On Thu, 8 Aug 2024 11:37:38 +0200 Paul Holzinger wrote:

On 08/08/2024 06:02, Stefano Brivio wrote:

...

Given that pasta supports specifying a command to be executed on the command line, even without the usual -- separator as long as there's no ambiguity, we shouldn't eat up options that are not meant for us.

Paul reports, for instance, that with:

pasta --config-net ip -6 route

-6 is taken by pasta to mean --ipv6-only, and we execute 'ip route'. That's because getopt_long(), by default, shuffles the argument list to shift non-option arguments at the end.

Avoid that by adding '+' at the beginning of 'optstring'.

Reported-by: Paul Holzinger Signed-off-by: Stefano Brivio --- v3: Use '+' in optstring and drop first non-option tracking

v2: Instead of overriding 'name' in the getopt_long() loop, to force exiting the loop, adjust the exit condition

conf.c | 4 ++-- util.c | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-)

I like this change but I like to point out that this is a breaking change for any user that sets options after the main argument, i.e. pid.

Oh, right, that actually happens to work, even if we never really supported that, the man page has only this form: pasta [OPTION]... PID I could go back to v2, and, on top of that, if we find a single non-option argument that looks like a PID (a number), we would push it to the end of argv and continue parsing. If we find any other number, including one that's after all the other options but before the presumed PID we just pushed, we'll report error. We have anyway the following problem, which we won't make any worse (it can't be done without an actual file with POSIX shell, Bash only): $ 1() { echo a; } $ pasta 1; echo Couldn't open user namespace /proc/1/ns/user: Permission denied $ pasta echo; 1 a

I can tell you that this will not effect podman but I don't know what other users exists out there...

As far as I know, the only other tool using pasta(1) at the moment is rootless-containers (Docker, Usernetes): https://github.com/rootless-containers/rootlesskit/blob/master/pkg/network/p... which is also fine. Other users are developers and people who try out network topologies and namespaces stuff without root, but I suppose adding the PID at the end is pretty natural anyway. On the other hand, if we can make sure we avoid this kind of breakage at a small cost, why not. I'll try.

I am not sure if it is worth the risk just to improve the UX for the command use case but I guess you already decided it is otherwise you would have not posted this patch.

No, not really, I wasn't actually aware of the fact that adding the PID before options worked. Thanks for pointing that out. -- Stefano

On Thu, 8 Aug 2024 13:29:09 +0200 Paul Holzinger wrote:

On 08/08/2024 12:16, Stefano Brivio wrote:

...

On Thu, 8 Aug 2024 11:37:38 +0200 Paul Holzinger wrote:

...

On 08/08/2024 06:02, Stefano Brivio wrote:

...

Given that pasta supports specifying a command to be executed on the command line, even without the usual -- separator as long as there's no ambiguity, we shouldn't eat up options that are not meant for us.

Paul reports, for instance, that with:

pasta --config-net ip -6 route

-6 is taken by pasta to mean --ipv6-only, and we execute 'ip route'. That's because getopt_long(), by default, shuffles the argument list to shift non-option arguments at the end.

Avoid that by adding '+' at the beginning of 'optstring'.

Reported-by: Paul Holzinger Signed-off-by: Stefano Brivio --- v3: Use '+' in optstring and drop first non-option tracking

v2: Instead of overriding 'name' in the getopt_long() loop, to force exiting the loop, adjust the exit condition

conf.c | 4 ++-- util.c | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-)

I like this change but I like to point out that this is a breaking change for any user that sets options after the main argument, i.e. pid. Oh, right, that actually happens to work, even if we never really supported that, the man page has only this form:

pasta [OPTION]... PID

I could go back to v2, and, on top of that, if we find a single non-option argument that looks like a PID (a number), we would push it to the end of argv and continue parsing.

If we find any other number, including one that's after all the other options but before the presumed PID we just pushed, we'll report error.

We have anyway the following problem, which we won't make any worse (it can't be done without an actual file with POSIX shell, Bash only):

$ 1() { echo a; } $ pasta 1; echo Couldn't open user namespace /proc/1/ns/user: Permission denied

$ pasta echo; 1

a

...

I can tell you that this will not effect podman but I don't know what other users exists out there... As far as I know, the only other tool using pasta(1) at the moment is rootless-containers (Docker, Usernetes):

https://github.com/rootless-containers/rootlesskit/blob/master/pkg/network/p...

which is also fine. Other users are developers and people who try out network topologies and namespaces stuff without root, but I suppose adding the PID at the end is pretty natural anyway.

