Skip to content

Latest commit

 

History

History
433 lines (334 loc) · 19.9 KB

prod-host-setup.md

File metadata and controls

433 lines (334 loc) · 19.9 KB

Production Host Setup Recommendations

Firecracker relies on KVM and on the processor virtualization features for workload isolation. The host and guest kernels and host microcode must be regularly patched in accordance with your distribution's security advisories such as ALAS for Amazon Linux.

Security guarantees and defense in depth can only be upheld, if the following list of recommendations are implemented in production.

Firecracker Configuration

Seccomp

Firecracker uses seccomp filters to limit the system calls allowed by the host OS to the required minimum.

By default, Firecracker uses the most restrictive filters, which is the recommended option for production usage.

Production usage of the --seccomp-filter or --no-seccomp parameters is not recommended.

8250 Serial Device

Firecracker implements the 8250 serial device, which is visible from the guest side and is tied to the Firecracker/non-daemonized jailer process stdout. Without proper handling, because the guest has access to the serial device, this can lead to unbound memory or storage usage on the host side. Firecracker does not offer users the option to limit serial data transfer, nor does it impose any restrictions on stdout handling. Users are responsible for handling the memory and storage usage of the Firecracker process stdout. We suggest using any upper-bounded forms of storage, such as fixed-size or ring buffers, using programs like journald or logrotate, or redirecting to /dev/null or a named pipe. Furthermore, we do not recommend that users enable the serial device in production. To disable it in the guest kernel, use the 8250.nr_uarts=0 boot argument when configuring the boot source. Please be aware that the device can be reactivated from within the guest even if it was disabled at boot.

If Firecracker's stdout buffer is non-blocking and full (assuming it has a bounded size), any subsequent writes will fail, resulting in data loss, until the buffer is freed.

Log files

Firecracker outputs logging data into a named pipe, socket, or file using the path specified in the log_path field of logger configuration. Firecracker can generate log data as a result of guest operations and therefore the guest can influence the volume of data written in the logs. Users are responsible for consuming and storing this data safely. We suggest using any upper-bounded forms of storage, such as fixed-size or ring buffers, programs like journald or logrotate, or redirecting to a named pipe.

Logging and performance

We recommend adding quiet loglevel=1 to the host kernel command line to limit the number of messages written to the serial console. This is because some host configurations can have an effect on Firecracker's performance as the process will generate host kernel logs during normal operations.

The most recent example of this was the addition of console=ttyAMA0 host kernel command line argument on one of our testing setups. This enabled console logging, which degraded the snapshot restore time from 3ms to 8.5ms on aarch64. In this case, creating the tap device for snapshot restore generated host kernel logs, which were very slow to write.

Logging and signal handlers

Firecracker installs custom signal handlers for some of the POSIX signals, such as SIGSEGV, SIGSYS, etc.

The custom signal handlers used by Firecracker are not async-signal-safe, since they write logs and flush the metrics, which use locks for synchronization. While very unlikely, it is possible that the handler will intercept a signal on a thread which is already holding a lock to the log or metrics buffer. This can result in a deadlock, where the specific Firecracker thread becomes unresponsive.

While there is no security impact caused by the deadlock, we recommend that customers have an overwatcher process on the host, that periodically looks for Firecracker processes that are unresponsive, and kills them, by SIGKILL.

Jailer Configuration

For assuring secure isolation in production deployments, Firecracker should be started using the jailer binary that's part of each Firecracker release, or executed under process constraints equal or more restrictive than those in the jailer. For more about Firecracker sandboxing please see Firecracker design

The Jailer process applies cgroup, namespace isolation and drops privileges of the Firecracker process.

To set up the jailer correctly, you'll need to:

  • Create a dedicated non-privileged POSIX user and group to run Firecracker under. Use the created POSIX user and group IDs in Jailer's --uid <uid> and --gid <gid> flags, respectively. This will run the Firecracker as the created non-privileged user and group. All file system resources used for Firecracker should be owned by this user and group. Apply least privilege to the resource files owned by this user and group to prevent other accounts from unauthorized file access. When running multiple Firecracker instances it is recommended that each runs with its unique uid and gid to provide an extra layer of security for their individually owned resources in the unlikely case where any one of the jails is broken out of.

Firecracker's customers are strongly advised to use the provided resource-limits and cgroup functionalities encapsulated within jailer, in order to control Firecracker's resource consumption in a way that makes the most sense to their specific workload. While aiming to provide as much control as possible, we cannot enforce aggressive default constraints resources such as memory or CPU because these are highly dependent on the workload type and usecase.

Here are some recommendations on how to limit the process's resources:

Disk

  • cgroup provides a Block IO Controller which allows users to control I/O operations through the following files:

    • blkio.throttle.io_serviced - bounds the number of I/Os issued to disk
    • blkio.throttle.io_service_bytes - sets a limit on the number of bytes transferred to/from the disk
  • Jailer's resource-limit provides control on the disk usage through:

    • fsize - limits the size in bytes for files created by the process
    • no-file - specifies a value greater than the maximum file descriptor number that can be opened by the process. If not specified, it defaults to 4096.

