Questions about the the pdump functionality

[Stephen Hemminger]

On Mon, 23 Oct 2023 18:32:48 +0300
Pavel Vazharov <freakpv at gmail.com> wrote:

> Hi there,
> 
> We've a DPDK based application from which we need to take packet dumps from
> time to time when problems arise. We are planning to use librte_pdump
> functions and the dpdk-dumppcap tool.
> I've few questions in related to the pdump functionality which we want to
> use:
> - Is calling `rte_pdump_init` at the startup of the main application
> causing some overhead for the packet processing during its run if there is
> no actual packet capturing enabled by the dpdk-dumppcap tool? I suppose
> there should be some check in place but is it something like a single `if`
> condition on a boolean flag or something heavier?
> - Is it possible then to call `rte_pdump_init` during the runtime of the
> main application only when I know that I'm about to start the dpdk-dumppcap
> tool? I mean, is it supported and safe to call `rte_pdump_init` and
> `rte_pdump_uninit` while the main application is running and processing
> packets or these functions are supposed to be called only at application
> startup and application stop.
> 
> Thanks in advance,
> Pavel.

Short answer:
The overhead only happens when the dump application is running.

Long answer:
Running rte_pdump_init() tells adds an additional service to the
multi-process (primary/secondary) communication mechanism.
Secondary and primary process connect with each other over Unix
domain socket, and the services are handled by the multi-process
thread in the primary process. Adding a service does not interact
with fast path at all.

The best way (as always) to discover this yourself is to read
the source code and follow along.