public inbox for passt-dev@passt.top
 help / color / mirror / code / Atom feed
From: Stefano Brivio <sbrivio@redhat.com>
To: Paul Holzinger <pholzing@redhat.com>
Cc: passt-dev@passt.top
Subject: Re: pasta does not correctly handle bind errors with port ranges
Date: Mon, 12 Feb 2024 15:13:31 +0100	[thread overview]
Message-ID: <20240212151331.277ef9fd@elisabeth> (raw)
In-Reply-To: <e1a5e9d0-9574-4787-8d47-23ed99d35453@redhat.com>

On Mon, 12 Feb 2024 12:45:00 +0100
Paul Holzinger <pholzing@redhat.com> wrote:

> Hi Stefano,
> 
> On 09/02/2024 22:09, Stefano Brivio wrote:
> > Hi Paul,
> >
> > On Fri, 9 Feb 2024 17:57:05 +0100
> > Paul Holzinger <pholzing@redhat.com> wrote:
> >  
> >> Hi all,
> >> I found some issues with the pasta port binding logic, it does not
> >> correctly handle errors when trying to bind a port range.
> >> Let's first bind a port so we can force an error condition it:
> >> $ nc -l -p 8080 &
> >> $ pasta -t 8080  true
> >> Failed to bind any port for '-t 8080', exiting <-- fails as expected
> >> $ pasta -t 8081 -t 8080  true
> >> Failed to bind any port for '-t 8080', exiting <-- here it also fails
> >> correctly
> >> $ pasta -t 8080-8081  true
> >> <-- no error even though pasta could not bind 8080  
> > This is actually intended: it only fails if it can't bind *any* port in
> > a given range, so that users don't have to explicitly exclude ports
> > from ranges in case some are already taken, knowingly or not. That's
> > why the error message says "any port".  
>
> Ok but this results in a very inconsistent behavior. I argue this is not 
> what a normal user expects at all.

This really depends on the user. The main usage behind the current
behaviour is the "all" forwarding mode, as originally designed for
KubeVirt, now available via libvirt in two forms: -t all / -u all, or
giving excluded ranges only (from the man page: "Specifying excluded
ranges only implies that all other ports are forwarded.").

At least this part, specifically, is an established behaviour and we
can't break it.

But that's not a substantial conflict with what you're asking for
Podman: we can keep this behaviour only if the user asks to forward
(almost) "all" the ports (in whatever way). And Podman never asks for
that, and probably never will.

> It is not even documented in the man page.

Right, yes, we'll need to describe it.

> > For two ports it probably makes no sense, but for larger ranges
> > excluding dozens of ports can get quite annoying for the user. And
> > warnings on failed bind() calls could get quite noisy, too.  
>
> At the same time this can result in a lot of unexpected problems for 
> users as it just hides potential problems, assuming 8001-9000 are 
> already in use then -t 8000-9000 would work and then one might 
> reasonably expect that if they connect to any of the given ports they 
> talk to the pasta namespace and not something else, this could be a 
> security concern. I would say in such case yes I want to see a logged 
> error for each individual port.

I see, but again, if just port 8005 is not available, you might say
that having pasta succeed is a security concern because somebody can
"steal" the port, or that having pasta fail is a nuisance because the
user probably just wanted to forward the remaining 1000 ports, and
graceful degradation would be desirable instead.

It's pretty impossible to draw a line in a different way that's not
considered inconsistent or inconvenient by some. So:

> > If it's a problem for Podman, I can think of two solutions. One would
> > be an option such as --strict-bind or suchlike (better names warmly
> > welcome).
>
> It is a big problem for podman as it does not do what users want us to 
> do. Also, because podman is smart enough to combine several ports into 
> ranges internally for performance reasons, something like -p 80:80 -p 
> 81:81 is always a range for podman thus there is no way for podman users 
> to avoid hitting this problem. An opt-in option works for me but then we 
> need to bump the minimum pasta requirement version for podman which is a 
> bit annoying as I would need to wait until the versions lands in our CI.

