From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jakub Kicinski Subject: Re: [PATCH net-next 9/9] nfp: eliminate an if statement in calculation of completed frames Date: Wed, 17 May 2017 10:36:37 -0700 Message-ID: <20170517103637.2dd8b49d@cakuba.netronome.com> References: <20170516005523.26124-1-jakub.kicinski@netronome.com> <20170516005523.26124-10-jakub.kicinski@netronome.com> <063D6719AE5E284EB5DD2968C1650D6DCFFF7B93@AcuExch.aculab.com> Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit Cc: "netdev@vger.kernel.org" , "oss-drivers@netronome.com" To: David Laight Return-path: Received: from mx4.wp.pl ([212.77.101.11]:9919 "EHLO mx4.wp.pl" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751348AbdEQRgt (ORCPT ); Wed, 17 May 2017 13:36:49 -0400 In-Reply-To: <063D6719AE5E284EB5DD2968C1650D6DCFFF7B93@AcuExch.aculab.com> Sender: netdev-owner@vger.kernel.org List-ID: On Wed, 17 May 2017 11:07:19 +0000, 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().