qcow2 -o extended_l2=on,cluster_size=128k img.qcow2 1T

And that’s all you need to do. Once the image is created all allocations will happen at the subcluster level.

More information

This work was presented at the 2020 edition of the KVM Forum. Here is the video recording of the presentation, where I cover all this in more detail:

You can also find the slides here.

Acknowledgments

This work has been possible thanks to Outscale, who have been sponsoring Igalia and my work in QEMU.

And thanks of course to the rest of the QEMU development team for their feedback and help with this!

Software Freedom Conservancy, the non-profit charity home of QEMU, Git, Inkscape, and many other free and open source software projects is running its annual fundraiser. They have announced a generous donation matching pledge so donations made until January 15th 2021 will be doubled! Full details are here.

What makes Software Freedom Conservancy important, besides being a home for numerous high-profile free and open source software projects, is that it is backed by individuals and works for the public interest. It is not a trade association funded by companies to represent their interests. With the success of free and open source software it's important we don't lose these freedoms or use them just to benefit businesses. That's why I support Software Freedom Conservancy.

Starting today, on December 1st, the first door of the QEMU Advent Calendar 2020 can now be opened! The advent calendar reveals a new disk image or something similar for download on each of the first 24 days in December 2020, to create a fun experience for the QEMU community, brightening your days in the winter season, and to provide some good images for testing various aspects of QEMU. There is also the possibility to contribute your own ideas for disk images - so if you’re interested in helping, please have a look at the announcement in the QEMU blog for details.

QEMU Advent Calendar publishes a disk image surprise each day from December 1-24. It's a QEMU community tradition that is back again this year!

If you want to contribute disk images to this year's advent calendar (puzzles, games, niche operating systems, retro computing fun, etc), please check out the call for disk images for details.

The recent generation of the mainframe (i.e. the z15) has the possibility to run protected KVM guests, so that the administrator of the host LPAR does not have the possibility anymore to read or alter the memory of a guest. This feature is called “Secure Execution” on the IBM Z.

IBM now published a nice introduction and demo of this new feature, which is in my opinion quite worthwhile to watch if you want to get a basic idea about running protected guests:

https://community.ibm.com/community/user/ibmz-and-linuxone/viewdocument/2020-11-ibm-secure-execution-for?CommunityKey=c1293167-6d93-448e-8854-3068846d3dfe&tab=librarydocuments

The video contains a short introduction into the concepts of the secure execution feature, followed by a demo that shows how to prepare a protected guest VM.

QEMU Advent Calendar 2020 is around the corner and we are looking for volunteers to contribute disk images that showcase something cool, bring back retro computing memories, or simply entertain with a puzzle or game.

QEMU Advent Calendar publishes a QEMU disk image each day from December 1-24. Each image is a surprise designed to delight an audience consisting of the QEMU community and beyond. You can see previous years at https://www.qemu-advent-calendar.org/

You can help us make this year’s calendar awesome by:

• Sending disk images
• Sending ideas for disk images

Here are the requirements for disk images:

• Content must be freely redistributable (i.e. no proprietary license that prevents distribution). For GPL based software, you need to provide the source code, too.
• Provide a name and a short description of the disk image (e.g. with hints on what to try)
• Provide a ./run shell script that prints out the name and description/hints and launches QEMU
• Provide a 320x240 screenshot/image/logo for the website
• Size should be ideally under 100 MB per disk image (but if some few images are bigger, that should be OK, too)

Check out this disk image as an example of how to distribute an image. Links to files over 25MB are appreciated in lieu of email attachments.

PS: QEMU Advent Calendar is a secular calendar (not religious). The idea is to create a fun experience for the QEMU community which can be shared with everyone. You don’t need to celebrate Christmas or another religious festival to participate!

 As, once again, a new QEMU release is around the corner, the time has come to list some s390x changes in there.

• TCG has gained emulation support for some additional instructions that had been introduced with the z14. More enhancements needed to be able to run distributions built for the z14 will likely come in the future.
• When running under KVM, QEMU now supports the diagnose 0x318 instruction. This can be used to set some diagnostic information (such as the operating system), which may be helpful when servicing the hardware. With this comes support for extended SCCBs; this is needed as the facility indication for diag318 encroaches into the control block used for reporting CPU information. A guest needs support for extended SCCBs to be able to see information for all CPUs if diag318 support is provided.
• You can now use virtiofs on s390x, thanks to some endianness fixes, and a vhost-user-fs-ccw device has been added.
• Up to now, both fully emulated PCI functions and PCI functions passed via vfio-pci reported the same values when the guest issued CLP instructions. However, the passed through functions may use different values for things such as the supported DMA range. If the host kernel supplies the respective capabilities for the vfio-pci device, QEMU can now provide the real values in the CLP queries.
• zPCI is now also able to honour vfio DMA limits, if passed via the vfio-pci device, and can trigger the guest to flush its DMA mappings when needed.
• The s390-ccw bios now tries harder to find a bootable device, if the first device is not suitable. This brings s390x booting a bit closer to what other architectures do.
• And the usual fixes and cleanups.

RHEL 8.3 is out, see the official announcement.
The unquestionable highlight from a KVM on Z perspective is certainly the addition of the Secure Execution functionality. Also, ECKD DASD can be transparently passed through to KVM guests, allowing full exploitation of all DASD features, including raw track access and IPL

For further details on the changes, see the release notes, and the respective blog entry in the Linux on Z blog here.

IBM Secure Execution for Linux allows to build a Trusted Execution Environment for IBM Z and LinuxONE that helps protect data in use.
This webcast gives an overview of the value and the key concepts of the technology, followed by a hands-on demo, outlining the steps needed to secure Linux workloads.

Audience: Clients, Business Partners, IT Architects, Systems Admins

Speaker: Viktor Mihajlovski, Linux on IBM Z Development, Product Owner for KVM on IBM Z

Date: November 18, 11:00 AM - 12:15 PM EST

Registration: Register here, and check system requirements here.

This article provides an overview of virtio-fs, a novel way for sharing the host file system with guests and OSv, a specialized, lightweight operating system (unikernel) for the cloud, as well as how these two fit together.

virtio-fs

Virtio-fs is a new host-guest shared filesystem, purpose-built for local file system semantics and performance. To that end, it takes full advantage of the host’s and the guest’s colocation on the same physical machine, unlike network-based efforts, like virtio-9p.

As the name suggests, virtio-fs builds on virtio for providing an efficient transport: it is included in the (currently draft, to become v1.2) virtio specification as a new device. The protocol used by the device is a slightly extended version of FUSE, providing a solid foundation for all file system operations native on Linux. Implementation-wise, on the QEMU side, it takes the approach of splitting between the guest interface (handled by QEMU) and the host file system interface (the device “backend”). The latter is handled by virtiofsd (“virtio-fs daemon”), running as a separate process, utilizing the vhost-user protocol to communicate with QEMU.

