]> Pileus Git - ~andy/linux/blobdiff - drivers/net/ethernet/sfc/ptp.c
Merge branch 'for-3.14-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq
[~andy/linux] / drivers / net / ethernet / sfc / ptp.c
index 3dd39dcfe36b92f3221119171b1e0c26729d85e8..d7a36829649a7eea1797057e0bb0e5773bc83c0e 100644 (file)
@@ -62,7 +62,7 @@
 #define        SYNCHRONISATION_GRANULARITY_NS  200
 
 /* Minimum permitted length of a (corrected) synchronisation time */
-#define        MIN_SYNCHRONISATION_NS          120
+#define        DEFAULT_MIN_SYNCHRONISATION_NS  120
 
 /* Maximum permitted length of a (corrected) synchronisation time */
 #define        MAX_SYNCHRONISATION_NS          1000
@@ -195,26 +195,29 @@ struct efx_ptp_event_rx {
 /**
  * struct efx_ptp_timeset - Synchronisation between host and MC
  * @host_start: Host time immediately before hardware timestamp taken
- * @seconds: Hardware timestamp, seconds
- * @nanoseconds: Hardware timestamp, nanoseconds
+ * @major: Hardware timestamp, major
+ * @minor: Hardware timestamp, minor
  * @host_end: Host time immediately after hardware timestamp taken
- * @waitns: Number of nanoseconds between hardware timestamp being read and
+ * @wait: Number of NIC clock ticks between hardware timestamp being read and
  *          host end time being seen
  * @window: Difference of host_end and host_start
  * @valid: Whether this timeset is valid
  */
 struct efx_ptp_timeset {
        u32 host_start;
-       u32 seconds;
-       u32 nanoseconds;
+       u32 major;
+       u32 minor;
        u32 host_end;
-       u32 waitns;
+       u32 wait;
        u32 window;     /* Derived: end - start, allowing for wrap */
 };
 
 /**
  * struct efx_ptp_data - Precision Time Protocol (PTP) state
- * @channel: The PTP channel
+ * @efx: The NIC context
+ * @channel: The PTP channel (Siena only)
+ * @rx_ts_inline: Flag for whether RX timestamps are inline (else they are
+ *     separate events)
  * @rxq: Receive queue (awaiting timestamps)
  * @txq: Transmit queue
  * @evt_list: List of MC receive events awaiting packets
@@ -231,41 +234,42 @@ struct efx_ptp_timeset {
  * @config: Current timestamp configuration
  * @enabled: PTP operation enabled
  * @mode: Mode in which PTP operating (PTP version)
+ * @time_format: Time format supported by this NIC
+ * @ns_to_nic_time: Function to convert from scalar nanoseconds to NIC time
+ * @nic_to_kernel_time: Function to convert from NIC to kernel time
+ * @min_synchronisation_ns: Minimum acceptable corrected sync window
+ * @ts_corrections.tx: Required driver correction of transmit timestamps
+ * @ts_corrections.rx: Required driver correction of receive timestamps
+ * @ts_corrections.pps_out: PPS output error (information only)
+ * @ts_corrections.pps_in: Required driver correction of PPS input timestamps
  * @evt_frags: Partly assembled PTP events
  * @evt_frag_idx: Current fragment number
  * @evt_code: Last event code
  * @start: Address at which MC indicates ready for synchronisation
  * @host_time_pps: Host time at last PPS
- * @last_sync_ns: Last number of nanoseconds between readings when synchronising
- * @base_sync_ns: Number of nanoseconds for last synchronisation.
- * @base_sync_valid: Whether base_sync_time is valid.
  * @current_adjfreq: Current ppb adjustment.
- * @phc_clock: Pointer to registered phc device
+ * @phc_clock: Pointer to registered phc device (if primary function)
  * @phc_clock_info: Registration structure for phc device
  * @pps_work: pps work task for handling pps events
  * @pps_workwq: pps work queue
  * @nic_ts_enabled: Flag indicating if NIC generated TS events are handled
  * @txbuf: Buffer for use when transmitting (PTP) packets to MC (avoids
  *         allocations in main data path).
- * @debug_ptp_dir: PTP debugfs directory
- * @missed_rx_sync: Number of packets received without syncrhonisation.
  * @good_syncs: Number of successful synchronisations.
- * @no_time_syncs: Number of synchronisations with no good times.
- * @bad_sync_durations: Number of synchronisations with bad durations.
+ * @fast_syncs: Number of synchronisations requiring short delay
  * @bad_syncs: Number of failed synchronisations.
- * @last_sync_time: Number of nanoseconds for last synchronisation.
  * @sync_timeouts: Number of synchronisation timeouts
- * @fast_syncs: Number of synchronisations requiring short delay
- * @min_sync_delta: Minimum time between event and synchronisation
- * @max_sync_delta: Maximum time between event and synchronisation
- * @average_sync_delta: Average time between event and synchronisation.
- *                      Modified moving average.
- * @last_sync_delta: Last time between event and synchronisation
- * @mc_stats: Context value for MC statistics
+ * @no_time_syncs: Number of synchronisations with no good times.
+ * @invalid_sync_windows: Number of sync windows with bad durations.
+ * @undersize_sync_windows: Number of corrected sync windows that are too small
+ * @oversize_sync_windows: Number of corrected sync windows that are too large
+ * @rx_no_timestamp: Number of packets received without a timestamp.
  * @timeset: Last set of synchronisation statistics.
  */
 struct efx_ptp_data {
+       struct efx_nic *efx;
        struct efx_channel *channel;
+       bool rx_ts_inline;
        struct sk_buff_head rxq;
        struct sk_buff_head txq;
        struct list_head evt_list;
@@ -282,14 +286,22 @@ struct efx_ptp_data {
        struct hwtstamp_config config;
        bool enabled;
        unsigned int mode;
+       unsigned int time_format;
+       void (*ns_to_nic_time)(s64 ns, u32 *nic_major, u32 *nic_minor);
+       ktime_t (*nic_to_kernel_time)(u32 nic_major, u32 nic_minor,
+                                     s32 correction);
+       unsigned int min_synchronisation_ns;
+       struct {
+               s32 tx;
+               s32 rx;
+               s32 pps_out;
+               s32 pps_in;
+       } ts_corrections;
        efx_qword_t evt_frags[MAX_EVENT_FRAGS];
        int evt_frag_idx;
        int evt_code;
        struct efx_buffer start;
        struct pps_event_time host_time_pps;
-       unsigned last_sync_ns;
-       unsigned base_sync_ns;
-       bool base_sync_valid;
        s64 current_adjfreq;
        struct ptp_clock *phc_clock;
        struct ptp_clock_info phc_clock_info;
@@ -297,6 +309,16 @@ struct efx_ptp_data {
        struct workqueue_struct *pps_workwq;
        bool nic_ts_enabled;
        MCDI_DECLARE_BUF(txbuf, MC_CMD_PTP_IN_TRANSMIT_LENMAX);
+
+       unsigned int good_syncs;
+       unsigned int fast_syncs;
+       unsigned int bad_syncs;
+       unsigned int sync_timeouts;
+       unsigned int no_time_syncs;
+       unsigned int invalid_sync_windows;
+       unsigned int undersize_sync_windows;
+       unsigned int oversize_sync_windows;
+       unsigned int rx_no_timestamp;
        struct efx_ptp_timeset
        timeset[MC_CMD_PTP_OUT_SYNCHRONIZE_TIMESET_MAXNUM];
 };
@@ -309,19 +331,263 @@ static int efx_phc_settime(struct ptp_clock_info *ptp,
 static int efx_phc_enable(struct ptp_clock_info *ptp,
                          struct ptp_clock_request *request, int on);
 
