Hi Joao,

On Fri, Nov 13, 2026 at 02:36:41PM -0300, Joao Morais wrote:
> Hi team,
> 
> Some of our products pin HAProxy to a specific version and backport
> CVE fixes instead of moving to the latest release of that HAProxy
> branch. To pick the right commits, we rely on the references in the
> CVE records.

Well, let me tell you that by doing this you ship the ABSOLUTE WORST
version of haproxy one can get. And this is not specific to haproxy.
That's a bad practice that some distros adopted two decades ago for
political reasons ("we can't be blamed for not picking something we
weren't told to and we won't be blamed for picking something that
was not necessary") and that's causing a lot of harm to their users.

The only thing you are certain about by doing this, is that your
versions are affected by exactly *all* of the known bugs that affect
the base version, which are listed on the bugs page. For example,
let's say that you start from 3.2.0, here's what affects you:

   https://www.haproxy.org/bugs/bugs-3.2.0.html

824 bugs, among which 1 critical, 32 major, 309 medium and 482 minor.

CVE are about vulnerability and exposure, but when a component sits at
the edge of the infrastructure, who knows how it is being used, and how
a seamingly innocuous bug can have stability impacts for certain users ?
So CVEs are filed on a reasonable rate for really exploitable stuff to
avoid drowning the distros who force their maintainers to do that, but
that doesn't mean that this is a right practice. In the case of compoents
like haproxy, there are far more serious bugs than those in CVE.

7 years ago I wanted to switch to a model by which we would automatically
file CVEs for each and every backported patch so that there wouldn't be
any doubt about what to pick. I was implored by certain distros not to do
that because that would make their task horrible, which I can easily
understand when they have to justify everything. Nowadays, the kernel
does that and distros had no solution but to adapt. Maybe we should
reconsider this option again so that this irresponsible practice stops
once for all.

> Two recent records seem to reference a commit other than the fix:
> 
> https://www.cve.org/CVERecord?id=CVE-2026-110353 (truncated H2 PING
> frame) references cec63e95f62a ("BUG/MEDIUM: mux-h1: only accept the
> H2 preface on the first request"). The fix matching the description
> seems to be eeafce8ea958 ("BUG/MAJOR: mux-h2: properly handle short
> reads on PING frames").
> 
> https://www.cve.org/CVERecord?id=CVE-2026-110354 (H1 chunked request
> loop) references 077906da140a ("MAJOR: mux-h1: Drain requests on
> client side before shut a stream down"), from February 2024. The fix
> matching the description seems to be 2e4ba6a57080 ("BUG/MAJOR: h1-htx:
> report an error on a too long chunk-size line"). Since 077906da140a
> added the draining the description mentions, maybe it references
> instead the commit that introduced the bug.

Sadly that's the result of dealing with embargoes, at some point you need
to have commit IDs to track your backports and these ones may change when
finally merged because in between, some more important fixes had to be
merged anyway. But yes, the descriptions look good.

> Could you confirm that eeafce8ea958 and 2e4ba6a57080 are the complete
> fixes for these two CVEs, with no other related commits we should
> backport?

Nobody can know. A fix is a change between one known state, and another one.
You're running in a state nobody else has ever tested the patch with. Who
knows if these patches require some of the other previous fixes to work
well ? All we know is that they work well together. Isolated, I don't know,
you don't know, nobody knows. Maybe they fix the visible symptom but when
conditions change a little bit the problem is still there. Maybe they make
the problem even worse.

The right way to think about a fix is to consider that it has for
dependencies all the patches that precede it, and that for stable
releases, we try hard to reduce this amount of dependencies (which is
why sometimes we adapt them or even backport some other harmless ones).
So if you pick an isolated fix, you're just missing the dependencies.

> If the references are wrong, could they be corrected in the
> records, or should we report that to the CNA ourselves?

I really don't know. I remember trying to contact MITRE in the past to
get some invalid reports fixed, and never got a response. Now it's got
worse, they are completely overloaded issuing tens of thousands of CVEs
a year to enhance the CV of some would-be AI-assisted researchers who
dream of having "their" CVE, and while doing this they're just no longer
able to catch up with specific demands. I even gave up asking for CVEs
because you have to attempt many times to hope for one response and I
cannot think about asking every single day for the same until one
arrives. I'd suggest that you be gentle with them and not add to the
burden they're dealing with, nobody cares about which patch it is, as
long as you take all the fixes we make the effort of carefully selecting
and test for you, you're safe. The rest is just garage forks nobody wants
to know about.

Hoping this helps,
Willy


Reply via email to