One prominent performance feature of virtio-fs is the DAX (Direct Access) window. It’s a shared memory window between the host and the guest, exposed as device memory (a PCI BAR) to the second. Upon request, the host (QEMU) maps file contents to the window for the guest to access directly. This bears performance gains due to taking VMEXITs out of the read/write data path and bypassing the guest page cache on Linux, while not counting against the VM’s memory (since it’s just device memory, managed on the host).

Virtio-fs is under active development, with its community focussing on a pair of device implementation in QEMU and device driver in Linux. Both components are already available upstream in their initial iterations, while upstreaming continues further e.g. with DAX window support.

OSv

OSv is a unikernel (framework). The two defining characteristics of a unikernel are:

• Application-specialized: a unikernel is an executable machine image, consisting of an application and supporting code (drivers, memory management, runtime etc.) linked together, running in a single address space (typically in guest “kernel mode”).
• Library OS: each unikernel only contains the functionality mandated by its application in terms of non-application code, i.e. no unused drivers, or even whole subsystems (e.g. networking, if the application doesn’t use the network).

OSv in particular strives for binary compatibility with Linux, using a dynamic linker. This means that applications built for Linux should run as OSv unikernels without requiring modifications or even rebuilding, at least most of the time. Of course, not the whole Linux ABI is supported, with system calls like fork() and relatives missing by design in all unikernels, which lack the notion of a process. Despite this limitation, OSv is quite full featured, with full SMP support, virtual memory, a virtual file system (and many filesystem implementations, including ZFS) as well as a mature networking stack, based on the FreeBSD sources.

At this point, one is sure to wonder “Why bother with unikernels?”. The problem they were originally introduced to solve is the bloated software stack in modern cloud computing. Running general-purpose operating systems as guests, typically for a single application/service, on top of a hypervisor which already takes care of isolation and provides a standard device model means duplication, as well as loss of efficiency. This is were unikernels come in, trying to be just enough to support a single application and as light-weight as possible, based on the assumption that they are executing inside a VM. Below is an illustration of the comparison between general-purpose OS, unikernels and containers (as another approach to the same problem, for completeness).

OSv, meet virtio-fs

As is apparent e.g. from the container world, it is very common for applications running in isolated environments (such as containers, or unikernels even more so) to require host file system access. Whereas containers sharing the host kernel thus have an obvious, controlled path to the host file system, with unikernels this has been more complex: all solutions were somewhat heavyweight, requiring a network link or indirection through network protocols. Virtio-fs then provided a significantly more attractive route: straight-forward mapping of fs operations (via FUSE), reusing the existing virtio transport and decent performance without high memory overhead.

The OSv community quickly identified the opportunity and came up with a read-only implementation on its side, when executing under QEMU. This emphasized being lightweight complexity-wise, while catering to many of its applications’ requirements (they are stateless, think e.g. serverless). Notably, it includes support for the DAX window (even before that’s merged in upstream QEMU), providing excellent performance, directly rivalling that of its local (non-shared) counterparts such as ZFS and ROFS (an OSv-specific read-only file system).

One central point is OSv’s support for booting from virtio-fs: this enables deploying a modified version or a whole new application without rebuilding the image, just by adjusting its root file system contents on the host. Last, owing to the DAX window practically providing low-overhead access to the host’s page cache, scalability is also expected to excel, with it being a common concern due to the potentially high density of unikernels per host.

For example, to build the cli OSv image, bootable from virtio-fs, using the core OSv build system:

scripts/build fs=virtiofs export=all image=cli

This results in a minimal image (just the initramfs), while the root fs contents are placed in a directory on the host (build/export here, by default).

Running the above image is just a step away (may want to use the virtio-fs development version of QEMU, e.g. for DAX window support):

scripts/run.py --virtio-fs-tag=myfs --virtio-fs-dir=$(pwd)/build/export

This orchestrates running both virtiofsd and QEMU, using the contents of build/export as the root file system. Any changes to this directory, directly from the host will be visible in the guest without re-running the previous build step.

Conclusion

OSv has gained a prominent new feature, powered by virtio-fs and its QEMU implementation. This allows efficient, lightweight and performant access to the host’s file system, thanks to the native virtio transport, usage of the FUSE protocol and the DAX window architecture. In turn, it enables use cases like rapid unikernel reconfiguration.

Get the latest news about the Linux exploitation and advantages of the IBM Z and LinuxONE platform in this technical 5 day workshop. Focusing also on new solutions and capabilities, such as Hybrid Cloud, Red Hat OpenShift Container Platform, Security, Performance, Networking and Virtualization presented by our development experts. We will start with introduction sessions on the first day, continue with 3 days of technical deep dive topics, and resume with the latest client experiences, and a panel discussion on the last day. You will also get a chance to interact directly with IBM developers and solution experts during the workshop.

Agenda highlights:

• Introduction day
• Linux on IBM Z and LinuxONE introduction and client stories
• Virtualization options with Linux on IBM Z and LinuxONERed Hat OpenShift Container Platform in the Hybrid Cloud Strategy for IBM Z & LinuxONE
• Cloud Paks for IBM Z overview
• Deep dive sessions
• What's new Linux on IBM Z and LinuxONE
• z/VM Platform update
• IBM Secure Execution for Linux - Introduction and Overview
• Hyper Protect Virtual Server onPrem - Differences to Cloud
• Red Hat OpenShift on IBM Z - Performance Experiences, Hints and Tips & Networking
• Securing the Workloads on Red Hat OpenShift on IBM Z
• IBM z15 Hardware Compression
• Preparing for Multifactor Authentication for z/VM
• Best Practices of Installing Red Hat OpenShift on IBM Z
• Competitive application performance on Red Hat OpenShift
• Cloud Paks for Data on IBM Z and LinuxONE
• KVM Network Performance - Best Practices and Tuning Recommendations
• Experiences and open discussion
• Client experiences and lessons learned
• Panel discussion for all
  

Date: November 9 - 13, 2020
Time schedule: 8:30 AM- 12:00 PM EDT / 2:30 - 6:00 PM CEST daily

Registration is open here till November 5.

Ubuntu Server 20.10 is out!
It ships

• Linux kernel 5.8
• QEMU v5.0
• libvirt v6.6

See the release notes here, and the blog entry at canonical with Z-specific highlights here.

In version 4.2 the microvm machine type was added to qemu. The initial commit describes it this way:

It's a minimalist machine type without PCI nor ACPI support, designed for short-lived guests. microvm also establishes a baseline for benchmarking and optimizing both QEMU and guest operating systems, since it is optimized for both boot time and footprint.

The initial code uses the minimal qboot firmware to initialize the guest, to load a linux kernel and boot it. For network/storage/etc devices virtio-mmio is used. The configuration is passed to the linux kernel on the command line so the guest is able to find the devices.

