Merge git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next

Pull networking updates from David Miller:
 "Another merge window, another pull full of stuff:

   1) Support alternative names for network devices, from Jiri Pirko.

   2) Introduce per-netns netdev notifiers, also from Jiri Pirko.

   3) Support MSG_PEEK in vsock/virtio, from Matias Ezequiel Vara
      Larsen.

   4) Allow compiling out the TLS TOE code, from Jakub Kicinski.

   5) Add several new tracepoints to the kTLS code, also from Jakub.

   6) Support set channels ethtool callback in ena driver, from Sameeh
      Jubran.

   7) New SCTP events SCTP_ADDR_ADDED, SCTP_ADDR_REMOVED,
      SCTP_ADDR_MADE_PRIM, and SCTP_SEND_FAILED_EVENT. From Xin Long.

   8) Add XDP support to mvneta driver, from Lorenzo Bianconi.

   9) Lots of netfilter hw offload fixes, cleanups and enhancements,
      from Pablo Neira Ayuso.

  10) PTP support for aquantia chips, from Egor Pomozov.

  11) Add UDP segmentation offload support to igb, ixgbe, and i40e. From
      Josh Hunt.

  12) Add smart nagle to tipc, from Jon Maloy.

  13) Support L2 field rewrite by TC offloads in bnxt_en, from Venkat
      Duvvuru.

  14) Add a flow mask cache to OVS, from Tonghao Zhang.

  15) Add XDP support to ice driver, from Maciej Fijalkowski.

  16) Add AF_XDP support to ice driver, from Krzysztof Kazimierczak.

  17) Support UDP GSO offload in atlantic driver, from Igor Russkikh.

  18) Support it in stmmac driver too, from Jose Abreu.

  19) Support TIPC encryption and auth, from Tuong Lien.

  20) Introduce BPF trampolines, from Alexei Starovoitov.

  21) Make page_pool API more numa friendly, from Saeed Mahameed.

  22) Introduce route hints to ipv4 and ipv6, from Paolo Abeni.

  23) Add UDP segmentation offload to cxgb4, Rahul Lakkireddy"

* git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next: (1857 commits)
  libbpf: Fix usage of u32 in userspace code
  mm: Implement no-MMU variant of vmalloc_user_node_flags
  slip: Fix use-after-free Read in slip_open
  net: dsa: sja1105: fix sja1105_parse_rgmii_delays()
  macvlan: schedule bc_work even if error
  enetc: add support Credit Based Shaper(CBS) for hardware offload
  net: phy: add helpers phy_(un)lock_mdio_bus
  mdio_bus: don't use managed reset-controller
  ax88179_178a: add ethtool_op_get_ts_info()
  mlxsw: spectrum_router: Fix use of uninitialized adjacency index
  mlxsw: spectrum_router: After underlay moves, demote conflicting tunnels
  bpf: Simplify __bpf_arch_text_poke poke type handling
  bpf: Introduce BPF_TRACE_x helper for the tracing tests
  bpf: Add bpf_jit_blinding_enabled for !CONFIG_BPF_JIT
  bpf, testing: Add various tail call test cases
  bpf, x86: Emit patchable direct jump as tail call
  bpf: Constant map key tracking for prog array pokes
  bpf: Add poke dependency tracking for prog array maps
  bpf: Add initial poke descriptor table for jit images
  bpf: Move owner type, jited info into array auxiliary data
  ...

1 files changed, 420 insertions, 12 deletions

diff --git a/drivers/net/dsa/sja1105/sja1105_tas.c b/drivers/net/dsa/sja1105/sja1105_tas.c
index 33eca6a82ec5..26b925b5dace 100644
--- a/drivers/net/dsa/sja1105/sja1105_tas.c
+++ b/drivers/net/dsa/sja1105/sja1105_tas.c
@@ -10,6 +10,11 @@
 #define SJA1105_TAS_MAX_DELTA		BIT(19)
 #define SJA1105_GATE_MASK		GENMASK_ULL(SJA1105_NUM_TC - 1, 0)
 
