From: Jose Abreu <Jose.Abreu@synopsys.com>
To: Jon Hunter <jonathanh@nvidia.com>,
Jose Abreu <Jose.Abreu@synopsys.com>,
"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
"netdev@vger.kernel.org" <netdev@vger.kernel.org>,
"linux-stm32@st-md-mailman.stormreply.com"
<linux-stm32@st-md-mailman.stormreply.com>,
"linux-arm-kernel@lists.infradead.org"
<linux-arm-kernel@lists.infradead.org>
Cc: Joao Pinto <Joao.Pinto@synopsys.com>,
"David S . Miller" <davem@davemloft.net>,
Giuseppe Cavallaro <peppe.cavallaro@st.com>,
Alexandre Torgue <alexandre.torgue@st.com>,
Maxime Coquelin <mcoquelin.stm32@gmail.com>,
Maxime Ripard <maxime.ripard@bootlin.com>,
Chen-Yu Tsai <wens@csie.org>, Robin Murphy <robin.murphy@arm.com>,
"wens@csie.org" <wens@csie.org>,
linux-tegra <linux-tegra@vger.kernel.org>,
"peppe.cavallaro@st.com" <peppe.cavallaro@st.com>
Subject: RE: [PATCH net-next 3/3] net: stmmac: Introducing support for Page Pool
Date: Thu, 25 Jul 2019 13:26:24 +0000 [thread overview]
Message-ID: <BYAPR12MB3269927AB1F67D46E150ED6BD3C10@BYAPR12MB3269.namprd12.prod.outlook.com> (raw)
In-Reply-To: <7a79be5d-7ba2-c457-36d3-1ccef6572181@nvidia.com>
From: Jon Hunter <jonathanh@nvidia.com>
Date: Jul/25/2019, 14:20:07 (UTC+00:00)
>
> On 03/07/2019 11:37, Jose Abreu wrote:
> > Mapping and unmapping DMA region is an high bottleneck in stmmac driver,
> > specially in the RX path.
> >
> > This commit introduces support for Page Pool API and uses it in all RX
> > queues. With this change, we get more stable troughput and some increase
> > of banwidth with iperf:
> > - MAC1000 - 950 Mbps
> > - XGMAC: 9.22 Gbps
> >
> > Signed-off-by: Jose Abreu <joabreu@synopsys.com>
> > Cc: Joao Pinto <jpinto@synopsys.com>
> > Cc: David S. Miller <davem@davemloft.net>
> > Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
> > Cc: Alexandre Torgue <alexandre.torgue@st.com>
> > Cc: Maxime Coquelin <mcoquelin.stm32@gmail.com>
> > Cc: Maxime Ripard <maxime.ripard@bootlin.com>
> > Cc: Chen-Yu Tsai <wens@csie.org>
> > ---
> > drivers/net/ethernet/stmicro/stmmac/Kconfig | 1 +
> > drivers/net/ethernet/stmicro/stmmac/stmmac.h | 10 +-
> > drivers/net/ethernet/stmicro/stmmac/stmmac_main.c | 196 ++++++----------------
> > 3 files changed, 63 insertions(+), 144 deletions(-)
>
> ...
>
> > @@ -3306,49 +3295,22 @@ static inline void stmmac_rx_refill(struct stmmac_priv *priv, u32 queue)
> > else
> > p = rx_q->dma_rx + entry;
> >
> > - if (likely(!rx_q->rx_skbuff[entry])) {
> > - struct sk_buff *skb;
> > -
> > - skb = netdev_alloc_skb_ip_align(priv->dev, bfsize);
> > - if (unlikely(!skb)) {
> > - /* so for a while no zero-copy! */
> > - rx_q->rx_zeroc_thresh = STMMAC_RX_THRESH;
> > - if (unlikely(net_ratelimit()))
> > - dev_err(priv->device,
> > - "fail to alloc skb entry %d\n",
> > - entry);
> > - break;
> > - }
> > -
> > - rx_q->rx_skbuff[entry] = skb;
> > - rx_q->rx_skbuff_dma[entry] =
> > - dma_map_single(priv->device, skb->data, bfsize,
> > - DMA_FROM_DEVICE);
> > - if (dma_mapping_error(priv->device,
> > - rx_q->rx_skbuff_dma[entry])) {
> > - netdev_err(priv->dev, "Rx DMA map failed\n");
> > - dev_kfree_skb(skb);
> > + if (!buf->page) {
> > + buf->page = page_pool_dev_alloc_pages(rx_q->page_pool);
> > + if (!buf->page)
> > break;
> > - }
> > -
> > - stmmac_set_desc_addr(priv, p, rx_q->rx_skbuff_dma[entry]);
> > - stmmac_refill_desc3(priv, rx_q, p);
> > -
> > - if (rx_q->rx_zeroc_thresh > 0)
> > - rx_q->rx_zeroc_thresh--;
> > -
> > - netif_dbg(priv, rx_status, priv->dev,
> > - "refill entry #%d\n", entry);
> > }
> > - dma_wmb();
> > +
> > + buf->addr = buf->page->dma_addr;
> > + stmmac_set_desc_addr(priv, p, buf->addr);
> > + stmmac_refill_desc3(priv, rx_q, p);
> >
> > rx_q->rx_count_frames++;
> > rx_q->rx_count_frames %= priv->rx_coal_frames;
> > use_rx_wd = priv->use_riwt && rx_q->rx_count_frames;
> >
> > - stmmac_set_rx_owner(priv, p, use_rx_wd);
> > -
> > dma_wmb();
> > + stmmac_set_rx_owner(priv, p, use_rx_wd);
> >
> > entry = STMMAC_GET_ENTRY(entry, DMA_RX_SIZE);
> > }
>
> I was looking at this change in a bit closer detail and one thing that
> stuck out to me was the above where the barrier had been moved from
> after the stmmac_set_rx_owner() call to before.
