How Microsoft made Linux faster, safer, and more reliable for everyone
At this time of Ignite, the spotlight is on innovations. New AI infrastructure, new silicon, and new services take center stage. This is also the right moment to highlight a different kind of innovation. It is the work Microsoft contributes to the open-source community, the work that rarely makes headlines but makes much of what you see at Ignite possible.
The Base OS (a.k.a LSG previously) at Microsoft builds operating-system technologies that help power Azure, including the Linux foundation for Azure Boost and capabilities spanning networking, virtualization, confidential computing, security, system lifecycle, and quality. The team works across the Linux kernel and projects such as Cloud Hypervisor, rust-vmm, systemd, DPDK, and LISA.
Our approach remains upstream first: solve engineering challenges in the open, collaborate with maintainers and partners, and contribute improvements to the shared projects on which the industry depends. In 2026, that approach helped strengthen Linux for high-performance cloud and AI workloads while improving reliability, security, serviceability, and validation across the ecosystem
Some of key Highlights Microsoft contributed upstream:
- Improved MANA networking efficiency, recovery, memory resilience, and diagnostics for large cloud and AI workloads.
- Expanded Arm64 support across the Microsoft Hypervisor, Hyper-V, Cloud Hypervisor, and confidential-computing scenarios.
- Completed the upstreaming of the mshv_vtl driver and continued enabling OpenHCL capabilities in mainline Linux.
- Hardened Linux networking for confidential VMs by validating data received from an untrusted host and correcting memory-safety issues.
- Advanced kernel memory-management and live-update foundations, including pre-merge testing for memory-management patches.
- Added systemd, kexec-tools, and OpenEmbedded-Core capabilities that improve controlled servicing and reproducible OS builds.
- Made LISA easier to install and extended it to bare-metal, IPv6, health-driven test selection, and new validation scenarios.
- Supported the open-source community through maintainership, code review, conference leadership, and Linux kernel meetups
Linux Kernel Contributions
Networking for Cloud and AI Scale:
High-performance AI and cloud workloads depend on networking that is fast, resilient, and diagnosable. In 2026, Microsoft contributed a broad set of improvements to the Microsoft Azure Network Adapter -MANA - driver for customer running their AI workloads on Azure.
Higher efficiency and throughput. Receive-completion coalescing and Dynamic Interrupt Moderation allow the driver to process more packets per interrupt and reduce CPU overhead. Extended coalescing support helped Linux reach line-rate performance on Azure Cobalt 200 in team testing. Microsoft also added SR-IOV support needed for Linux-hosted, multi-VM scenarios on next-generation AI platforms.
Recovery without a VM reboot. New transmit-timeout and hardware-channel health handling allow the driver to detect stalled networking and initiate recovery automatically. The goal is to turn failures that once required a full VM restart into recoverable device events.
Reliability under memory pressure. New queue-access abstractions and a fallback from large contiguous allocations to smaller scattered pages allow MANA queues to initialize when sufficient memory is available but physically contiguous memory is fragmented. Additional changes improved interrupt placement on smaller VMs and reserved completion-queue capacity for a corner case that could otherwise stall networking.
Better diagnostic evidence. The driver now preserves the actual cause of receive-queue setup failures instead of reporting every error as an out-of-memory condition. A diagnostic control can preserve device and queue state after a transmit timeout, while normal production recovery remains enabled by default. WALinuxAgent also gained bounded soft-lockup stack and kernel-log capture, creating more actionable evidence when a Linux VM stops making progress.
Microsoft also contributed DPDK changes that improve behavior during device servicing, PCI unplug, hardware errors, and multi-port testing -important capabilities for low-latency, HPC, and AI environments.
- MANA: Dynamic Interrupt Moderation
- MANA: SR-IOV handler
- MANA: transmit-timeout diagnostic control
- MANA: receive-queue error reporting
- WALinuxAgent: soft-lockup diagnostic capture
- DPDK: simplified multi-port testing
Virtualization and Arm64 Enablement:
In 2026, Microsoft expanded Arm64 support across the Linux virtualization stack. Contributions to the Linux kernel and Cloud Hypervisor enabled direct virtualization with the Microsoft Hypervisor on Arm64, including hypervisor interfaces, firmware boot support, interrupt-controller fixes, and broader continuous-integration coverage.
A major milestone was the completion of the multi-year effort to merge the mshv_vtl driver into mainline Linux. This reduces the need for long-lived downstream code and creates an upstream foundation for future OpenHCL improvements. Follow-on work introduced an Arm64 enablement series and resolved kernel and block-layer issues that were preventing adoption of Linux 6.18.
We also made virtual PCI probing asynchronous so devices can be discovered in parallel rather than serially, reducing enumeration time in the tested Hyper-V scenario. Other fixes improved interrupt handling during CPU hot-unplug, mapped hypercall failures to meaningful Linux errors, removed boot blockers, and added clearer warnings when a guest is waiting too long for a host PCI response.
Across Cloud Hypervisor and rust-vmm, Microsoft reported more than 170 upstream commits during the second and third quarters of 2026. The work covered bug fixes, MSHV releases, confidential-VM fixes, storage correctness, guest-facing input validation, Arm64 testing, and test-framework integration. Together, these contributions help keep the virtualization stack stable, testable, and ready for wider adoption..
