armvirt: add SFP support patches for NXP Layerscape DPAA2 platforms
authorMathew McBride <matt@traverse.com.au>
Fri, 2 Jun 2023 04:14:14 +0000 (04:14 +0000)
committerPetr Štetiar <ynezz@true.cz>
Fri, 2 Jun 2023 09:36:23 +0000 (11:36 +0200)
This series resolves a long term issue with Phylink and SFPs on
DPAA2 hardware (LS1088,LS2088,LX2160) where a locking problem
prevented the system from rebooting. Patches solving this
issue existed prior to 6.2 but were not accepted upstream.

See the original series on patchwork:
https://patchwork.kernel.org/project/netdevbpf/cover/20221129141221.872653-1-vladimir.oltean@nxp.com/

And this thread on the Traverse forum:
https://forum.traverse.com.au/t/reboot-command-regression-from-5-10-to-5-15-kernel/133/12

Signed-off-by: Mathew McBride <matt@traverse.com.au>
12 files changed:
target/linux/armvirt/patches-6.1/701-v6.2-0001-net-dpaa2-eth-don-t-use-ENOTSUPP-error-code.patch [new file with mode: 0644]
target/linux/armvirt/patches-6.1/701-v6.2-0002-net-dpaa2-replace-dpaa2_mac_is_type_fixed-with-dpaa2.patch [new file with mode: 0644]
target/linux/armvirt/patches-6.1/701-v6.2-0003-net-dpaa2-mac-absorb-phylink_start-call-into-dpaa2_m.patch [new file with mode: 0644]
target/linux/armvirt/patches-6.1/701-v6.2-0004-net-dpaa2-mac-remove-defensive-check-in-dpaa2_mac_di.patch [new file with mode: 0644]
target/linux/armvirt/patches-6.1/701-v6.2-0005-net-dpaa2-eth-assign-priv-mac-after-dpaa2_mac_connec.patch [new file with mode: 0644]
target/linux/armvirt/patches-6.1/701-v6.2-0006-net-dpaa2-switch-assign-port_priv-mac-after-dpaa2_ma.patch [new file with mode: 0644]
target/linux/armvirt/patches-6.1/701-v6.2-0007-net-dpaa2-publish-MAC-stringset-to-ethtool-S-even-if.patch [new file with mode: 0644]
target/linux/armvirt/patches-6.1/701-v6.2-0008-net-dpaa2-switch-replace-direct-MAC-access-with-dpaa.patch [new file with mode: 0644]
target/linux/armvirt/patches-6.1/701-v6.2-0009-net-dpaa2-eth-connect-to-MAC-before-requesting-the-e.patch [new file with mode: 0644]
target/linux/armvirt/patches-6.1/701-v6.2-0010-net-dpaa2-eth-serialize-changes-to-priv-mac-with-a-m.patch [new file with mode: 0644]
target/linux/armvirt/patches-6.1/701-v6.2-0011-net-dpaa2-switch-serialize-changes-to-priv-mac-with.patch [new file with mode: 0644]
target/linux/armvirt/patches-6.1/701-v6.2-0012-net-dpaa2-mac-move-rtnl_lock-only-around-phylink.patch [new file with mode: 0644]

diff --git a/target/linux/armvirt/patches-6.1/701-v6.2-0001-net-dpaa2-eth-don-t-use-ENOTSUPP-error-code.patch b/target/linux/armvirt/patches-6.1/701-v6.2-0001-net-dpaa2-eth-don-t-use-ENOTSUPP-error-code.patch
new file mode 100644 (file)
index 0000000..6238c0a
--- /dev/null
@@ -0,0 +1,44 @@
+From f3763a0c1b07273218cbf5886bdf8df9df501111 Mon Sep 17 00:00:00 2001
+From: Vladimir Oltean <vladimir.oltean@nxp.com>
+Date: Tue, 29 Nov 2022 16:12:10 +0200
+Subject: [PATCH 03/14] net: dpaa2-eth: don't use -ENOTSUPP error code
+
+dpaa2_eth_setup_dpni() is called from the probe path and
+dpaa2_eth_set_link_ksettings() is propagated to user space.
