| Message ID | 20210412101713.15161-1-kurt@linutronix.de |
|---|---|
| State | RFC |
| Headers | show |
| Series | [RFC,net] igb: Fix XDP with PTP enabled | expand |
On Mon Apr 12 2021, Jesper Dangaard Brouer wrote: > On Mon, 12 Apr 2021 12:17:13 +0200 > Kurt Kanzenbach <kurt@linutronix.de> wrote: > >> When using native XDP with the igb driver, the XDP frame data doesn't point to >> the beginning of the packet. It's off by 16 bytes. Everything works as expected >> with XDP skb mode. >> >> Actually these 16 bytes are used to store the packet timestamps. Therefore, pull >> the timestamp before executing any XDP operations and adjust all other code >> accordingly. The igc driver does it like that as well. > > (Cc. Alexander Duyck) Thanks. > > Do we have enough room for the packet page-split tricks when these 16 > bytes are added? I think so. AFAICT the timestamp header is accounted. There is IGB_2K_TOO_SMALL_WITH_PADDING. If 2k isn't sufficient, then 3k buffers are used. The only thing this patch does, is adjusting the xdp->data pointer before executing igb_run_xdp() instead of doing it afterwards. So, that in the eBPF program `data' points to the packet data, and not to the timestamp. Thanks, Kurt
On Mon, Apr 12, 2021 at 7:29 AM Jesper Dangaard Brouer <brouer@redhat.com> wrote: > > > On Mon, 12 Apr 2021 12:17:13 +0200 > Kurt Kanzenbach <kurt@linutronix.de> wrote: > > > When using native XDP with the igb driver, the XDP frame data doesn't point to > > the beginning of the packet. It's off by 16 bytes. Everything works as expected > > with XDP skb mode. > > > > Actually these 16 bytes are used to store the packet timestamps. Therefore, pull > > the timestamp before executing any XDP operations and adjust all other code > > accordingly. The igc driver does it like that as well. > > (Cc. Alexander Duyck) > > Do we have enough room for the packet page-split tricks when these 16 > bytes are added? > > AFAIK this driver like ixgbe+i40e split the page in two 2048 bytes packets. > > The XDP headroom is reduced to 192 bytes. > The skb_shared_info is 320 bytes in size. > > 2048-192-320 = 1536 bytes > > MTU(L3) 1500 > Ethernet (L2) headers 14 bytes > VLAN 4 bytes, but Q-in-Q vlan 8 bytes. > > Single VLAN: 1536-1500-14-4 = 18 bytes left > Q-in-q VLAN: 1536-1500-14-8 = 14 bytes left So the Q-in-q case should kick us over to jumbo frames since we have to add the extra size into the final supported frame size. So the size itself should work. > > diff --git a/drivers/net/ethernet/intel/igb/igb_ptp.c b/drivers/net/ethernet/intel/igb/igb_ptp.c > > index 86a576201f5f..0cbdf48285d3 100644 > > --- a/drivers/net/ethernet/intel/igb/igb_ptp.c > > +++ b/drivers/net/ethernet/intel/igb/igb_ptp.c > > @@ -863,23 +863,22 @@ static void igb_ptp_tx_hwtstamp(struct igb_adapter *adapter) > > * igb_ptp_rx_pktstamp - retrieve Rx per packet timestamp > > * @q_vector: Pointer to interrupt specific structure > > * @va: Pointer to address containing Rx buffer > > - * @skb: Buffer containing timestamp and packet > > * > > * This function is meant to retrieve a timestamp from the first buffer of an > > * incoming frame. The value is stored in little endian format starting on > > * byte 8 > > * > > - * Returns: 0 if success, nonzero if failure > > + * Returns: 0 on failure, timestamp on success > > **/ > > -int igb_ptp_rx_pktstamp(struct igb_q_vector *q_vector, void *va, > > - struct sk_buff *skb) > > +ktime_t igb_ptp_rx_pktstamp(struct igb_q_vector *q_vector, void *va) > > { > > struct igb_adapter *adapter = q_vector->adapter; > > + struct skb_shared_hwtstamps ts; > > __le64 *regval = (__le64 *)va; > > int adjust = 0; > > > > if (!(adapter->ptp_flags & IGB_PTP_ENABLED)) > > - return IGB_RET_PTP_DISABLED; > > + return 0; > > > > /* The timestamp is recorded in little endian format. > > * DWORD: 0 1 2 3 > > @@ -888,10 +887,9 @@ int igb_ptp_rx_pktstamp(struct igb_q_vector *q_vector, void *va, > > > > /* check reserved dwords are zero, be/le doesn't matter for zero */ > > if (regval[0]) > > - return IGB_RET_PTP_INVALID; > > + return 0; > > One thing that needs to be cleaned up in the patch is that if it is going to drop these return values it should probably drop the defines for them since I don't think they are used anywhere else.
