> From: lihuisong (C) [mailto:lihuis...@huawei.com]
> Sent: Tuesday, 31 October 2023 03.58
>
> 在 2023/10/31 2:48, Stephen Hemminger 写道:
> > On Mon, 30 Oct 2023 09:25:34 +0800
> > "lihuisong (C)" <lihuis...@huawei.com> wrote:
> >
> >>>
> >>>> The "min_rx_bufsize" in struct rte_eth_dev_info stands for the
> >>>> minimum Rx buffer size supported by hardware. Actually, some
> >>>> engines also have the maximum Rx buffer specification, like, hns3.
> >>>>
> >>>> If mbuf data room size in mempool is greater then the maximum Rx
> >>>> buffer size supported by HW, the data size application used in
> >>>> each mbuf is just as much as the maximum Rx buffer size supported
> >>>> by HW instead of the whole data room size.
> >>>>
> >>>> So introduce maximum Rx buffer size which is not enforced just to
> >>>> report user to avoid memory waste.
> >>> I am not convinced this is really necessary.
> >>> Your device will use up to 4K of buffer size, not sure why an
> >>> application would want to use much larger than that because it
> >>> would be wasting a lot of buffer space (most packets are smaller)
> >>> anyway.
> >>>
> >>> The only case where it might be useful is if application is using
> >>> jumbo frames (9K) and the application was not able to handle multi
> >>> segment packets.
> >> Yeah, it is useful if user want a large packet (like, 6K) is in a
> >> mbuf. But, in current layer, user don't know what the maximum buffer
> >> size per descriptor supported by HW is.
> >>> Not handling multi segment packets in SW is just programmer
> >>> laziness.
> >> User do decide their implement based on their cases in project.
> >> May it be a point for this that user don't want to do memcpy for
> >> multi segment packets and just use the first mbuf memory.
> >>
> >> Now that there is the "min_rx_bufsize" to report in ethdev layer.
> >> Anyway, DPDK is indeed the lack of the way to report the maximum Rx
> >> buffer size per hw descriptor.
> > My concern is that you are creating a special case for one driver.
> understand your concern.
> > And other drivers probably have similar upper bound.
> Yes, they also have similar upper bound.
> From the codes, the max buffer size of Most PMDs are 16K and bnxt is
> 9600Byte.
> Do we need to report this size? It's a common feature for all PMDs.
I think this could be a useful feature for applications not wanting to deal
with scattered packets. I don't consider such applications exotic, so I support
adding this feature.
> >
> > Could the warning be better handled in the driver specific configure
> > routine rather than updating the ethdev API. Something like:
> >
> > if (multi-segment-flag off) {
Many drivers ignore the RTE_ETH_RX_OFFLOAD_SCATTER configuration, and only use
RTE_ETH_RX_OFFLOAD_SCATTER for capabilities reporting. Perhaps this should be
reported in Bugzilla?
> > if (mtu > driver max buf size) {
> > return error;
> > } else {
> > if (mtu > driver max buf size &&
> > mtu < mempool_buf_size(mp)) {
> > warn that packet maybe segmented ??
> > }
> > }
Such a log message would also be useful.
