Re: [PATCH iwl-next,v1 1/1] igc: Add Tx hardware timestamp request for AF_XDP zero-copy packet
From: Vinicius Costa Gomes
Date: Tue Jan 02 2024 - 09:51:05 EST
Song Yoong Siang <yoong.siang.song@xxxxxxxxx> writes:
> This patch adds support to per-packet Tx hardware timestamp request to
> AF_XDP zero-copy packet via XDP Tx metadata framework. Please note that
> user needs to enable Tx HW timestamp capability via igc_ioctl() with
> SIOCSHWTSTAMP cmd before sending xsk Tx timestamp request.
>
> Same as implementation in RX timestamp XDP hints kfunc metadata, Timer 0
> (adjustable clock) is used in xsk Tx hardware timestamp. i225/i226 have
> four sets of timestamping registers. A pointer named "xsk_pending_ts"
> is introduced to indicate the timestamping register is already occupied.
> Furthermore, the mentioned pointer also being used to hold the transmit
> completion until the tx hardware timestamp is ready. This is because for
> i225/i226, the timestamp notification comes some time after the transmit
> completion event. The driver will retrigger hardware irq to clean the
> packet after retrieve the tx hardware timestamp.
>
> Besides, a pointer named "xsk_meta" is added into igc_tx_timestamp_request
> structure as a hook to the metadata location of the transmit packet. When
> a Tx timestamp interrupt happens, the interrupt handler will copy the
> value of Tx timestamp into metadata via xsk_tx_metadata_complete().
>
> This patch is tested with tools/testing/selftests/bpf/xdp_hw_metadata
> on Intel ADL-S platform. Below are the test steps and results.
>
> Command on DUT:
> sudo ./xdp_hw_metadata <interface name>
> sudo hwstamp_ctl -i <interface name> -t 1 -r 1
> sudo ./testptp -d /dev/ptp0 -s
>
> Command on Link Partner:
> echo -n xdp | nc -u -q1 <destination IPv4 addr> 9091
>
> Result:
> xsk_ring_cons__peek: 1
> 0x555b112ae958: rx_desc[6]->addr=86110 addr=86110 comp_addr=86110 EoP
> rx_hash: 0xBFDEC36E with RSS type:0x1
> HW RX-time: 1677762429190040955 (sec:1677762429.1900) delta to User RX-time sec:0.0001 (100.124 usec)
> XDP RX-time: 1677762429190123163 (sec:1677762429.1901) delta to User RX-time sec:0.0000 (17.916 usec)
> 0x555b112ae958: ping-pong with csum=404e (want c59e) csum_start=34 csum_offset=6
> 0x555b112ae958: complete tx idx=6 addr=6010
> HW TX-complete-time: 1677762429190173323 (sec:1677762429.1902) delta to User TX-complete-time sec:0.0100 (10035.884 usec)
> XDP RX-time: 1677762429190123163 (sec:1677762429.1901) delta to User TX-complete-time sec:0.0101 (10086.044 usec)
> HW RX-time: 1677762429190040955 (sec:1677762429.1900) delta to HW TX-complete-time sec:0.0001 (132.368 usec)
> 0x555b112ae958: complete rx idx=134 addr=86110
>
> Signed-off-by: Song Yoong Siang <yoong.siang.song@xxxxxxxxx>
> ---
> drivers/net/ethernet/intel/igc/igc.h | 15 ++++
> drivers/net/ethernet/intel/igc/igc_main.c | 88 ++++++++++++++++++++++-
> drivers/net/ethernet/intel/igc/igc_ptp.c | 42 ++++++++---
> 3 files changed, 134 insertions(+), 11 deletions(-)
>
> diff --git a/drivers/net/ethernet/intel/igc/igc.h b/drivers/net/ethernet/intel/igc/igc.h
> index ac7c861e83a0..c831dde01662 100644
> --- a/drivers/net/ethernet/intel/igc/igc.h
> +++ b/drivers/net/ethernet/intel/igc/igc.h
> @@ -79,6 +79,9 @@ struct igc_tx_timestamp_request {
> u32 regl; /* which TXSTMPL_{X} register should be used */
> u32 regh; /* which TXSTMPH_{X} register should be used */
> u32 flags; /* flags that should be added to the tx_buffer */
> + u8 xsk_queue_index; /* Tx queue which requesting timestamp */
> + bool *xsk_pending_ts; /* ref to tx ring for waiting timestamp event */
I think that this indirection level to xsk_pending_ts in the tx_buffer is a
bit too hard to follow. What I am thinking is keeping a pointer to
tx_buffer here in igc_tx_timestamp_request, perhaps even in a union with
the skb, and use a similar logic, if that pointer is valid the timestamp
request is in use.
Do you think it could work?