+#define work_to_sja1105_tas(d) \
+	container_of((d), struct sja1105_tas_data, tas_work)
+#define tas_to_sja1105(d) \
+	container_of((d), struct sja1105_private, tas_data)
+
 /* This is not a preprocessor macro because the "ns" argument may or may not be
  * s64 at caller side. This ensures it is properly type-cast before div_s64.
  */
@@ -18,6 +23,100 @@ static s64 ns_to_sja1105_delta(s64 ns)
 	return div_s64(ns, 200);
 }
 
+static s64 sja1105_delta_to_ns(s64 delta)
+{
+	return delta * 200;
+}
+
+/* Calculate the first base_time in the future that satisfies this
+ * relationship:
+ *
+ * future_base_time = base_time + N x cycle_time >= now, or
+ *
+ *      now - base_time
+ * N >= ---------------
+ *         cycle_time
+ *
+ * Because N is an integer, the ceiling value of the above "a / b" ratio
+ * is in fact precisely the floor value of "(a + b - 1) / b", which is
+ * easier to calculate only having integer division tools.
+ */
+static s64 future_base_time(s64 base_time, s64 cycle_time, s64 now)
+{
+	s64 a, b, n;
+
+	if (base_time >= now)
+		return base_time;
+
+	a = now - base_time;
+	b = cycle_time;
+	n = div_s64(a + b - 1, b);
+
+	return base_time + n * cycle_time;
+}
+
+static int sja1105_tas_set_runtime_params(struct sja1105_private *priv)
+{
+	struct sja1105_tas_data *tas_data = &priv->tas_data;
+	struct dsa_switch *ds = priv->ds;
+	s64 earliest_base_time = S64_MAX;
+	s64 latest_base_time = 0;
+	s64 its_cycle_time = 0;
+	s64 max_cycle_time = 0;
+	int port;
+
+	tas_data->enabled = false;
+
+	for (port = 0; port < SJA1105_NUM_PORTS; port++) {
+		const struct tc_taprio_qopt_offload *offload;
+
+		offload = tas_data->offload[port];
+		if (!offload)
+			continue;
+
+		tas_data->enabled = true;
+
+		if (max_cycle_time < offload->cycle_time)
+			max_cycle_time = offload->cycle_time;
+		if (latest_base_time < offload->base_time)
+			latest_base_time = offload->base_time;
+		if (earliest_base_time > offload->base_time) {
+			earliest_base_time = offload->base_time;
+			its_cycle_time = offload->cycle_time;
+		}
+	}
+
+	if (!tas_data->enabled)
+		return 0;
+
+	/* Roll the earliest base time over until it is in a comparable
+	 * time base with the latest, then compare their deltas.
+	 * We want to enforce that all ports' base times are within
+	 * SJA1105_TAS_MAX_DELTA 200ns cycles of one another.
+	 */
+	earliest_base_time = future_base_time(earliest_base_time,
+					      its_cycle_time,
+					      latest_base_time);
+	while (earliest_base_time > latest_base_time)
+		earliest_base_time -= its_cycle_time;
+	if (latest_base_time - earliest_base_time >
+	    sja1105_delta_to_ns(SJA1105_TAS_MAX_DELTA)) {
+		dev_err(ds->dev,
+			"Base times too far apart: min %llu max %llu\n",
+			earliest_base_time, latest_base_time);
+		return -ERANGE;
+	}
+
+	tas_data->earliest_base_time = earliest_base_time;
+	tas_data->max_cycle_time = max_cycle_time;
+
+	dev_dbg(ds->dev, "earliest base time %lld ns\n", earliest_base_time);
+	dev_dbg(ds->dev, "latest base time %lld ns\n", latest_base_time);
+	dev_dbg(ds->dev, "longest cycle time %lld ns\n", max_cycle_time);
+
+	return 0;
+}
+
 /* Lo and behold: the egress scheduler from hell.
  *
  * At the hardware level, the Time-Aware Shaper holds a global linear arrray of
@@ -99,7 +198,11 @@ static int sja1105_init_scheduling(struct sja1105_private *priv)
 	int num_cycles = 0;
 	int cycle = 0;
 	int i, k = 0;
-	int port;
+	int port, rc;
+
+	rc = sja1105_tas_set_runtime_params(priv);
+	if (rc < 0)
+		return rc;
 
