Re: [RESEND net-next 1/2] net: hns3: add support for PTP

On 2021/6/3 21:22, Richard Cochran wrote:

On Wed, Jun 02, 2021 at 10:57:43AM +0800, Guangbin Huang wrote:

quoted

diff --git a/drivers/net/ethernet/hisilicon/hns3/hns3pf/hclge_ptp.c b/drivers/net/ethernet/hisilicon/hns3/hns3pf/hclge_ptp.c
new file mode 100644
index 000000000000..b133b5984584
--- /dev/null
+++ b/drivers/net/ethernet/hisilicon/hns3/hns3pf/hclge_ptp.c

@@ -0,0 +1,520 @@
+// SPDX-License-Identifier: GPL-2.0+
+// Copyright (c) 2021 Hisilicon Limited.
+
+#include <linux/skbuff.h>
+#include "hclge_main.h"
+#include "hnae3.h"
+
+static int hclge_ptp_adjfreq(struct ptp_clock_info *ptp, s32 ppb)
+{
+	struct hclge_dev *hdev = hclge_ptp_get_hdev(ptp);
+	u64 adj_val, adj_base, diff;
+	bool is_neg = false;
+	u32 quo, numerator;
+
+	if (ppb < 0) {
+		ppb = -ppb;
+		is_neg = true;
+	}
+
+	adj_base = HCLGE_PTP_CYCLE_ADJ_BASE * HCLGE_PTP_CYCLE_ADJ_UNIT;
+	adj_val = adj_base * ppb;
+	diff = div_u64(adj_val, 1000000000ULL);
+
+	if (is_neg)
+		adj_val = adj_base - diff;
+	else
+		adj_val = adj_base + diff;
+
+	/* This clock cycle is defined by three part: quotient, numerator
+	 * and denominator. For example, 2.5ns, the quotient is 2,
+	 * denominator is fixed to HCLGE_PTP_CYCLE_ADJ_UNIT, and numerator
+	 * is 0.5 * HCLGE_PTP_CYCLE_ADJ_UNIT.
+	 */
+	quo = div_u64_rem(adj_val, HCLGE_PTP_CYCLE_ADJ_UNIT, &numerator);
+	writel(quo, hdev->ptp->io_base + HCLGE_PTP_CYCLE_QUO_REG);
+	writel(numerator, hdev->ptp->io_base + HCLGE_PTP_CYCLE_NUM_REG);
+	writel(HCLGE_PTP_CYCLE_ADJ_UNIT,
+	       hdev->ptp->io_base + HCLGE_PTP_CYCLE_DEN_REG);
+	writel(HCLGE_PTP_CYCLE_ADJ_EN,
+	       hdev->ptp->io_base + HCLGE_PTP_CYCLE_CFG_REG);

Need mutex or spinlock to protest against concurrent access.

Ok, thank you. Is it better to add mutex or spinlock in the public place, like ptp_clock_adjtime()?
Then there's no need to add mutex or spinlock in all the drives.

quoted

+
+	return 0;
+}
+
+bool hclge_ptp_set_tx_info(struct hnae3_handle *handle, struct sk_buff *skb)
+{
+	struct hclge_vport *vport = hclge_get_vport(handle);
+	struct hclge_dev *hdev = vport->back;
+	struct hclge_ptp *ptp = hdev->ptp;
+
+	if (!test_bit(HCLGE_PTP_FLAG_TX_EN, &ptp->flags) ||
+	    test_and_set_bit(HCLGE_STATE_PTP_TX_HANDLING, &hdev->state)) {
+		ptp->tx_skipped++;
+		return false;
+	}
+
+	ptp->tx_start = jiffies;
+	ptp->tx_skb = skb_get(skb);
+	ptp->tx_cnt++;
+
+	return true;
+}
+
+void hclge_ptp_clean_tx_hwts(struct hclge_dev *hdev)
+{
+	struct sk_buff *skb = hdev->ptp->tx_skb;
+	struct skb_shared_hwtstamps hwts;
+	u32 hi, lo;
+	u64 ns;
+
+	ns = readl(hdev->ptp->io_base + HCLGE_PTP_TX_TS_NSEC_REG) &
+	     HCLGE_PTP_TX_TS_NSEC_MASK;
+	lo = readl(hdev->ptp->io_base + HCLGE_PTP_TX_TS_SEC_L_REG);
+	hi = readl(hdev->ptp->io_base + HCLGE_PTP_TX_TS_SEC_H_REG) &
+	     HCLGE_PTP_TX_TS_SEC_H_MASK;
+	hdev->ptp->last_tx_seqid = readl(hdev->ptp->io_base +
+		HCLGE_PTP_TX_TS_SEQID_REG);
+
+	if (skb) {
+		hdev->ptp->tx_skb = NULL;
+		hdev->ptp->tx_cleaned++;
+
+		ns += (((u64)hi) << 32 | lo) * NSEC_PER_SEC;
+		hwts.hwtstamp = ns_to_ktime(ns);
+		skb_tstamp_tx(skb, &hwts);
+		dev_kfree_skb_any(skb);
+	}
+
+	clear_bit(HCLGE_STATE_PTP_TX_HANDLING, &hdev->state);
+}
+
+void hclge_ptp_get_rx_hwts(struct hnae3_handle *handle, struct sk_buff *skb,
+			   u32 nsec, u32 sec)
+{
+	struct hclge_vport *vport = hclge_get_vport(handle);
+	struct hclge_dev *hdev = vport->back;
+	u64 ns = nsec;
+	u32 sec_h;
+
+	if (!test_bit(HCLGE_PTP_FLAG_RX_EN, &hdev->ptp->flags))
+		return;
+
+	/* Since the BD does not have enough space for the higher 16 bits of
+	 * second, and this part will not change frequently, so read it
+	 * from register.
+	 */
+	sec_h = readl(hdev->ptp->io_base + HCLGE_PTP_CUR_TIME_SEC_H_REG);