That works for direct kernel boot, using qemu -kernel vmlinuz, because qemu can easily patch the kernel command line then. But what if you want - for example - boot the Fedora Cloud image? Using the Fedora kernel stored within the image?

A better plan for device discovery

When not using direct kernel boot patching the kernel command line for device discovery isn't going to fly, so we need something else. There are two established standard ways to do that in modern systems. One is device trees. The other is ACPI. Both have support for virtio-mmio.

A device tree entry for virtio-mmio looks like this:

virtio_block@3000 {
	compatible = "virtio,mmio";
	reg = <0x3000 0x100>;
	interrupts = <41>;
}

And this is the ACPI DSDT version:

Device (VR06)
{
    Name (_HID, "LNRO0005")  // _HID: Hardware ID
    Name (_UID, 0x06)  // _UID: Unique ID
    Name (_CCA, One)  // _CCA: Cache Coherency Attribute
    Name (_CRS, ResourceTemplate ()  // _CRS: Current Resource Settings
    {
        Memory32Fixed (ReadWrite,
            0xFEB00C00,         // Address Base
            0x00000200,         // Address Length
            )
        Interrupt (ResourceConsumer, Level, ActiveHigh, Exclusive, ,, )
        {
            0x00000016,
        }
    })
}

Both carry essentially the same information: What kind of device that is and which resources (registers & interrupt) it uses.

On the arm platform both are established, with device trees being common for small board computers like the raspberry pi and ACPI being used in the arm server space. On the x86 platform we don't have much of a choice though. There are some niche attempts to establish device trees, the google android goldfish platform for example. But for widespread support there is no way around using ACPI.

The nice thing about arm server using ACPI too is that this paved the way for us. The linux kernel supports both device tree and ACPI for the discovery of virtio-mmio devices:

root@fedora-bios ~/acpi# modinfo virtio-mmio | grep alias
alias:          of:N*T*Cvirtio,mmioC*
alias:          of:N*T*Cvirtio,mmio
alias:          acpi*:LNRO0005:*

So linux kernel support is a solved problem already. Yay!

virtio-mmio support for seabios

If we want load a kernel from the disk image the firmware must be able to find and read the disk. We already have virtio-pci support for blk and scsi in seabios. The differences between virtio-pci and virtio-mmio transports are not that big. Also some infrastructure for different transport modes was already there to deal with legacy vs. modern virtio-pci. So adding virtio-mmio support to the drivers wasn't much of a problem.

But of course seabios also has the problem that it must discover the devices before it can initialize the driver. Various approaches to find virtio-mmio devices using the available information sources where tried. All of them had the one or the other non-working corner case, except using ACPI. So seabios ended up getting a simple DSDT parser for device discovery.

While being at it some other small fixes where added to seabios too to make it work better with microvm. The hard dependency on the RTC CMOS has been removed for example, so latest seabios works fine with qemu -M microvm,rtc=off.

This ships with seabios version 1.14.

While speaking about seabios: When using a serial console I'd strongly recommend to run with qemu -M microvm,graphics=off. That will enable serial console support in seabios. This is one of the tweaks done by the qemu -nographic shortcut. The machine option works with the pc and q35 machine types too.

ACPI cleanups in qemu

Hooking up ACPI support for microvm on the qemu side turned out to be surprisingly difficuilt due to some historical baggage.

Years ago qemu used to have a static ACPI DSDT table. All ISA devices (serial & parallel ports, floppy, ...) are declared there, but they might not be actually present depending on qemu configuration. The LPC/ISA bride has some bits in pci config space saying whenever a device is actually present or not (qemu emulation follows physical hardware behavior here). So the devices have a _STA method looking up those bits and returning the device status. The guest had to run the method using AML interpreter to figure whenever the declared device is actually there.

// this is the qemu q35 ISA bridge at 00:1f.0
Device (ISA)
{
    Name (_ADR, 0x001F0000)  // _ADR: Address
    // ... snip ...
    OperationRegion (LPCE, PCI_Config, 0x82, 0x02)
    Field (LPCE, AnyAcc, NoLock, Preserve)
    {
        CAEN,   1,  // serial port #1 enable bit
        CBEN,   1,  // serial port #2 enable bit
        LPEN,   1   // parallel port enable bit
    }
}

// serial port #1
Device (COM1)
{
    Name (_HID, EisaId ("PNP0501") /* 16550A-compatible COM Serial Port */)
    Name (_UID, One)  // _UID: Unique ID
    Method (_STA, 0, NotSerialized)  // _STA: Status
    {
        Local0 = CAEN /* \_SB_.PCI0.ISA_.CAEN */
        If ((Local0 == Zero))
        {
            // serial port #1 is disabled
            Return (Zero)
        }
        Else
        {
            // serial port #1 is enabled
            Return (0x0F)
        }
    }
    // ... snip ...
}

The microvm machine type simply has no PCI support, so that approach isn't going to fly. Also these days all ACPI tables are dynamically generated anyway, so there is no reason to have the guests AML interpreter go dig into pci config space. Instead we can handle that in qemu when generating the DSDT table. Disabled devices are simply not listed. For enabled devices this is enough:

// serial port #1
Device (COM1)
{
    Name (_HID, EisaId ("PNP0501") /* 16550A-compatible COM Serial Port */)
    Name (_UID, One)  // _UID: Unique ID
    Name (_STA, 0x0F)  // _STA: Status
    // ... snip ...
}

So I've ended up reorganizing and simplifying the code which creates the DSDT entries for ISA devices. This landed in qemu version 5.1.

ACPI support for microvm

Now with the roadblocks out of the way it was finally possible to add acpi support to microvm. There is little reason to worry about backward compatibility to historic x86 platforms here, old guests wouldn't be able to handle virtio-mmio anyway. So this takes a rather modern approach and looks more like an arm virt machines than a x86 q35 machine. Like arm it uses the generic event device for power management support.

ACPI support for microvm is switchable, simliar to the other machine types, using the acpi=on|off machine option. The -no-acpi switch works too. By default ACPI support is enabled.

With ACPI enabled qemu uses virtio-mmio enabled seabios as firmware and doesn't bother patching the linux kernel command line for device discovery.

With ACPI disabled qemu continues to use qboot as firmware like older qemu versions do. Likewise it continues to add virtio-mmio devices to the linux kernel command line.

This will be available in qemu version 5.2. It is already merged in the master branch.

ACPI advantages

• Number one is device discovery obviously, this is why we started all this in the first place. seabios and linux kernel find virtio-mmio devices automatically. You can boot Fedora cloud images in microvm without needing any tricks. Probably other distros too, even though I didn't try that. Compiling the linux kernel with CONFIG_VIRTIO_MMIO=y (or =m & adding the module to initramfs) is pretty much the only requirement for this to work.

