Proxmox dated the proxmox-kernel-7.0 (7.0.14-30) entry 16 November, 22:55. Its first bullet is the one everybody quotes: it brings "fixes for about 130 published security issues". The one that matters here is the ninth, at the end of the VXLAN paragraph: "Since 7.0.14-25, traffic to such MAC addresses was flooded to all remote VTEPs". 7.0.14-25 shipped on 17 October. That is thirty days. The same document that closes 130 holes is telling you, in passing, that you spent a month flooding traffic to every remote VTEP you have. Nobody warned you, because that line was the warning.
We run Proxmox VE clusters in production and we read these bulletins because we are the ones who sign off on them. November's was worth more than an internal note, so here is the whole thing: what it brings, what broke along the way, and the order in which we put a new kernel into a cluster that cannot stop.
What 7.0.14-30 brings (and why you want it)
Sources move to Ubuntu-7.0.0-44.44 plus the stable changes from v7.3.3 and v6.18.62, and that is where the roughly 130 published issues come from. In the same entry, three paragraphs change the conversation when what sits underneath is a hypervisor rather than a laptop:
- 1On AMD CPUs with broadcast TLB invalidation, "two processes possibly getting the same global TLB address space ID, letting them read or corrupt each other's memory". On a virtualisation host, "two processes" are almost never two processes: they are two customers.
- 2A "host crash that guests could cause through the UDP tunnel segmentation offload of virtio-net", which QEMU offers on machine versions 10.2 and newer. One virtual machine takes down the node, and the node has other virtual machines on it.
- 3
pidfdnow restricts namespace ioctls the wayprocfsdoes, "so a container can no longer get namespace file descriptors of processes outside of it". If you lean on unprivileged containers as a boundary, that line is yours.
There is a fourth item that is not a fix but a decision somebody has to make: the Transient Scheduler Attack mitigation on AMD Zen 5. With microcode that supports it, CPU buffers are cleared when returning to user space and when entering a guest, and the changelog says it plainly: "which costs some performance". There is a tsa= parameter to limit or disable it. So the patch does not merely fix things: it puts a lever with a price tag in front of you. If nobody looks at it, the default decides, and the default is the right option even though nobody chose it.
So there is no doubt where this post is heading: you install this. What follows is about timing, ordering and knowing what you are moving.
And in the same stretch, the eight it introduced itself
The count is checkable, so here is its exact definition: across the five bulletins published between 22 October and 16 November —7.0.14-26 through 7.0.14-30— there are eight fixes carrying a "since" with a version number. These are not inherited Linux bugs. They are things that worked in this very 7.0 series and stopped working because of an earlier kernel in it. Ordered by exposure time:
Read that left-hand column with an operator's eyes. None of those eight faults shows up as an outage. The NFS one is "it's slow". The amd-pstate one is "this box has been performing oddly for a few weeks". The btrfs one is syslog noise somebody will eventually mute. And the VXLAN one is invisible until you look at traffic counters on the inter-site link and the numbers refuse to add up. Faults that arrive via a patch almost never look like an incident: they look like things being slightly worse.
The ZFS case: a fix, a revert, and a revert that was not one
This one deserves its own section because it explains better than anything why "am I on the latest patch?" is a poor question. On 22 October, 7.0.14-26 fixes bug #7653: memory-mapped files served stale data after a rollback or a receive into their mounted dataset, "which could affect containers after storage replication or migration". It sounds like an important fix, because it is.
On 31 October, 7.0.14-27 reverts it: the fix "could deadlock a receive or rollback into a mounted dataset while programs accessed its files through memory mappings, hanging all access to the dataset and blocking reboots". A data-correctness fix that, in the right combination, leaves you unable to reboot the node. So far the system is working, and working well: somebody caught the side effect in nine days and backed it out.
And on 2 November, 7.0.14-28 opens with a sentence worth reading slowly: "ZFS: actually revert the fix… The 7.0.14-27 package still contained it, as its build reused outdated ZFS sources". For two days, the changelog said the thing was reverted and the package still shipped it. And from the fix going in to it actually coming out: eleven days.
This is not bad faith or sloppiness: it is how fast-moving software behaves, and moving fast is exactly what you want from the people sending you 130 security fixes in one evening. But it has a consequence that is yours. If your only health signal is "I applied the latest patch", that signal cannot tell these three states apart — fix in, fix out, and fix the changelog declared out while it was still in. All three look identical from apt.
Why you did not notice: the kernel does not land until you reboot
An installed kernel is not a running kernel. apt leaves the package on disk; what is executing is still the old one until the node boots again. Which is why the useful question goes to the machine, not the package manager: uname -r tells you what is running, pveversion -v tells you what is installed. When they disagree, the first one wins. And it is worth remembering that this kernel is Proxmox's, not Debian's: we already covered what that means when a named bug lands and Debian's advisory is not yours.
Two consequences follow, and they are almost never stated together. First: you can be running a six-week-old kernel on a system marked "up to date", and the eight faults in that table apply to you or not depending on which one you are executing, not on what apt list --upgradable says.
The second one changes how we plan. When you finally reboot, you do not move one kernel: you move the whole stretch. Between 17 October and 16 November, five shipped —26, 27, 28, 29 and 30—. If something starts behaving differently that night, you are left with five changelogs of candidates and no cheap way to separate them. Which is why deferring the reboot window is not the prudent option it looks like: the longer you wait, the bigger the jump and the harder it is to attribute whatever happens next. Prudence here means rebooting more often. And if the node reboots on its own, the reboot also takes the evidence with it.
