| Message ID | 20230420121152.2737625-1-horatiu.vultur@microchip.com (mailing list archive) |
|---|---|
| State | Changes Requested |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | [net-next] net: lan966x: Don't use xdp_frame when action is XDP_TX | expand |
From: Horatiu Vultur <horatiu.vultur@microchip.com> Date: Thu, 20 Apr 2023 14:11:52 +0200 > When the action of an xdp program was XDP_TX, lan966x was creating > a xdp_frame and use this one to send the frame back. But it is also > possible to send back the frame without needing a xdp_frame, because > it possible to send it back using the page. > And then once the frame is transmitted is possible to use directly > page_pool_recycle_direct as lan966x is using page pools. > This would save some CPU usage on this path. > > Signed-off-by: Horatiu Vultur <horatiu.vultur@microchip.com> [...] > @@ -702,6 +704,7 @@ static void lan966x_fdma_tx_start(struct lan966x_tx *tx, int next_to_use) > int lan966x_fdma_xmit_xdpf(struct lan966x_port *port, > struct xdp_frame *xdpf, > struct page *page, > + u32 len, > bool dma_map) I think you can cut the number of arguments by almost a half: int lan966x_fdma_xmit_xdpf(struct lan966x_port *port, void *ptr, u32 len) { if (len) { /* XDP_TX, ptr is page */ page = ptr; dma_sync_here(page, len); } else { /* XDP_REDIR, ptr is xdp_frame */ xdpf = ptr; dma_map_here(xdpf->data, xdpf->len); } @page and @xdpf are mutually exclusive. When @xdpf is non-null, @len is excessive (xdpf->len is here), so you can use @len as logical `len * !dma_map`, i.e. zero for REDIR and the actual frame length for TX. I generally enjoy seeing how you constantly improve stuff in your driver :) > { > struct lan966x *lan966x = port->lan966x; > @@ -722,6 +725,15 @@ int lan966x_fdma_xmit_xdpf(struct lan966x_port *port, > goto out; > } [...] Thanks, Olek
On Thu, Apr 20, 2023 at 02:11:52PM +0200, Horatiu Vultur wrote: 'net: ' in patch subject is excessive to me > When the action of an xdp program was XDP_TX, lan966x was creating > a xdp_frame and use this one to send the frame back. But it is also > possible to send back the frame without needing a xdp_frame, because > it possible to send it back using the page. s/it/it is > And then once the frame is transmitted is possible to use directly > page_pool_recycle_direct as lan966x is using page pools. > This would save some CPU usage on this path. i remember this optimization gave me noticeable perf improvement, would you mind sharing it in % on your side? > > Signed-off-by: Horatiu Vultur <horatiu.vultur@microchip.com> > --- > .../ethernet/microchip/lan966x/lan966x_fdma.c | 35 +++++++++++-------- > .../ethernet/microchip/lan966x/lan966x_main.h | 2 ++ > .../ethernet/microchip/lan966x/lan966x_xdp.c | 11 +++--- > 3 files changed, 27 insertions(+), 21 deletions(-) > > diff --git a/drivers/net/ethernet/microchip/lan966x/lan966x_fdma.c b/drivers/net/ethernet/microchip/lan966x/lan966x_fdma.c > index 2ed76bb61a731..7947259e67e4e 100644 > --- a/drivers/net/ethernet/microchip/lan966x/lan966x_fdma.c > +++ b/drivers/net/ethernet/microchip/lan966x/lan966x_fdma.c > @@ -390,6 +390,7 @@ static void lan966x_fdma_stop_netdev(struct lan966x *lan966x) > static void lan966x_fdma_tx_clear_buf(struct lan966x *lan966x, int