+
+include/linux/errno.h says that ENOTSUPP is "Defined for the NFSv3
+protocol". Conventional wisdom has it to not use it in networking
+drivers. Replace it with -EOPNOTSUPP.
+
+Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
+Reviewed-by: Andrew Lunn <andrew@lunn.ch>
+Reviewed-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Tested-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Signed-off-by: Paolo Abeni <pabeni@redhat.com>
+---
+ drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c     | 2 +-
+ drivers/net/ethernet/freescale/dpaa2/dpaa2-ethtool.c | 2 +-
+ 2 files changed, 2 insertions(+), 2 deletions(-)
+
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
+@@ -3613,7 +3613,7 @@ static int dpaa2_eth_setup_dpni(struct f
+               dev_err(dev, "DPNI version %u.%u not supported, need >= %u.%u\n",
+                       priv->dpni_ver_major, priv->dpni_ver_minor,
+                       DPNI_VER_MAJOR, DPNI_VER_MINOR);
+-              err = -ENOTSUPP;
++              err = -EOPNOTSUPP;
+               goto close;
+       }
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-ethtool.c
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-ethtool.c
+@@ -118,7 +118,7 @@ dpaa2_eth_set_link_ksettings(struct net_
+       struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
+       if (!dpaa2_eth_is_type_phy(priv))
+-              return -ENOTSUPP;
++              return -EOPNOTSUPP;
+       return phylink_ethtool_ksettings_set(priv->mac->phylink, link_settings);
+ }
diff --git a/target/linux/armvirt/patches-6.1/701-v6.2-0002-net-dpaa2-replace-dpaa2_mac_is_type_fixed-with-dpaa2.patch b/target/linux/armvirt/patches-6.1/701-v6.2-0002-net-dpaa2-replace-dpaa2_mac_is_type_fixed-with-dpaa2.patch
new file mode 100644 (file)
index 0000000..501eaf4
--- /dev/null
@@ -0,0 +1,99 @@
+From 022a11062261dc4703da846d3bf4d194ef6bebf5 Mon Sep 17 00:00:00 2001
+From: Vladimir Oltean <vladimir.oltean@nxp.com>
+Date: Tue, 29 Nov 2022 16:12:11 +0200
+Subject: [PATCH 04/14] net: dpaa2: replace dpaa2_mac_is_type_fixed() with
+ dpaa2_mac_is_type_phy()
+
+dpaa2_mac_is_type_fixed() is a header with no implementation and no
+callers, which is referenced from the documentation though. It can be
+deleted.
+
+On the other hand, it would be useful to reuse the code between
+dpaa2_eth_is_type_phy() and dpaa2_switch_port_is_type_phy(). That common
+code should be called dpaa2_mac_is_type_phy(), so let's create that.
+
+The removal and the addition are merged into the same patch because,
+in fact, is_type_phy() is the logical opposite of is_type_fixed().
+
+Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
+Reviewed-by: Andrew Lunn <andrew@lunn.ch>
+Reviewed-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Tested-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Signed-off-by: Paolo Abeni <pabeni@redhat.com>
+---
+ .../ethernet/freescale/dpaa2/mac-phy-support.rst       |  9 ++++++---
+ drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h       |  7 +------
+ drivers/net/ethernet/freescale/dpaa2/dpaa2-mac.h       | 10 ++++++++--
+ drivers/net/ethernet/freescale/dpaa2/dpaa2-switch.h    |  7 +------
+ 4 files changed, 16 insertions(+), 17 deletions(-)
+
+--- a/Documentation/networking/device_drivers/ethernet/freescale/dpaa2/mac-phy-support.rst
++++ b/Documentation/networking/device_drivers/ethernet/freescale/dpaa2/mac-phy-support.rst
+@@ -181,10 +181,13 @@ when necessary using the below listed AP
+  - int dpaa2_mac_connect(struct dpaa2_mac *mac);
+  - void dpaa2_mac_disconnect(struct dpaa2_mac *mac);
+-A phylink integration is necessary only when the partner DPMAC is not of TYPE_FIXED.