>
> So I moved this back and I no longer saw the crash. However, then I
> recalled I had the patch to enable the debug prints for the buffer
> address applied but after reverting that, the crash occurred again.
>
> In other words, what works for me is moving the above barrier and adding
> the debug print, which appears to suggest that there is some
> timing/coherency issue here. Anyway, maybe this is clue to what is going
> on?
>
> diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
> index a7486c2f3221..2f016397231b 100644
> --- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
> +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
> @@ -3303,8 +3303,8 @@ static inline void stmmac_rx_refill(struct stmmac_priv *priv, u32 queue)
> rx_q->rx_count_frames %= priv->rx_coal_frames;
> use_rx_wd = priv->use_riwt && rx_q->rx_count_frames;
>
> - dma_wmb();
> stmmac_set_rx_owner(priv, p, use_rx_wd);
> + dma_wmb();
>
> entry = STMMAC_GET_ENTRY(entry, DMA_RX_SIZE);
> }
> @@ -3438,6 +3438,10 @@ static int stmmac_rx(struct stmmac_priv *priv, int limit, u32 queue)
> dma_sync_single_for_device(priv->device, buf->addr,
> frame_len, DMA_FROM_DEVICE);
>
> + pr_info("%s: paddr=0x%llx, vaddr=0x%llx, len=%d", __func__,
> + buf->addr, page_address(buf->page),
> + frame_len);
> +
> if (netif_msg_pktdata(priv)) {
> netdev_dbg(priv->dev, "frame received (%dbytes)",
> frame_len);
>
> Cheers
> Jon
>
> --
> nvpublic
Well, I wasn't expecting that :/
Per documentation of barriers I think we should set descriptor fields
and then barrier and finally ownership to HW so that remaining fields
are coherent before owner is set.
Anyway, can you also add a dma_rmb() after the call to
stmmac_rx_status() ?
---
Thanks,
Jose Miguel Abreu
next prev parent reply other threads:[~2019-07-25 13:26 UTC|newest]
Thread overview: 87+ messages / expand[flat|nested] mbox.gz Atom feed top
2019-07-03 10:37 [PATCH net-next 0/3] net: stmmac: Some performance improvements and a fix Jose Abreu
2019-07-03 10:37 ` [PATCH net-next 1/3] net: stmmac: Implement RX Coalesce Frames setting Jose Abreu
2019-07-03 20:41 ` Jakub Kicinski
2019-07-03 10:37 ` [PATCH net-next 2/3] net: stmmac: Fix descriptors address being in > 32 bits address space Jose Abreu
2019-07-03 10:37 ` [PATCH net-next 3/3] net: stmmac: Introducing support for Page Pool Jose Abreu
2019-07-03 10:40 ` Jose Abreu
2019-07-04 9:39 ` Jesper Dangaard Brouer
2019-07-04 14:45 ` Jose Abreu
2019-07-04 15:09 ` Jesper Dangaard Brouer
2019-07-04 15:18 ` Jose Abreu
2019-07-04 15:33 ` Jesper Dangaard Brouer
2019-07-04 9:48 ` Jesper Dangaard Brouer
2019-07-04 10:00 ` Jesper Dangaard Brouer
2019-07-04 10:13 ` Jose Abreu
2019-07-04 11:11 ` Ilias Apalodimas
2019-07-04 11:54 ` Jesper Dangaard Brouer
2019-07-04 12:04 ` Ilias Apalodimas
2019-07-04 12:59 ` Jose Abreu