weight) > { > struct lan966x_tx *tx = &lan966x->tx; > + struct lan966x_rx *rx = &lan966x->rx; > struct lan966x_tx_dcb_buf *dcb_buf; > struct xdp_frame_bulk bq; > struct lan966x_db *db; > @@ -432,7 +433,8 @@ static void lan966x_fdma_tx_clear_buf(struct lan966x *lan966x, int weight) > if (dcb_buf->xdp_ndo) > xdp_return_frame_bulk(dcb_buf->data.xdpf, &bq); > else > - xdp_return_frame_rx_napi(dcb_buf->data.xdpf); > + page_pool_recycle_direct(rx->page_pool, > + dcb_buf->data.page); > } > > clear = true; > @@ -702,6 +704,7 @@ static void lan966x_fdma_tx_start(struct lan966x_tx *tx, int next_to_use) > int lan966x_fdma_xmit_xdpf(struct lan966x_port *port, > struct xdp_frame *xdpf, > struct page *page, > + u32 len, agreed with Olek regarding arguments reduction here > bool dma_map) > { > struct lan966x *lan966x = port->lan966x; > @@ -722,6 +725,15 @@ int lan966x_fdma_xmit_xdpf(struct lan966x_port *port, > goto out; > } > > + /* Fill up the buffer */ > + next_dcb_buf = &tx->dcbs_buf[next_to_use]; > + next_dcb_buf->use_skb = false; > + next_dcb_buf->xdp_ndo = dma_map; a bit misleading that xdp_ndo is a bool :P > + next_dcb_buf->len = len + IFH_LEN_BYTES; > + next_dcb_buf->used = true; > + next_dcb_buf->ptp = false; > + next_dcb_buf->dev = port->dev; > + > /* Generate new IFH */ > if (dma_map) { > if (xdpf->headroom < IFH_LEN_BYTES) { > @@ -736,16 +748,18 @@ int lan966x_fdma_xmit_xdpf(struct lan966x_port *port, > > dma_addr = dma_map_single(lan966x->dev, > xdpf->data - IFH_LEN_BYTES, > - xdpf->len + IFH_LEN_BYTES, > + len + IFH_LEN_BYTES, > DMA_TO_DEVICE); > if (dma_mapping_error(lan966x->dev, dma_addr)) { > ret = NETDEV_TX_OK; > goto out; > } > > + next_dcb_buf->data.xdpf = xdpf; > + > /* Setup next dcb */ > lan966x_fdma_tx_setup_dcb(tx, next_to_use, > - xdpf->len + IFH_LEN_BYTES, > + len + IFH_LEN_BYTES, > dma_addr); > } else { > ifh = page_address(page) + XDP_PACKET_HEADROOM; > @@ -756,25 +770,18 @@ int lan966x_fdma_xmit_xdpf(struct lan966x_port *port, > dma_addr = page_pool_get_dma_addr(page); > dma_sync_single_for_device(lan966x->dev, > dma_addr + XDP_PACKET_HEADROOM, > - xdpf->len + IFH_LEN_BYTES, > + len + IFH_LEN_BYTES, > DMA_TO_DEVICE); > > + next_dcb_buf->data.page = page; > + > /* Setup next dcb */ > lan966x_fdma_tx_setup_dcb(tx, next_to_use, > - xdpf->len + IFH_LEN_BYTES, > + len + IFH_LEN_BYTES, > dma_addr + XDP_PACKET_HEADROOM); > } > > - /* Fill up the buffer */ > - next_dcb_buf = &tx->dcbs_buf[next_to_use]; > - next_dcb_buf->use_skb = false; > - next_dcb_buf->data.xdpf = xdpf; > - next_dcb_buf->xdp_ndo = dma_map; > - next_dcb_buf->len = xdpf->len + IFH_LEN_BYTES; > next_dcb_buf->dma_addr = dma_addr; > - next_dcb_buf->used = true; > - next_dcb_buf->ptp = false; > - next_dcb_buf->dev = port->dev; > > /* Start the transmission */ > lan966x_fdma_tx_start(tx, next_to_use); > diff --git a/drivers/net/ethernet/microchip/lan966x/lan966x_main.h b/drivers/net/ethernet/microchip/lan966x/lan966x_main.h > index 851afb0166b19..59da35a2c93d4 100644 > --- a/drivers/net/ethernet/microchip/lan966x/lan966x_main.h > +++ b/drivers/net/ethernet/microchip/lan966x/lan966x_main.h > @@ -243,6 +243,7 @@ struct lan966x_tx_dcb_buf { > union { > struct sk_buff *skb; > struct xdp_frame *xdpf; > + struct page *page; > } data; > u32 len; > u32 