AlmaLinux OS 10.3 and 9.9 nu tillgänglig!  |  Hitta oss på SC26 November 16 till 19:e, stånd #1704  |  Berätta för oss vad AlmaLinux bör göra härnäst
Alla blogg inlägg

NoisyTCP (CVE-2026-99288): Patched kernels available in testing

profile
Andrew Lukoshko

AlmaLinux Lead Architect

NoisyTCP (CVE-2026-99288): Patched. Call for testing

A new Linux kernel vulnerability named NoisyTCP, tracked as CVE-2026-99288, is a race condition in the TCP code that lets a remote attacker crash the kernel by sending TCP packets to a listening port. No account and no authentication are needed. Red Hat rates it Important. AlmaLinux 8, 9, and 10 are affected. Patched kernels are available in the testing repository today, and we would like your help verifying them before we release them to production.

More information about the vulnerability:

Summary

When a TCP packet arrives, the kernel has to find the socket it belongs to. For speed, that lookup takes no locks: it finds a matching entry in a hash table, takes a reference to it, and then checks again that the entry still matches the packet.

The memory behind these entries can be reused for a new socket while a lookup on another CPU is still holding the old address. The second check was meant to catch that, but it only compared addresses and ports. It did not check whether the object was actually in the hash table. If the memory was reused for a new connection request with the same addresses and ports, the lookup returned an object that was never meant to be found that way.

The researcher showed this with SYN cookies. A connection request rebuilt from a SYN cookie is not in the table and has no listening socket attached. When a stale lookup returns it, the TCP receive path reads through that missing listener and the kernel panics.

The bug was introduced in Linux 4.4 in 2015 and went unnoticed for about 11 years. The fix makes the lookup reject entries that are not in the hash table and try again, and makes the affected error paths release connection requests safely.

Impact

This is a remote denial of service. An attacker who can send TCP packets to a listening port can crash the whole machine. Ordinary services with default settings, such as a web server or SSH, are enough. On a container host, the crash takes down the host and every container on it.

The attack is a race, so it has conditions and takes time. It needs the server to process incoming packets on more than one CPU, which means a network card with two or more receive queues, or Receive Packet Steering (RPS) enabled. That describes most servers and cloud instances. The researcher reports crashing virtual machines in between five seconds and two minutes, while a quiet bare-metal server took about 14 hours.

The researcher also describes a second variant that leads to a use-after-free rather than a crash, and calls it generally hard to trigger. Red Hat’s assessment is that the flaw disrupts availability and does not allow data alteration or code execution. We are not aware of attacks in the wild at the time of writing.

Patching ahead of our upstream

Security is a top priority at AlmaLinux. At the time of writing, no fixed RHEL kernel is available, and Red Hat lists the RHEL 8, 9, and 10 kernels as affected. Our core team has backported the upstream fix to every supported AlmaLinux branch. The decision to ship ahead of a CentOS Stream / RHEL update was made by our technical steering committee, ALESCo.

For AlmaLinux 9 and 10 the upstream fix applies almost as is. The AlmaLinux 8 kernel (4.18) is older: it does not have the request migration feature that part of the fix deals with, and its lookup code is written differently. We adapted the fix to that kernel, so testing on AlmaLinux 8 is especially welcome.

Patched kernel versions

  • AlmaLinux 8: kernel-4.18.0-553.186.2.el8_10
  • AlmaLinux 9: kernel-5.14.0-741.22.2.el9_9
  • AlmaLinux 10: kernel-6.12.0-264.16.2.el10_3

Installing from the testing repository

sudo dnf install -y almalinux-release-testing
sudo dnf update 'kernel*' --enablerepo=almalinux-testing
sudo reboot

Confirm with uname -r or rpm -q kernel against the versions above.

If you do not want the testing repo enabled afterwards (and on a production box you almost certainly do not), disable it once you have rebooted:

sudo dnf config-manager --disable almalinux-testing

If you run into anything, please let us know in AlmaLinux chat or on bugs.almalinux.org.

Is there a workaround?

There are two temporary mitigations, but neither is a substitute for the fix. Each one weakens or slows down the server in its own way, and each has to be applied and kept in place on every affected system and container. Treat them as a stopgap for machines you cannot reboot right away, and plan to install the patched kernel as soon as you can.

Disable SYN cookies. This closes the path used to crash the kernel:

sudo sysctl -w net.ipv4.tcp_syncookies=0

The setting is per network namespace, so it has to be applied in every container that has its own network namespace as well as on the host. It also removes the kernel’s protection against SYN floods, so on an internet-facing server you trade one way to be knocked offline for another. It covers the crash described above; the researcher’s second, harder to trigger variant does not go through SYN cookies.

Process incoming packets on a single CPU. The race needs packets to be handled on different CPUs at the same time. Reducing the network interface to one receive queue and disabling RPS and RFS removes that, at the price of network throughput, which a busy server may not be able to afford.

For each interface that receives traffic from outside (replace eth0 with its name), check how many queues it has and set it to one:

ethtool -l eth0
sudo ethtool -L eth0 combined 1

Some drivers have separate receive and transmit queues instead of combined ones; for those use sudo ethtool -L eth0 rx 1. Then make sure RPS and RFS are off, which is the default:

echo 0 | sudo tee /sys/class/net/eth0/queues/rx-*/rps_cpus
echo 0 | sudo tee /sys/class/net/eth0/queues/rx-*/rps_flow_cnt
sudo sysctl -w net.core.rps_sock_flow_entries=0

These settings are lost on reboot and may be reset when the interface is reconfigured, and not every network card or virtual NIC allows changing its queue count.

Only the patched kernel removes the bug itself. If you can update and reboot, do that instead.

Thanks

Thanks to Hyunwoo Kim for finding and reporting this vulnerability and for writing the upstream fix, to Eric Dumazet and Kuniyuki Iwashima for reviewing it, and to the Linux kernel CVE team.

Thanks to the AlmaLinux core team for turning around patched builds for every supported release, and to ALESCo for moving quickly to approve shipping ahead of upstream.

Stay informed

Remaining aware of these vulnerabilities and acting quickly can keep your systems and data safe. Follow the AlmaLinux Blog, join the Mattermost Community Chat, and subscribe to the Announce and Security mailing lists to stay informed.

Changelog

  • 2026-11-16: Initial post published. Patched kernels available in almalinux-testing for AlmaLinux 8, 9, and 10.

Håll dig uppdaterad!