Zen 5 Gets Its TSA Patch: Check Your Linux Box
AMD says Zen 5 is hit by Transient Scheduler Attacks. Here is what Linux 7.3.4 changed, how VERW_CLEAR works, and how to check your servers and VMs today.
BY ASHLIN DARIUS GOVINDASAMY
Plot twist for anyone who bought a shiny Zen 5 server because it was 'not affected': on 10 November 2026 AMD published bulletin AMD-SB-7073 saying Zen 5 processors are affected by Transient Scheduler Attacks after all. If you run AMD Zen 5 hardware on Linux, here is what changed in the kernel, how the fix detects itself, and a five-minute check for your fleet.
What are Transient Scheduler Attacks?
Transient Scheduler Attacks (TSA) are a family of speculative side channels AMD disclosed in 2025 for earlier Zen generations. The idea is a cousin of the Spectre and MDS bugs: under certain conditions the CPU's scheduler speculatively lets an instruction proceed with data it should not have, and timing differences can leak it. The variants AMD names are TSA-SQ (store queue) and TSA-L1 (L1 data cache).
AMD's new Zen 5 bulletin lists the same two CVEs, CVE-2024-36350 for TSA-SQ and CVE-2024-36357 for TSA-L1, each with a CVSS 3.1 base score of 5.6. A medium score, so no need to panic, but note that fully mitigating it takes two things: platform firmware from your vendor and an operating system update.
What changed in the Linux kernel
The Zen 5 patch, titled x86/bugs: Mitigate TSA on AMD Zen5, was written by Borislav Petkov of AMD. It landed in the 7.4 release candidates (rc3, tagged 15 November) and was also released in stable 7.3.4 on 10 November. Per reports on the commit, Zen 5 is affected by a third variant, TSA_FILL, as well.
The clever part is detection. The mitigation reuses the VERW instruction in its memory form to clear CPU buffers on privilege transitions. AMD exposes a CPUID bit, VERW_CLEAR, at leaf 0x80000021, register EAX, bit 5. When updated microcode sets the bit, the kernel knows the mitigation is usable. So no giant table of microcode revisions needs to live in the kernel.
Why the scheduler leaks at all
Modern cores keep many loads and stores in flight. The scheduler decides when each can issue, and it sometimes guesses that a load has everything it needs before the hardware has finished checking. When the guess is wrong the result is thrown away, but the load may briefly have seen data from the store queue or the L1 cache that belonged to another privilege level. An attacker cannot read that value directly. They measure timing, over thousands of tries, to reconstruct it. This is why the fix is not a clever patch to one function: the kernel has to flush the relevant buffers whenever execution crosses a trust boundary.
That also explains the cost model. The flush happens on transitions, system calls and VM exits, not on every instruction. A compute-bound number cruncher that rarely enters the kernel barely notices. A database, proxy or hypervisor that lives on syscalls notices more. The honest way to know is to run your own load test with and without the mitigation on a spare machine, not to quote someone else's percentage.
Check your machines
First, see what your kernel thinks. The TSA entry sits beside the other CPU vulnerability files, and I'd rather you read the real output than trust any blog post, including this one. Expect a line starting with Mitigation or Vulnerable, plus a reason.
uname -r
grep -m1 'model name' /proc/cpuinfo
grep -H . /sys/devices/system/cpu/vulnerabilities/tsa
echo '--- everything else ---'
grep -H . /sys/devices/system/cpu/vulnerabilities/* | sed 's#/sys/devices/system/cpu/vulnerabilities/##'Second, check whether the CPUID bit is actually set, which tells you whether your firmware or microcode delivers the VERW_CLEAR capability. The cpuid tool decodes leaf 0x80000021, and you want bit 5 of EAX.
sudo apt install -y cpuid # or: dnf install cpuid
cpuid -1 -l 0x80000021 -r | awk '{ for (i=1;i<=NF;i++) if ($i ~ /^eax=/) { v=substr($i,5); print "eax=" v } }'
# Alternative without the cpuid tool: read the cpuid device on core 0
sudo modprobe cpuid
sudo python3 - <<'PY'
import os, struct
fd = os.open('/dev/cpu/0/cpuid', os.O_RDONLY)
# the cpuid device encodes leaf as file offset
data = os.pread(fd, 16, 0x80000021)
eax, ebx, ecx, edx = struct.unpack('<4I', data)
print(hex(eax), 'VERW_CLEAR =', bool(eax >> 5 & 1))
PYRun the check on one host per hardware model, not just one per datacentre, because BIOS versions differ between SKUs and even between racks bought a few months apart. If bit 5 is zero on a Zen 5 host, you are waiting on a BIOS or microcode update from your vendor. Updating the kernel alone is not enough, and AMD cautions the bit is not set merely because a vulnerable system received a patch, so verify rather than assume.