• Number two is device discovery too. ACPI will also tell the kernel which devices are not there. So with acpi=on the kernel simply skips the PS/2 probe in case the DSDT doesn't list a keyboard controller. With acpi=off the kernel assumes legacy hardware, goes into probe-harder mode and needs one second to figure that there really is no keyboard controller:

  [    0.414840] i8042: PNP: No PS/2 controller found.
  [    0.415287] i8042: Probing ports directly.
  [    1.454723] i8042: No controller found
  

  We have an simliar effect with the real time clock. With acpi=off the kernel goes register an IRQ line for the RTC even in case the device isn't there.

• Number three is (basic) power management. ACPI provides a virtual power button, so the guest will honor shutdown requests sent that way. ACPI also provides S5 (aka poweroff) support, so qemu gets a notification from the guest when the shutdown is done and can exit.

• Number four is better IRQ routing. The linux kernel finds the IO-APIC declared in the APIC table and uses it for IRQ routing. It is possible to use lines 16-23 for the virtio-mmio devices, avoiding IRQ sharing. Also we can refine the configuration using IRQ flags in the DSDT table.

  With acpi=off this does not work not reliable. I've seen the kernel ignore the IO-APIC in the past. Doesn't always happen though. Not clear which factors play a role here, I didn't investigate that in detail. Maybe newer kernel versions are a bit more clever here and find the IO-APIC even without ACPI.

Bottom line: ACPI helps moving the microvm machine type forward towards a world without legacy x86 platform devices.

But isn't ACPI bloated and slow?

Well, on my microvm test guest all ACPI tables combined are less than 1k in size:

root@fedora-bios /sys/firmware/acpi/tables# find -type f | xargs ls -l
-r--------. 1 root root  78 Oct  2 09:36 ./APIC
-r--------. 1 root root 482 Oct  2 09:36 ./DSDT
-r--------. 1 root root 268 Oct  2 09:36 ./FACP

I wouldn't call that bloated. This is a rather small virtual machine, with larger configurations (more CPUs, more devices) the tables will grow a bit of course.

When testing boot times I figured it is pretty hard to find any differences due to ACPI initialization. The noise (differences when doing 2-3 runs with identical configuration) is larger than the acpi=on/off difference. Seems to be at most a handful of milliseconds.

When trying that yourself take care to boot the kernel with 'quiet'. This is a good idea anyway if you want boot as fast as possible. The kernel prints more boot information with acpi=on, so slow console logging can skew your numbers if you let the kernel print out everything.

Runtime differences should be zero. There is only one AML method in the DSDT table. It toggles the power button when a notification comes in from the generic event device. It runs only on generic event device interrupts.

USB support for microvm

qemu just got a sysbus (non-pci) version of the xhci host adapter. It is needed for some arm boards. Nice thing is now that we have ACPI we can just wire that up in microvm too, add it in the DSDT table, then linux will find and use it:

Device (XHCI)
{
    Name (_HID, EisaId ("PNP0D10") /* XHCI USB Controller */)
    Name (_CRS, ResourceTemplate ()  // _CRS: Current Resource Settings
    {
        Memory32Fixed (ReadWrite,
            0xFE900000,         // Address Base
            0x00004000,         // Address Length
            )
        Interrupt (ResourceConsumer, Level, ActiveHigh, Exclusive, ,, )
        {
            0x0000000A,
        }
    })
}

USB support will be disabled by default, it can be enabled using the usual machine option: qemu -M microvm,usb=on.

Patches for qemu are in flight, should land in version 5.2
Patches for seabios are merged, will be available in version 1.15

PCIe support for microvm

There is one more arm platform thing we can reuse in microvm: The PCI Express host bridge. Again the same approach: Wire everything up, declare it in the ACPI DSDT. Linux kernel finds and uses it.

Not adding an asl snippet this time. The PCIe host bridge is a complex device so the description is a bit larger. It has GSI subdevices, IRQ routing information for each PCI slot, mmconfig configuration etc. Also shows in the DSDT size (even though that is still less than half the size q35 has):

root@fedora-bios /sys/firmware/acpi/tables# ll DSDT
-r--------. 1 root root 3130 Oct  2 10:20 DSDT

PCIe support will be disabled by default, it can be enabled using the new pcie machine option: qemu -M microvm,pcie=on.

This will be available in qemu version 5.2. It is already merged in the master branch.

Future work

My TODO list for qemu isn't very long:

• Add second IO-APIC, allowing more IRQ lines for more virtio-mmio devices. Experimental patches exist.

• IOMMU support, using virtio-iommu. Depends on ACPI spec update for virtio-iommu being finalized and support being merged into qemu and linux kernel. The actual wireup for microvm should be easy once all this is done.

Outside qemu there are a few more items:

• Investigate microvm PCIe support in seabios. Experimental patches exist. I'm not sure yet whenever seabios should care though.

  The linux kernel can initialize the PCIe bus just fine on its own, whereas proper seabios support has its challenges. When running in real mode the mmconfig space can not be reached. Legacy pci config space access via port 0xcf8 is not available. Which breaks pcibios support. Probably fixable with a temporary switch to 32bit mode, at the cost of triggering a bunch of extra vmexits. Beside that seabios will spend more time at boot, initializing the pci devices.

  So, is it worth the effort? The benefit would be that seabios could support booting from pci devices on microvm then.

• Maybe add microvm support to edk2/ovmf.

  Looks not that easy on a quick glance. ArmVirtPkg depends on device trees for virtio-mmio detection, so while we can re-use the virtio-mmio drivers we can not re-use device discovery code. Unless we maybe have qemu provide both ACPI tables and a device tree, even if ovmf happens to be the only device tree user.

  It also is not clear what other dragons (dependencies on classic x86 platform devices) are lurking in the ovmf codebase.

• Support the new microvm features (possibly adding microvm support first) in other projects.

  Candidate number one is of course libvirt because it is the foundation for many other projects. Beside that microvm support is probably mostly useful for cloud/container-style workloads, i.e. kata and kubevirt.

Over the past months I have participated in discussions about out-of-process device emulation. This post describes the requirements that have become apparent. I hope this will be a useful guide to understanding the big picture about out-of-process device emulation.

What is out-of-process device emulation?

Device emulation is traditionally implemented in the program that executes guest code. This approach is natural because accesses to device registers are trapped as part of the CPU run loop that sits at the core of an emulator or virtual machine monitor (VMM).

In some use cases it is advantageous to perform device emulation in separate processes. For example, software-defined network switches can minimize data copies by emulating network cards directly in the switch process. Out-of-process device emulation also enables privilege separation and tighter sandboxing for security.

Why are these requirements important?