(Perhaps we would need to also store the buffer type in the request, but
I don't think that would be too weird)
> + struct xsk_tx_metadata_compl xsk_meta; /* ref to xsk Tx metadata */
> };
>
> struct igc_inline_rx_tstamps {
> @@ -319,6 +322,9 @@ void igc_disable_tx_ring(struct igc_ring *ring);
> void igc_enable_tx_ring(struct igc_ring *ring);
> int igc_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags);
>
> +/* AF_XDP TX metadata operations */
> +extern const struct xsk_tx_metadata_ops igc_xsk_tx_metadata_ops;
> +
> /* igc_dump declarations */
> void igc_rings_dump(struct igc_adapter *adapter);
> void igc_regs_dump(struct igc_adapter *adapter);
> @@ -528,6 +534,7 @@ struct igc_tx_buffer {
> DEFINE_DMA_UNMAP_ADDR(dma);
> DEFINE_DMA_UNMAP_LEN(len);
> u32 tx_flags;
> + bool xsk_pending_ts;
> };
>
> struct igc_rx_buffer {
> @@ -553,6 +560,14 @@ struct igc_xdp_buff {
> struct igc_inline_rx_tstamps *rx_ts; /* data indication bit IGC_RXDADV_STAT_TSIP */
> };
>
> +struct igc_metadata_request {
> + struct xsk_tx_metadata *meta;
> + struct igc_adapter *adapter;
If you have access to the tx_ring, you have access to the adapter, no
need to have it here.
> + struct igc_ring *tx_ring;
> + bool *xsk_pending_ts;
> + u32 *cmd_type;
I think this also would be clearer if here you had a pointer to the
tx_buffer instead of only 'xsk_pending_ts'.
I guess for cmd_type, no need for it to be a pointer, we can affort the
extra copy.
> +};
> +
> struct igc_q_vector {
> struct igc_adapter *adapter; /* backlink */
> void __iomem *itr_register;
> diff --git a/drivers/net/ethernet/intel/igc/igc_main.c b/drivers/net/ethernet/intel/igc/igc_main.c
> index 61db1d3bfa0b..311c85f2d82d 100644
> --- a/drivers/net/ethernet/intel/igc/igc_main.c
> +++ b/drivers/net/ethernet/intel/igc/igc_main.c
> @@ -1553,7 +1553,7 @@ static bool igc_request_tx_tstamp(struct igc_adapter *adapter, struct sk_buff *s
> for (i = 0; i < IGC_MAX_TX_TSTAMP_REGS; i++) {
> struct igc_tx_timestamp_request *tstamp = &adapter->tx_tstamp[i];
>
> - if (tstamp->skb)
> + if (tstamp->skb || tstamp->xsk_pending_ts)
> continue;
>
> tstamp->skb = skb_get(skb);
> @@ -2878,6 +2878,71 @@ static void igc_update_tx_stats(struct igc_q_vector *q_vector,
> q_vector->tx.total_packets += packets;
> }
>
> +static void igc_xsk_request_timestamp(void *_priv)
> +{
> + struct igc_metadata_request *meta_req = _priv;
> + struct igc_ring *tx_ring = meta_req->tx_ring;
> + struct igc_tx_timestamp_request *tstamp;
> + u32 *cmd_type = meta_req->cmd_type;
> + u32 tx_flags = IGC_TX_FLAGS_TSTAMP;
> + struct igc_adapter *adapter;
> + unsigned long lock_flags;
> + bool found = 0;
> + int i;
> +
> + if (test_bit(IGC_RING_FLAG_TX_HWTSTAMP, &tx_ring->flags)) {
> + adapter = meta_req->adapter;
> +
> + spin_lock_irqsave(&adapter->ptp_tx_lock, lock_flags);
> +
> + for (i = 0; i < IGC_MAX_TX_TSTAMP_REGS; i++) {
> + tstamp = &adapter->tx_tstamp[i];
> +
> + if (tstamp->skb || tstamp->xsk_pending_ts)
> + continue;
> +
> + found = 1;
nitpick: found is a bool, 'true' would read better.
> + break;
> + }
> +
> + if (!found) {
> + adapter->tx_hwtstamp_skipped++;
I think this is one those cases, that an early return or a goto would
make the code easier to understand.