QUICK CHECK
On a Zen 5 server running a patched kernel, /sys/.../vulnerabilities/tsa still says Vulnerable. What is the most likely missing piece?
Tuning: the tsa= boot parameter
Since the original TSA work in 2025 the kernel has had a tsa= boot parameter, reported as controlling whether the mitigation is always on, applied only at user/kernel transitions, applied only at guest/host transitions, or disabled. A Kconfig option, MITIGATION_TSA, gates the code. Check the exact accepted values in your kernel's admin-guide kernel-parameters.txt before you set it.
# 1. Read what your kernel documents
zgrep -i 'tsa' /usr/share/doc/linux-doc*/kernel-parameters.txt* 2>/dev/null | head
# 2. Add the parameter you decided on, then regenerate GRUB
sudo sed -i 's/^GRUB_CMDLINE_LINUX="/GRUB_CMDLINE_LINUX="tsa=on /' /etc/default/grub
sudo update-grub # or: sudo grub2-mkconfig -o /boot/grub2/grub.cfg
sudo reboot
# 3. After reboot
cat /proc/cmdline
cat /sys/devices/system/cpu/vulnerabilities/tsa›Why not just turn it off for performance?
Buffer clearing with VERW happens on privilege transitions, so syscall-heavy and virtualisation-heavy workloads pay most. One Windows user reported about 9 percent lower multi-core performance after an update on a Ryzen 7 9850X3D, but that report is unconfirmed and not about Linux, so do not treat it as a benchmark. Measure your own workload with perf stat or your usual load test before and after, and only relax the mitigation on single-tenant machines where you trust all code on the box.
Virtual machines: the part people forget
A guest can only use VERW_CLEAR if the hypervisor passes the CPUID bit through. QEMU has recent patches adding a verw-clear feature flag for exactly this. If you run KVM guests, update QEMU and libvirt, then expose the CPU model with the flag. Otherwise a perfectly patched guest kernel inside a perfectly patched host will still report that it cannot mitigate.
grep -H . /sys/devices/system/cpu/vulnerabilities/tsa
grep -o 'verw_clear' /proc/cpuinfo | sort -u || echo 'flag not exposed to this guest'Takeaway
- AMD-SB-7073 (10 November 2026) says Zen 5 is affected by TSA-SQ and TSA-L1, CVSS 5.6 each.
- Linux has the fix in stable 7.3.4 and the 7.4 release candidates; the new kernel alone is not enough.
- You also need vendor firmware or microcode that sets VERW_CLEAR (CPUID 0x80000021 EAX bit 5).
- Check /sys/devices/system/cpu/vulnerabilities/tsa on every host, then every guest.
- Prioritise multi-tenant machines; measure before relaxing anything.
SOURCES
- AMD Transient Scheduler Attacks for Zen 5 Processors (AMD-SB-7073) — AMD
- Technical guidance for mitigating transient scheduler attacks — AMD
- Linux Patched For Transient Scheduler Attacks TSA Impacting AMD CPUs — Phoronix
- Linux 7.4-rc3 Released — Phoronix
- Xen Security Advisory 471 (CVE-2024-36350, CVE-2024-36357) — oss-security
- target/i386: Add TSA feature flag verw-clear — qemu-devel
Frequently asked questions
Is AMD Zen 5 affected by Transient Scheduler Attacks?
Yes. AMD bulletin AMD-SB-7073 of 10 November 2026 lists Zen 5 as affected by TSA-SQ (CVE-2024-36350) and TSA-L1 (CVE-2024-36357), both CVSS 5.6.
Does a kernel update alone fix TSA on Zen 5?
No. AMD says full mitigation needs platform firmware from your system vendor plus an operating system update. The kernel relies on microcode setting the VERW_CLEAR CPUID bit.
How do I check TSA status on Linux?
Read /sys/devices/system/cpu/vulnerabilities/tsa. A line starting with Mitigation means the kernel is protecting you; Vulnerable means you still need firmware, microcode or a kernel update.
Which kernels contain the Zen 5 TSA fix?
Stable 7.3.4, released on 10 November 2026, and the 7.4 release candidates from rc3 on. Check your distribution for backports to older kernels.
NEWSLETTER
Get the engineering notes in your inbox.
Kernels, GPUs, networks, AI infrastructure — and the things we broke along the way. No spam, one click to leave.
We store only your email address, to send you new articles. See our privacy policy.
TAGS
RELATED
Linux 7.4-rc1: Test a Record-Size Kernel Safely
Linux 7.4-rc1 closed the biggest merge window ever. Here is how to measure it yourself, boot an rc kernel safely, and bisect a regression to one commit.
Linux CPU Isolation: isolcpus, nohz_full, rcu_nocbs
Stop kernel noise from hitting your latency-critical Linux service. Learn isolcpus, nohz_full, rcu_nocbs, IRQ affinity and cpuset partitions with real commands.