When emulated devices are implemented in the VMM they use common VMM APIs. Adding new devices is relatively easy because the APIs are already there and the developer can focus on the device specifics. Out-of-process device emulation potentially leaves developers without APIs since the device emulation program is a separate program that literally starts from main(). Developers want to focus on implementing their specific device, not on solving general problems related to out-of-process device emulation infrastructure.

It is not only a lot of work to implement an out-of-process device completely from scratch, but there is also a risk of developing the wrong solution because some subtleties of device emulation are not obvious at first glance.

I hope sharing these requirements will help in the creation of common infrastructure so it's easy to implement high-quality out-of-process devices.

Not all use cases have the full set of requirements. Therefore it's best if requirements are addressed in separate, reusable libraries so that device implementors can pick the ones that are relevant to them.

Device emulation

Device resources

Devices provide resources that drivers interact with such as hardware registers, memory, or interrupts. The fundamental requirement of out-of-process device emulation is exposing device resources.

The following types of device resources are needed:

Synchronous MMIO/PIO accesses

The most basic device emulation operation is the hardware register access. This is a memory-mapped I/O (MMIO) or programmed I/O (PIO) access to the device. A read loads a value from a device register. A write stores a value to a device register. These operations are synchronous because the vCPU is paused until completion.

Asynchronous doorbells

Devices often have doorbell registers, allowing the driver to inform the device that new requests are ready for processing. The vCPU does not need to wait since the access is a posted write.

The kvm.ko ioeventfd mechanism can be used to implement asynchronous doorbells.

Shared device memory

Devices may have memory-like regions that the CPU can access (such as PCI Memory BARs). The device emulation process therefore needs to share a region of its memory space with the VMM so the guest can access it. This mechanism also allows device emulation to busy wait (poll) instead of using synchronous MMIO/PIO accesses or asynchronous doorbells for notifications.

Direct Memory Access (DMA)

Devices often require read and write access to a memory address space belonging to the CPU. This allows network cards to transmit packet payloads that are located in guest RAM, for example.

Early out-of-process device emulation interfaces simply shared guest RAM. The allowed DMA to any guest physical memory address. More advanced IOMMU and address space identifier mechanisms are now becoming ubiquitous. Therefore, new out-of-process device emulation interfaces should incorporate IOMMU functionality.

The key requirement for IOMMU mechanisms is allowing the VMM to grant access to a region of memory so the device emulation process can read from and/or write to it.

Interrupts

Devices notify the CPU using interrupts. An interrupt is simply a message sent by the device emulation process to the VMM. Interrupt configuration is flexible on modern devices, meaning the driver may be able to select the number of interrupts and a mapping (using one interrupt with multiple event sources). This can be implemented using the Linux eventfd mechanism or via in-band device emulation protocol messages, for example.

Extensibility for new bus types

It should be possible to support multiple bus types. vhost-user only supports vhost devices. VFIO is more extensible but currently focussed on PCI devices. It is likely that QEMU SysBus devices will be desirable for implementing ad-hoc out-of-process devices (especially for System-on-Chip target platforms).

Bus-level APIs, not protocol bindings

Developers should not need to learn the out-of-process device emulation protocol (vfio-user, etc). APIs should focus on bus-level concepts such as defining VIRTIO or PCI devices rather than protocol bindings for dealing with protocol messages, file descriptor passing, and shared memory.

In other words, developers should be thinking in terms of the problem domain, not worrying about how out-of-process device emulation is implemented. The protocol should be hidden behind bus-level APIs.

Multi-threading support from the beginning

Threading issues arise often in device emulation because asynchronous requests or multi-queue devices can be implemented using threads. Therefore it is necessary to clearly document what threading models are supported and how device lifecycle operations like reset interact with in-flight requests.

Live migration, live upgrade, and crash recovery

There are several related issues around device state and restarting the device emulation program without disrupting the guest.

Live migration

Live migration transfers the state of a device from one device emulation process to another (typically running on another host). This requires the following functionality:

Quiescing the device

Some devices can be live migrated at any point in time without any preparation, while others must be put into a quiescent state to avoid issues. An example is a storage controller that has a write request in flight. It is not safe to live migration until the write request has completed or been canceled. Failure to wait might result in data corruption if the write takes effect after the destination has resumed execution.

Therefore it is necessary to quiesce a device. After this point there is no further device activity and no guest-visible changes will be made by the device.

Saving/loading device state

It must be possible to save and load device state. Device state includes the contents of hardware registers as well as device-internal state necessary for resuming operation.

It is typically necessary to determine whether the device emulation processes on the migration source and destination are compatible before attempting migration. This avoids migration failure when the destination tries to load the device state and discovers it doesn't support it. It may be desirable to support loading device state that was generated by a different implementation of the same device type (for example, two virtio-net implementations).

Dirty memory logging

Pre-copy live migration starts with an iterative phase where dirty memory pages are copied from the migration source to the destination host. Devices need to participate in dirty memory logging so that all written pages are transferred to the destination and no pages are "missed".

Crash recovery

If the device emulation process crashes it should be possible to restart it and resume device emulation without disrupting the guest (aside from a possible pause during reconnection).

Doing this requires maintaining device state (contents of hardware registers, etc) outside the device emulation process. This way the state remains even if the process crashes and it can be resume when a new process starts.

Live upgrade

It must be possible to upgrade the device emulation process and the VMM without disrupting the guest. Upgrading the device emulation process is similar to crash recovery in that the process terminates and a new one resumes with the previous state.

Device versioning

The guest-visible aspects of the device must be versioned. In the simplest case the device emulation program would have a --compat-version=Ncommand-line option that controls which version of the device the guest sees. When guest-visible changes are made to the program the version number must be increased.

By giving control of the guest-visible device behavior it is possible to save/load and live migrate reliably. Otherwise loading device state in a newer device emulation program could affect the running guest. Guest drivers typically are not prepared for the device to change underneath them and doing so could result in guest crashes or data corruption.

Security

The trust model

The VMM must not trust the device emulation program. This is key to implementing privilege separation and the principle of least privilege. If a compromised device emulation program is able to gain control of the VMM then out-of-process device emulation has failed to provide isolation between devices.

The device emulation program must not trust the VMM to the extent that this is possible. For example, it must validate inputs so that the VMM cannot gain control of the device emulation process through memory corruptions or other bugs. This makes it so that even if the VMM has been compromised, access to device resources and associated system calls still requires further compromising the device emulation process.

Unprivileged operation

The device emulation program should run unprivileged to the extent that this is possible. If special permissions are required to access hardware resources then these resources can sometimes be provided via file descriptor passing by a more privileged parent process.

Sandboxing

Operating system sandboxing mechanisms can be applied to device emulation processes more effectively than monolithic VMMs. Seccomp can limit the Linux system calls that may be invoked. SELinux can restrict access to system resources.