+#define PTP_SW_STAT(ext_name, field_name)                              \
+       { #ext_name, 0, offsetof(struct efx_ptp_data, field_name) }
+#define PTP_MC_STAT(ext_name, mcdi_name)                               \
+       { #ext_name, 32, MC_CMD_PTP_OUT_STATUS_STATS_ ## mcdi_name ## _OFST }
+static const struct efx_hw_stat_desc efx_ptp_stat_desc[] = {
+       PTP_SW_STAT(ptp_good_syncs, good_syncs),
+       PTP_SW_STAT(ptp_fast_syncs, fast_syncs),
+       PTP_SW_STAT(ptp_bad_syncs, bad_syncs),
+       PTP_SW_STAT(ptp_sync_timeouts, sync_timeouts),
+       PTP_SW_STAT(ptp_no_time_syncs, no_time_syncs),
+       PTP_SW_STAT(ptp_invalid_sync_windows, invalid_sync_windows),
+       PTP_SW_STAT(ptp_undersize_sync_windows, undersize_sync_windows),
+       PTP_SW_STAT(ptp_oversize_sync_windows, oversize_sync_windows),
+       PTP_SW_STAT(ptp_rx_no_timestamp, rx_no_timestamp),
+       PTP_MC_STAT(ptp_tx_timestamp_packets, TX),
+       PTP_MC_STAT(ptp_rx_timestamp_packets, RX),
+       PTP_MC_STAT(ptp_timestamp_packets, TS),
+       PTP_MC_STAT(ptp_filter_matches, FM),
+       PTP_MC_STAT(ptp_non_filter_matches, NFM),
+};
+#define PTP_STAT_COUNT ARRAY_SIZE(efx_ptp_stat_desc)
+static const unsigned long efx_ptp_stat_mask[] = {
+       [0 ... BITS_TO_LONGS(PTP_STAT_COUNT) - 1] = ~0UL,
+};
+
+size_t efx_ptp_describe_stats(struct efx_nic *efx, u8 *strings)
+{
+       if (!efx->ptp_data)
+               return 0;
+
+       return efx_nic_describe_stats(efx_ptp_stat_desc, PTP_STAT_COUNT,
+                                     efx_ptp_stat_mask, strings);
+}
+
+size_t efx_ptp_update_stats(struct efx_nic *efx, u64 *stats)
+{
+       MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_STATUS_LEN);
+       MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_STATUS_LEN);
+       size_t i;
+       int rc;
+
+       if (!efx->ptp_data)
+               return 0;
+
+       /* Copy software statistics */
+       for (i = 0; i < PTP_STAT_COUNT; i++) {
+               if (efx_ptp_stat_desc[i].dma_width)
+                       continue;
+               stats[i] = *(unsigned int *)((char *)efx->ptp_data +
+                                            efx_ptp_stat_desc[i].offset);
+       }
+
+       /* Fetch MC statistics.  We *must* fill in all statistics or
+        * risk leaking kernel memory to userland, so if the MCDI
+        * request fails we pretend we got zeroes.
+        */
+       MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_STATUS);
+       MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
+       rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
+                         outbuf, sizeof(outbuf), NULL);
+       if (rc) {
+               netif_err(efx, hw, efx->net_dev,
+                         "MC_CMD_PTP_OP_STATUS failed (%d)\n", rc);
+               memset(outbuf, 0, sizeof(outbuf));
+       }
+       efx_nic_update_stats(efx_ptp_stat_desc, PTP_STAT_COUNT,
+                            efx_ptp_stat_mask,
+                            stats, _MCDI_PTR(outbuf, 0), false);
+
+       return PTP_STAT_COUNT;
+}
+
+/* For Siena platforms NIC time is s and ns */
+static void efx_ptp_ns_to_s_ns(s64 ns, u32 *nic_major, u32 *nic_minor)
+{
+       struct timespec ts = ns_to_timespec(ns);
+       *nic_major = ts.tv_sec;
+       *nic_minor = ts.tv_nsec;
+}
+
+static ktime_t efx_ptp_s_ns_to_ktime_correction(u32 nic_major, u32 nic_minor,
+                                               s32 correction)
+{
+       ktime_t kt = ktime_set(nic_major, nic_minor);
+       if (correction >= 0)
+               kt = ktime_add_ns(kt, (u64)correction);
+       else
+               kt = ktime_sub_ns(kt, (u64)-correction);
+       return kt;
+}
+
+/* To convert from s27 format to ns we multiply then divide by a power of 2.
+ * For the conversion from ns to s27, the operation is also converted to a
+ * multiply and shift.
+ */
+#define S27_TO_NS_SHIFT        (27)
+#define NS_TO_S27_MULT (((1ULL << 63) + NSEC_PER_SEC / 2) / NSEC_PER_SEC)
+#define NS_TO_S27_SHIFT        (63 - S27_TO_NS_SHIFT)
+#define S27_MINOR_MAX  (1 << S27_TO_NS_SHIFT)
+
+/* For Huntington platforms NIC time is in seconds and fractions of a second
+ * where the minor register only uses 27 bits in units of 2^-27s.
+ */
+static void efx_ptp_ns_to_s27(s64 ns, u32 *nic_major, u32 *nic_minor)
+{
+       struct timespec ts = ns_to_timespec(ns);
+       u32 maj = ts.tv_sec;
+       u32 min = (u32)(((u64)ts.tv_nsec * NS_TO_S27_MULT +
+                        (1ULL << (NS_TO_S27_SHIFT - 1))) >> NS_TO_S27_SHIFT);
+
+       /* The conversion can result in the minor value exceeding the maximum.
+        * In this case, round up to the next second.
+        */
+       if (min >= S27_MINOR_MAX) {
+               min -= S27_MINOR_MAX;
+               maj++;
+       }
+
+       *nic_major = maj;
+       *nic_minor = min;
+}
+
+static inline ktime_t efx_ptp_s27_to_ktime(u32 nic_major, u32 nic_minor)
+{
+       u32 ns = (u32)(((u64)nic_minor * NSEC_PER_SEC +
+                       (1ULL << (S27_TO_NS_SHIFT - 1))) >> S27_TO_NS_SHIFT);
+       return ktime_set(nic_major, ns);
+}
+
+static ktime_t efx_ptp_s27_to_ktime_correction(u32 nic_major, u32 nic_minor,
+                                              s32 correction)
+{
+       /* Apply the correction and deal with carry */
+       nic_minor += correction;
+       if ((s32)nic_minor < 0) {
+               nic_minor += S27_MINOR_MAX;
+               nic_major--;
+       } else if (nic_minor >= S27_MINOR_MAX) {
+               nic_minor -= S27_MINOR_MAX;
+               nic_major++;
+       }
+
+       return efx_ptp_s27_to_ktime(nic_major, nic_minor);
+}
+
+/* Get PTP attributes and set up time conversions */
+static int efx_ptp_get_attributes(struct efx_nic *efx)
+{
+       MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_ATTRIBUTES_LEN);
+       MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN);
+       struct efx_ptp_data *ptp = efx->ptp_data;
+       int rc;
+       u32 fmt;
+       size_t out_len;
+
+       /* Get the PTP attributes. If the NIC doesn't support the operation we
+        * use the default format for compatibility with older NICs i.e.
+        * seconds and nanoseconds.
+        */
+       MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_GET_ATTRIBUTES);
+       MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
+       rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
+                         outbuf, sizeof(outbuf), &out_len);
+       if (rc == 0)
+               fmt = MCDI_DWORD(outbuf, PTP_OUT_GET_ATTRIBUTES_TIME_FORMAT);
+       else if (rc == -EINVAL)
+               fmt = MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS;
+       else
+               return rc;
+
+       if (fmt == MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_27FRACTION) {
+               ptp->ns_to_nic_time = efx_ptp_ns_to_s27;
+               ptp->nic_to_kernel_time = efx_ptp_s27_to_ktime_correction;
+       } else if (fmt == MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS) {
+               ptp->ns_to_nic_time = efx_ptp_ns_to_s_ns;
+               ptp->nic_to_kernel_time = efx_ptp_s_ns_to_ktime_correction;
+       } else {
+               return -ERANGE;
+       }
+
+       ptp->time_format = fmt;
+
+       /* MC_CMD_PTP_OP_GET_ATTRIBUTES is an extended version of an older
+        * operation MC_CMD_PTP_OP_GET_TIME_FORMAT that also returns a value
+        * to use for the minimum acceptable corrected synchronization window.
+        * If we have the extra information store it. For older firmware that
+        * does not implement the extended command use the default value.
+        */
+       if (rc == 0 && out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN)
+               ptp->min_synchronisation_ns =
+                       MCDI_DWORD(outbuf,
+                                  PTP_OUT_GET_ATTRIBUTES_SYNC_WINDOW_MIN);
+       else
+               ptp->min_synchronisation_ns = DEFAULT_MIN_SYNCHRONISATION_NS;
+
+       return 0;
+}
+
+/* Get PTP timestamp corrections */
+static int efx_ptp_get_timestamp_corrections(struct efx_nic *efx)
+{
+       MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_TIMESTAMP_CORRECTIONS_LEN);
+       MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_LEN);
+       int rc;
+
+       /* Get the timestamp corrections from the NIC. If this operation is
+        * not supported (older NICs) then no correction is required.
+        */
+       MCDI_SET_DWORD(inbuf, PTP_IN_OP,
+                      MC_CMD_PTP_OP_GET_TIMESTAMP_CORRECTIONS);
+       MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
+
+       rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
+                         outbuf, sizeof(outbuf), NULL);
+       if (rc == 0) {
+               efx->ptp_data->ts_corrections.tx = MCDI_DWORD(outbuf,
+                       PTP_OUT_GET_TIMESTAMP_CORRECTIONS_TRANSMIT);
+               efx->ptp_data->ts_corrections.rx = MCDI_DWORD(outbuf,
+                       PTP_OUT_GET_TIMESTAMP_CORRECTIONS_RECEIVE);
+               efx->ptp_data->ts_corrections.pps_out = MCDI_DWORD(outbuf,
+                       PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_OUT);
+               efx->ptp_data->ts_corrections.pps_in = MCDI_DWORD(outbuf,
+                       PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_IN);
+       } else if (rc == -EINVAL) {
+               efx->ptp_data->ts_corrections.tx = 0;
+               efx->ptp_data->ts_corrections.rx = 0;
+               efx->ptp_data->ts_corrections.pps_out = 0;
+               efx->ptp_data->ts_corrections.pps_in = 0;
+       } else {
+               return rc;
+       }
+
+       return 0;
+}
+
 /* Enable MCDI PTP support. */
 static int efx_ptp_enable(struct efx_nic *efx)
 {
        MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ENABLE_LEN);
+       MCDI_DECLARE_BUF_OUT_OR_ERR(outbuf, 0);
+       int rc;
 
        MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ENABLE);
        MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
        MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_QUEUE,
-                      efx->ptp_data->channel->channel);
+                      efx->ptp_data->channel ?
+                      efx->ptp_data->channel->channel : 0);
        MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_MODE, efx->ptp_data->mode);
 
-       return efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
-                           NULL, 0, NULL);
+       rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
+                               outbuf, sizeof(outbuf), NULL);
+       rc = (rc == -EALREADY) ? 0 : rc;
+       if (rc)
+               efx_mcdi_display_error(efx, MC_CMD_PTP,
+                                      MC_CMD_PTP_IN_ENABLE_LEN,
+                                      outbuf, sizeof(outbuf), rc);
+       return rc;
 }
 
 /* Disable MCDI PTP support.
@@ -332,11 +598,19 @@ static int efx_ptp_enable(struct efx_nic *efx)
 static int efx_ptp_disable(struct efx_nic *efx)
 {
        MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_DISABLE_LEN);
+       MCDI_DECLARE_BUF_OUT_OR_ERR(outbuf, 0);
+       int rc;
 
        MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_DISABLE);
        MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
-       return efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
-                           NULL, 0, NULL);
+       rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
+                               outbuf, sizeof(outbuf), NULL);
+       rc = (rc == -EALREADY) ? 0 : rc;
+       if (rc)
+               efx_mcdi_display_error(efx, MC_CMD_PTP,
+                                      MC_CMD_PTP_IN_DISABLE_LEN,
+                                      outbuf, sizeof(outbuf), rc);
+       return rc;
 }
 
 static void efx_ptp_deliver_rx_queue(struct sk_buff_head *q)
@@ -404,11 +678,10 @@ static void efx_ptp_read_timeset(MCDI_DECLARE_STRUCT_PTR(data),
        unsigned start_ns, end_ns;
 
        timeset->host_start = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTSTART);
-       timeset->seconds = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_SECONDS);
-       timeset->nanoseconds = MCDI_DWORD(data,
-                                        PTP_OUT_SYNCHRONIZE_NANOSECONDS);
+       timeset->major = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MAJOR);
+       timeset->minor = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MINOR);
        timeset->host_end = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTEND),
-       timeset->waitns = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_WAITNS);
+       timeset->wait = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_WAITNS);
 
        /* Ignore seconds */
        start_ns = timeset->host_start & MC_NANOSECOND_MASK;
@@ -437,62 +710,73 @@ efx_ptp_process_times(struct efx_nic *efx, MCDI_DECLARE_STRUCT_PTR(synch_buf),
                MCDI_VAR_ARRAY_LEN(response_length,
                                   PTP_OUT_SYNCHRONIZE_TIMESET);
        unsigned i;
-       unsigned total;
        unsigned ngood = 0;
        unsigned last_good = 0;
        struct efx_ptp_data *ptp = efx->ptp_data;
        u32 last_sec;
        u32 start_sec;
        struct timespec delta;
+       ktime_t mc_time;
 
        if (number_readings == 0)
                return -EAGAIN;
 
