Never drop socket syscalls to ensure we have fdinfo for subsequent binds. SMAGENT-1270

[thom-sd]

No description provided.

[thom-sd]

@gianlucaborello, we've encountered some situations where we're missing listening ports from bind() after the socket() calls are dropped. So, we're marking socket as NEVER_DROP here.

However, we're not quite sure what the performance implications would be of marking socket calls as NEVER_DROP.
Would it make more sense to have the driver pass the fdinfo with the bind instead?

[gianlucaborello]

You could probably patch the bind code and return some information about the socket that you'd need in order to recreate it. My perspective about the dropping vs not dropping, completely ignorant to the specific details, is that you are already marking close as non-droppable, so at a minimum the driver is always hit by at least one single system call per connection.

So, if you also keep socket I don't expect you'll get hit with 10X the events, but more like 2X in the very worst case, and this is for a system call for which I'd be curious to see a workload that generates more than a few thousands per second, since these are network connections. If we were talking about an actual I/O event (e.g. write), then my feedback would be different, as in those cases it would be pretty easy to generate 1000X the amount of events necessary to create the socket.

Please also notice that I didn't initially write that code, so my opinion might not be too valuable here, and I might be missing something that you're already considering (e.g. actual cost of the userspace dragent process required to create and add the fd to the table, but it seems like you'll want that anyway).