Memory

  • cgroup provides a Memory Resource Controller to allow setting upper limits to memory usage:
    • memory.limit_in_bytes - bounds the memory usage
    • memory.memsw.limit_in_bytes - limits the memory+swap usage
    • memory.soft_limit_in_bytes - enables flexible sharing of memory. Under normal circumstances, control groups are allowed to use as much of the memory as needed, constrained only by their hard limits set with the memory.limit_in_bytes parameter. However, when the system detects memory contention or low memory, control groups are forced to restrict their consumption to their soft limits.

vCPU

  • cgroup’s CPU Controller can guarantee a minimum number of CPU shares when a system is busy and provides CPU bandwidth control through:
    • cpu.shares - limits the amount of CPU that each group it is expected to get. The percentage of CPU assigned is the value of shares divided by the sum of all shares in all cgroups in the same level
    • cpu.cfs_period_us - bounds the duration in us of each scheduler period, for bandwidth decisions. This defaults to 100ms
    • cpu.cfs_quota_us - sets the maximum time in microseconds during each cfs_period_us for which the current group will be allowed to run
    • cpuacct.usage_percpu - limits the CPU time, in ns, consumed by the process in the group, separated by CPU

Additional details of Jailer features can be found in the Jailer documentation.

Host Security Configuration

Constrain CPU overhead caused by kvm-pit kernel threads

The current implementation results in host CPU usage increase on x86 CPUs when a guest injects timer interrupts with the help of kvm-pit kernel thread. kvm-pit kthread is by default part of the root cgroup.

To mitigate the CPU overhead we recommend two system level configurations.

  1. Use an external agent to move the kvm-pit/<pid of firecracker> kernel thread in the microVM’s cgroup (e.g., created by the Jailer). This cannot be done by Firecracker since the thread is created by the Linux kernel after guest start, at which point Firecracker is de-privileged.
  2. Configure the kvm limit to a lower value. This is a system-wide configuration available to users without Firecracker or Jailer changes. However, the same limit applies to APIC timer events, and users will need to test their workloads in order to apply this mitigation.

To modify the kvm limit for interrupts that can be injected in a second.

  1. sudo modprobe -r (kvm_intel|kvm_amd) kvm
  2. sudo modprobe kvm min_timer_period_us={new_value}
  3. sudo modprobe (kvm_intel|kvm_amd)

To have this change persistent across boots we can append the option to /etc/modprobe.d/kvm.conf:

echo "options kvm min_timer_period_us=" >> /etc/modprobe.d/kvm.conf

Mitigating Network flooding issues

Network can be flooded by creating connections and sending/receiving a significant amount of requests. This issue can be mitigated either by configuring rate limiters for the network interface as explained within Network Interface documentation, or by using one of the tools presented below:

  • tc qdisc - manipulate traffic control settings by configuring filters.

When traffic enters a classful qdisc, the filters are consulted and the packet is enqueued into one of the classes within. Besides containing other qdiscs, most classful qdiscs perform rate control.

  • netnamespace and iptables
    • --pid-owner - can be used to match packets based on the PID that was responsible for them
    • connlimit - restricts the number of connections for a destination IP address/from a source IP address, as well as limit the bandwidth

Mitigating Noisy-Neighbour Storage Device Contention

Data written to storage devices is managed in Linux with a page cache. Updates to these pages are written through to their mapped storage devices asynchronously at the host operating system's discretion. As a result, high storage output can result in this cache being filled quickly resulting in a backlog which can slow down I/O of other guests on the host.

To protect the resource access of the guests, make sure to tune each Firecracker process via the following tools:

  • Jailer: A wrapper environment designed to contain Firecracker and strictly control what the process and its guest has access to. Take note of the jailer operations guide, paying particular note to the --resource-limit parameter.
  • Rate limiting: Rate limiting functionality is supported for both networking and storage devices and is configured by the operator of the environment that launches the Firecracker process and its associated guest. See the block device documentation for examples of calling the API to configure rate limiting.

Disabling swapping to disk or enabling secure swap

Memory pressure on a host can cause memory to be written to drive storage when swapping is enabled. Disabling swap mitigates data remanence issues related to having guest memory contents on microVM storage devices.

Verify that swap is disabled by running:

grep -q "/dev" /proc/swaps && \
echo "swap partitions present (Recommendation: no swap)" \
|| echo "no swap partitions (OK)"

Mitigating hardware vulnerabilities

Caution

Firecracker is not able to mitigate host's hardware vulnerabilities. Adequate mitigations need to be put in place when configuring the host.

Caution

Firecracker is designed to provide isolation boundaries between microVMs running in different Firecracker processes. It is strongly recommended that each Firecracker process corresponds to a workload of a single tenant.

Caution

For security and stability reasons it is highly recommended to load updated microcode as soon as possible. Aside from keeping the system firmware up-to-date, when the kernel is used to load updated microcode of the CPU this should be done as early as possible in the boot process.

Side channel attacks

For the purposes of this document we assume a workload that involves arbitrary code execution in a multi-tenant context where each Firecracker process corresponds to a single tenant.