Sandboxing is a common task that most device emulation programs need. Therefore it is a good candidate for a library or launcher tool that is shared by device emulation programs.

Management

Command-line interface

A common command-line interface should be defined where possible. For example, vhost-user's standard --socket-path=PATH argument makes it easy to launch any vhost-user device backend. Protocol-specific options (e.g. socket path) and device type-specific options (e.g. virtio-net) can be standardized.

Some options are necessarily specific to the device emulation program and therefore cannot be standardized.

The advantage of standard options is that management tools like libvirt can launch the device emulation programs without further user configuration.

RPC interface

It may be necessary to issue commands at runtime. Examples include adjusting throttling limits, enabling/disabling logging, etc. These operations can be performed over an RPC interface.

Various RPC interfaces are used throughout open source virtualization software. Adopting a widely-used RPC protocol and standardizing commands is beneficial because it makes it easy to communicate with the software and management tools can support them relatively easily.

Conclusion

This was largely a brain dump but I hope it is useful food for thought as out-of-process device emulation interfaces are designed and developed. There is a lot more to it than simply implementing a protocol for device register accesses and guest RAM DMA. Developing open source libraries in Rust and C that can be used as needed will ensure that out-of-process devices are high-quality and easy for users to deploy.

The recent development of a Linux driver framework called VIRTIO Data Path Acceleration (vDPA) has laid the groundwork for exciting new vhost-user features. The implications of vDPA have not yet rippled through the community so I want to share my thoughts on how the vhost-user protocol can take advantage of new ideas from vDPA.

This post is aimed at developers and assumes familiarity with the vhost-user protocol and VIRTIO. No knowledge of vDPA is required.

vDPA helps with writing drivers for hardware that supports VIRTIO offload. Its goal is to enable hybrid hardware/software VIRTIO devices, but as a nice side-effect it has overcome limitations in the kernel vhost interface. It turns out that applying ideas from vDPA to the vhost-user protocol solves the same issues there. In this article I'll show how extending the vhost-user protocol with vDPA has the following benefits:

• Allows existing VIRTIO device emulation code to act as a vhost-user device backend.
• Removes the need for shim devices in the virtual machine monitor (VMM).
• Replaces undocumented conventions with a well-defined device model.

These things can be done while reusing existing vhost-user and VIRTIO code. In fact, this is especially good news for existing codebases like QEMU because they already have a wealth of vhost-user and VIRTIO code that can now finally be connected together!

Let's look at the advantages of extending vhost-user with vDPA first and then discuss how to do it.

Why extend vhost-user with vDPA?

Reusing VIRTIO emulation code for vhost-user backends

It is a common misconception that a vhost device is a VIRTIO device. VIRTIO devices are defined in the VIRTIO specification and consist of a configuration space, virtqueues, and a device lifecycle that includes feature negotiation. A vhost device is a subset of the corresponding VIRTIO device. The exact subset depends on the device type, and some vhost devices are closer to the full functionality of their corresponding VIRTIO device than others. The most well-known example is that vhost-net devices have rx/tx virtqueues and but lack the virtio-net control virtqueue. Also, the configuration space and device lifecycle are only partially available to vhost devices.

This difference makes it impossible to use a VIRTIO device as a vhost-user device and vice versa. There is an impedance mismatch and missing functionality. That's a shame because existing VIRTIO device emulation code is mature and duplicating it to provide vhost-user backends creates additional work.

If there was a way to reuse existing VIRTIO device emulation code it would be easier to move to a multi-process architecture in QEMU. Want to run --netdev user,id=netdev0 --device virtio-net-pci,netdev=netdev0 in a separate, sandboxed process? Easy, run it as a vhost-user-net device instead of as virtio-net.

Making VMM device shims optional

Today each new vhost device type requires a shim device in the VMM. QEMU has --device vhost-user-blk-pci, --device vhost-user-input-pci, and so on. Why can't there be a single --device vhost-user device?

This limitation is a consequence of the fact that vhost devices are not full VIRTIO devices. In fact, a vhost device does not even have a way to report its device type (net, blk, scsi, etc). Therefore it is impossible for today's VMMs to offer a generic device. Each vhost device type requires a shim device.

In some cases a shim device is desirable because it allows the VMM to handle some aspects of the device instead of passing everything through to the vhost device backend. But requiring shims by default involves lots of tedious boilerplate code and prevents new device types from being used by older VMMs.

Providing a well-defined device model in vhost-user

Although vhost-user works well for users, it is difficult for developers to learn and extend. The protocol does not have a well-defined device model. Each device type has its own undocumented set of protocol messages that are used. For example, the vhost-user-blk device uses the configuration space whereas most other device types do not use the configuration space at all.

Since protocol use is not fully documented in the specification, developers might resort to reading Linux, QEMU, and DPDK code in order to figure out how to make their devices work. They typically have to debug vhost-user protocol messages and adjust their code until it appears to work. Hopefully the critical bugs are caught before the code ships. This is problematic because it's hard to produce high-quality vhost-user implementations.

Although the protocol specification can certainly be cleaned up, the problem is more fundamental. vhost-user badly needs a well-defined device model so that protocol usage is clear and uniform for all device types. The best way to do that is to simply adopt the VIRTIO device model. The VIRTIO specification already defines the device lifecycle and details of the device types. By making vhost-user devices full VIRTIO devices there is no need for additional vhost device specifications. The vhost-user specification just becomes a transport for the established VIRTIO device model. Luckily that is effectively what vDPA has done for kernel vhost ioctls.

How to do this in QEMU

The following QEMU changes are needed to implement vhost-user vDPA support. Below I will focus on vhost-user-net but most of the work is generic and benefits all device types.

Import vDPA ioctls into vhost-user

vDPA extends the Linux vhost ioctl interface. It uses a subset of vhost ioctls and adds new vDPA-specific ioctls that are implemented in the vhost_vdpa.ko kernel module. These new ioctls enable the full VIRTIO device model, including device IDs, the status register, configuration space, and so on.

In theory vhost-user could be fixed without vDPA, but it would involve effectively adding the same set of functionality that vDPA has already added onto kernel vhost. Reusing the vDPA ioctls allows VMMs to support both kernel vDPA and vhost-user with minimal code duplication.

This can be done by adding a VHOST_USER_PROTOCOL_F_VDPA feature bit to the vhost-user protocol. If both the vhost-user frontend and backend support vDPA then all vDPA messages are available. Otherwise they can either fall back on legacy vhost-user behavior or drop the connection.

The vhost-user specification could be split into a legacy section and a modern vDPA-enabled section. The modern protocol will remove vhost-user messages that are not needed by vDPA, simplifying the protocol for new developers while allowing existing implementations to support both with minimal changes.