On the other hand, if we can make sure we avoid this kind of breakage at a small cost, why not. I'll try.

...

I am not sure if it is worth the risk just to improve the UX for the command use case but I guess you already decided it is otherwise you would have not posted this patch. No, not really, I wasn't actually aware of the fact that adding the PID before options worked. Thanks for pointing that out.

Well not just pid, it works the same with a command:

$ pasta ip addr --config-net

With this patch this no longer works, as --config-net is now passed to the ip command. I don't think using it like that makes any sense and it is super confusing so I like the new way but whoever does use such a syntax will get broken. Thus it is a trade off to be made.

Ah, sure, I see, but that's a completely different level of unreasonableness. I don't think anybody ever dreamt of using it like that. That something I would happily "break". -- Stefano

On Thu, Aug 08, 2024 at 08:33:42PM +0200, Stefano Brivio wrote:

On Thu, 8 Aug 2024 12:16:58 +0200 Stefano Brivio wrote:

...

On Thu, 8 Aug 2024 11:37:38 +0200 Paul Holzinger wrote:

...

On 08/08/2024 06:02, Stefano Brivio wrote:

...

Given that pasta supports specifying a command to be executed on the command line, even without the usual -- separator as long as there's no ambiguity, we shouldn't eat up options that are not meant for us.

Paul reports, for instance, that with:

pasta --config-net ip -6 route

-6 is taken by pasta to mean --ipv6-only, and we execute 'ip route'. That's because getopt_long(), by default, shuffles the argument list to shift non-option arguments at the end.

Avoid that by adding '+' at the beginning of 'optstring'.

Reported-by: Paul Holzinger Signed-off-by: Stefano Brivio --- v3: Use '+' in optstring and drop first non-option tracking

v2: Instead of overriding 'name' in the getopt_long() loop, to force exiting the loop, adjust the exit condition

conf.c | 4 ++-- util.c | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-)

I like this change but I like to point out that this is a breaking change for any user that sets options after the main argument, i.e. pid.

Oh, right, that actually happens to work, even if we never really supported that, the man page has only this form:

pasta [OPTION]... PID

I could go back to v2, and, on top of that, if we find a single non-option argument that looks like a PID (a number), we would push it to the end of argv and continue parsing.

Please no, this sounds horribly complicated. My experience is that when you try to intuit what the user wants this way, you're more likely to just behave unpredictably in other cases.

...

If we find any other number, including one that's after all the other options but before the presumed PID we just pushed, we'll report error.

We have anyway the following problem, which we won't make any worse (it can't be done without an actual file with POSIX shell, Bash only):

$ 1() { echo a; } $ pasta 1; echo Couldn't open user namespace /proc/1/ns/user: Permission denied

$ pasta echo; 1

a

...

I can tell you that this will not effect podman but I don't know what other users exists out there...

As far as I know, the only other tool using pasta(1) at the moment is rootless-containers (Docker, Usernetes):

https://github.com/rootless-containers/rootlesskit/blob/master/pkg/network/p...

which is also fine. Other users are developers and people who try out network topologies and namespaces stuff without root, but I suppose adding the PID at the end is pretty natural anyway.

On the other hand, if we can make sure we avoid this kind of breakage at a small cost, why not. I'll try.

...no, not really a small cost. The logic itself is spectacularly complicated, something on the lines of:

case 1: if (c->mode != MODE_PASTA) usage(argv[0], stderr, EXIT_FAILURE);

/* There can be just one PID in the middle of options */ errno = 0; strtol(argv[optind - 1], NULL, 10); if (errno) { /* Not a PID, stop parsing here */ name = -1; break; }

if (!pid_found) { SWAP(argv[optind - 1], argv[argc - 1]); pid_found = true; }

if (optind <= argc) optind--;

if (optind == argc) name = -1;

break;

but this would still be acceptable, I guess. However, we have multiple getopt_long() calls, and we would need to either handle the '1' case for each one of them, or modify optstring, or extract, here, some logic from conf_pasta_ns()... which I would really like to avoid.

Let me go ahead with v3, which anyway takes care of all the cases that are documented and we actually meant to support. If somebody really needs to insert a PID in the middle of the option list, we'll "fix" this.

Phew. -- David Gibson (he or they) | I'll have my music baroque, and my code david AT gibson.dropbear.id.au | minimalist, thank you, not the other way | around. http://www.ozlabs.org/~dgibson