 	/* Discard previous Schedule Table */
 	table = &priv->static_config.tables[BLK_IDX_SCHEDULE];
@@ -184,11 +287,13 @@ static int sja1105_init_scheduling(struct sja1105_private *priv)
 	schedule_entry_points = table->entries;
 
 	/* Finally start populating the static config tables */
-	schedule_entry_points_params->clksrc = SJA1105_TAS_CLKSRC_STANDALONE;
+	schedule_entry_points_params->clksrc = SJA1105_TAS_CLKSRC_PTP;
 	schedule_entry_points_params->actsubsch = num_cycles - 1;
 
 	for (port = 0; port < SJA1105_NUM_PORTS; port++) {
 		const struct tc_taprio_qopt_offload *offload;
+		/* Relative base time */
+		s64 rbt;
 
 		offload = tas_data->offload[port];
 		if (!offload)
@@ -196,15 +301,21 @@ static int sja1105_init_scheduling(struct sja1105_private *priv)
 
 		schedule_start_idx = k;
 		schedule_end_idx = k + offload->num_entries - 1;
-		/* TODO this is the base time for the port's subschedule,
-		 * relative to PTPSCHTM. But as we're using the standalone
-		 * clock source and not PTP clock as time reference, there's
-		 * little point in even trying to put more logic into this,
-		 * like preserving the phases between the subschedules of
-		 * different ports. We'll get all of that when switching to the
-		 * PTP clock source.
+		/* This is the base time expressed as a number of TAS ticks
+		 * relative to PTPSCHTM, which we'll (perhaps improperly) call
+		 * the operational base time.
+		 */
+		rbt = future_base_time(offload->base_time,
+				       offload->cycle_time,
+				       tas_data->earliest_base_time);
+		rbt -= tas_data->earliest_base_time;
+		/* UM10944.pdf 4.2.2. Schedule Entry Points table says that
+		 * delta cannot be zero, which is shitty. Advance all relative
+		 * base times by 1 TAS delta, so that even the earliest base
+		 * time becomes 1 in relative terms. Then start the operational
+		 * base time (PTPSCHTM) one TAS delta earlier than planned.
 		 */
-		entry_point_delta = 1;
+		entry_point_delta = ns_to_sja1105_delta(rbt) + 1;
 
 		schedule_entry_points[cycle].subschindx = cycle;
 		schedule_entry_points[cycle].delta = entry_point_delta;
@@ -352,7 +463,7 @@ int sja1105_setup_tc_taprio(struct dsa_switch *ds, int port,
 		if (rc < 0)
 			return rc;
 
-		return sja1105_static_config_reload(priv);
+		return sja1105_static_config_reload(priv, SJA1105_SCHEDULING);
 	}
 
 	/* The cycle time extension is the amount of time the last cycle from
@@ -400,11 +511,306 @@ int sja1105_setup_tc_taprio(struct dsa_switch *ds, int port,
 	if (rc < 0)
 		return rc;
 