On Tue Apr 13 2021, Alexander Duyck wrote: > On Mon, Apr 12, 2021 at 7:29 AM Jesper Dangaard Brouer > <brouer@redhat.com> wrote: >> > +ktime_t igb_ptp_rx_pktstamp(struct igb_q_vector *q_vector, void *va) >> > { >> > struct igb_adapter *adapter = q_vector->adapter; >> > + struct skb_shared_hwtstamps ts; >> > __le64 *regval = (__le64 *)va; >> > int adjust = 0; >> > >> > if (!(adapter->ptp_flags & IGB_PTP_ENABLED)) >> > - return IGB_RET_PTP_DISABLED; >> > + return 0; >> > >> > /* The timestamp is recorded in little endian format. >> > * DWORD: 0 1 2 3 >> > @@ -888,10 +887,9 @@ int igb_ptp_rx_pktstamp(struct igb_q_vector *q_vector, void *va, >> > >> > /* check reserved dwords are zero, be/le doesn't matter for zero */ >> > if (regval[0]) >> > - return IGB_RET_PTP_INVALID; >> > + return 0; >> > > > One thing that needs to be cleaned up in the patch is that if it is > going to drop these return values it should probably drop the defines > for them since I don't think they are used anywhere else. Yes, of course. I'll clean it up before sending a non RFC version. Thanks, Kurt
>-----Original Message----- >From: Intel-wired-lan <intel-wired-lan-bounces@osuosl.org> On Behalf Of >Kurt Kanzenbach >Sent: Monday, April 12, 2021 3:47 PM >To: Brandeburg, Jesse <jesse.brandeburg@intel.com>; Nguyen, Anthony L ><anthony.l.nguyen@intel.com>; David S. Miller <davem@davemloft.net>; >Jakub Kicinski <kuba@kernel.org> >Cc: Kurt Kanzenbach <kurt@linutronix.de>; Jesper Dangaard Brouer ><hawk@kernel.org>; Daniel Borkmann <daniel@iogearbox.net>; >netdev@vger.kernel.org; Sebastian Andrzej Siewior <bigeasy@linutronix.de>; >Ilias Apalodimas <ilias.apalodimas@linaro.org>; Alexei Starovoitov ><ast@kernel.org>; intel-wired-lan@lists.osuosl.org; bpf@vger.kernel.org; >Lorenzo Bianconi <lorenzo@kernel.org> >Subject: [Intel-wired-lan] [PATCH RFC net] igb: Fix XDP with PTP enabled > >When using native XDP with the igb driver, the XDP frame data doesn't point >to the beginning of the packet. It's off by 16 bytes. Everything works as >expected with XDP skb mode. > >Actually these 16 bytes are used to store the packet timestamps. Therefore, >pull the timestamp before executing any XDP operations and adjust all other >code accordingly. The igc driver does it like that as well. > >Tested with Intel i210 card and AF_XDP sockets. > >Fixes: 9cbc948b5a20 ("igb: add XDP support") >Signed-off-by: Kurt Kanzenbach <kurt@linutronix.de> >--- > drivers/net/ethernet/intel/igb/igb.h | 3 +- > drivers/net/ethernet/intel/igb/igb_main.c | 45 ++++++++++++----------- >drivers/net/ethernet/intel/igb/igb_ptp.c | 18 ++++----- > 3 files changed, 32 insertions(+), 34 deletions(-) > Tested-by: Sandeep Penigalapati <sandeep.penigalapati@intel.com>
diff --git a/drivers/net/ethernet/intel/igb/igb.h b/drivers/net/ethernet/intel/igb/igb.h index 7bda8c5edea5..72cf967c1a00 100644 --- a/drivers/net/ethernet/intel/igb/igb.h +++ b/drivers/net/ethernet/intel/igb/igb.h @@ -748,8 +748,7 @@ void igb_ptp_suspend(struct igb_adapter *adapter); void igb_ptp_rx_hang(struct igb_adapter *adapter); void igb_ptp_tx_hang(struct igb_adapter *adapter); void igb_ptp_rx_rgtstamp(struct igb_q_vector *q_vector, struct sk_buff *skb); -int igb_ptp_rx_pktstamp(struct igb_q_vector *q_vector, void *va, - struct sk_buff *skb); +ktime_t igb_ptp_rx_pktstamp(struct