used : 1; > @@ -544,6 +545,7 @@ int lan966x_fdma_xmit(struct sk_buff *skb, __be32 *ifh, struct net_device *dev); > int lan966x_fdma_xmit_xdpf(struct lan966x_port *port, > struct xdp_frame *frame, > struct page *page, > + u32 len, > bool dma_map); > int lan966x_fdma_change_mtu(struct lan966x *lan966x); > void lan966x_fdma_netdev_init(struct lan966x *lan966x, struct net_device *dev); > diff --git a/drivers/net/ethernet/microchip/lan966x/lan966x_xdp.c b/drivers/net/ethernet/microchip/lan966x/lan966x_xdp.c > index 2e6f486ec67d7..a8ad1f4e431cb 100644 > --- a/drivers/net/ethernet/microchip/lan966x/lan966x_xdp.c > +++ b/drivers/net/ethernet/microchip/lan966x/lan966x_xdp.c > @@ -62,7 +62,7 @@ int lan966x_xdp_xmit(struct net_device *dev, > struct xdp_frame *xdpf = frames[i]; > int err; > > - err = lan966x_fdma_xmit_xdpf(port, xdpf, NULL, true); > + err = lan966x_fdma_xmit_xdpf(port, xdpf, NULL, xdpf->len, true); > if (err) > break; > > @@ -76,7 +76,6 @@ int lan966x_xdp_run(struct lan966x_port *port, struct page *page, u32 data_len) > { > struct bpf_prog *xdp_prog = port->xdp_prog; > struct lan966x *lan966x = port->lan966x; > - struct xdp_frame *xdpf; > struct xdp_buff xdp; > u32 act; > > @@ -90,11 +89,9 @@ int lan966x_xdp_run(struct lan966x_port *port, struct page *page, u32 data_len) > case XDP_PASS: > return FDMA_PASS; > case XDP_TX: > - xdpf = xdp_convert_buff_to_frame(&xdp); > - if (!xdpf) > - return FDMA_DROP; > - > - return lan966x_fdma_xmit_xdpf(port, xdpf, page, false) ? > + return lan966x_fdma_xmit_xdpf(port, NULL, page, > + data_len - IFH_LEN_BYTES, > + false) ? > FDMA_DROP : FDMA_TX; > case XDP_REDIRECT: > if (xdp_do_redirect(port->dev, &xdp, xdp_prog)) > -- > 2.38.0 >
The 04/20/2023 16:49, Alexander Lobakin wrote: > > From: Horatiu Vultur <horatiu.vultur@microchip.com> > Date: Thu, 20 Apr 2023 14:11:52 +0200 Hi Olek, > > > When the action of an xdp program was XDP_TX, lan966x was creating > > a xdp_frame and use this one to send the frame back. But it is also > > possible to send back the frame without needing a xdp_frame, because > > it possible to send it back using the page. > > And then once the frame is transmitted is possible to use directly > > page_pool_recycle_direct as lan966x is using page pools. > > This would save some CPU usage on this path. > > > > Signed-off-by: Horatiu Vultur <horatiu.vultur@microchip.com> > > [...] > > > @@ -702,6 +704,7 @@ static void lan966x_fdma_tx_start(struct lan966x_tx *tx, int next_to_use) > > int lan966x_fdma_xmit_xdpf(struct lan966x_port *port, > > struct xdp_frame *xdpf, > > struct page *page, > > + u32 len, > > bool dma_map) > > I think you can cut the number of arguments by almost a half: > > int lan966x_fdma_xmit_xdpf(struct lan966x_port *port, > void *ptr, u32 len) > { > if (len) { > /* XDP_TX, ptr is page */ > page = ptr; > > dma_sync_here(page, len); > } else { > /* XDP_REDIR, ptr is xdp_frame */ > xdpf = ptr; > > dma_map_here(xdpf->data, xdpf->len); > } > > @page and @xdpf are mutually exclusive. When @xdpf is non-null, @len is > excessive (xdpf->len is here), so you can use @len as logical > `len * !dma_map`, i.e. zero for REDIR and the actual frame length for TX. Thanks for the review. You are right, I can reduce number of arguments, the reason why I have done it like this, I thought it is a little bit more clear this way. But I will update as you propose in the next version > > I generally enjoy seeing how you constantly improve stuff in your driver :) > > > { > > struct lan966x *lan966x = port->lan966x; > > @@ -722,6 +725,15 @@ int lan966x_fdma_xmit_xdpf(struct lan966x_port *port, > > goto out; > > } > [...] > > Thanks, > Olek
