Re: [dpdk-dev] [PATCH v6 3/5] telemetry: use unique socket paths for in-memory mode

[Bruce Richardson <bruce.richardson@intel.com>]

On Tue, Oct 05, 2021 at 03:41:17PM +0100, Kevin Traynor wrote:
> Hi Bruce, I started reviewing v5 before v6 was pushed, so these are just
> comments from v5,
> 

No problem. Changes V6-v5 are fairly small anyway. Thanks for the review.

> On 05/10/2021 14:59, Bruce Richardson wrote:
> > When DPDK is run using "in-memory" flag, multiple processes can be run
> > using the same file-prefix and hence the same runtime directory. To
> > avoid problems with conflicting telemetry unix socket paths, we can
> > put the pid of the process into the socket name. As with the existing
> > telemetry socket files, these sockets are removed on normal program
> > exit.
> > 
> > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> > ---
> >   doc/guides/howto/telemetry.rst     | 17 ++++++++++++++++-
> >   lib/eal/freebsd/eal.c              |  1 +
> >   lib/eal/linux/eal.c                |  1 +
> >   lib/telemetry/telemetry.c          | 15 ++++++++++++---
> >   lib/telemetry/telemetry_internal.h |  3 ++-
> >   5 files changed, 32 insertions(+), 5 deletions(-)
> > 
> > diff --git a/doc/guides/howto/telemetry.rst b/doc/guides/howto/telemetry.rst
> > index 8f4fa1a510..8a61302459 100644
> > --- a/doc/guides/howto/telemetry.rst
> > +++ b/doc/guides/howto/telemetry.rst
> > @@ -13,12 +13,27 @@ ethdev port list, and eal parameters.
> >   Telemetry Interface
> >   -------------------
> > -The :doc:`../prog_guide/telemetry_lib` opens a socket with path
> > +For applications run normally, i.e. without the `--in-memory` EAL flag,
> > +the :doc:`../prog_guide/telemetry_lib` opens a socket with path
> >   *<runtime_directory>/dpdk_telemetry.<version>*. The version represents the
> >   telemetry version, the latest is v2. For example, a client would connect to a
> >   socket with path  */var/run/dpdk/\*/dpdk_telemetry.v2* (when the primary process
> >   is run by a root user).
> > +For applications run with the `--in-memory` EAL flag,
> > +the socket file is created with an additional suffix of the process PID.
> > +This is because multiple independent DPDK processes can be run simultaneously
> > +using the same runtime directory when *in-memory* mode is used.
> > +For example, when a user with UID 1000 runs processes with in-memory mode,
> > +we would find sockets available such as::
> > +
> > +  /run/user/1000/dpdk/rte/dpdk_telemetry.v2.1982
> > +  /run/user/1000/dpdk/rte/dpdk_telemetry.v2.1935
> > +
> 
> It seems an unnecessary step unless there is multiple process. As a
> suggestion, how about "simplifying" by always adding a check for an active
> socket with the default name. If it is not found, create it with the default
> (pre patches) name. If it is found and active, create a new one with process
> id appended. e.g.
> 
> First:
> /run/user/1000/dpdk/rte/dpdk_telemetry.v2
> 
> Next:
> /run/user/1000/dpdk/rte/dpdk_telemetry.v2.1982
> /run/user/1000/dpdk/rte/dpdk_telemetry.v2.1935
> 
> It means that existing socket can still be used by anything using telemetry.
> I think it is a nice default to keep as it doesn't require any changes for
> anything that will connect (e.g. collectd?)
> 
> The downside is that one will have a different name but it seems an
> acceptable trade-off to keep compatibility for single process case.
> 
> WDYT?
> 

Yes, that is an interesting idea, and probably not a bad one.

Taking things further, I wonder if using the pid is the best choice for a
suffix for the non-single-process case. If we used .1, .2 etc. it would
make things more predictable, perhaps easing integration for other tools.
Each process starting up would use the lowest free suffix, and any
external monitoring tools could just be set up to monitor the
first/second/third instance, with reproducable names across process
restarts.

> > +Where `/run/user/<uid>` is the runtime directory for the user given by the
> > +`$XDG_RUNTIME_DIR` environment variable,
> > +and `rte` is the default DPDK file prefix used for a runtime directory.
> > +
> >   Telemetry Initialization
> >   ------------------------
> > diff --git a/lib/eal/freebsd/eal.c b/lib/eal/freebsd/eal.c
> > index b06a2c1662..ed39d10b4e 100644
> > --- a/lib/eal/freebsd/eal.c
> > +++ b/lib/eal/freebsd/eal.c
> > @@ -952,6 +952,7 @@ rte_eal_init(int argc, char **argv)
> >   		if (tlog < 0)
> >   			tlog = RTE_LOGTYPE_EAL;
> >   		if (rte_telemetry_init(rte_eal_get_runtime_dir(),
> > +				internal_conf->in_memory | internal_conf->no_shconf,
> >   				rte_version(),
> >   				&internal_conf->ctrl_cpuset, rte_log, tlog) != 0)
> >   			return -1;
> > diff --git a/lib/eal/linux/eal.c b/lib/eal/linux/eal.c
> > index 0d0fc66668..9db4eb7913 100644
> > --- a/lib/eal/linux/eal.c
> > +++ b/lib/eal/linux/eal.c
> > @@ -1326,6 +1326,7 @@ rte_eal_init(int argc, char **argv)
> >   		if (tlog < 0)
> >   			tlog = RTE_LOGTYPE_EAL;
> >   		if (rte_telemetry_init(rte_eal_get_runtime_dir(),
> > +				internal_conf->in_memory | internal_conf->no_shconf,
> 
> should be logical OR
> 

I don't think it actually matters since any non-zero value will come
through as "true". However, will change it in v7.

/Bruce