On Fri, Jul 19, 2024 at 09:59:50AM +0100, Ferruh Yigit wrote:
> On 7/11/2024 1:35 PM, Bruce Richardson wrote:
> > When allocating memory for an ethdev, the rte_malloc_socket call used
> > only allocates memory on the NUMA node/socket local to the device. This
> > means that even if the user wanted to, they could never use a remote NIC
> > without also having memory on that NIC's socket.
> >
> > For example, if we change examples/skeleton/basicfwd.c to have
> > SOCKET_ID_ANY as the socket_id parameter for Rx and Tx rings, we should
> > be able to run the app cross-numa e.g. as below, where the two PCI
> > devices are on socket 1, and core 1 is on socket 0:
> >
> > ./build/examples/dpdk-skeleton -l 1 --legacy-mem --socket-mem=1024,0 \
> > -a a8:00.0 -a b8:00.0
> >
> > This fails however, with the error:
> >
> > ETHDEV: failed to allocate private data
> > PCI_BUS: Requested device 0000:a8:00.0 cannot be used
> >
> > We can remove this restriction by doing a fallback call to general
> > rte_malloc after a call to rte_malloc_socket fails. This should be safe
> > to do because the later ethdev calls to setup Rx/Tx queues all take a
> > socket_id parameter, which can be used by applications to enforce the
> > requirement for local-only memory for a device, if so desired. [If
> > device-local memory is present it will be used as before, while if not
> > present the rte_eth_dev_configure call will now pass, but the subsequent
> > queue setup calls requesting local memory will fail].
> >
> > Fixes: e489007a411c ("ethdev: add generic create/destroy ethdev APIs")
> > Fixes: dcd5c8112bc3 ("ethdev: add PCI driver helpers")
> > Cc: sta...@dpdk.org
> >
> > Signed-off-by: Bruce Richardson <bruce.richard...@intel.com>
> > Signed-off-by: Padraig Connolly <padraig.j.conno...@intel.com>
> >
>
> Hi Bruce,
>
> If device-local memory is present, behavior will be same, so I agree
> this is low impact.
>
> But for the case device-local memory is NOT present, should we enforce
> the HW setup is the question. This can be beneficial for users new to DPDK.
>
> Probably 'dev_private' on its own has small impact on the performance
> (although it may depend if these fields used in datapath), but it may be
> vehicle to enforce local memory.
>
Actually, thinking about it further, I actually think that using rte_malloc
itself may even be a better solution in general, because the dev_private
structure is never accessed by the NIC itself, but only by the cores.
Therefore, logically, it makes more sense to allocate the memory for it
according to where the process is running rather than where the NIC is
actually located. rte_malloc does this, by first attempting to allocate
based on the calling core socket, and then falling back to allocating on
any other socket on failure of the initial socket-local attempt.
/Bruce