The 04/20/2023 22:52, Maciej Fijalkowski wrote: > [Some people who received this message don't often get email from maciej.fijalkowski@intel.com. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ] > Hi Maciej, > > On Thu, Apr 20, 2023 at 02:11:52PM +0200, Horatiu Vultur wrote: > > 'net: ' in patch subject is excessive to me I usually have set this in the subject. I can remove this in the next version and I will try to keep in mind for other patches for lan966x. > > > When the action of an xdp program was XDP_TX, lan966x was creating > > a xdp_frame and use this one to send the frame back. But it is also > > possible to send back the frame without needing a xdp_frame, because > > it possible to send it back using the page. > > s/it/it is > > > And then once the frame is transmitted is possible to use directly > > page_pool_recycle_direct as lan966x is using page pools. > > This would save some CPU usage on this path. > > i remember this optimization gave me noticeable perf improvement, would > you mind sharing it in % on your side? The way I have done the measurements, is to measure actually how much more traffic can be send back. I tried with different frame sizes, frame size improvement 64 ~8% 256 ~11% 512 ~8% 1000 ~0% 1500 ~0% I will make sure do add this to the comments in the next version. > > > > > Signed-off-by: Horatiu Vultur <horatiu.vultur@microchip.com> > > --- > > .../ethernet/microchip/lan966x/lan966x_fdma.c | 35 +++++++++++-------- > > .../ethernet/microchip/lan966x/lan966x_main.h | 2 ++ > > .../ethernet/microchip/lan966x/lan966x_xdp.c | 11 +++--- > > 3 files changed, 27 insertions(+), 21 deletions(-) > > > > diff --git a/drivers/net/ethernet/microchip/lan966x/lan966x_fdma.c b/drivers/net/ethernet/microchip/lan966x/lan966x_fdma.c > > index 2ed76bb61a731..7947259e67e4e 100644 > > --- a/drivers/net/ethernet/microchip/lan966x/lan966x_fdma.c > > +++ b/drivers/net/ethernet/microchip/lan966x/lan966x_fdma.c > > @@ -390,6 +390,7 @@ static void lan966x_fdma_stop_netdev(struct lan966x *lan966x) > > static void lan966x_fdma_tx_clear_buf(struct lan966x *lan966x, int weight) > > { > > struct lan966x_tx *tx = &lan966x->tx; > > + struct lan966x_rx *rx = &lan966x->rx; > > struct lan966x_tx_dcb_buf *dcb_buf; > > struct xdp_frame_bulk bq; > > struct lan966x_db *db; > > @@ -432,7 +433,8 @@ static void lan966x_fdma_tx_clear_buf(struct lan966x *lan966x, int weight) > > if (dcb_buf->xdp_ndo) > > xdp_return_frame_bulk(dcb_buf->data.xdpf, &bq); > > else > > - xdp_return_frame_rx_napi(dcb_buf->data.xdpf); > > + page_pool_recycle_direct(rx->page_pool, > > + dcb_buf->data.page); > > } > > > > clear = true; > > @@ -702,6 +704,7 @@ static void lan966x_fdma_tx_start(struct lan966x_tx *tx, int next_to_use) > > int lan966x_fdma_xmit_xdpf(struct lan966x_port *port, > > struct xdp_frame *xdpf, > > struct page *page, > > + u32 len, > > agreed