-	return sja1105_static_config_reload(priv);
+	return sja1105_static_config_reload(priv, SJA1105_SCHEDULING);
+}
+
+static int sja1105_tas_check_running(struct sja1105_private *priv)
+{
+	struct sja1105_tas_data *tas_data = &priv->tas_data;
+	struct dsa_switch *ds = priv->ds;
+	struct sja1105_ptp_cmd cmd = {0};
+	int rc;
+
+	rc = sja1105_ptp_commit(ds, &cmd, SPI_READ);
+	if (rc < 0)
+		return rc;
+
+	if (cmd.ptpstrtsch == 1)
+		/* Schedule successfully started */
+		tas_data->state = SJA1105_TAS_STATE_RUNNING;
+	else if (cmd.ptpstopsch == 1)
+		/* Schedule is stopped */
+		tas_data->state = SJA1105_TAS_STATE_DISABLED;
+	else
+		/* Schedule is probably not configured with PTP clock source */
+		rc = -EINVAL;
+
+	return rc;
+}
+
+/* Write to PTPCLKCORP */
+static int sja1105_tas_adjust_drift(struct sja1105_private *priv,
+				    u64 correction)
+{
+	const struct sja1105_regs *regs = priv->info->regs;
+	u32 ptpclkcorp = ns_to_sja1105_ticks(correction);
+
+	return sja1105_xfer_u32(priv, SPI_WRITE, regs->ptpclkcorp,
+				&ptpclkcorp, NULL);
+}
+
+/* Write to PTPSCHTM */
+static int sja1105_tas_set_base_time(struct sja1105_private *priv,
+				     u64 base_time)
+{
+	const struct sja1105_regs *regs = priv->info->regs;
+	u64 ptpschtm = ns_to_sja1105_ticks(base_time);
+
+	return sja1105_xfer_u64(priv, SPI_WRITE, regs->ptpschtm,
+				&ptpschtm, NULL);
+}
+
+static int sja1105_tas_start(struct sja1105_private *priv)
+{
+	struct sja1105_tas_data *tas_data = &priv->tas_data;
+	struct sja1105_ptp_cmd *cmd = &priv->ptp_data.cmd;
+	struct dsa_switch *ds = priv->ds;
+	int rc;
+
+	dev_dbg(ds->dev, "Starting the TAS\n");
+
+	if (tas_data->state == SJA1105_TAS_STATE_ENABLED_NOT_RUNNING ||
+	    tas_data->state == SJA1105_TAS_STATE_RUNNING) {
+		dev_err(ds->dev, "TAS already started\n");
+		return -EINVAL;
+	}
+
+	cmd->ptpstrtsch = 1;
+	cmd->ptpstopsch = 0;
+
+	rc = sja1105_ptp_commit(ds, cmd, SPI_WRITE);
+	if (rc < 0)
+		return rc;
+
+	tas_data->state = SJA1105_TAS_STATE_ENABLED_NOT_RUNNING;
+
+	return 0;
+}
+
+static int sja1105_tas_stop(struct sja1105_private *priv)
+{
+	struct sja1105_tas_data *tas_data = &priv->tas_data;
+	struct sja1105_ptp_cmd *cmd = &priv->ptp_data.cmd;
+	struct dsa_switch *ds = priv->ds;
+	int rc;
+
+	dev_dbg(ds->dev, "Stopping the TAS\n");
+
+	if (tas_data->state == SJA1105_TAS_STATE_DISABLED) {
+		dev_err(ds->dev, "TAS already disabled\n");
+		return -EINVAL;
+	}
+
+	cmd->ptpstopsch = 1;
+	cmd->ptpstrtsch = 0;
+
+	rc = sja1105_ptp_commit(ds, cmd, SPI_WRITE);
+	if (rc < 0)
+		return rc;
+
+	tas_data->state = SJA1105_TAS_STATE_DISABLED;
+
+	return 0;
+}
+
+/* The schedule engine and the PTP clock are driven by the same oscillator, and
+ * they run in parallel. But whilst the PTP clock can keep an absolute
+ * time-of-day, the schedule engine is only running in 'ticks' (25 ticks make
+ * up a delta, which is 200ns), and wrapping around at the end of each cycle.
+ * The schedule engine is started when the PTP clock reaches the PTPSCHTM time
+ * (in PTP domain).
+ * Because the PTP clock can be rate-corrected (accelerated or slowed down) by
+ * a software servo, and the schedule engine clock runs in parallel to the PTP
+ * clock, there is logic internal to the switch that periodically keeps the
+ * schedule engine from drifting away. The frequency with which this internal
