RE: [PATCH net-next 9/9] nfp: eliminate an if statement in calculation of completed frames
From: Jakub Kicinski
> Sent: 16 May 2017 01:55
> Given that our rings are always a power of 2, we can simplify the
> calculation of number of completed TX descriptors by using masking
> instead of if statement based on whether the index have wrapped
> or not.
>
> Signed-off-by: Jakub Kicinski
> ---
> drivers/net/ethernet/netronome/nfp/nfp_net_common.c | 10 ++
> 1 file changed, 2 insertions(+), 8 deletions(-)
>
> diff --git a/drivers/net/ethernet/netronome/nfp/nfp_net_common.c
> b/drivers/net/ethernet/netronome/nfp/nfp_net_common.c
> index c64514f8ee65..da83e17b8b20 100644
> --- a/drivers/net/ethernet/netronome/nfp/nfp_net_common.c
> +++ b/drivers/net/ethernet/netronome/nfp/nfp_net_common.c
> @@ -940,10 +940,7 @@ static void nfp_net_tx_complete(struct nfp_net_tx_ring
> *tx_ring)
> if (qcp_rd_p == tx_ring->qcp_rd_p)
> return;
>
> - if (qcp_rd_p > tx_ring->qcp_rd_p)
> - todo = qcp_rd_p - tx_ring->qcp_rd_p;
> - else
> - todo = qcp_rd_p + tx_ring->cnt - tx_ring->qcp_rd_p;
> + todo = D_IDX(tx_ring, qcp_rd_p + tx_ring->cnt - tx_ring->qcp_rd_p);
I'm not sure you need to add tx_ring->cnt here.
I bet D_IDX() masks it away.
> while (todo--) {
> idx = D_IDX(tx_ring, tx_ring->rd_p++);
That '++' looks suspicious.
I think you need to decide whether you are incrementing pointers into the ring
or indexes into it.
Sometimes it is safer to use a non-wrapping index and mask when accessing the
entry.
entry_ptr = &ring[idx & (RING_SIZE - 1)]
Ring full is then (read_idx == write_idx + RING_SIZE),
ring empty (read_idx == write_idx).
So the index just wrap at (probably)_2^32.
David
Re: [PATCH net-next 9/9] nfp: eliminate an if statement in calculation of completed frames
On Wed, 17 May 2017 11:07:19 +, David Laight wrote:
> From: Jakub Kicinski
> > Sent: 16 May 2017 01:55
> > Given that our rings are always a power of 2, we can simplify the
> > calculation of number of completed TX descriptors by using masking
> > instead of if statement based on whether the index have wrapped
> > or not.
> >
> > Signed-off-by: Jakub Kicinski
> > ---
> > drivers/net/ethernet/netronome/nfp/nfp_net_common.c | 10 ++
> > 1 file changed, 2 insertions(+), 8 deletions(-)
> >
> > diff --git a/drivers/net/ethernet/netronome/nfp/nfp_net_common.c
> > b/drivers/net/ethernet/netronome/nfp/nfp_net_common.c
> > index c64514f8ee65..da83e17b8b20 100644
> > --- a/drivers/net/ethernet/netronome/nfp/nfp_net_common.c
> > +++ b/drivers/net/ethernet/netronome/nfp/nfp_net_common.c
> > @@ -940,10 +940,7 @@ static void nfp_net_tx_complete(struct nfp_net_tx_ring
> > *tx_ring)
> > if (qcp_rd_p == tx_ring->qcp_rd_p)
> > return;
> >
> > - if (qcp_rd_p > tx_ring->qcp_rd_p)
> > - todo = qcp_rd_p - tx_ring->qcp_rd_p;
> > - else
> > - todo = qcp_rd_p + tx_ring->cnt - tx_ring->qcp_rd_p;
> > + todo = D_IDX(tx_ring, qcp_rd_p + tx_ring->cnt - tx_ring->qcp_rd_p);
>
> I'm not sure you need to add tx_ring->cnt here.
> I bet D_IDX() masks it away.
True, feel free to send a fix, or I will queue up a correction after
other work I have pending.
> > while (todo--) {
> > idx = D_IDX(tx_ring, tx_ring->rd_p++);
>
> That '++' looks suspicious.
> I think you need to decide whether you are incrementing pointers into the ring
> or indexes into it.
> Sometimes it is safer to use a non-wrapping index and mask when accessing the
> entry.
> entry_ptr = &ring[idx & (RING_SIZE - 1)]
> Ring full is then (read_idx == write_idx + RING_SIZE),
> ring empty (read_idx == write_idx).
> So the index just wrap at (probably)_2^32.
I may be missing the point. I use a mix of the two, actually, the
software pointers are free running (non-wrapping) but the HW QCP
pointers wrap. Because HW pointers wrap I always keep one entry on
the rings empty, see nfp_net_tx_full().
RE: [PATCH net-next 9/9] nfp: eliminate an if statement in calculation of completed frames
From: Jakub Kicinski
> Sent: 17 May 2017 18:37
..
> > > while (todo--) {
> > > idx = D_IDX(tx_ring, tx_ring->rd_p++);
> >
> > That '++' looks suspicious.
> > I think you need to decide whether you are incrementing pointers into the
> > ring
> > or indexes into it.
> > Sometimes it is safer to use a non-wrapping index and mask when accessing
> > the entry.
> > entry_ptr = &ring[idx & (RING_SIZE - 1)]
> > Ring full is then (read_idx == write_idx + RING_SIZE),
> > ring empty (read_idx == write_idx).
> > So the index just wrap at (probably)_2^32.
>
> I may be missing the point. I use a mix of the two, actually, the
> software pointers are free running (non-wrapping) but the HW QCP
> pointers wrap. Because HW pointers wrap I always keep one entry on
> the rings empty, see nfp_net_tx_full().
Ah, I'd assumed that rd_p was a pointer, not an index.
David