igb_q_vector *q_vector, void *va); int igb_ptp_set_ts_config(struct net_device *netdev, struct ifreq *ifr); int igb_ptp_get_ts_config(struct net_device *netdev, struct ifreq *ifr); void igb_set_flag_queue_pairs(struct igb_adapter *, const u32); diff --git a/drivers/net/ethernet/intel/igb/igb_main.c b/drivers/net/ethernet/intel/igb/igb_main.c index a45cd2b416c8..8fab55fd18fc 100644 --- a/drivers/net/ethernet/intel/igb/igb_main.c +++ b/drivers/net/ethernet/intel/igb/igb_main.c @@ -8281,7 +8281,7 @@ static void igb_add_rx_frag(struct igb_ring *rx_ring, static struct sk_buff *igb_construct_skb(struct igb_ring *rx_ring, struct igb_rx_buffer *rx_buffer, struct xdp_buff *xdp, - union e1000_adv_rx_desc *rx_desc) + ktime_t timestamp) { #if (PAGE_SIZE < 8192) unsigned int truesize = igb_rx_pg_size(rx_ring) / 2; @@ -8301,12 +8301,8 @@ static struct sk_buff *igb_construct_skb(struct igb_ring *rx_ring, if (unlikely(!skb)) return NULL; - if (unlikely(igb_test_staterr(rx_desc, E1000_RXDADV_STAT_TSIP))) { - if (!igb_ptp_rx_pktstamp(rx_ring->q_vector, xdp->data, skb)) { - xdp->data += IGB_TS_HDR_LEN; - size -= IGB_TS_HDR_LEN; - } - } + if (timestamp) + skb_hwtstamps(skb)->hwtstamp = timestamp; /* Determine available headroom for copy */ headlen = size; @@ -8337,7 +8333,7 @@ static struct sk_buff *igb_construct_skb(struct igb_ring *rx_ring, static struct sk_buff *igb_build_skb(struct igb_ring *rx_ring, struct igb_rx_buffer *rx_buffer, struct xdp_buff *xdp, - union e1000_adv_rx_desc *rx_desc) + ktime_t timestamp) { #if (PAGE_SIZE < 8192) unsigned int truesize = igb_rx_pg_size(rx_ring) / 2; @@ -8364,11 +8360,8 @@ static struct sk_buff *igb_build_skb(struct igb_ring *rx_ring, if (metasize) skb_metadata_set(skb, metasize); - /* pull timestamp out of packet data */ - if (igb_test_staterr(rx_desc, E1000_RXDADV_STAT_TSIP)) { - if (!igb_ptp_rx_pktstamp(rx_ring->q_vector, skb->data, skb)) - __skb_pull(skb, IGB_TS_HDR_LEN); - } + if (timestamp) + skb_hwtstamps(skb)->hwtstamp = timestamp; /* update buffer offset */ #if (PAGE_SIZE < 8192) @@ -8683,7 +8676,10 @@ static int igb_clean_rx_irq(struct igb_q_vector *q_vector, const int budget) while (likely(total_packets < budget)) { union e1000_adv_rx_desc *rx_desc; struct igb_rx_buffer *rx_buffer; + ktime_t timestamp = 0; + int pkt_offset = 0; unsigned int size; + void *pktbuf; /* return some buffers to hardware, one at a time is too slow */ if (cleaned_count >= IGB_RX_BUFFER_WRITE) { @@ -8703,15 +8699,21 @@ static int igb_clean_rx_irq(struct igb_q_vector *q_vector, const int budget) dma_rmb(); rx_buffer = igb_get_rx_buffer(rx_ring, size, &rx_buf_pgcnt); + pktbuf = page_address(rx_buffer->page) + rx_buffer->page_offset; + + if (igb_test_staterr(rx_desc, E1000_RXDADV_STAT_TSIP)) { + timestamp = igb_ptp_rx_pktstamp(rx_ring->q_vector, + pktbuf); + pkt_offset += IGB_TS_HDR_LEN; + size -= IGB_TS_HDR_LEN; + } /* retrieve a buffer from the ring */ if (!skb) { - unsigned int offset = igb_rx_offset(rx_ring); - unsigned char *hard_start; - - hard_start = page_address(rx_buffer->page) + - rx_buffer->page_offset - offset; - xdp_prepare_buff(&xdp, hard_start, offset, size, true); + xdp.data = pktbuf + pkt_offset; + xdp.data_end = xdp.data + size; + xdp.data_meta = xdp.data; + xdp.data_hard_start = pktbuf - igb_rx_offset(rx_ring); #if (PAGE_SIZE > 4096) /* At larger PAGE_SIZE, frame_sz depend on len size */ xdp.frame_sz = igb_rx_frame_truesize(rx_ring, size); @@ -8733,10 +8735,11 @@ static int igb_clean_rx_irq(struct igb_q_vector *q_vector, const int budget) } else if (skb) igb_add_rx_frag(rx_ring, rx_buffer, skb, size); else if (ring_uses_build_skb(rx_ring)) - skb = igb_build_skb(rx_ring, rx_buffer, &xdp, rx_desc); + skb = igb_build_skb(rx_ring, rx_buffer, &xdp, + timestamp); else skb = igb_construct_skb(rx_ring, rx_buffer, - &xdp, rx_desc); + &xdp, timestamp); /* exit if we failed to retrieve a buffer */ if (!skb) { diff --git a/drivers/net/ethernet/intel/igb/igb_ptp.c b/drivers/net/ethernet/intel/igb/igb_ptp.c index 86a576201f5f..0cbdf48285d3 100644 --- a/drivers/net/ethernet/intel/igb/igb_ptp.c +++ b/drivers/net/ethernet/intel/igb/igb_ptp.c @@ -863,23 +863,22 @@ static void igb_ptp_tx_hwtstamp(struct igb_adapter *adapter) * igb_ptp_rx_pktstamp - retrieve Rx per packet timestamp * @q_vector: Pointer to interrupt specific structure * @va: Pointer to address containing Rx buffer - * @skb: Buffer containing timestamp and packet * * This function is meant to retrieve a timestamp from the first buffer of an * incoming frame. The value is stored in little endian format starting on * byte 8 * - * Returns: 0 if success, nonzero if failure + * Returns: 0 on failure, timestamp on success **/ -int igb_ptp_rx_pktstamp(struct igb_q_vector *q_vector, void *va, - struct sk_buff *skb) +ktime_t igb_ptp_rx_pktstamp(struct igb_q_vector *q_vector, void *va) { struct igb_adapter *adapter = q_vector->adapter; + struct skb_shared_hwtstamps ts; __le64 *regval = (__le64 *)va; int adjust = 0; if (!(adapter->ptp_flags & IGB_PTP_ENABLED)) - return IGB_RET_PTP_DISABLED; + return 0; /* The timestamp is recorded in little endian format. * DWORD: 0 1 2 3 @@ -888,10 +887,9 @@ int igb_ptp_rx_pktstamp(struct igb_q_vector *q_vector, void *va, /* check reserved dwords are zero, be/le doesn't matter for zero */ if (regval[0]) - return IGB_RET_PTP_INVALID; + return 0; - igb_ptp_systim_to_hwtstamp(adapter, skb_hwtstamps(skb), - le64_to_cpu(regval[1])); + igb_ptp_systim_to_hwtstamp(adapter, &ts, le64_to_cpu(regval[1])); /* adjust timestamp for the RX latency based on link speed */ if (adapter->hw.mac.type == e1000_i210) { @@ -907,10 +905,8 @@ int igb_ptp_rx_pktstamp(struct igb_q_vector *q_vector, void *va, break; } } - skb_hwtstamps(skb)->hwtstamp = - ktime_sub_ns(skb_hwtstamps(skb)->hwtstamp, adjust); - return 0; + return ktime_sub_ns(ts.hwtstamp, adjust); } /**
When using native XDP with the igb driver, the XDP frame data doesn't point to the beginning of the packet. It's off by 16 bytes. Everything works as expected with XDP skb mode. Actually these 16 bytes are used to store the packet timestamps. Therefore, pull the timestamp before executing any XDP operations and adjust all other code accordingly. The igc driver does it like that as well. Tested with Intel i210 card and AF_XDP sockets. Fixes: 9cbc948b5a20 ("igb: add XDP support") Signed-off-by: Kurt Kanzenbach <kurt@linutronix.de> --- drivers/net/ethernet/intel/igb/igb.h | 3 +- drivers/net/ethernet/intel/igb/igb_main.c | 45 ++++++++++++----------- drivers/net/ethernet/intel/igb/igb_ptp.c | 18 ++++----- 3 files changed, 32 insertions(+), 34 deletions(-)