-       /* Read the set of results and increment stats for any results that
-        * appera to be erroneous.
+       /* Read the set of results and find the last good host-MC
+        * synchronization result. The MC times when it finishes reading the
+        * host time so the corrected window time should be fairly constant
+        * for a given platform. Increment stats for any results that appear
+        * to be erroneous.
         */
        for (i = 0; i < number_readings; i++) {
+               s32 window, corrected;
+               struct timespec wait;
+
                efx_ptp_read_timeset(
                        MCDI_ARRAY_STRUCT_PTR(synch_buf,
                                              PTP_OUT_SYNCHRONIZE_TIMESET, i),
                        &ptp->timeset[i]);
-       }
 
-       /* Find the last good host-MC synchronization result. The MC times
-        * when it finishes reading the host time so the corrected window time
-        * should be fairly constant for a given platform.
-        */
-       total = 0;
-       for (i = 0; i < number_readings; i++)
-               if (ptp->timeset[i].window > ptp->timeset[i].waitns) {
-                       unsigned win;
-
-                       win = ptp->timeset[i].window - ptp->timeset[i].waitns;
-                       if (win >= MIN_SYNCHRONISATION_NS &&
-                           win < MAX_SYNCHRONISATION_NS) {
-                               total += ptp->timeset[i].window;
-                               ngood++;
-                               last_good = i;
-                       }
+               wait = ktime_to_timespec(
+                       ptp->nic_to_kernel_time(0, ptp->timeset[i].wait, 0));
+               window = ptp->timeset[i].window;
+               corrected = window - wait.tv_nsec;
+
+               /* We expect the uncorrected synchronization window to be at
+                * least as large as the interval between host start and end
+                * times. If it is smaller than this then this is mostly likely
+                * to be a consequence of the host's time being adjusted.
+                * Check that the corrected sync window is in a reasonable
+                * range. If it is out of range it is likely to be because an
+                * interrupt or other delay occurred between reading the system
+                * time and writing it to MC memory.
+                */
+               if (window < SYNCHRONISATION_GRANULARITY_NS) {
+                       ++ptp->invalid_sync_windows;
+               } else if (corrected >= MAX_SYNCHRONISATION_NS) {
+                       ++ptp->oversize_sync_windows;
+               } else if (corrected < ptp->min_synchronisation_ns) {
+                       ++ptp->undersize_sync_windows;
+               } else {
+                       ngood++;
+                       last_good = i;
                }
+       }
 
        if (ngood == 0) {
                netif_warn(efx, drv, efx->net_dev,
-                          "PTP no suitable synchronisations %dns\n",
-                          ptp->base_sync_ns);
+                          "PTP no suitable synchronisations\n");
                return -EAGAIN;
        }
 
-       /* Average minimum this synchronisation */
-       ptp->last_sync_ns = DIV_ROUND_UP(total, ngood);
-       if (!ptp->base_sync_valid || (ptp->last_sync_ns < ptp->base_sync_ns)) {
-               ptp->base_sync_valid = true;
-               ptp->base_sync_ns = ptp->last_sync_ns;
-       }
+       /* Convert the NIC time into kernel time. No correction is required-
+        * this time is the output of a firmware process.
+        */
+       mc_time = ptp->nic_to_kernel_time(ptp->timeset[last_good].major,
+                                         ptp->timeset[last_good].minor, 0);
 
        /* Calculate delay from actual PPS to last_time */
-       delta.tv_nsec =
-               ptp->timeset[last_good].nanoseconds +
+       delta = ktime_to_timespec(mc_time);
+       delta.tv_nsec +=
                last_time->ts_real.tv_nsec -
                (ptp->timeset[last_good].host_start & MC_NANOSECOND_MASK);
 
@@ -553,6 +837,11 @@ static int efx_ptp_synchronize(struct efx_nic *efx, unsigned int num_readings)
                loops++;
        }
 
+       if (loops <= 1)
+               ++ptp->fast_syncs;
+       if (!time_before(jiffies, timeout))
+               ++ptp->sync_timeouts;
+
        if (ACCESS_ONCE(*start))
                efx_ptp_send_times(efx, &last_time);
 
@@ -561,9 +850,20 @@ static int efx_ptp_synchronize(struct efx_nic *efx, unsigned int num_readings)
                                 MC_CMD_PTP_IN_SYNCHRONIZE_LEN,
                                 synch_buf, sizeof(synch_buf),
                                 &response_length);
-       if (rc == 0)
+       if (rc == 0) {
                rc = efx_ptp_process_times(efx, synch_buf, response_length,
                                           &last_time);
+               if (rc == 0)
+                       ++ptp->good_syncs;
+               else
+                       ++ptp->no_time_syncs;
+       }
+
+       /* Increment the bad syncs counter if the synchronize fails, whatever
+        * the reason.
+        */
+       if (rc != 0)
+               ++ptp->bad_syncs;
 
        return rc;
 }
@@ -602,9 +902,10 @@ static int efx_ptp_xmit_skb(struct efx_nic *efx, struct sk_buff *skb)
                goto fail;
 
        memset(&timestamps, 0, sizeof(timestamps));
-       timestamps.hwtstamp = ktime_set(
-               MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_SECONDS),
-               MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_NANOSECONDS));
+       timestamps.hwtstamp = ptp_data->nic_to_kernel_time(
+               MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MAJOR),
+               MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MINOR),
+               ptp_data->ts_corrections.tx);
 
        skb_tstamp_tx(skb, &timestamps);
 
@@ -622,6 +923,9 @@ static void efx_ptp_drop_time_expired_events(struct efx_nic *efx)
        struct list_head *cursor;
        struct list_head *next;
 
+       if (ptp->rx_ts_inline)
+               return;
+
        /* Drop time-expired events */
        spin_lock_bh(&ptp->evt_lock);
        if (!list_empty(&ptp->evt_list)) {
@@ -655,6 +959,8 @@ static enum ptp_packet_state efx_ptp_match_rx(struct efx_nic *efx,
        struct efx_ptp_match *match;
        enum ptp_packet_state rc = PTP_PACKET_STATE_UNMATCHED;
 
+       WARN_ON_ONCE(ptp->rx_ts_inline);
+
        spin_lock_bh(&ptp->evt_lock);
        evts_waiting = !list_empty(&ptp->evt_list);
        spin_unlock_bh(&ptp->evt_lock);
@@ -696,13 +1002,10 @@ static enum ptp_packet_state efx_ptp_match_rx(struct efx_nic *efx,
 /* Process any queued receive events and corresponding packets
  *
  * q is returned with all the packets that are ready for delivery.
- * true is returned if at least one of those packets requires
- * synchronisation.
  */
-static bool efx_ptp_process_events(struct efx_nic *efx, struct sk_buff_head *q)
+static void efx_ptp_process_events(struct efx_nic *efx, struct sk_buff_head *q)
 {
        struct efx_ptp_data *ptp = efx->ptp_data;
-       bool rc = false;
        struct sk_buff *skb;
 
        while ((skb = skb_dequeue(&ptp->rxq))) {
@@ -713,13 +1016,10 @@ static bool efx_ptp_process_events(struct efx_nic *efx, struct sk_buff_head *q)
                        __skb_queue_tail(q, skb);
                } else if (efx_ptp_match_rx(efx, skb) ==
                           PTP_PACKET_STATE_MATCHED) {
-                       rc = true;
                        __skb_queue_tail(q, skb);
                } else if (time_after(jiffies, match->expiry)) {
                        match->state = PTP_PACKET_STATE_TIMED_OUT;
-                       if (net_ratelimit())
-                               netif_warn(efx, rx_err, efx->net_dev,
-                                          "PTP packet - no timestamp seen\n");
+                       ++ptp->rx_no_timestamp;
                        __skb_queue_tail(q, skb);
                } else {
                        /* Replace unprocessed entry and stop */
@@ -727,8 +1027,6 @@ static bool efx_ptp_process_events(struct efx_nic *efx, struct sk_buff_head *q)
                        break;
                }
        }
-
-       return rc;
 }
 