+ * syntonization happens is the PTP clock correction period (PTPCLKCORP). It is
+ * a value also in the PTP clock domain, and is also rate-corrected.
+ * To be precise, during a correction period, there is logic to determine by
+ * how many scheduler clock ticks has the PTP clock drifted. At the end of each
+ * correction period/beginning of new one, the length of a delta is shrunk or
+ * expanded with an integer number of ticks, compared with the typical 25.
+ * So a delta lasts for 200ns (or 25 ticks) only on average.
+ * Sometimes it is longer, sometimes it is shorter. The internal syntonization
+ * logic can adjust for at most 5 ticks each 20 ticks.
+ *
+ * The first implication is that you should choose your schedule correction
+ * period to be an integer multiple of the schedule length. Preferably one.
+ * In case there are schedules of multiple ports active, then the correction
+ * period needs to be a multiple of them all. Given the restriction that the
+ * cycle times have to be multiples of one another anyway, this means the
+ * correction period can simply be the largest cycle time, hence the current
+ * choice. This way, the updates are always synchronous to the transmission
+ * cycle, and therefore predictable.
+ *
+ * The second implication is that at the beginning of a correction period, the
+ * first few deltas will be modulated in time, until the schedule engine is
+ * properly phase-aligned with the PTP clock. For this reason, you should place
+ * your best-effort traffic at the beginning of a cycle, and your
+ * time-triggered traffic afterwards.
+ *
+ * The third implication is that once the schedule engine is started, it can
+ * only adjust for so much drift within a correction period. In the servo you
+ * can only change the PTPCLKRATE, but not step the clock (PTPCLKADD). If you
+ * want to do the latter, you need to stop and restart the schedule engine,
+ * which is what the state machine handles.
+ */
+static void sja1105_tas_state_machine(struct work_struct *work)
+{
+	struct sja1105_tas_data *tas_data = work_to_sja1105_tas(work);
+	struct sja1105_private *priv = tas_to_sja1105(tas_data);
+	struct sja1105_ptp_data *ptp_data = &priv->ptp_data;
+	struct timespec64 base_time_ts, now_ts;
+	struct dsa_switch *ds = priv->ds;
+	struct timespec64 diff;
+	s64 base_time, now;
+	int rc = 0;
+
+	mutex_lock(&ptp_data->lock);
+
+	switch (tas_data->state) {
+	case SJA1105_TAS_STATE_DISABLED:
+		/* Can't do anything at all if clock is still being stepped */
+		if (tas_data->last_op != SJA1105_PTP_ADJUSTFREQ)
+			break;
+
+		rc = sja1105_tas_adjust_drift(priv, tas_data->max_cycle_time);
+		if (rc < 0)
+			break;
+
+		rc = __sja1105_ptp_gettimex(ds, &now, NULL);
+		if (rc < 0)
+			break;
+
+		/* Plan to start the earliest schedule first. The others
+		 * will be started in hardware, by way of their respective
+		 * entry points delta.
+		 * Try our best to avoid fringe cases (race condition between
+		 * ptpschtm and ptpstrtsch) by pushing the oper_base_time at
+		 * least one second in the future from now. This is not ideal,
+		 * but this only needs to buy us time until the
+		 * sja1105_tas_start command below gets executed.
+		 */
+		base_time = future_base_time(tas_data->earliest_base_time,
+					     tas_data->max_cycle_time,