with Olek regarding arguments reduction here Yes, I will change this in the next version. > > > bool dma_map) > > { > > struct lan966x *lan966x = port->lan966x; > > @@ -722,6 +725,15 @@ int lan966x_fdma_xmit_xdpf(struct lan966x_port *port, > > goto out; > > } > > > > + /* Fill up the buffer */ > > + next_dcb_buf = &tx->dcbs_buf[next_to_use]; > > + next_dcb_buf->use_skb = false; > > + next_dcb_buf->xdp_ndo = dma_map; > > a bit misleading that xdp_ndo is a bool :P There are few other variables that are misleading :), I need to get to this and clean it a little bit. > > > + next_dcb_buf->len = len + IFH_LEN_BYTES; > > + next_dcb_buf->used = true; > > + next_dcb_buf->ptp = false; > > + next_dcb_buf->dev = port->dev; > > + > > /* Generate new IFH */ > > if (dma_map) { > > if (xdpf->headroom < IFH_LEN_BYTES) {
diff --git a/drivers/net/ethernet/microchip/lan966x/lan966x_fdma.c b/drivers/net/ethernet/microchip/lan966x/lan966x_fdma.c index 2ed76bb61a731..7947259e67e4e 100644 --- a/drivers/net/ethernet/microchip/lan966x/lan966x_fdma.c +++ b/drivers/net/ethernet/microchip/lan966x/lan966x_fdma.c @@ -390,6 +390,7 @@ static void lan966x_fdma_stop_netdev(struct lan966x *lan966x) static void lan966x_fdma_tx_clear_buf(struct lan966x *lan966x, int weight) { struct lan966x_tx *tx = &lan966x->tx; + struct lan966x_rx *rx = &lan966x->rx; struct lan966x_tx_dcb_buf *dcb_buf; struct xdp_frame_bulk bq; struct lan966x_db *db; @@ -432,7 +433,8 @@ static void lan966x_fdma_tx_clear_buf(struct lan966x *lan966x, int weight) if (dcb_buf->xdp_ndo) xdp_return_frame_bulk(dcb_buf->data.xdpf, &bq); else - xdp_return_frame_rx_napi(dcb_buf->data.xdpf); + page_pool_recycle_direct(rx->page_pool, + dcb_buf->data.page); } clear = true; @@ -702,6 +704,7 @@ static void lan966x_fdma_tx_start(struct lan966x_tx *tx, int next_to_use) int lan966x_fdma_xmit_xdpf(struct lan966x_port *port, struct xdp_frame *xdpf, struct page *page, + u32 len, bool dma_map) { struct lan966x *lan966x = port->lan966x; @@ -722,6 +725,15 @@ int lan966x_fdma_xmit_xdpf(struct lan966x_port *port, goto out; } + /* Fill up the buffer */ + next_dcb_buf = &tx->dcbs_buf[next_to_use]; + next_dcb_buf->use_skb = false; + next_dcb_buf->xdp_ndo = dma_map; + next_dcb_buf->len = len + IFH_LEN_BYTES; + next_dcb_buf->used = true; + next_dcb_buf->ptp = false; + next_dcb_buf->dev = port->dev; + /* Generate new IFH */ if (dma_map) { if (xdpf->headroom < IFH_LEN_BYTES) { @@ -736,16 +748,18 @@ int lan966x_fdma_xmit_xdpf(struct lan966x_port *port, dma_addr = dma_map_single(lan966x->dev, xdpf->data - IFH_LEN_BYTES, - xdpf->len + IFH_LEN_BYTES, + len + IFH_LEN_BYTES, DMA_TO_DEVICE); if (dma_mapping_error(lan966x->dev, dma_addr)) { ret = NETDEV_TX_OK; goto out; } + next_dcb_buf->data.xdpf = xdpf; + /* Setup next dcb */ lan966x_fdma_tx_setup_dcb(tx, next_to_use, - xdpf->len + IFH_LEN_BYTES, + len + IFH_LEN_BYTES, dma_addr); } else { ifh = page_address(page) + XDP_PACKET_HEADROOM; @@ -756,25 +770,18 @@ int lan966x_fdma_xmit_xdpf(struct lan966x_port *port, dma_addr = page_pool_get_dma_addr(page); dma_sync_single_for_device(lan966x->dev, dma_addr + XDP_PACKET_HEADROOM, - xdpf->len + IFH_LEN_BYTES, + len + IFH_LEN_BYTES, DMA_TO_DEVICE); + next_dcb_buf->data.page = page; + /* Setup next dcb */ lan966x_fdma_tx_setup_dcb(tx, next_to_use, - xdpf->len + IFH_LEN_BYTES, + len + IFH_LEN_BYTES, dma_addr + XDP_PACKET_HEADROOM); } - /* Fill up the buffer */ - next_dcb_buf = &tx->dcbs_buf[next_to_use]; - next_dcb_buf->use_skb = false; - next_dcb_buf->data.xdpf = xdpf; - next_dcb_buf->xdp_ndo = dma_map; - next_dcb_buf->len = xdpf->len + IFH_LEN_BYTES; next_dcb_buf->dma_addr = dma_addr; - next_dcb_buf->used = true; - next_dcb_buf->ptp = false; - next_dcb_buf->dev = port->dev; /* Start the transmission */ lan966x_fdma_tx_start(tx, next_to_use); diff --git a/drivers/net/ethernet/microchip/lan966x/lan966x_main.h b/drivers/net/ethernet/microchip/lan966x/lan966x_main.h index 851afb0166b19..59da35a2c93d4 100644 --- a/drivers/net/ethernet/microchip/lan966x/lan966x_main.h +++ b/drivers/net/ethernet/microchip/lan966x/lan966x_main.h @@ -243,6 +243,7 @@ struct lan966x_tx_dcb_buf { union { struct sk_buff *skb; struct xdp_frame *xdpf; + struct page *page; } data; u32 len; u32 used : 1; @@ -544,6 +545,7 @@ int lan966x_fdma_xmit(struct sk_buff *skb, __be32 *ifh, struct net_device *dev); int lan966x_fdma_xmit_xdpf(struct lan966x_port *port, struct xdp_frame *frame, struct page *page, + u32 len, bool dma_map); int lan966x_fdma_change_mtu(struct lan966x *lan966x); void lan966x_fdma_netdev_init(struct lan966x *lan966x, struct net_device *dev); diff --git a/drivers/net/ethernet/microchip/lan966x/lan966x_xdp.c b/drivers/net/ethernet/microchip/lan966x/lan966x_xdp.c index 2e6f486ec67d7..a8ad1f4e431cb 100644 --- a/drivers/net/ethernet/microchip/lan966x/lan966x_xdp.c +++ b/drivers/net/ethernet/microchip/lan966x/lan966x_xdp.c @@ -62,7 +62,7 @@ int lan966x_xdp_xmit(struct net_device *dev, struct xdp_frame *xdpf = frames[i]; int err; - err = lan966x_fdma_xmit_xdpf(port, xdpf, NULL, true); + err = lan966x_fdma_xmit_xdpf(port, xdpf, NULL, xdpf->len, true); if (err) break; @@ -76,7 +76,6 @@ int lan966x_xdp_run(struct lan966x_port *port, struct page *page, u32 data_len) { struct bpf_prog *xdp_prog = port->xdp_prog; struct lan966x *lan966x = port->lan966x; - struct xdp_frame *xdpf; struct xdp_buff xdp; u32 act; @@ -90,11 +89,9 @@ int lan966x_xdp_run(struct lan966x_port *port, struct page *page, u32 data_len) case XDP_PASS: return FDMA_PASS; case XDP_TX: - xdpf = xdp_convert_buff_to_frame(&xdp); - if (!xdpf) - return FDMA_DROP; - - return lan966x_fdma_xmit_xdpf(port, xdpf, page, false) ? + return lan966x_fdma_xmit_xdpf(port, NULL, page, + data_len - IFH_LEN_BYTES, + false) ? FDMA_DROP : FDMA_TX; case XDP_REDIRECT: if (xdp_do_redirect(port->dev, &xdp, xdp_prog))
When the action of an xdp program was XDP_TX, lan966x was creating a xdp_frame and use this one to send the frame back. But it is also possible to send back the frame without needing a xdp_frame, because it possible to send it back using the page. And then once the frame is transmitted is possible to use directly page_pool_recycle_direct as lan966x is using page pools. This would save some CPU usage on this path. Signed-off-by: Horatiu Vultur <horatiu.vultur@microchip.com> --- .../ethernet/microchip/lan966x/lan966x_fdma.c | 35 +++++++++++-------- .../ethernet/microchip/lan966x/lan966x_main.h | 2 ++ .../ethernet/microchip/lan966x/lan966x_xdp.c | 11 +++--- 3 files changed, 27 insertions(+), 21 deletions(-)