 /* Complete processing of a received packet */
@@ -739,13 +1037,27 @@ static inline void efx_ptp_process_rx(struct efx_nic *efx, struct sk_buff *skb)
        local_bh_enable();
 }
 
-static int efx_ptp_start(struct efx_nic *efx)
+static void efx_ptp_remove_multicast_filters(struct efx_nic *efx)
+{
+       struct efx_ptp_data *ptp = efx->ptp_data;
+
+       if (ptp->rxfilter_installed) {
+               efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
+                                         ptp->rxfilter_general);
+               efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
+                                         ptp->rxfilter_event);
+               ptp->rxfilter_installed = false;
+       }
+}
+
+static int efx_ptp_insert_multicast_filters(struct efx_nic *efx)
 {
        struct efx_ptp_data *ptp = efx->ptp_data;
        struct efx_filter_spec rxfilter;
        int rc;
 
-       ptp->reset_required = false;
+       if (!ptp->channel || ptp->rxfilter_installed)
+               return 0;
 
        /* Must filter on both event and general ports to ensure
         * that there is no packet re-ordering.
@@ -778,23 +1090,37 @@ static int efx_ptp_start(struct efx_nic *efx)
                goto fail;
        ptp->rxfilter_general = rc;
 
+       ptp->rxfilter_installed = true;
+       return 0;
+
+fail:
+       efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
+                                 ptp->rxfilter_event);
+       return rc;
+}
+
+static int efx_ptp_start(struct efx_nic *efx)
+{
+       struct efx_ptp_data *ptp = efx->ptp_data;
+       int rc;
+
+       ptp->reset_required = false;
+
+       rc = efx_ptp_insert_multicast_filters(efx);
+       if (rc)
+               return rc;
+
        rc = efx_ptp_enable(efx);
        if (rc != 0)
-               goto fail2;
+               goto fail;
 
        ptp->evt_frag_idx = 0;
        ptp->current_adjfreq = 0;
-       ptp->rxfilter_installed = true;
 
        return 0;
 
-fail2:
-       efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
-                                 ptp->rxfilter_general);
 fail:
-       efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
-                                 ptp->rxfilter_event);
-
+       efx_ptp_remove_multicast_filters(efx);
        return rc;
 }
 
@@ -810,13 +1136,7 @@ static int efx_ptp_stop(struct efx_nic *efx)
 
        rc = efx_ptp_disable(efx);
 
-       if (ptp->rxfilter_installed) {
-               efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
-                                         ptp->rxfilter_general);
-               efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
-                                         ptp->rxfilter_event);
-               ptp->rxfilter_installed = false;
-       }
+       efx_ptp_remove_multicast_filters(efx);
 
        /* Make sure RX packets are really delivered */
        efx_ptp_deliver_rx_queue(&efx->ptp_data->rxq);
@@ -844,7 +1164,7 @@ static void efx_ptp_pps_worker(struct work_struct *work)
 {
        struct efx_ptp_data *ptp =
                container_of(work, struct efx_ptp_data, pps_work);
-       struct efx_nic *efx = ptp->channel->efx;
+       struct efx_nic *efx = ptp->efx;
        struct ptp_clock_event ptp_evt;
 
        if (efx_ptp_synchronize(efx, PTP_SYNC_ATTEMPTS))
@@ -855,13 +1175,11 @@ static void efx_ptp_pps_worker(struct work_struct *work)
        ptp_clock_event(ptp->phc_clock, &ptp_evt);
 }
 
-/* Process any pending transmissions and timestamp any received packets.
- */
 static void efx_ptp_worker(struct work_struct *work)
 {
        struct efx_ptp_data *ptp_data =
                container_of(work, struct efx_ptp_data, work);
-       struct efx_nic *efx = ptp_data->channel->efx;
+       struct efx_nic *efx = ptp_data->efx;
        struct sk_buff *skb;
        struct sk_buff_head tempq;
 
@@ -874,42 +1192,50 @@ static void efx_ptp_worker(struct work_struct *work)
        efx_ptp_drop_time_expired_events(efx);
 
        __skb_queue_head_init(&tempq);
-       if (efx_ptp_process_events(efx, &tempq) ||
-           !skb_queue_empty(&ptp_data->txq)) {
+       efx_ptp_process_events(efx, &tempq);
 
-               while ((skb = skb_dequeue(&ptp_data->txq)))
-                       efx_ptp_xmit_skb(efx, skb);
-       }
+       while ((skb = skb_dequeue(&ptp_data->txq)))
+               efx_ptp_xmit_skb(efx, skb);
 
        while ((skb = __skb_dequeue(&tempq)))
                efx_ptp_process_rx(efx, skb);
 }
 
-/* Initialise PTP channel and state.
- *
- * Setting core_index to zero causes the queue to be initialised and doesn't
- * overlap with 'rxq0' because ptp.c doesn't use skb_record_rx_queue.
- */
-static int efx_ptp_probe_channel(struct efx_channel *channel)
+static const struct ptp_clock_info efx_phc_clock_info = {
+       .owner          = THIS_MODULE,
+       .name           = "sfc",
+       .max_adj        = MAX_PPB,
+       .n_alarm        = 0,
+       .n_ext_ts       = 0,
+       .n_per_out      = 0,
+       .pps            = 1,
+       .adjfreq        = efx_phc_adjfreq,
+       .adjtime        = efx_phc_adjtime,
+       .gettime        = efx_phc_gettime,
+       .settime        = efx_phc_settime,
+       .enable         = efx_phc_enable,
+};
+
+/* Initialise PTP state. */
+int efx_ptp_probe(struct efx_nic *efx, struct efx_channel *channel)
 {
-       struct efx_nic *efx = channel->efx;
        struct efx_ptp_data *ptp;
        int rc = 0;
        unsigned int pos;
 
-       channel->irq_moderation = 0;
-       channel->rx_queue.core_index = 0;
-
        ptp = kzalloc(sizeof(struct efx_ptp_data), GFP_KERNEL);
        efx->ptp_data = ptp;
        if (!efx->ptp_data)
                return -ENOMEM;
 
+       ptp->efx = efx;
+       ptp->channel = channel;
+       ptp->rx_ts_inline = efx_nic_rev(efx) >= EFX_REV_HUNT_A0;
+
        rc = efx_nic_alloc_buffer(efx, &ptp->start, sizeof(int), GFP_KERNEL);
        if (rc != 0)
                goto fail1;
 
