Bind to device *and IP*

[orgads]

If a device has several IP addresses, it should be possible to force one of them.

We have a patch that implements this as an additional option in libcurl (and a command-line argument).

While preparing it for PR, I was wondering if it would be better to extend the existing if!/host! notation, and do something like if!eth0!10.2.3.4 instead.

Current implementation (in our patched curl):

    curl_easy_setopt(curl, CURLOPT_INTERFACE, "eth0"); /* or "if!eth0" */
    curl_easy_setopt(curl, CURLOPT_INTERFACE_IP, "10.2.3.4");

Proposed change:

    curl_easy_setopt(curl, CURLOPT_INTERFACE, "if!eth0!10.2.3.4");

If the second option is preferred, then I can also drop the extra command-line argument.

Will you accept such a feature? If you will, which of these options do you prefer?

[orgads]

@bagder @vszakats ?

[vszakats]

With the single option, perhaps a distinct new prefix, e.g. ifhost!eth1!10.2.3.4, would make it more unambiguous to write/read and easier to parse?

Regardless of the prefix chosen, when passed to older curl without support for this, it will result in an unexpected failure, while with the extra CURLOPT option the additional IP/host constraint would just be a no-op. Though perhaps a failure is the better outcome here.

As for CURLOPT_INTERFACE_IP, assuming it also support hosts, would CURLOPT_INTERFACE_HOST be a better name? (rhyming with the host! prefix). The downside of the extra option is that some might assume this also works standalone to bind to an IP.

I tend to prefer the the single option with a distinct new prefix.

[orgads]

I like ifhost!. @bagder do you have a preference?

[bagder]

How exactly does this differ from binding to the IP address alone? Surely a host cannot have the same IP on multiple network interfaces?

[orgads]

1. See --interface and CURLOPT_INTERFACE works different for Ip Address and Interface #13613 for an example why IP address is not enough.
2. You'll be surprised, but we do have customers with the same IP address on different network interfaces, don't ask 😅

[bagder]

I don't think #13613 explained it, hence me asking. And now I am wiser! 😄

[orgads]

These are 2 different reasons. Regarding #13613, even if you don't have the same IP on different devices, the kernel can route outgoing packets from a different device, not necessarily the one with the bound address.

[orgads]

Draft PR here: #13719

There are 2 commits there. The first one adds an option in the library, and the second one implements ifhost! (and reverts most of the first one).

Once you state your preference, I'll either squash them or leave only the first one.

[orgads]

Ok, I implemented ifhost! notation in #13719. Closing this discussion.