- PCI Hyper-V: asynchronous virtual PCI probing
- PCI and IRQ: CPU hot-unplug interrupt handling
- Hyper-V: meaningful hypercall error mapping
- Cloud Hypervisor
- rust-vmm community
Confidential Computing, Security, and Compliance
Confidential VMs change the trust boundary: the guest must treat values supplied by the host as untrusted. Microsoft’s 2026 MANA hardening work added validation for queue counts, MTU values, doorbell identifiers, request indexes, and other device responses. Also corrected time-of-check/time-of-use, use-after-free, double-free, and out-of-bounds issues that could surface during shutdown, live migration, hardware reset, or dynamic reconfiguration.
For Arm confidential computing, Microsoft contributed Hyper-V support for Arm Confidential Compute Architecture Realm guests and updated networking buffers so they can be shared correctly in confidential-VM scenarios. These changes address both guest enablement and the networking path needed to make the scenario practical.
Microsoft also advanced FIPS 140 readiness in the Linux kernel by prohibiting non-compliant PKCS#1 SHA-1 hashing while the kernel is operating in FIPS mode. In parallel, Microsoft continued upstream maintainership and review across the Linux Security Module framework, SELinux, Audit, labeled networking, libseccomp, and related user-space projects.
- Arm64 and Hyper-V: Realm guest support series
- Hyper-V networking buffers for confidential VMs
- RDMA: MANA capability-field hardening
- Kernel crypto: FIPS-mode RSA and SHA-1 change
Memory Management and Live Update Foundations
Last year, we celebrated Kexec HandOver (KHO) landing in Linux 6.16. In 2026, Microsoft engineers continued as upstream maintainers for memory management, KHO, and LiveUpdate, reviewing and merging community patches and sending pull requests upstream. We also started a memory-management CI project that tests patches before they enter the memory-management tree, so problems are caught earlier.
On the code side, we replaced older allocation patterns with kmalloc() where it makes sense, closing long-standing technical debt. We refactored userfaultfd to prepare for guest_memfd support and improved how the kernel manages executable memory, including the caches used by BPF.
We also added noiommu mode support to iommufd, so devices in that mode can now take part in live update. It is a building block that expands the device scenarios live update can support.
Developer Tools and OS Lifecycle
More Controlled System Management with systemd
Microsoft contributed a new set of systemd features and fixes in 2026. These included support for restarting delegated services more reliably, an EncryptKDF= option for LUKS2 partitions, a RouteTable= option for DHCPv6, improved LUKS1 handling in systemd-dissect, and more resilient behavior when systemd-gpt-auto-generator is built without libcryptsetup.
A new networkctl reload --no-configure option separates loading network configuration from applying it. Operators can stage updated settings and then selectively reconfigure interfaces, providing finer control over when changes take effect.
- systemd: delegated-service restart fix
- systemd: EncryptKDF= for LUKS2
- systemd: RouteTable= for DHCPv6
- systemd: networkctl reload --no-configure
Signed-Kernel Servicing and Reproducible Builds
A kexec-tools fix now preserves extended attributes such as IMA signatures when a new kernel image is loaded. This enables systems that depend on signed kernel images to adopt newer kexec-tools releases without losing the metadata required by their security policies.
Microsoft also contributed five reproducibility fixes to OpenEmbedded-Core. The changes remove build-machine details and nondeterministic ordering from generated artifacts across libtool, test Makefiles, Python wheels, Rust data, and Cargo metadata. Another contribution improved the upstream mechanism for retaining build-failure diagnostics. Together, this work helps ensure that build results reflect the intended source and configuration rather than differences in the machine that performed the build.
- OpenEmbedded-Core
- OpenEmbedded-Core: libtool reproducibility improvement
- OpenEmbedded-Core: Rust-data ordering improvement
- OpenEmbedded-Core: build-failure retention controls
Empowering Linux Quality at Scale
LISA Linux Integration Services Automation is Microsoft’s open-source framework for validating Linux kernels, distributions, and images. In 2026, Microsoft made LISA easier to adopt and expanded what it can validate:
- Python packaging and publishing support for simpler installation.
- Quick-start scripts for WSL and local development environments.
- Bare-metal platform support with pipeline integration and documentation.
- Health-driven test selection that uses platform signals to target validation.
- Broader IPv6 compatibility across networking code and tests.
- New validation for Hyper-V Dynamic Memory and BBR3 TCP congestion control.
- Automated pull-request validation and AI-assisted log analysis.
- Broader coverage for newer Linux distributions and Arm64, along with fixes that reduce false failures.
Because LISA is open source, distribution partners and other contributors can run the same tests, propose improvements, and collaborate on shared quality goals. This shortens the distance between upstream development, partner validation, and reliable Linux experiences in the cloud.
Community Engagement and Leadership
Open source is sustained by more than code. In 2026, Microsoft engineers continued reviewing changes across Linux memory management, Hyper-V, VMBus, PCI, IOMMU, and other shared components. That review work helps prevent regressions and brings platform experience into community design discussions.
Microsoft engineers also served on the program committee for Open Source Summit India 2026 and supported the April and September 2026 Bangalore Linux Kernel Meetups, including co-hosting the September event at Microsoft’s Bengaluru campus. The team also shared systemd storage-management work at SCALE. These activities help create space for technical exchange, strengthen local contributor communities, and encourage students and early-career engineers to participate in Linux development.
Closing Thoughts
At Ignite 2026, we are celebrating innovation but open-source innovation is never a single-company effort. It is built through shared design, public review, sustained maintainership, and collaboration across companies, distributions, silicon partners, and individual contributors.
Microsoft will continue to develop upstream first, work with distribution partners so improvements reach supported kernels, and invest openly in the Linux technologies that build cloud and AI infrastructure. None of this happens alone. Thank you to the maintainers, reviewers, and contributors across the Linux community who reviewed our patches and pushed us to do better.