Firecracker relies on KVM and on the processor virtualization features for workload isolation. Security guarantees and defense in depth can only be upheld, if the following list of recommendations are implemented in production.
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.
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.
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.
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.
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.
For assuring secure isolation in production deployments, Firecracker should
must 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 uniqueuid
andgid
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:
-
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 diskblkio.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 processno-file
- specifies a value one greater than the maximum file descriptor number that can be opened by the process. If not specified, it defaults to 4096.
cgroup
provides a Memory Resource Controller to allow setting upper limits to memory usage:memory.limit_in_bytes
- bounds the memory usagememory.memsw.limit_in_bytes
- limits the memory+swap usagememory.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 thememory.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.
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 allcgroups
in the same levelcpu.cfs_period_us
- bounds the duration in us of each scheduler period, for bandwidth decisions. This defaults to 100mscpu.cfs_quota_us
- sets the maximum time in microseconds during eachcfs_period_us
for which the current group will be allowed to runcpuacct.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.
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.
- 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. - 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.
sudo modprobe -r (kvm_intel|kvm_amd) kvm
sudo modprobe kvm min_timer_period_us={new_value}
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
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 qdisk
- 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
andiptables
--pid-owner
- can be used to match packets based on the PID that was responsible for themconnlimit
- restricts the number of connections for a destination IP address/from a source IP address, as well as limit the bandwidth
When deploying Firecracker microVMs to handle multi-tenant workloads, the following host environment configurations are strongly recommended to guard against side-channel security issues.
Some of the mitigations are platform specific. When applicable, this information will be specified between brackets.
Disabling SMT will help mitigate side-channels issues between sibling threads on the same physical core.
SMT can be disabled by adding the following Kernel boot parameter to the host:
nosmt=force
Verification can be done by running:
(grep -q "^forceoff$" /sys/devices/system/cpu/smt/control && \
echo "Hyperthreading: DISABLED (OK)") || \
(grep -q "^notsupported$\|^notimplemented$" \
/sys/devices/system/cpu/smt/control && \
echo "Hyperthreading: Not Supported (OK)") || \
echo "Hyperthreading: ENABLED (Recommendation: DISABLED)"
Note There are some newer aarch64 CPUs that also implement SMT, however AWS Graviton processors do not implement it.
KPTI is used to prevent certain side-channel issues that allow access to protected kernel memory pages that are normally inaccessible to guests. Some variants of Meltdown can be mitigated by enabling this feature.
Verification can be done by running:
(grep -q "^Mitigation: PTI$" /sys/devices/system/cpu/vulnerabilities/meltdown \
&& echo "KPTI: SUPPORTED (OK)") || \
(grep -q "^Not affected$" /sys/devices/system/cpu/vulnerabilities/meltdown \
&& echo "KPTI: Not Affected (OK)") || \
echo "KPTI: NOT SUPPORTED (Recommendation: SUPPORTED)"
A full list of the ARM processors that are vulnerable to side-channel attacks and the mechanisms of these attacks can be found here. KPTI is implemented for ARM in version 4.16 and later of the Linux kernel.
Note Graviton-enabled hardware is not affected by this.
Disabling KSM mitigates side-channel issues which rely on de-duplication to reveal what memory line was accessed by another process.
KSM can be disabled by executing the following as root:
echo "0" > /sys/kernel/mm/ksm/run
Verification can be done by running:
(grep -q "^0$" /sys/kernel/mm/ksm/run && echo "KSM: DISABLED (OK)") || \
echo "KSM: ENABLED (Recommendation: DISABLED)"
We recommend using a kernel compiled with eIBRS or RETPOLINE
, together with
microcode supporting conditional Indirect Branch Prediction Barriers (IBPB).
Verification can be done by running:
cat /sys/devices/system/cpu/vulnerabilities/spectre_v2
The output should mention the following mitigations being in use:
Enhanced IBRS
orRetpolines
IBPB
at leastconditional
We recommend using a kernel compiled with MITIGATE_SPECTRE_BRANCH_HISTORY
.