-       ptp->channel = channel;
        skb_queue_head_init(&ptp->rxq);
        skb_queue_head_init(&ptp->txq);
        ptp->workwq = create_singlethread_workqueue("sfc_ptp");
@@ -929,33 +1255,32 @@ static int efx_ptp_probe_channel(struct efx_channel *channel)
                list_add(&ptp->rx_evts[pos].link, &ptp->evt_free_list);
        ptp->evt_overflow = false;
 
-       ptp->phc_clock_info.owner = THIS_MODULE;
-       snprintf(ptp->phc_clock_info.name,
-                sizeof(ptp->phc_clock_info.name),
-                "%pm", efx->net_dev->perm_addr);
-       ptp->phc_clock_info.max_adj = MAX_PPB;
-       ptp->phc_clock_info.n_alarm = 0;
-       ptp->phc_clock_info.n_ext_ts = 0;
-       ptp->phc_clock_info.n_per_out = 0;
-       ptp->phc_clock_info.pps = 1;
-       ptp->phc_clock_info.adjfreq = efx_phc_adjfreq;
-       ptp->phc_clock_info.adjtime = efx_phc_adjtime;
-       ptp->phc_clock_info.gettime = efx_phc_gettime;
-       ptp->phc_clock_info.settime = efx_phc_settime;
-       ptp->phc_clock_info.enable = efx_phc_enable;
-
-       ptp->phc_clock = ptp_clock_register(&ptp->phc_clock_info,
-                                           &efx->pci_dev->dev);
-       if (IS_ERR(ptp->phc_clock)) {
-               rc = PTR_ERR(ptp->phc_clock);
+       /* Get the NIC PTP attributes and set up time conversions */
+       rc = efx_ptp_get_attributes(efx);
+       if (rc < 0)
                goto fail3;
-       }
 
-       INIT_WORK(&ptp->pps_work, efx_ptp_pps_worker);
-       ptp->pps_workwq = create_singlethread_workqueue("sfc_pps");
-       if (!ptp->pps_workwq) {
-               rc = -ENOMEM;
-               goto fail4;
+       /* Get the timestamp corrections */
+       rc = efx_ptp_get_timestamp_corrections(efx);
+       if (rc < 0)
+               goto fail3;
+
+       if (efx->mcdi->fn_flags &
+           (1 << MC_CMD_DRV_ATTACH_EXT_OUT_FLAG_PRIMARY)) {
+               ptp->phc_clock_info = efx_phc_clock_info;
+               ptp->phc_clock = ptp_clock_register(&ptp->phc_clock_info,
+                                                   &efx->pci_dev->dev);
+               if (IS_ERR(ptp->phc_clock)) {
+                       rc = PTR_ERR(ptp->phc_clock);
+                       goto fail3;
+               }
+
+               INIT_WORK(&ptp->pps_work, efx_ptp_pps_worker);
+               ptp->pps_workwq = create_singlethread_workqueue("sfc_pps");
+               if (!ptp->pps_workwq) {
+                       rc = -ENOMEM;
+                       goto fail4;
+               }
        }
        ptp->nic_ts_enabled = false;
 
@@ -976,14 +1301,27 @@ fail1:
        return rc;
 }
 
-static void efx_ptp_remove_channel(struct efx_channel *channel)
+/* Initialise PTP channel.
+ *
+ * Setting core_index to zero causes the queue to be initialised and doesn't
+ * overlap with 'rxq0' because ptp.c doesn't use skb_record_rx_queue.
+ */
+static int efx_ptp_probe_channel(struct efx_channel *channel)
 {
        struct efx_nic *efx = channel->efx;
 
+       channel->irq_moderation = 0;
+       channel->rx_queue.core_index = 0;
+
+       return efx_ptp_probe(efx, channel);
+}
+
+void efx_ptp_remove(struct efx_nic *efx)
+{
        if (!efx->ptp_data)
                return;
 
-       (void)efx_ptp_disable(channel->efx);
+       (void)efx_ptp_disable(efx);
 
        cancel_work_sync(&efx->ptp_data->work);
        cancel_work_sync(&efx->ptp_data->pps_work);
@@ -991,15 +1329,22 @@ static void efx_ptp_remove_channel(struct efx_channel *channel)
        skb_queue_purge(&efx->ptp_data->rxq);
        skb_queue_purge(&efx->ptp_data->txq);
 
-       ptp_clock_unregister(efx->ptp_data->phc_clock);
+       if (efx->ptp_data->phc_clock) {
+               destroy_workqueue(efx->ptp_data->pps_workwq);
+               ptp_clock_unregister(efx->ptp_data->phc_clock);
+       }
 
        destroy_workqueue(efx->ptp_data->workwq);
-       destroy_workqueue(efx->ptp_data->pps_workwq);
 
        efx_nic_free_buffer(efx, &efx->ptp_data->start);
        kfree(efx->ptp_data);
 }
 
+static void efx_ptp_remove_channel(struct efx_channel *channel)
+{
+       efx_ptp_remove(channel->efx);
+}
+
 static void efx_ptp_get_channel_name(struct efx_channel *channel,
                                     char *buf, size_t len)
 {
@@ -1080,14 +1425,8 @@ static bool efx_ptp_rx(struct efx_channel *channel, struct sk_buff *skb)
 
        /* Does this packet require timestamping? */
        if (ntohs(*(__be16 *)&skb->data[PTP_DPORT_OFFSET]) == PTP_EVENT_PORT) {
-               struct skb_shared_hwtstamps *timestamps;
-
                match->state = PTP_PACKET_STATE_UNMATCHED;
 
-               /* Clear all timestamps held: filled in later */
-               timestamps = skb_hwtstamps(skb);
-               memset(timestamps, 0, sizeof(*timestamps));
-
                /* We expect the sequence number to be in the same position in
                 * the packet for PTP V1 and V2
                 */
@@ -1132,8 +1471,13 @@ int efx_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
        return NETDEV_TX_OK;
 }
 
-static int efx_ptp_change_mode(struct efx_nic *efx, bool enable_wanted,
-                              unsigned int new_mode)
+int efx_ptp_get_mode(struct efx_nic *efx)
+{
+       return efx->ptp_data->mode;
+}
+
+int efx_ptp_change_mode(struct efx_nic *efx, bool enable_wanted,
+                       unsigned int new_mode)
 {
        if ((enable_wanted != efx->ptp_data->enabled) ||
            (enable_wanted && (efx->ptp_data->mode != new_mode))) {
@@ -1177,8 +1521,6 @@ static int efx_ptp_change_mode(struct efx_nic *efx, bool enable_wanted,
 
 static int efx_ptp_ts_init(struct efx_nic *efx, struct hwtstamp_config *init)
 {
-       bool enable_wanted = false;
-       unsigned int new_mode;
        int rc;
 
        if (init->flags)
@@ -1188,63 +1530,20 @@ static int efx_ptp_ts_init(struct efx_nic *efx, struct hwtstamp_config *init)
            (init->tx_type != HWTSTAMP_TX_ON))
                return -ERANGE;
 