+					     now + 1ull * NSEC_PER_SEC);
+		base_time -= sja1105_delta_to_ns(1);
+
+		rc = sja1105_tas_set_base_time(priv, base_time);
+		if (rc < 0)
+			break;
+
+		tas_data->oper_base_time = base_time;
+
+		rc = sja1105_tas_start(priv);
+		if (rc < 0)
+			break;
+
+		base_time_ts = ns_to_timespec64(base_time);
+		now_ts = ns_to_timespec64(now);
+
+		dev_dbg(ds->dev, "OPER base time %lld.%09ld (now %lld.%09ld)\n",
+			base_time_ts.tv_sec, base_time_ts.tv_nsec,
+			now_ts.tv_sec, now_ts.tv_nsec);
+
+		break;
+
+	case SJA1105_TAS_STATE_ENABLED_NOT_RUNNING:
+		if (tas_data->last_op != SJA1105_PTP_ADJUSTFREQ) {
+			/* Clock was stepped.. bad news for TAS */
+			sja1105_tas_stop(priv);
+			break;
+		}
+
+		/* Check if TAS has actually started, by comparing the
+		 * scheduled start time with the SJA1105 PTP clock
+		 */
+		rc = __sja1105_ptp_gettimex(ds, &now, NULL);
+		if (rc < 0)
+			break;
+
+		if (now < tas_data->oper_base_time) {
+			/* TAS has not started yet */
+			diff = ns_to_timespec64(tas_data->oper_base_time - now);
+			dev_dbg(ds->dev, "time to start: [%lld.%09ld]",
+				diff.tv_sec, diff.tv_nsec);
+			break;
+		}
+
+		/* Time elapsed, what happened? */
+		rc = sja1105_tas_check_running(priv);
+		if (rc < 0)
+			break;
+
+		if (tas_data->state != SJA1105_TAS_STATE_RUNNING)
+			/* TAS has started */
+			dev_err(ds->dev,
+				"TAS not started despite time elapsed\n");
+
+		break;
+
+	case SJA1105_TAS_STATE_RUNNING:
+		/* Clock was stepped.. bad news for TAS */
+		if (tas_data->last_op != SJA1105_PTP_ADJUSTFREQ) {
+			sja1105_tas_stop(priv);
+			break;
+		}
+
+		rc = sja1105_tas_check_running(priv);
+		if (rc < 0)
+			break;
+
+		if (tas_data->state != SJA1105_TAS_STATE_RUNNING)
+			dev_err(ds->dev, "TAS surprisingly stopped\n");
+
+		break;
+
+	default:
+		if (net_ratelimit())
+			dev_err(ds->dev, "TAS in an invalid state (incorrect use of API)!\n");
+	}
+
+	if (rc && net_ratelimit())
+		dev_err(ds->dev, "An operation returned %d\n", rc);
+
+	mutex_unlock(&ptp_data->lock);
+}
+
+void sja1105_tas_clockstep(struct dsa_switch *ds)
+{
+	struct sja1105_private *priv = ds->priv;
+	struct sja1105_tas_data *tas_data = &priv->tas_data;
+
+	if (!tas_data->enabled)
+		return;
+
+	tas_data->last_op = SJA1105_PTP_CLOCKSTEP;
+	schedule_work(&tas_data->tas_work);
+}
+
+void sja1105_tas_adjfreq(struct dsa_switch *ds)
+{
+	struct sja1105_private *priv = ds->priv;
+	struct sja1105_tas_data *tas_data = &priv->tas_data;
+
+	if (!tas_data->enabled)
+		return;
+
+	/* No reason to schedule the workqueue, nothing changed */
+	if (tas_data->state == SJA1105_TAS_STATE_RUNNING)
+		return;
+
+	tas_data->last_op = SJA1105_PTP_ADJUSTFREQ;
+	schedule_work(&tas_data->tas_work);
 }
 
 void sja1105_tas_setup(struct dsa_switch *ds)
 {
+	struct sja1105_private *priv = ds->priv;
+	struct sja1105_tas_data *tas_data = &priv->tas_data;
+
+	INIT_WORK(&tas_data->tas_work, sja1105_tas_state_machine);
+	tas_data->state = SJA1105_TAS_STATE_DISABLED;
+	tas_data->last_op = SJA1105_PTP_NONE;
 }
 
 void sja1105_tas_teardown(struct dsa_switch *ds)
@@ -413,6 +819,8 @@ void sja1105_tas_teardown(struct dsa_switch *ds)
 	struct tc_taprio_qopt_offload *offload;
 	int port;
 
+	cancel_work_sync(&priv->tas_data.tas_work);
+
 	for (port = 0; port < SJA1105_NUM_PORTS; port++) {
 		offload = priv->tas_data.offload[port];
 		if (!offload)