On Tue, Oct 22, 2019 at 11:56 AM Shahaf Shuler <shah...@mellanox.com> wrote:
>
> Thursday, October 17, 2019 8:19 PM, Jerin Jacob:
> > Subject: Re: [dpdk-dev] [RFC PATCH 20.02] mbuf: hint PMD not to inline
> > packet
> >
> > On Thu, Oct 17, 2019 at 4:30 PM Shahaf Shuler <shah...@mellanox.com>
> > wrote:
> > >
> > > Thursday, October 17, 2019 11:17 AM, Jerin Jacob:
> > > > Subject: Re: [dpdk-dev] [RFC PATCH 20.02] mbuf: hint PMD not to
> > > > inline packet
> > > >
> > > > On Thu, Oct 17, 2019 at 12:57 PM Shahaf Shuler
> > > > <shah...@mellanox.com>
> > > > wrote:
> > > > >
> > > > > Some PMDs inline the mbuf data buffer directly to device. This is
> > > > > in order to save the overhead of the PCI headers involved when the
> > > > > device DMA read the buffer pointer. For some devices it is
> > > > > essential in order to reach the pick BW.
> > > > >
> > > > > However, there are cases where such inlining is in-efficient. For
> > > > > example when the data buffer resides on other device memory (like
> > > > > GPU or storage device). attempt to inline such buffer will result
> > > > > in high PCI overhead for reading and copying the data from the remote
> > device.
> > > >
> > > > Some questions to understand the use case # Is this use case where
> > > > CPU, local DRAM, NW card and GPU memory connected on the coherent
> > > > bus
> > >
> > > Yes. For example one can allocate GPU memory and map it to the GPU bar,
> > make it accessible from the host CPU through LD/ST.
> > >
> > > > # Assuming the CPU needs to touch the buffer prior to Tx, In that
> > > > case, it will be useful?
> > >
> > > If the CPU needs to modify the data then no. it will be more efficient to
> > copy the data to CPU and then send it.
> > > However there are use cases where the data is DMA w/ zero copy to the
> > GPU (for example) , GPU perform the processing on the data, and then CPU
> > send the mbuf (w/o touching the data).
> >
> > OK. If I understanding it correctly it is for offloading the Network/Compute
> > functions to GPU from NW card and/or CPU.
>
> Mostly the compute. The networking on this model is expected to be done by
> the CPU.
> Note this is only one use case.
>
> >
> > >
> > > > # How the application knows, The data buffer is in GPU memory in
> > > > order to use this flag efficiently?
> > >
> > > Because it made it happen. For example it attached the mbuf external
> > buffer from the other device memory.
> > >
> > > > # Just an random thought, Does it help, if we create two different
> > > > mempools one from local DRAM and one from GPU memory so that the
> > > > application can work transparently.
> > >
> > > But you will still need to teach the PMD which pool it can inline and
> > > which
> > cannot.
> > > IMO it is more generic to have it per mbuf. Moreover, application has this
> > info.
> >
> > IMO, we can not use PKT_TX_DONT_INLINE_HINT flag for generic
> > applications, The application usage will be tightly coupled with the
> > platform
> > and capabilities of GPU or Host CPU etc.
> >
> > I think, pushing this logic to the application is bad idea. But if you are
> > writing
> > some custom application and the per packet-level you need to control then
> > this flag may be the only way.
>
> Yes. This flag is for custom application who do unique acceleration (by doing
> Zero copy for compute/compression/encryption accelerators) on specific
> platforms.
> Such application is fully aware to the platform and the location where the
> data resides hence it is very simple for it to know how to set this flag.
# if it is per packet, it will be an implicit requirement to add it mbuf.
If so,
# Does it makes sense to add through dynamic mbuf? Maybe it is not
worth it for a single bit.
Since we have only 17 bits (40 - 23) remaining for Rx and Tx and it is
custom application requirement,
how about adding PKT_PMD_CUSTOM1 flags so that similar requirement by other PMDs
can leverage the same bit for such custom applications.(We have a
similar use case for smart NIC (not so make much sense for generic
applications) but needed for per packet)
>
> Note, This flag is 0 by default - meaning no hint and generic application
> works same as today.
>
> >
> >
> > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > > > >
> > > > > To support a mixed traffic pattern (some buffers from local DRAM,
> > > > > some buffers from other devices) with high BW, a hint flag is
> > > > > introduced in the mbuf.
> > > > > Application will hint the PMD whether or not it should try to
> > > > > inline the given mbuf data buffer. PMD should do best effort to
> > > > > act upon this request.
> > > > >
> > > > > Signed-off-by: Shahaf Shuler <shah...@mellanox.com>
> > > > > ---
> > > > > lib/librte_mbuf/rte_mbuf.h | 9 +++++++++
> > > > > 1 file changed, 9 insertions(+)
> > > > >
> > > > > diff --git a/lib/librte_mbuf/rte_mbuf.h
> > > > > b/lib/librte_mbuf/rte_mbuf.h index 98225ec80b..5934532b7f 100644
> > > > > --- a/lib/librte_mbuf/rte_mbuf.h
> > > > > +++ b/lib/librte_mbuf/rte_mbuf.h
> > > > > @@ -203,6 +203,15 @@ extern "C" {
> > > > > /* add new TX flags here */
> > > > >
> > > > > /**
> > > > > + * Hint to PMD to not inline the mbuf data buffer to device
> > > > > + * rather let the device use its DMA engine to fetch the data
> > > > > +with the
> > > > > + * provided pointer.
> > > > > + *
> > > > > + * This flag is a only a hint. PMD should enforce it as best effort.
> > > > > + */
> > > > > +#define PKT_TX_DONT_INLINE_HINT (1ULL << 39)
> > > > > +
> > > > > +/**
> > > > > * Indicate that the metadata field in the mbuf is in use.
> > > > > */
> > > > > #define PKT_TX_METADATA (1ULL << 40)
> > > > > --
> > > > > 2.12.0
> > > > >