-       new_mode = efx->ptp_data->mode;
-       /* Determine whether any PTP HW operations are required */
-       switch (init->rx_filter) {
-       case HWTSTAMP_FILTER_NONE:
-               break;
-       case HWTSTAMP_FILTER_PTP_V1_L4_EVENT:
-       case HWTSTAMP_FILTER_PTP_V1_L4_SYNC:
-       case HWTSTAMP_FILTER_PTP_V1_L4_DELAY_REQ:
-               init->rx_filter = HWTSTAMP_FILTER_PTP_V1_L4_EVENT;
-               new_mode = MC_CMD_PTP_MODE_V1;
-               enable_wanted = true;
-               break;
-       case HWTSTAMP_FILTER_PTP_V2_L4_EVENT:
-       case HWTSTAMP_FILTER_PTP_V2_L4_SYNC:
-       case HWTSTAMP_FILTER_PTP_V2_L4_DELAY_REQ:
-       /* Although these three are accepted only IPV4 packets will be
-        * timestamped
-        */
-               init->rx_filter = HWTSTAMP_FILTER_PTP_V2_L4_EVENT;
-               new_mode = MC_CMD_PTP_MODE_V2_ENHANCED;
-               enable_wanted = true;
-               break;
-       case HWTSTAMP_FILTER_PTP_V2_EVENT:
-       case HWTSTAMP_FILTER_PTP_V2_SYNC:
-       case HWTSTAMP_FILTER_PTP_V2_DELAY_REQ:
-       case HWTSTAMP_FILTER_PTP_V2_L2_EVENT:
-       case HWTSTAMP_FILTER_PTP_V2_L2_SYNC:
-       case HWTSTAMP_FILTER_PTP_V2_L2_DELAY_REQ:
-               /* Non-IP + IPv6 timestamping not supported */
-               return -ERANGE;
-               break;
-       default:
-               return -ERANGE;
-       }
-
-       if (init->tx_type != HWTSTAMP_TX_OFF)
-               enable_wanted = true;
-
-       /* Old versions of the firmware do not support the improved
-        * UUID filtering option (SF bug 33070).  If the firmware does
-        * not accept the enhanced mode, fall back to the standard PTP
-        * v2 UUID filtering.
-        */
-       rc = efx_ptp_change_mode(efx, enable_wanted, new_mode);
-       if ((rc != 0) && (new_mode == MC_CMD_PTP_MODE_V2_ENHANCED))
-               rc = efx_ptp_change_mode(efx, enable_wanted, MC_CMD_PTP_MODE_V2);
-       if (rc != 0)
+       rc = efx->type->ptp_set_ts_config(efx, init);
+       if (rc)
                return rc;
 
        efx->ptp_data->config = *init;
-
        return 0;
 }
 
 void efx_ptp_get_ts_info(struct efx_nic *efx, struct ethtool_ts_info *ts_info)
 {
        struct efx_ptp_data *ptp = efx->ptp_data;
+       struct efx_nic *primary = efx->primary;
+
+       ASSERT_RTNL();
 
        if (!ptp)
                return;
@@ -1252,18 +1551,14 @@ void efx_ptp_get_ts_info(struct efx_nic *efx, struct ethtool_ts_info *ts_info)
        ts_info->so_timestamping |= (SOF_TIMESTAMPING_TX_HARDWARE |
                                     SOF_TIMESTAMPING_RX_HARDWARE |
                                     SOF_TIMESTAMPING_RAW_HARDWARE);
-       ts_info->phc_index = ptp_clock_index(ptp->phc_clock);
+       if (primary && primary->ptp_data && primary->ptp_data->phc_clock)
+               ts_info->phc_index =
+                       ptp_clock_index(primary->ptp_data->phc_clock);
        ts_info->tx_types = 1 << HWTSTAMP_TX_OFF | 1 << HWTSTAMP_TX_ON;
-       ts_info->rx_filters = (1 << HWTSTAMP_FILTER_NONE |
-                              1 << HWTSTAMP_FILTER_PTP_V1_L4_EVENT |
-                              1 << HWTSTAMP_FILTER_PTP_V1_L4_SYNC |
-                              1 << HWTSTAMP_FILTER_PTP_V1_L4_DELAY_REQ |
-                              1 << HWTSTAMP_FILTER_PTP_V2_L4_EVENT |
-                              1 << HWTSTAMP_FILTER_PTP_V2_L4_SYNC |
-                              1 << HWTSTAMP_FILTER_PTP_V2_L4_DELAY_REQ);
+       ts_info->rx_filters = ptp->efx->type->hwtstamp_filters;
 }
 
-int efx_ptp_ioctl(struct efx_nic *efx, struct ifreq *ifr, int cmd)
+int efx_ptp_set_ts_config(struct efx_nic *efx, struct ifreq *ifr)
 {
        struct hwtstamp_config config;
        int rc;
@@ -1283,6 +1578,15 @@ int efx_ptp_ioctl(struct efx_nic *efx, struct ifreq *ifr, int cmd)
                ? -EFAULT : 0;
 }
 
+int efx_ptp_get_ts_config(struct efx_nic *efx, struct ifreq *ifr)
+{
+       if (!efx->ptp_data)
+               return -EOPNOTSUPP;
+
+       return copy_to_user(ifr->ifr_data, &efx->ptp_data->config,
+                           sizeof(efx->ptp_data->config)) ? -EFAULT : 0;
+}
+
 static void ptp_event_failure(struct efx_nic *efx, int expected_frag_len)
 {
        struct efx_ptp_data *ptp = efx->ptp_data;
@@ -1302,6 +1606,9 @@ static void ptp_event_rx(struct efx_nic *efx, struct efx_ptp_data *ptp)
 {
        struct efx_ptp_event_rx *evt = NULL;
 
+       if (WARN_ON_ONCE(ptp->rx_ts_inline))
+               return;
+
        if (ptp->evt_frag_idx != 3) {
                ptp_event_failure(efx, 3);
                return;
@@ -1320,9 +1627,10 @@ static void ptp_event_rx(struct efx_nic *efx, struct efx_ptp_data *ptp)
                                              MCDI_EVENT_SRC) << 8) |
                             (EFX_QWORD_FIELD(ptp->evt_frags[0],
                                              MCDI_EVENT_SRC) << 16));
-               evt->hwtimestamp = ktime_set(
+               evt->hwtimestamp = efx->ptp_data->nic_to_kernel_time(
                        EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA),
-                       EFX_QWORD_FIELD(ptp->evt_frags[1], MCDI_EVENT_DATA));
+                       EFX_QWORD_FIELD(ptp->evt_frags[1], MCDI_EVENT_DATA),
+                       ptp->ts_corrections.rx);
                evt->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
                list_add_tail(&evt->link, &ptp->evt_list);
 
@@ -1360,6 +1668,13 @@ void efx_ptp_event(struct efx_nic *efx, efx_qword_t *ev)
        struct efx_ptp_data *ptp = efx->ptp_data;
        int code = EFX_QWORD_FIELD(*ev, MCDI_EVENT_CODE);
 
+       if (!ptp) {
+               if (net_ratelimit())
+                       netif_warn(efx, drv, efx->net_dev,
+                                  "Received PTP event but PTP not set up\n");
+               return;
+       }
+
        if (!ptp->enabled)
                return;
 
@@ -1397,12 +1712,99 @@ void efx_ptp_event(struct efx_nic *efx, efx_qword_t *ev)
        }
 }
 