More information on the processors vulnerable to this type of attack and detailed information on the mitigations can be found in the ARM security documentation.
Verification can be done by running:
grep -q "^(Mitigation: CSV2, BHB|Not affected)$" \
/sys/devices/system/cpu/vulnerabilities/spectre_v2 && \
echo "SPECTRE V2 -> OK" || echo "SPECTRE V2 -> NOT OK"
Verification for mitigation against Spectre V1 can be done:
grep -q "^(Mitigation:|Not affected)$" \
/sys/devices/system/cpu/vulnerabilities/spectre_v1 && \
echo "SPECTRE V1 -> OK" || echo "SPECTRE V1 -> NOT OK"
These features provide mitigation for Foreshadow/L1TF side-channel issue on affected hardware. They can be enabled by adding the following Linux kernel boot parameter:
l1tf=full,force
which will also implicitly disable SMT. This will apply the mitigation when execution context switches into microVMs. Verification can be done by running:
declare -a CONDITIONS=("Mitigation: PTE Inversion" "VMX: cache flushes")
for cond in "${CONDITIONS[@]}"; \
do (grep -q "$cond" /sys/devices/system/cpu/vulnerabilities/l1tf && \
echo "$cond: ENABLED (OK)") || \
echo "$cond: DISABLED (Recommendation: ENABLED)"; done
See more details here.
This will mitigate variants of Spectre side-channel issues such as Speculative Store Bypass (Spectre v4) and SpectreNG.
We recommend applying SSBD to Firecracker and the host kernel.
On x86_64 systems, this can be done using the following kernel cmdline parameter:
spec_store_bypass_disable=on
Unfortunately, this applies SSBD to all the other processes running on the host as well.
On aarch64 systems, SSBD can be applied to the kernel by using the following kernel cmdline parameter:
ssbd=kernel
SSBD is applied to Firecracker by using the prctl
interface.
However, this is only available on host kernels Linux >=4.17 and also Amazon
Linux 4.14. Alternatively, a global mitigation can be enabled by adding the
following Linux kernel cmdline parameter:
ssbd=force-on
The following command can be used to check if SSBD is applied to Firecracker:
cat /proc/$(pgrep firecracker | head -n1)/status | grep Speculation_Store_Bypass
Output shows one of the following:
- vulnerable
- not vulnerable
- thread mitigated
- thread force mitigated
- globally mitigated
For any process running on the host that communicates with Firecracker and handles sensitive data, we recommend hardening it against spectre-like attacks by:
- compiling it with speculative load hardening
- compiling it with retpolines
- applying SSBD to it
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.
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)"
General recommendation: Keep the host and the guest kernels up to date.
In a Linux KVM guest that has PV TLB enabled, a process in the guest kernel may be able to read memory locations from another process in the same guest.
Under certain conditions the TLB will contain invalid entries. A malicious attacker running on the guest can get access to the memory of other running process on that guest.
The vulnerability affects systems where all the following conditions are present:
- the host kernel >= 4.10.
- the guest kernel >= 4.16.
- the
KVM_FEATURE_PV_TLB_FLUSH
is set in the CPUID of the guest. This is theEAX
bit 9 in theKVM_CPUID_FEATURES (0x40000001)
entry.
This can be checked by running
cpuid -r
and by searching for the entry corresponding to the leaf 0x40000001
.
Example output:
0x40000001 0x00: eax=0x200 ebx=0x00000000 ecx=0x00000000 edx=0x00000000
EAX 010004fb = 0010 0000 0000
EAX Bit 9: KVM_FEATURE_PV_TLB_FLUSH = 1
The vulnerability is fixed by the following host kernel patches.
The fix was integrated in the mainline kernel and in 4.19.103, 5.4.19, 5.5.3 stable kernel releases. Please follow kernel.org and once the fix is available in your stable release please update the host kernel. If you are not using a vanilla kernel, please check with Linux distro provider.
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.