GDS - Gather Data Sampling

Gather Data Sampling is a hardware vulnerability which allows unprivileged speculative access to data which was previously stored in vector registers.


When a gather instruction performs loads from memory, different data elements are merged into the destination vector register. However, when a gather instruction that is transiently executed encounters a fault, stale data from architectural or internal vector registers may get transiently forwarded to the destination vector register instead. This will allow a malicious attacker to infer stale data using typical side channel techniques like cache timing attacks. GDS is a purely sampling-based attack.

The attacker uses gather instructions to infer the stale vector register data. The victim does not need to do anything special other than use the vector registers. The victim does not need to use gather instructions to be vulnerable.

Because the buffers are shared between Hyper-Threads cross Hyper-Thread attacks are possible.

Attack scenarios

Without mitigation, GDS can infer stale data across virtually all permission boundaries:

Non-enclaves can infer SGX enclave data Userspace can infer kernel data Guests can infer data from hosts Guest can infer guest from other guests Users can infer data from other users

Because of this, it is important to ensure that the mitigation stays enabled in lower-privilege contexts like guests and when running outside SGX enclaves.

The hardware enforces the mitigation for SGX. Likewise, VMMs should ensure that guests are not allowed to disable the GDS mitigation. If a host erred and allowed this, a guest could theoretically disable GDS mitigation, mount an attack, and re-enable it.

Mitigation mechanism

This issue is mitigated in microcode. The microcode defines the following new bits:



Enumerates GDS vulnerability and mitigation support.



Processor is not vulnerable.



Disables the mitigation 0 by default.



Locks GDS_MITG_DIS=0. Writes to GDS_MITG_DIS are ignored Can't be cleared once set.

GDS can also be mitigated on systems that don't have updated microcode by disabling AVX. This can be done by setting gather_data_sampling="force" or "clearcpuid=avx" on the kernel command-line.

If used, these options will disable AVX use by turning off XSAVE YMM support. However, the processor will still enumerate AVX support. Userspace that does not follow proper AVX enumeration to check both AVX and XSAVE YMM support will break.

Mitigation control on the kernel command line

The mitigation can be disabled by setting "gather_data_sampling=off" or "mitigations=off" on the kernel command line. Not specifying either will default to the mitigation being enabled. Specifying "gather_data_sampling=force" will use the microcode mitigation when available or disable AVX on affected systems where the microcode hasn't been updated to include the mitigation.

GDS System Information

The kernel provides vulnerability status information through sysfs. For GDS this can be accessed by the following sysfs file:


The possible values contained in this file are:

Not affected

Processor not vulnerable.


Processor vulnerable and mitigation disabled.

Vulnerable: No microcode

Processor vulnerable and microcode is missing mitigation.

Mitigation: AVX disabled,

no microcode

Processor is vulnerable and microcode is missing mitigation. AVX disabled as mitigation.

Mitigation: Microcode

Processor is vulnerable and mitigation is in effect.

Mitigation: Microcode (locked)

Processor is vulnerable and mitigation is in effect and cannot be disabled.

Unknown: Dependent on

hypervisor status

Running on a virtual guest processor that is affected but with no way to know if host processor is mitigated or vulnerable.

GDS Default mitigation

The updated microcode will enable the mitigation by default. The kernel's default action is to leave the mitigation enabled.