Need mutex or spinlock to protest against concurrent acces. >

Same above.

quoted

+	ns += (((u64)sec_h) << HCLGE_PTP_SEC_H_OFFSET | sec) * NSEC_PER_SEC;
+	skb_hwtstamps(skb)->hwtstamp = ns_to_ktime(ns);
+	hdev->ptp->last_rx = jiffies;
+	hdev->ptp->rx_cnt++;
+}
+
+static int hclge_ptp_gettimex(struct ptp_clock_info *ptp, struct timespec64 *ts,
+			      struct ptp_system_timestamp *sts)
+{
+	struct hclge_dev *hdev = hclge_ptp_get_hdev(ptp);
+	u32 hi, lo;
+	u64 ns;
+
+	ns = readl(hdev->ptp->io_base + HCLGE_PTP_CUR_TIME_NSEC_REG);
+	hi = readl(hdev->ptp->io_base + HCLGE_PTP_CUR_TIME_SEC_H_REG);
+	lo = readl(hdev->ptp->io_base + HCLGE_PTP_CUR_TIME_SEC_L_REG);

Need mutex or spinlock to protest against concurrent access.

Same above.

quoted

+	ns += (((u64)hi) << HCLGE_PTP_SEC_H_OFFSET | lo) * NSEC_PER_SEC;
+	*ts = ns_to_timespec64(ns);
+
+	return 0;
+}
+
+static int hclge_ptp_settime(struct ptp_clock_info *ptp,
+			     const struct timespec64 *ts)
+{
+	struct hclge_dev *hdev = hclge_ptp_get_hdev(ptp);
+
+	writel(ts->tv_nsec, hdev->ptp->io_base + HCLGE_PTP_TIME_NSEC_REG);
+	writel(ts->tv_sec >> HCLGE_PTP_SEC_H_OFFSET,
+	       hdev->ptp->io_base + HCLGE_PTP_TIME_SEC_H_REG);
+	writel(ts->tv_sec & HCLGE_PTP_SEC_L_MASK,
+	       hdev->ptp->io_base + HCLGE_PTP_TIME_SEC_L_REG);
+	/* synchronize the time of phc */
+	writel(HCLGE_PTP_TIME_SYNC_EN,
+	       hdev->ptp->io_base + HCLGE_PTP_TIME_SYNC_REG);

Need mutex or spinlock to protest against concurrent access.

Same above.

quoted

+	return 0;
+}
+
+static int hclge_ptp_adjtime(struct ptp_clock_info *ptp, s64 delta)
+{
+	struct hclge_dev *hdev = hclge_ptp_get_hdev(ptp);
+	bool is_neg = false;
+	u32 adj_val = 0;
+
+	if (delta < 0) {
+		adj_val |= HCLGE_PTP_TIME_NSEC_NEG;
+		delta = -delta;
+		is_neg = true;
+	}
+
+	if (delta > HCLGE_PTP_TIME_NSEC_MASK) {
+		struct timespec64 ts;
+		s64 ns;
+
+		hclge_ptp_gettimex(ptp, &ts, NULL);
+		ns = timespec64_to_ns(&ts);
+		ns = is_neg ? ns - delta : ns + delta;
+		ts = ns_to_timespec64(ns);
+		return hclge_ptp_settime(ptp, &ts);
+	}
+
+	adj_val |= delta & HCLGE_PTP_TIME_NSEC_MASK;
+	writel(adj_val, hdev->ptp->io_base + HCLGE_PTP_TIME_NSEC_REG);
+	writel(HCLGE_PTP_TIME_ADJ_EN,
+	       hdev->ptp->io_base + HCLGE_PTP_TIME_ADJ_REG);

Need mutex or spinlock to protest against concurrent access.

Same above.

quoted

+
+	return 0;
+}

Thanks,
Richard

.