One detail is that vDPA does not use the memory table mechanism for sharing memory. Instead it relies on the richer IOMMU message family that is optional in vhost-user today. This approach can be adopted in vhost-user too, making the IOMMU code path standard for all implementations and dropping the memory table mechanism.

Add vhost-user vDPA to the vhost code

QEMU's hw/virtio/vhost*.c code supports kernel vhost, vhost-user, and kernel vDPA. A vhost-user vDPA mode must be added to implement the new protocol. It can be implemented as a combination of the vhost-user and kernel vDPA code already in QEMU. Most likely the existing vhost-user code can simply be extended to enable vDPA support if the backend supports it.

Only small changes to hw/net/virtio-net.c and net/vhost-user.c are needed to use vhost-user vDPA with net devices. At that point QEMU can connect to a vhost-user-net device backend and use vDPA extensions.

Add a vhost-user vDPA VIRTIO transport

Next a vhost-user-net device backend can be put together using QEMU's virtio-net emulation. A translation layer is needed between the vhost-user protocol and the VirtIODevice type in QEMU. This can be done by implementing a new VIRTIO transport alongside the existing pci, mmio, and ccw transports. The transport processes vhost-user protocol messages from the UNIX domain socket and invokes VIRTIO device emulation code inside QEMU. It acts as a VIRTIO bus so that virtio-net-device, virtio-blk-device, and other device types can be plugged in.

This piece is the most involved but the vhost-user protocol communication part was already implemented in the virtio-vhost-user prototype that I worked on previously. Most of the communication code can be reused and the remainder is implementing the VirtioBusClass interface.

To summarize, a new transport acts as the vhost-user device backend and invokes QEMU VirtIODevice methods in response to vhost-user protocol messages. The syntax would be something like --device virtio-net-device,id=virtio-net0 --device vhost-user-backend,device=virtio-net0,addr.type=unix,addr.path=/tmp/vhost-user-net.sock.

Where this converges with multi-process QEMU

At this point QEMU can run ad-hoc vhost-user backends using existing VIRTIO device models. It is possible to go further by creating a qemu-dev launcher executable that implements the vhost-user spec's "Backend program conventions". This way a minimal device emulator executable hosts the device instead of a full system emulator.

The requirements for this are similar to the multi-process QEMU effort, which aims to run QEMU devices as separate processes. One of the main open questions is how to design build system and Kconfig support for building minimal device emulator executables.

In the case of vhost-user-net the qemu-dev-vhost-user-net executable would contain virtio-net-device, vhost-user-backend, any netdevs the user wishes to include, a QMP monitor, and a vhost-user backend command-line interface.

Where does this leave us? QEMU's existing VIRTIO device models can be used as vhost-user devices and run in a separate processes from the VMM. It's a great way of reusing code and having the flexibility to deploy it in the way that makes most sense for the intended use case.

Conclusion

The vhost-user protocol can be simplified by adopting the vhost vDPA ioctls that have recently been introduced in Linux. This turns vhost-user into a VIRTIO transport and vhost-user devices become full VIRTIO devices. Existing VIRTIO device emulation code can then be reused in vhost-user device backends.

As part of GSOC 2019, Athina Plaskasoviti implemented --cloud-init support for virt-install. This post provides a bit more info about the feature.

Why cloud-init

For a long while, most mainstream Linux distros have shipped 'cloud images': raw or qcow2 formatted disk images with the distro minimally pre-installed on it. These images typically have cloud-init set to run on VM bootup. cloud-init can do a variety of things, like add users, change passwords, register ssh keys, and generally perform any desired action on the VM OS. This only works when cloud-init is passed the right configuration by whatever platform is starting the VM, like OpenStack or virt-install. cloud-init supports many different datasources for getting configuration outside the VM.

Historically though the problem is that slapping these images into virt-install or virt-manager gives crappy results, because these tools were not providing any datasource. In this case, cloud-init reverts to its distro default configured behavior, which in most cases is unusable. For example on Fedora, the result was: hang waiting for cloud-init data, time out, drop to login prompt with all accounts locked.

Prior to virt-install --cloud-init support, the simplest workaround was to use libguestfs, specifically virt-customize, to set a root password and disable cloud-init:

$ virt-customize -a MY-CLOUD-IMAGE.qcow2 \
    --root-password password:SUPER-SECRET-PASSWORD \
    --uninstall cloud-init

--cloud-init option

The --cloud-init option will tell virt-install to set up a nocloud datasource via a specially formatted .iso file that is generated on the fly, and only used for the first VM bootup.

The default behavior when --cloud-init is specified with no suboptions will do the following:

• Generate a random root password and print it to the terminal.
• Default to VM serial console access rather than graphical console access. Makes it easier to paste the password and gives more ssh-like behavior.
• Sets the root password to expire on first login. So the temporary password is only used once.
• Disables cloud-init for subsequent VM startups. Otherwise on the next VM boot you'd face locked accounts again.

The --cloud-init also has suboptions to specify your own behavior, like transfer in a host ssh public key, pass in raw cloud-init user-data/meta-data, etc. See the virt-install --cloud-init man page section for the specifics.

Room for improvement

This is all a step in the right direction but there's lots more we can do here:

• Extend virt-install's --install option to learn how to fetch cloud images for the specified distro. libosinfo and osinfo-db already track cloud image download links for many distros so the info we need is already in place. We could make virt-install --install fedora32,cloud=yes a single way to pull down a cloud image, generate a cloud-init datasource, and create the VM in one shot.
• Use --cloud-init by default when the user passes us a cloud-init enabled disk image. virt-customize has a lot of disk image detection smarts already, but we aren't using that in virt-install yet.
• virt-manager UI support. There's an issue tracking this with some more thoughts in it.
• Similar support for CoreOS Ignition which fulfills a similar purpose as cloud-init for distros like Fedora CoreOS. There's an issue tracking this too.

I personally don't have plans to work on these any time soon, but I'm happy to provide guidance if anyone is interested in helping out!

Abstract
Join us for this webinar session to get an overview of the different networking configuration choices running KVM guests on the IBM Z platforms and details to setup and capabilities. It also provides tuning recommendations for the KVM host and KVM guest environment to achieve greater network performance.

Speaker
Dr. Jürgen Dölle, Linux on IBM Z Performance Evaluation

Registration
Register here. You can check the system requirements here.
After registering, you will receive a confirmation email containing information about joining the webinar.

Replay & Archive
All sessions are recorded. For the archive as well as a replay and handout of this session and all previous webcasts see here.

QEMU is participating in the Outreachy December-March open source internship program. The internships are 12 weeks of full-time paid remote work contributing to QEMU/KVM. For more information about eligibility and what the internships are like, please see the Outreachy website.

The initial application deadline is September 20 and it only takes 5-30 minutes to complete: https://www.outreachy.org/apply/

If you or someone you know are considering doing an internship with QEMU, Linux kernel, or another participating organization, please remember to submit an initial application by September 20th!