The order we use to land a kernel
There is nothing sophisticated about it, and that is the point. What usually fails here is not the technique: it is that nobody writes anything down before touching.
- 1Know where you are starting from.
uname -ron every node, not on one. In a cluster whose nodes last rebooted in different months, the stretch to read is different for each of them. - 2Read the stretch, not the latest bulletin, and search it for one word:
since. That is where the vendor tells you, version and all, what it broke. Mind how you search: in the November bulletin one of the eight is split across a line break (broken since/7.0.14-17) and a naivegrepwalks straight past it. It is the best-spent minute of the whole window. - 3Measure BEFORE, in the places that stretch touches. If you run EVPN, the forwarding table and the inter-site link counters. If you run AMD, frequency. If you run Ceph,
ceph -sand latency. Four numbers written down beforehand beat any dashboard you look at afterwards, because afterwards you no longer know what it looked like. - 4One node, the least important one, then wait. A cluster is the only honest lab you own: same hardware, same network, same load. Between the first node and the rest we leave days, not minutes — the faults in that table take days to become visible.
- 5Actually drain it before rebooting: migrate the machines, and if you run Ceph,
ceph osd set nooutso the cluster does not start rebalancing hundreds of gigabytes over a node that is back in four minutes. Andceph osd unset nooutwhen it returns, which is the half everybody forgets. - 6Leave the previous one bootable.
proxmox-boot-tool kernel listfor the exact version strings,proxmox-boot-tool kernel pin <version> --next-bootif you want to try one for a single boot, andproxmox-boot-tool refreshafterwards to sync the boot partitions. The way back has to be decided before you reboot, not improvised at three in the morning. - 7Compare the same numbers from step 3. If you do not compare, you have not tested anything: you have waited.
Step 3 is the most skipped and the highest-yielding. We have seen elsewhere that two machines on the same kernel and the same version of everything behave differently because of details that appear in no bulletin. Without the before, the after tells you nothing.
When not to take any of this advice
There is one specific way to misread this post, so let us close it ourselves: none of the above is an argument for delaying a patch. What 7.0.14-30 fixes includes a virtual machine capable of taking down its host and two processes reading each other's memory. Against that, a month of VXLAN flooding is a bad month. The ordering we propose is there so that when something changes, you know what changed it.
Nor should you turn the pin into a policy. Proxmox says it in its own bootloader documentation, and it is the sentence every pinned kernel ought to carry printed on it: "Such a pin should be removed as soon as possible so that all current security patches of the latest kernel are also applied to the system." A pin with an expiry date is a tool; one without is debt with interest. Same with tsa=: you do not switch off a mitigation to claw back performance without measuring it and without knowing who else shares that host.
And if you have a single node, none of this applies to you yet — but not for good reasons: you have nowhere to test and nowhere to move the machines while you reboot. What is missing there is not a patching policy but somewhere for it to fit. That gets fixed before this does.
Fair is fair: this post exists because they write it down
It would be easy to end this with a complaint about Proxmox, and it would be unfair. This post only exists because they document in plain prose, with bug numbers, naming the version that broke each thing — the kind of material most vendors dispose of with "various bug fixes and improvements". The short turnarounds are genuinely good —sixteen days for the ipip tunnel, fourteen for the Windows clients—; the long ones rather less so, and the same table shows them: forty-three for NFS, sixty-seven for SCTP. But the document is there, with its countable days, and that is the rare part. What is missing, on the side of whoever operates the machine, is somebody reading it.
So we close on a question that either gets answered in ten seconds or does not get answered: could you say right now, without logging in, which kernel each node of your cluster is running and what date it is from? If the answer is no, the risk is not the 130 issues in November's bulletin. It is the ones you have been carrying since September.
Sources (verified at origin): every quote, version, date and bug number in this article comes from the pve-kernel changelog, read on 18/11/2026: the entries for 7.0.14-30 (16/11/2026, 22:55), 7.0.14-29 (06/11), 7.0.14-28 (02/11), 7.0.14-27 (31/10), 7.0.14-26 (22/10), 7.0.14-25 (17/10), 7.0.14-20 (24/09) and 7.0.14-17 (10/09). That is the source of "about 130 published security issues", Ubuntu-7.0.0-44.44, the AMD TLB bug, the virtio-net host crash on QEMU machine versions 10.2+, the pidfd restriction, the TSA mitigation on Zen 5 and its tsa=, the eight regressions with their "since", and the whole ZFS #7653 sequence. The syntax of proxmox-boot-tool kernel list / pin / unpin, the --next-boot option, proxmox-boot-tool refresh and the sentence "Such a pin should be removed as soon as possible…" — Proxmox VE wiki, Host Bootloader. That the 7.0 series is the default stable kernel of Proxmox VE 9.2 — Proxmox VE roadmap. The definition of the count of eight (fixes carrying a "since" plus version number across the five bulletins from 22/10 to 16/11), the day arithmetic and the operational reading —the stretch jump, the reboot order and the "when not to"— are ours. Photograph: NOIRLab/NSF/AURA, CC BY 4.0 via Wikimedia Commons, cropped. We are not Proxmox resellers, nor VMware's, and we do not sell anybody's licences.
Who reads your hypervisor's changelog?
At everyWAN we read the whole stretch and reboot nodes in order, with the way back decided in advance. It is part of the 24x7 IT support with which we run our customers' infrastructure and cloud. If you want to know which kernel is running on each of your nodes today, get in touch.
Talk to everyWAN