FSCL - FPU Side-Channel Leakage¶
FPU Side-Channel Leakage (FSCL) is a hardware vulnerability on LoongArch processors. Operations involving FPU/vector registers may cause data leakage through the high bits of vector registers.
Affected processors¶
The following processors are not affected:
Processors that do not support LSX and LASX (CPUCFG2.LSX=0, CPUCFG2.LASX=0)
Processors with hardware fixes: CPUCFG7.FP_CLEAN=1 or CPUCFG7.FP_NO_L1DLEAK=1
Whether the current processor is affected can be queried via sysfs, see FSCL system information. Not all processors are affected by all FSCL variants, but the mitigations for all variants are identical, so the kernel treats them as a single vulnerability.
Problem¶
The FPU/LSX/LASX instructions on LoongArch processors share a unified encoding. Different instructions operate on registers of different bit widths, which may cause data leakage through the high bits of vector registers. The source of leaked data may be the L1 data cache or residual register state from other processes. The leaked data is random in nature and does not cause program execution errors or state changes.
Since buffers may be shared between hyper-threads, cross-hyper-thread attacks are possible.
Attack scenarios¶
FSCL attacks can be launched by unprivileged user-space applications. The attack requires the attacker and victim to be scheduled on the same physical core. The FSCL vulnerability does not allow the attacker to control the memory target address. Therefore, the attack is purely sample-based. Malicious operations can forward this speculative data to leakage components, from which the value can be inferred through cache side-channel attacks.
FSCL system information¶
The Linux kernel reports the current FSCL status via the sysfs interface. The relevant file is:
/sys/devices/system/cpu/vulnerabilities/fscl
The possible values for this file are:
‘Not affected’
The processor is not affected
‘Vulnerable’
The processor is vulnerable, but no mitigation enabled
‘Mitigation: FPU Side Channel Leakage bug was eliminated by flushing context’
AUTO mode: mitigates the issue by flushing the L1 cache and floating-point registers when the kernel returns to user space
‘Mitigation: FPU Side Channel Leakage bug was eliminated by disable SMT and flushing context’
FLUSH_NOSMT mode: mitigates the issue by flushing the L1 cache and floating-point registers when the kernel returns to user space, and disabling Simultaneous Multi-Threading
Mitigation mechanism¶
At boot time, the kernel detects affected CPUs via the CPUCFG7 register. If the CPU is vulnerable and mitigations are not globally disabled, the kernel selects a mitigation strategy based on hardware capabilities and kernel command-line parameters.
Residual data flushing¶
During kernel-to-user and user-to-kernel transitions, the kernel conditionally flushes the L1D cache and/or vector registers based on context:
If exiting from FPU/LSX/LASX exception handling: flush L1D and vector registers.
If exiting from a non-FPU exception: disable FPU/LSX/LASX to prevent future leakage.
Virtualization mitigation¶
Before entering a guest, perform residual data cleanup operations. This only covers the host-to-guest transition, i.e., preventing data leakage from the host to the guest, but does not protect within the guest. The guest must have its own protection measures.
Three mitigation modes:
FSCL_MITIGATION_OFF¶
No mitigation. The system is vulnerable to cross-thread FPU side-channel attacks.
FSCL_MITIGATION_AUTO (default)¶
During kernel-to-user and user-to-kernel transitions, the kernel conditionally flushes the L1D cache and/or vector registers based on context:
If exiting from FPU/LSX/LASX exception handling: flush L1D and vector registers.
If exiting from a non-FPU exception: disable FPU/LSX/LASX to prevent future leakage.
FSCL_MITIGATION_FLUSH_NOSMT (disable SMT + force flush)¶
Based on FSCL_MITIGATION_AUTO, with SMT disabled:
This mode provides full security guarantees at the cost of potential performance loss from disabling SMT.
SMT control¶
All FSCL variants can be exploited via cross-hyper-thread attacks. This means that on affected CPUs, SMT must be disabled to achieve full protection.
Disabling SMT may have a significant impact on performance, though the degree depends on the workload type.
Mitigation control on the kernel command line¶
The kernel command line allows controlling FSCL mitigations at boot time via the following options:
fscl=offDisable the mitigation entirely. The system is vulnerable.
fscl=autoEnable AUTO mode (default). SMT-aware dynamic mitigation.
fscl=auto,nosmtForce FLUSH_NOSMT mode. Disables SMT and performs unconditional flushing on every kernel-to-user transition.
The global mitigations= parameter also affects FSCL:
mitigations=offEquivalent to
fscl=off.mitigations=autoEquivalent to
fscl=auto.mitigations=auto,nosmtEquivalent to
fscl=auto,nosmt.
Mitigation selection guide¶
1. Trusted userspace¶
If all user-space applications come from trusted sources and do not execute externally-provided untrusted code, the mitigation can be disabled.
2. Virtualization with trusted guests¶
The same considerations apply as for trusted userspace above.
3. Virtualization with untrusted guests¶
If FSCL mitigation is enabled and SMT is disabled, guest-to-host and guest-to-guest attacks can be prevented.
Default mitigations¶
The default mitigation for vulnerable processors is:
Enable CPU buffer flushing
The kernel does not force-disable SMT by default, which means SMT systems remain vulnerable when running untrusted code.