> + } else {
> + tstamp->start = jiffies;
> + tstamp->xsk_queue_index = tx_ring->queue_index;
> +
> + tstamp->xsk_pending_ts = meta_req->xsk_pending_ts;
> + *tstamp->xsk_pending_ts = true;
> +
> + xsk_tx_metadata_to_compl(meta_req->meta,
> + &tstamp->xsk_meta);
> +
> + /* set timestamp bit based on the _TSTAMP(_X) bit. */
> + tx_flags |= tstamp->flags;
> + *cmd_type |= IGC_SET_FLAG(tx_flags, IGC_TX_FLAGS_TSTAMP,
> + (IGC_ADVTXD_MAC_TSTAMP));
> + *cmd_type |= IGC_SET_FLAG(tx_flags, IGC_TX_FLAGS_TSTAMP_1,
> + (IGC_ADVTXD_TSTAMP_REG_1));
> + *cmd_type |= IGC_SET_FLAG(tx_flags, IGC_TX_FLAGS_TSTAMP_2,
> + (IGC_ADVTXD_TSTAMP_REG_2));
> + *cmd_type |= IGC_SET_FLAG(tx_flags, IGC_TX_FLAGS_TSTAMP_3,
> + (IGC_ADVTXD_TSTAMP_REG_3));
> + }
> +
> + spin_unlock_irqrestore(&adapter->ptp_tx_lock, lock_flags);
> + }
> +}
> +
> +static u64 igc_xsk_fill_timestamp(void *_priv)
> +{
> + return *(u64 *)_priv;
> +}
> +
> +const struct xsk_tx_metadata_ops igc_xsk_tx_metadata_ops = {
> + .tmo_request_timestamp = igc_xsk_request_timestamp,
> + .tmo_fill_timestamp = igc_xsk_fill_timestamp,
> +};
> +
> static void igc_xdp_xmit_zc(struct igc_ring *ring)
> {
> struct xsk_buff_pool *pool = ring->xsk_pool;
> @@ -2899,6 +2964,8 @@ static void igc_xdp_xmit_zc(struct igc_ring *ring)
> budget = igc_desc_unused(ring);
>
> while (xsk_tx_peek_desc(pool, &xdp_desc) && budget--) {
> + struct igc_metadata_request meta_req;
> + struct xsk_tx_metadata *meta = NULL;
> u32 cmd_type, olinfo_status;
> struct igc_tx_buffer *bi;
> dma_addr_t dma;
> @@ -2909,14 +2976,23 @@ static void igc_xdp_xmit_zc(struct igc_ring *ring)
> olinfo_status = xdp_desc.len << IGC_ADVTXD_PAYLEN_SHIFT;
>
> dma = xsk_buff_raw_get_dma(pool, xdp_desc.addr);
> + meta = xsk_buff_get_metadata(pool, xdp_desc.addr);
> xsk_buff_raw_dma_sync_for_device(pool, dma, xdp_desc.len);
> + bi = &ring->tx_buffer_info[ntu];
> +
> + meta_req.adapter = netdev_priv(ring->netdev);
> + meta_req.tx_ring = ring;
> + meta_req.meta = meta;
> + meta_req.cmd_type = &cmd_type;
> + meta_req.xsk_pending_ts = &bi->xsk_pending_ts;
> + xsk_tx_metadata_request(meta, &igc_xsk_tx_metadata_ops,
> + &meta_req);
>
> tx_desc = IGC_TX_DESC(ring, ntu);
> tx_desc->read.cmd_type_len = cpu_to_le32(cmd_type);
> tx_desc->read.olinfo_status = cpu_to_le32(olinfo_status);
> tx_desc->read.buffer_addr = cpu_to_le64(dma);
>
> - bi = &ring->tx_buffer_info[ntu];
> bi->type = IGC_TX_BUFFER_TYPE_XSK;
> bi->protocol = 0;
> bi->bytecount = xdp_desc.len;
> @@ -2979,6 +3055,13 @@ static bool igc_clean_tx_irq(struct igc_q_vector *q_vector, int napi_budget)
> if (!(eop_desc->wb.status & cpu_to_le32(IGC_TXD_STAT_DD)))
> break;
>
> + /* Hold the completions while there's a pending tx hardware
> + * timestamp request from XDP Tx metadata.
> + */
> + if (tx_buffer->type == IGC_TX_BUFFER_TYPE_XSK &&
> + tx_buffer->xsk_pending_ts)
> + break;
> +
One scenario that I am worried about the completion part is when tstamp
and non-tstamp packets are mixed in the same queue.
For example, when the user sends a 1 tstamp packet followed by 1
non-tstamp packet. Some other ratios might be interesting to test as
well, 1:10 for example. I guess a simple bandwith test would be enough,
comparing "non-tstamp only" with mixed traffic.
Perhaps are some bad recollections from the past, but I remember that
the hardware takes a bit of time when generating the timestamp
interrupts, and so those types of mixed traffic would have wasted
bandwidth.