...would probing for the option be a possibility? That is, run 'pasta
--strict-bind', and if it fails, log a warning and continue without the
option.

If that's too cumbersome, then I would skip the new option and change
it to simply fail if we can't bind a single port (and report the single
error right away), unless the user requests to forward "almost all" the
ports, in which case we could probably warn() with a list of the ports
that we couldn't bind.

> > Another idea would be that the back-end in Podman passes ranges as
> > single ports... but then the command line might explode and that's
> > not ideal for users, either. I'd rather favour the extra option.  
>
> Yeah that works but for large ranges it would not be as performant and 
> there is the risk of hitting ARG_MAX (although I understand it is 
> unlikely to be real problem given one would need to give millions of 
> ports to hit it).
> >  
> >> Also besides this I find the error message less than ideal. It missing
> >> the errno from the bind syscall so important context gets lost (i.e.
> >> Address already in use vs Permission denied).  
> > The problem is that we might fail to bind multiple ports, so there
> > isn't necessarily a single bind() error. But if we go with
> > --strict-bind, we could report the first error (including return code
> > from the system call) and exit right away.  
> 
> I am fine with this, exit on the first error is what all our other 
> network tools do.
> But the behavior also exists for even a single port today, for reference 
> the error messages today:
> 
> rootlessport:
> 
> $ podman run -p 53:53 --rm quay.io/libpod/testimage:20221018
> Error: rootlessport cannot expose privileged port 53, you can add 
> 'net.ipv4.ip_unprivileged_port_start=53' to /etc/sysctl.conf (currently 
> 1024), or choose a larger port number (>= 1024): listen tcp 0.0.0.0:53: 
> bind: permission denied
> $ podman run -p 8000:8000 --network slirp4netns  --rm 
> quay.io/libpod/testimage:20221018
> Error: rootlessport listen tcp 0.0.0.0:8000: bind: address already in use
> 
> pasta:
> 
> $ podman run -p 53:53 --network pasta  --rm 
> quay.io/libpod/testimage:20221018
> Error: pasta failed with exit code 1:
> No external routable interface for IPv6

Reading [1] below, this string is considered confusing too, even though
I feel that hiding it conceptually clashes with the rationale of your
very request. That is, one might argue that the user wanted IPv6 too,
but it's not working, and we should tell the user.

Anyway, we can change it to info() and run pasta with -q / --quiet.
Would that help?

> Failed to bind any port for '-t 53-53:53-53', exiting
> $ podman run -p 8000:8000 --network pasta  --rm 
> quay.io/libpod/testimage:20221018
> Error: pasta failed with exit code 1:
> No external routable interface for IPv6
> Failed to bind any port for '-t 8000-8000:8000-8000', exiting
> 
> I think the answer of what is more helpful to users is obvious and that 
> isn't just my opinion[1].

Sure, I understand, especially in case of a single port. This is just
the first time somebody reports that.

> > Let me know if any of this would address your problem, I can write a
> > patch in the next days in case (or feel free to submit one).
>
> [1] https://github.com/containers/podman/pull/21563#issuecomment-1937024642

-- 
Stefano


  reply	other threads:[~2024-02-12 14:14 UTC|newest]

Thread overview: 8+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2024-02-09 16:57 pasta does not correctly handle bind errors with port ranges Paul Holzinger
2024-02-09 21:09 ` Stefano Brivio
2024-02-12 11:45   ` Paul Holzinger
2024-02-12 14:13     ` Stefano Brivio [this message]
2024-02-12 14:43       ` Paul Holzinger
2024-02-12 16:56         ` Stefano Brivio
2024-02-14  9:15           ` Stefano Brivio
2024-02-14 10:24             ` Paul Holzinger

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20240212151331.277ef9fd@elisabeth \
    --to=sbrivio@redhat.com \
    --cc=passt-dev@passt.top \
    --cc=pholzing@redhat.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
Code repositories for project(s) associated with this public inbox

	https://passt.top/passt

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for IMAP folder(s).