+void efx_time_sync_event(struct efx_channel *channel, efx_qword_t *ev)
+{
+       channel->sync_timestamp_major = MCDI_EVENT_FIELD(*ev, PTP_TIME_MAJOR);
+       channel->sync_timestamp_minor =
+               MCDI_EVENT_FIELD(*ev, PTP_TIME_MINOR_26_19) << 19;
+       /* if sync events have been disabled then we want to silently ignore
+        * this event, so throw away result.
+        */
+       (void) cmpxchg(&channel->sync_events_state, SYNC_EVENTS_REQUESTED,
+                      SYNC_EVENTS_VALID);
+}
+
+/* make some assumptions about the time representation rather than abstract it,
+ * since we currently only support one type of inline timestamping and only on
+ * EF10.
+ */
+#define MINOR_TICKS_PER_SECOND 0x8000000
+/* Fuzz factor for sync events to be out of order with RX events */
+#define FUZZ (MINOR_TICKS_PER_SECOND / 10)
+#define EXPECTED_SYNC_EVENTS_PER_SECOND 4
+
+static inline u32 efx_rx_buf_timestamp_minor(struct efx_nic *efx, const u8 *eh)
+{
+#if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
+       return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_ts_offset));
+#else
+       const u8 *data = eh + efx->rx_packet_ts_offset;
+       return (u32)data[0]       |
+              (u32)data[1] << 8  |
+              (u32)data[2] << 16 |
+              (u32)data[3] << 24;
+#endif
+}
+
+void __efx_rx_skb_attach_timestamp(struct efx_channel *channel,
+                                  struct sk_buff *skb)
+{
+       struct efx_nic *efx = channel->efx;
+       u32 pkt_timestamp_major, pkt_timestamp_minor;
+       u32 diff, carry;
+       struct skb_shared_hwtstamps *timestamps;
+
+       pkt_timestamp_minor = (efx_rx_buf_timestamp_minor(efx,
+                                                         skb_mac_header(skb)) +
+                              (u32) efx->ptp_data->ts_corrections.rx) &
+                             (MINOR_TICKS_PER_SECOND - 1);
+
+       /* get the difference between the packet and sync timestamps,
+        * modulo one second
+        */
+       diff = (pkt_timestamp_minor - channel->sync_timestamp_minor) &
+               (MINOR_TICKS_PER_SECOND - 1);
+       /* do we roll over a second boundary and need to carry the one? */
+       carry = channel->sync_timestamp_minor + diff > MINOR_TICKS_PER_SECOND ?
+               1 : 0;
+
+       if (diff <= MINOR_TICKS_PER_SECOND / EXPECTED_SYNC_EVENTS_PER_SECOND +
+                   FUZZ) {
+               /* packet is ahead of the sync event by a quarter of a second or
+                * less (allowing for fuzz)
+                */
+               pkt_timestamp_major = channel->sync_timestamp_major + carry;
+       } else if (diff >= MINOR_TICKS_PER_SECOND - FUZZ) {
+               /* packet is behind the sync event but within the fuzz factor.
+                * This means the RX packet and sync event crossed as they were
+                * placed on the event queue, which can sometimes happen.
+                */
+               pkt_timestamp_major = channel->sync_timestamp_major - 1 + carry;
+       } else {
+               /* it's outside tolerance in both directions. this might be
+                * indicative of us missing sync events for some reason, so
+                * we'll call it an error rather than risk giving a bogus
+                * timestamp.
+                */
+               netif_vdbg(efx, drv, efx->net_dev,
+                         "packet timestamp %x too far from sync event %x:%x\n",
+                         pkt_timestamp_minor, channel->sync_timestamp_major,
+                         channel->sync_timestamp_minor);
+               return;
+       }
+
+       /* attach the timestamps to the skb */
+       timestamps = skb_hwtstamps(skb);
+       timestamps->hwtstamp =
+               efx_ptp_s27_to_ktime(pkt_timestamp_major, pkt_timestamp_minor);
+}
+
 static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta)
 {
        struct efx_ptp_data *ptp_data = container_of(ptp,
                                                     struct efx_ptp_data,
                                                     phc_clock_info);
-       struct efx_nic *efx = ptp_data->channel->efx;
+       struct efx_nic *efx = ptp_data->efx;
        MCDI_DECLARE_BUF(inadj, MC_CMD_PTP_IN_ADJUST_LEN);
        s64 adjustment_ns;
        int rc;
@@ -1432,18 +1834,20 @@ static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta)
 
 static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta)
 {
+       u32 nic_major, nic_minor;
        struct efx_ptp_data *ptp_data = container_of(ptp,
                                                     struct efx_ptp_data,
                                                     phc_clock_info);
-       struct efx_nic *efx = ptp_data->channel->efx;
-       struct timespec delta_ts = ns_to_timespec(delta);
+       struct efx_nic *efx = ptp_data->efx;
        MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ADJUST_LEN);
 
+       efx->ptp_data->ns_to_nic_time(delta, &nic_major, &nic_minor);
+
        MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
        MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
        MCDI_SET_QWORD(inbuf, PTP_IN_ADJUST_FREQ, ptp_data->current_adjfreq);
-       MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_SECONDS, (u32)delta_ts.tv_sec);
-       MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_NANOSECONDS, (u32)delta_ts.tv_nsec);
+       MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MAJOR, nic_major);
+       MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MINOR, nic_minor);
        return efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
                            NULL, 0, NULL);
 }
@@ -1453,10 +1857,11 @@ static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec *ts)
        struct efx_ptp_data *ptp_data = container_of(ptp,
                                                     struct efx_ptp_data,
                                                     phc_clock_info);
-       struct efx_nic *efx = ptp_data->channel->efx;
+       struct efx_nic *efx = ptp_data->efx;
        MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_READ_NIC_TIME_LEN);
        MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_READ_NIC_TIME_LEN);
        int rc;
+       ktime_t kt;
 
        MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_READ_NIC_TIME);
        MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
@@ -1466,8 +1871,10 @@ static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec *ts)
        if (rc != 0)
                return rc;
 
-       ts->tv_sec = MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_SECONDS);
-       ts->tv_nsec = MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_NANOSECONDS);
+       kt = ptp_data->nic_to_kernel_time(
+               MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MAJOR),
+               MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MINOR), 0);
+       *ts = ktime_to_timespec(kt);
        return 0;
 }
 
@@ -1519,7 +1926,7 @@ static const struct efx_channel_type efx_ptp_channel_type = {
        .keep_eventq            = false,
 };
 
-void efx_ptp_probe(struct efx_nic *efx)
+void efx_ptp_defer_probe_with_channel(struct efx_nic *efx)
 {
        /* Check whether PTP is implemented on this NIC.  The DISABLE
         * operation will succeed if and only if it is implemented.
@@ -1533,9 +1940,15 @@ void efx_ptp_start_datapath(struct efx_nic *efx)
 {
        if (efx_ptp_restart(efx))
                netif_err(efx, drv, efx->net_dev, "Failed to restart PTP.\n");
+       /* re-enable timestamping if it was previously enabled */
+       if (efx->type->ptp_set_ts_sync_events)
+               efx->type->ptp_set_ts_sync_events(efx, true, true);
 }
 
 void efx_ptp_stop_datapath(struct efx_nic *efx)
 {
+       /* temporarily disable timestamping */
+       if (efx->type->ptp_set_ts_sync_events)
+               efx->type->ptp_set_ts_sync_events(efx, false, true);
        efx_ptp_stop(efx);
 }