Yesterday I released virt-manager 3.0.0. Despite the major version number bump, things shouldn't look too different from the previous release. For me the major version number bump reflects certain feature removals (like dropping virt-convert), and the large amount of internal code changes that were done, though there's a few long awaited features sprinkled in like virt-install --cloud-init support which I plan to write more about later.

This release includes:

• virt-install --cloud-init support (Athina Plaskasoviti, Cole Robinson)
• The virt-convert tool has been removed. Please use virt-v2v instead
• A handful of UI XML configuration options have been removed, in an effort to reduce maintenance ongoing maintenance burden, and to be more consistent with what types of UI knobs we expose. The XML editor is an alternative in most cases. For a larger discussion see this thread and virt-manager's DESIGN.md file.
• The 'New VM' UI now has a 'Manual Install' option which creates a VM without any required install media
• In the 'New VM' UI, the network/pxe install option has been removed. If you need network boot, choose 'Manual Install' and set the boot device after initial VM creation
• 'Clone VM' UI has been reworked and simplified
• 'Migrate VM' UI now has an XML editor for the destination VM
• Global and per-vm option to disable graphical console autoconnect. This makes it easier to use virt-manager alongside another client like virt-viewer
• virt-manager: set guest time after VM restore (Michael Weiser)
• virt-manager: option to delete storage when removing disk device (Lily Nie)
• virt-manager: show warnings if snapshot operation is unsafe (Michael Weiser)
• Unattended install improvements (Fabiano Fidêncio)
• cli: new --xml XPATH=VAL option for making direct XML changes
• virt-install: new --reinstall=DOMAIN option
• virt-install: new --autoconsole text|graphical|none option
• virt-install: new --os-variant detect=on,require=on suboptions
• cli: --clock, --keywrap, --blkiotune, --cputune additions (Athina Plaskasoviti)
• cli: add --features kvm.hint-dedicated.state= (Menno Lageman)
• cli: add --iommu option (Menno Lageman)
• cli: Add --graphics websocket= support (Petr Benes)
• cli: Add --disk type=nvme source.* suboptions
• cli: Fill in all --filesystem suboptions
• Translation string improvements (Pino Toscano)
• Convert from .pod to .rst for man pages
• Switch to pytest as our test runner
• Massively improved unittest and uitest code coverage
• Now using github issues as our bug tracker

This article introduces QEMU storage concepts including disk images, emulated storage controllers, block jobs, the qemu-img utility, and qemu-storage-daemon. If you are new to QEMU or want an overview of storage functionality in QEMU then this article explains how things fit together.

Storage technologies

Persistently storing data and retrieving it later is the job of storage devices such as hard disks, solid state drives (SSDs), USB flash drives, network attached storage, and many others. Technologies vary in their storage capacity (disk size), access speed, price, and other factors but most of them follow the same block device model.

Block devices are accessed in storage units called blocks. It is not possible to access individual bytes, instead an entire block must be transferred. Block sizes vary between devices with 512 bytes and 4KB block sizes being the most common.

As an emulator and virtualizer of computer systems, QEMU naturally has to offer block device functionality. QEMU is capable of emulating hard disks, solid state drives (SSDs), USB flash drives, SD cards, and more.

Storage for virtual machines

There is more to storage than just persisting data on behalf of a virtual machine. The lifecycle of a disk image includes several operations that are briefly covered below.

Virtual machines consist of device configuration (how much RAM, which graphics card, etc) and the contents of their disks. Transferring virtual machines either to migrate them between hosts or to distribute them to users is an important workflow that QEMU and its utilities support.

Much like ISO files are used to distribute operating system installer images, QEMU supports disk image file formats that are more convenient for transferring disk images than the raw contents of a disk. In fact, disk image file formats offer many other features such as the ability to import/export disks from other hypervisors, snapshots, and instantiating new disk images from a backing file.

Finally, managing disk images also involves the ability to take backups and restore them should it be necessary to roll back after the current disk contents have been lost or corrupted.

Emulated storage controllers

The virtual machine accesses block devices through storage controllers. These are the devices that the guest talks to in order to read or write blocks. Some storage controllers facilitate access to multiple block devices, such as a SCSI Host Bus Adapter that provides access to many SCSI disks.

Storage controllers vary in their features, performance, and guest operating system support. They expose a storage interface such as virtio-blk, NVMe, or SCSI. Virtual machines program storage controller registers to transfer data between memory buffers in RAM and block devices. Modern storage controllers support multiple request queues so that I/O can processed in parallel at high rates.

The most common storage controllers in QEMU are virtio-blk, virtio-scsi, AHCI (SATA), IDE for legacy systems, and SD Card controllers on embedded or smaller boards.

Disk image file formats

Disk image file formats handle the layout of blocks within a host file or device. The simplest format is the raw format where each block is located at its Logical Block Address (LBA) in the host file. This simple scheme does not offer much in the way of features.

QEMU’s native disk image format is QCOW2 and it offers a number of features:

• Compactness - the host file grows as blocks are written so a sparse disk image can be much smaller than the virtual disk size.
• Backing files - disk images can be based on a parent image so that a master image can be shared by virtual machines.
• Snapshots - the state of the disk image can be saved and later reverted.
• Compression - block compression reduces the image size.
• Encryption - the disk image can be encrypted to protect data at rest.
• Dirty bitmaps - backup applications can track changed blocks so that efficient incremental backups are possible.

A number of other disk image file formats are available for importing/exporting disk images for use with other software including VMware and Hyper-V.

Block jobs

Block jobs are background operations that manipulate disk images:

• Commit - merging backing files to shorten a backing file chain.
• Backup - copying out a point-in-time snapshot of a disk.
• Mirror - copying an image to a new destination while the virtual machine can still write to it.
• Stream - populating a disk image from its backing file.
• Create - creating new disk image files.

These background operations are powerful tools for building storage migration and backup workflows.

Some operations like mirror and stream can take a long time because they copy large amounts of data. Block jobs support throttling to limit the performance impact on virtual machines.

qemu-img and qemu-storage-daemon

The qemu-img utility manipulates disk images. It can create, resize, snapshot, repair, and inspect disk images. It has both human-friendly and JSON output formats, making it suitable for manual use as well as scripting.

qemu-storage-daemon exposes QEMU’s storage functionality in a server process without running a virtual machine. It can export disk images over the Network Block Device (NBD) protocol as well as run block jobs and other storage commands. This makes qemu-storage-daemon useful for applications that want to automate disk image manipulation.

Conclusion

QEMU presents block devices to virtual machines via emulated storage controllers. On the host side the disk image file format, block jobs, and qemu-img/qemu-storage-daemon utilities provide functionality for working with disk images. Future blog posts will dive deeper into some of these areas and describe best practices for configuring storage.