2019-07-04 13:06 ` Ilias Apalodimas
2019-07-04 10:30 ` Ilias Apalodimas
2019-07-04 12:14 ` Arnd Bergmann
2019-07-04 12:49 ` Ilias Apalodimas
2019-07-17 18:58 ` Jon Hunter
2019-07-18 7:29 ` Jose Abreu
2019-07-18 7:48 ` Jose Abreu
2019-07-18 9:16 ` Jon Hunter
2019-07-19 7:51 ` Jose Abreu
2019-07-19 8:37 ` Jon Hunter
2019-07-19 8:44 ` Jose Abreu
2019-07-19 8:49 ` Jon Hunter
2019-07-19 10:25 ` Jose Abreu
2019-07-19 12:28 ` Jose Abreu
2019-07-19 13:33 ` Jon Hunter
2019-07-19 12:30 ` Jon Hunter
2019-07-19 12:32 ` Jose Abreu
2019-07-19 13:35 ` Jon Hunter
2019-07-22 7:23 ` Jose Abreu
2019-07-22 9:37 ` Jon Hunter
2019-07-22 9:47 ` Jose Abreu
2019-07-22 9:57 ` Jose Abreu
2019-07-22 10:27 ` Jon Hunter
2019-07-22 10:18 ` Ilias Apalodimas
2019-07-22 11:11 ` Lars Persson
2019-07-22 11:39 ` Jose Abreu
2019-07-22 12:05 ` Jon Hunter
2019-07-22 14:04 ` Jose Abreu
2019-07-23 8:14 ` Jose Abreu
2019-07-23 10:01 ` Jon Hunter
2019-07-23 10:07 ` Jose Abreu
2019-07-23 10:29 ` Robin Murphy
2019-07-23 11:22 ` Jose Abreu
2019-07-23 12:09 ` Jon Hunter
2019-07-23 13:19 ` Robin Murphy
2019-07-23 21:39 ` Jon Hunter
2019-07-24 10:03 ` Robin Murphy
2019-07-23 18:51 ` David Miller
2019-07-24 8:54 ` Ilias Apalodimas
2019-07-24 9:43 ` Jose Abreu
2019-07-24 9:53 ` Ilias Apalodimas
2019-07-24 10:04 ` Jose Abreu
2019-07-24 11:10 ` Jon Hunter
2019-07-24 11:34 ` Jose Abreu
2019-07-24 11:58 ` Jon Hunter
2019-07-25 7:44 ` Jose Abreu
2019-07-25 9:45 ` Jon Hunter
2019-07-25 11:39 ` Ilias Apalodimas
2019-07-23 10:38 ` Jon Hunter
2019-07-23 10:49 ` Jose Abreu
2019-07-23 11:58 ` Jon Hunter
2019-07-23 12:51 ` Jose Abreu
2019-07-23 13:34 ` Jon Hunter
2019-07-29 9:45 ` Mikko Perttunen
2019-07-25 13:20 ` Jon Hunter
2019-07-25 13:26 ` Jose Abreu [this message]
2019-07-25 14:25 ` Jon Hunter
2019-07-25 15:12 ` Jose Abreu
2019-07-26 14:11 ` Jon Hunter
2019-07-27 15:56 ` Jose Abreu
2019-07-29 8:16 ` Jose Abreu
2019-07-29 10:55 ` Jon Hunter
2019-07-29 11:29 ` Jose Abreu
2019-07-29 11:52 ` Robin Murphy
2019-07-29 14:08 ` Jose Abreu
2019-07-29 21:33 ` Jon Hunter
2019-07-30 9:39 ` Jose Abreu
2019-07-30 13:36 ` Jon Hunter
2019-07-30 13:58 ` Jose Abreu
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=BYAPR12MB3269927AB1F67D46E150ED6BD3C10@BYAPR12MB3269.namprd12.prod.outlook.com \
--to=jose.abreu@synopsys.com \
--cc=Joao.Pinto@synopsys.com \
--cc=alexandre.torgue@st.com \
--cc=davem@davemloft.net \
--cc=jonathanh@nvidia.com \
--cc=linux-arm-kernel@lists.infradead.org \
--cc=linux-kernel@vger.kernel.org \
--cc=linux-stm32@st-md-mailman.stormreply.com \
--cc=linux-tegra@vger.kernel.org \
--cc=maxime.ripard@bootlin.com \
--cc=mcoquelin.stm32@gmail.com \
--cc=netdev@vger.kernel.org \
--cc=peppe.cavallaro@st.com \
--cc=robin.murphy@arm.com \
--cc=wens@csie.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).