> /* clear next_to_watch to prevent false hangs */
> tx_buffer->next_to_watch = NULL;
>
> @@ -6819,6 +6902,7 @@ static int igc_probe(struct pci_dev *pdev,
>
> netdev->netdev_ops = &igc_netdev_ops;
> netdev->xdp_metadata_ops = &igc_xdp_metadata_ops;
> + netdev->xsk_tx_metadata_ops = &igc_xsk_tx_metadata_ops;
> igc_ethtool_set_ops(netdev);
> netdev->watchdog_timeo = 5 * HZ;
>
> diff --git a/drivers/net/ethernet/intel/igc/igc_ptp.c b/drivers/net/ethernet/intel/igc/igc_ptp.c
> index 885faaa7b9de..b722bca40309 100644
> --- a/drivers/net/ethernet/intel/igc/igc_ptp.c
> +++ b/drivers/net/ethernet/intel/igc/igc_ptp.c
> @@ -11,6 +11,7 @@
> #include <linux/ktime.h>
> #include <linux/delay.h>
> #include <linux/iopoll.h>
> +#include <net/xdp_sock.h>
>
> #define INCVALUE_MASK 0x7fffffff
> #define ISGN 0x80000000
> @@ -555,8 +556,15 @@ static void igc_ptp_clear_tx_tstamp(struct igc_adapter *adapter)
> for (i = 0; i < IGC_MAX_TX_TSTAMP_REGS; i++) {
> struct igc_tx_timestamp_request *tstamp = &adapter->tx_tstamp[i];
>
> - dev_kfree_skb_any(tstamp->skb);
> - tstamp->skb = NULL;
> + if (tstamp->skb) {
> + dev_kfree_skb_any(tstamp->skb);
> + tstamp->skb = NULL;
> + } else if (tstamp->xsk_pending_ts) {
> + *tstamp->xsk_pending_ts = false;
> + tstamp->xsk_pending_ts = NULL;
> + igc_xsk_wakeup(adapter->netdev, tstamp->xsk_queue_index,
> + 0);
> + }
> }
>
> spin_unlock_irqrestore(&adapter->ptp_tx_lock, flags);
> @@ -657,8 +665,15 @@ static int igc_ptp_set_timestamp_mode(struct igc_adapter *adapter,
> static void igc_ptp_tx_timeout(struct igc_adapter *adapter,
> struct igc_tx_timestamp_request *tstamp)
> {
> - dev_kfree_skb_any(tstamp->skb);
> - tstamp->skb = NULL;
> + if (tstamp->skb) {
> + dev_kfree_skb_any(tstamp->skb);
> + tstamp->skb = NULL;
> + } else if (tstamp->xsk_pending_ts) {
> + *tstamp->xsk_pending_ts = false;
> + tstamp->xsk_pending_ts = NULL;
> + igc_xsk_wakeup(adapter->netdev, tstamp->xsk_queue_index, 0);
> + }
> +
> adapter->tx_hwtstamp_timeouts++;
>
> netdev_warn(adapter->netdev, "Tx timestamp timeout\n");
> @@ -677,7 +692,7 @@ void igc_ptp_tx_hang(struct igc_adapter *adapter)
> for (i = 0; i < IGC_MAX_TX_TSTAMP_REGS; i++) {
> tstamp = &adapter->tx_tstamp[i];
>
> - if (!tstamp->skb)
> + if (!tstamp->skb && !tstamp->xsk_pending_ts)
> continue;
>
> if (time_is_after_jiffies(tstamp->start + IGC_PTP_TX_TIMEOUT))
> @@ -705,7 +720,7 @@ static void igc_ptp_tx_reg_to_stamp(struct igc_adapter *adapter,
> int adjust = 0;
>
> skb = tstamp->skb;
> - if (!skb)
> + if (!skb && !tstamp->xsk_pending_ts)
> return;
>
> if (igc_ptp_systim_to_hwtstamp(adapter, &shhwtstamps, regval))
> @@ -729,10 +744,19 @@ static void igc_ptp_tx_reg_to_stamp(struct igc_adapter *adapter,
> shhwtstamps.hwtstamp =
> ktime_add_ns(shhwtstamps.hwtstamp, adjust);
>
> - tstamp->skb = NULL;
> + if (skb) {
> + tstamp->skb = NULL;
> + skb_tstamp_tx(skb, &shhwtstamps);
> + dev_kfree_skb_any(skb);
> + } else {
> + xsk_tx_metadata_complete(&tstamp->xsk_meta,
> + &igc_xsk_tx_metadata_ops,
> + &shhwtstamps.hwtstamp);
>
> - skb_tstamp_tx(skb, &shhwtstamps);
> - dev_kfree_skb_any(skb);
> + *tstamp->xsk_pending_ts = false;
> + tstamp->xsk_pending_ts = NULL;
> + igc_xsk_wakeup(adapter->netdev, tstamp->xsk_queue_index, 0);
> + }
> }
>
> /**
> --
> 2.34.1
>
--
Vinicius