Specific mitigations for side channel issues are constantly evolving as researchers find additional issues on a regular basis. Firecracker itself has no control over many lower-level software and hardware behaviors and capabilities and is not able to mitigate all these issues. Thus, it is strongly recommended that users follow the very latest Linux kernel documentation on hardware vulnerabilities as well as hardware/processor-specific recommendations and firmware updates (see vendor-specific recommendations below) when configuring mitigations against side channel attacks including "Spectre" and "Meltdown" attacks.

However, some generic recommendations are also provided in what follows.

Disable SMT

Simultaneous Multi-Threading (SMT) is frequently a precondition for speculation issues utilized in side channel attacks such as Spectre variants, MDS, and others, where one tenant could leak information to another tenant or the host. As such, our recommendation is to disable SMT in production scenarios that require tenant separation.

Disable Kernel Samepage Merging

Users should disable Kernel Samepage Merging to mitigate side channel issues that rely on page deduplication for revealing what memory pages are accessed by another process.

Use memory with Rowhammer mitigation support

Rowhammer is a memory side-channel issue that can lead to unauthorized cross- process memory changes.

Using DDR4 memory that supports Target Row Refresh (TRR) with error-correcting code (ECC) is recommended. Use of pseudo target row refresh (pTRR) for systems with pTRR-compliant DDR3 memory can help mitigate the issue, but it also incurs a performance penalty.

Vendor-specific recommendations

For vendor-specific recommendations, please consult the resources below:

[ARM only] Physical counter directly passed through to the guest

On ARM, the physical counter (i.e CNTPCT) it is returning the actual EL1 physical counter value of the host. From the discussions before merging this change upstream, this seems like a conscious design decision of the ARM code contributors, giving precedence to performance over the ability to trap and control this in the hypervisor.

Verification

spectre-meltdown-checker script can be used to assess host's resilience against several transient execution CVEs and receive guidance on how to mitigate them.

The script is used in integration tests by the Firecracker team. It can be downloaded and executed like:

# Read https://meltdown.ovh before running it.
wget -O - https://meltdown.ovh | bash

Linux 6.1 boot time regressions

Linux 6.1 introduced some regressions in the time it takes to boot a VM, for the x86_64 architecture. They can be mitigated depending on the CPU and the version of cgroups in use.

Explanation

The regression happens in the KVM_CREATE_VM ioctl and there are two factors that cause the issue:

  1. In the implementation of the mitigation for the iTLB multihit vulnerability, KVM creates a worker thread called kvm-nx-lpage-recovery. This thread is responsible for recovering huge pages split when the mitigation kicks-in. In the process of creating this thread, KVM calls cgroup_attach_task_all() to move it to the same cgroup used by the hypervisor thread
  2. In kernel v4.4, upstream converted a cgroup per process read-write semaphore into a per-cpu read-write semaphore to allow to perform operations across multiple processes (commit). It was found that this conversion introduced high latency for write paths, which mainly includes moving tasks between cgroups. This was fixed in kernel v4.9 by commit which chose to favor writers over readers since moving tasks between cgroups is a common operation for Android. However, In kernel 6.0, upstream decided to revert back again and favor readers over writers re-introducing the original behavior of the rw semaphore (commit). At the same time, this commit provided an option called favordynmods to favor writers over readers.
  3. Since the kvm-nx-lpage-recovery thread creation and its cgroup change is done in the KVM_CREATE_VM call, the high latency we observe in 6.1 is due to the upstream decision to favor readers over writers for this per-cpu rw semaphore. While the 4.14 and 5.10 kernels favor writers over readers.

The first step is to check if the host is vulnerable to iTLB multihit. Look at the value of cat /sys/devices/system/cpu/vulnerabilities/itlb_multihit. If it does says Not affected, the host is not vulnerable and you can apply mitigation 2, and optionally 1 for best results. Otherwise it is vulnerable and you can only apply mitigation 1.

Mitigation 1: favordynmods

The mitigation in this case is to enable favordynmods in cgroupsv1 or cgroupsv2. This changes the behavior of all cgroups in the host, and makes it closer to the performance of Linux 5.10 and 4.14.

For cgroupsv2, run this command:

sudo mount -o remount,favordynmods /sys/fs/cgroup

For cgroupsv1, remounting with favordynmods is not supported, so it has to be done at boot time, through a kernel command line option. Add cgroup_favordynmods=true to your kernel command line in GRUB. Refer to your distribution's documentation for where to make this change1

Mitigation 2: kvm.nx_huge_pages=never

This mitigation is preferred to the previous one as it is less invasive (it doesn't affect other cgroups), but it can also be combined with the cgroups mitigation.

KVM_VENDOR_MOD=$(lsmod |grep -P "^kvm_(amd|intel)" | awk '{print $1}')
sudo modprobe -r $KVM_VENDOR_MOD kvm
sudo modprobe kvm nx_huge_pages=never
sudo modprobe $KVM_VENDOR_MOD

To validate that the change took effect, the file /sys/module/kvm/parameters/nx_huge_pages should say never.

Footnotes

  1. Look for GRUB_CMDLINE_LINUX in file /etc/default/grub in RPM-based systems, and this doc for Ubuntu.