+-One can check for this condition using the below API::
++A phylink integration is necessary only when the partner DPMAC is not of
++``TYPE_FIXED``. This means it is either of ``TYPE_PHY``, or of
++``TYPE_BACKPLANE`` (the difference being the two that in the ``TYPE_BACKPLANE``
++mode, the MC firmware does not access the PCS registers). One can check for
++this condition using the following helper::
+- - bool dpaa2_mac_is_type_fixed(struct fsl_mc_device *dpmac_dev,struct fsl_mc_io *mc_io);
++ - static inline bool dpaa2_mac_is_type_phy(struct dpaa2_mac *mac);
+ Before connection to a MAC, the caller must allocate and populate the
+ dpaa2_mac structure with the associated net_device, a pointer to the MC portal
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h
+@@ -733,12 +733,7 @@ static inline unsigned int dpaa2_eth_rx_
+ static inline bool dpaa2_eth_is_type_phy(struct dpaa2_eth_priv *priv)
+ {
+-      if (priv->mac &&
+-          (priv->mac->attr.link_type == DPMAC_LINK_TYPE_PHY ||
+-           priv->mac->attr.link_type == DPMAC_LINK_TYPE_BACKPLANE))
+-              return true;
+-
+-      return false;
++      return dpaa2_mac_is_type_phy(priv->mac);
+ }
+ static inline bool dpaa2_eth_has_mac(struct dpaa2_eth_priv *priv)
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-mac.h
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-mac.h
+@@ -30,8 +30,14 @@ struct dpaa2_mac {
+       struct phy *serdes_phy;
+ };
+-bool dpaa2_mac_is_type_fixed(struct fsl_mc_device *dpmac_dev,
+-                           struct fsl_mc_io *mc_io);
++static inline bool dpaa2_mac_is_type_phy(struct dpaa2_mac *mac)
++{
++      if (!mac)
++              return false;
++
++      return mac->attr.link_type == DPMAC_LINK_TYPE_PHY ||
++             mac->attr.link_type == DPMAC_LINK_TYPE_BACKPLANE;
++}
+ int dpaa2_mac_open(struct dpaa2_mac *mac);
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-switch.h
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-switch.h
+@@ -230,12 +230,7 @@ static inline bool dpaa2_switch_supports
+ static inline bool
+ dpaa2_switch_port_is_type_phy(struct ethsw_port_priv *port_priv)
+ {
+-      if (port_priv->mac &&
+-          (port_priv->mac->attr.link_type == DPMAC_LINK_TYPE_PHY ||
+-           port_priv->mac->attr.link_type == DPMAC_LINK_TYPE_BACKPLANE))
+-              return true;
+-
+-      return false;
++      return dpaa2_mac_is_type_phy(port_priv->mac);
+ }
+ static inline bool dpaa2_switch_port_has_mac(struct ethsw_port_priv *port_priv)
diff --git a/target/linux/armvirt/patches-6.1/701-v6.2-0003-net-dpaa2-mac-absorb-phylink_start-call-into-dpaa2_m.patch b/target/linux/armvirt/patches-6.1/701-v6.2-0003-net-dpaa2-mac-absorb-phylink_start-call-into-dpaa2_m.patch
new file mode 100644 (file)
index 0000000..e84195f
--- /dev/null
@@ -0,0 +1,88 @@
+From 97c07369ab8bf9895e05d4b468f18e6567263154 Mon Sep 17 00:00:00 2001
+From: Vladimir Oltean <vladimir.oltean@nxp.com>
+Date: Tue, 29 Nov 2022 16:12:12 +0200
+Subject: [PATCH 05/14] net: dpaa2-mac: absorb phylink_start() call into
+ dpaa2_mac_start()
+
+The phylink handling is intended to be hidden inside the dpaa2_mac
+object. Move the phylink_start() call into dpaa2_mac_start(), and
+phylink_stop() into dpaa2_mac_stop().
+
+Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
+Reviewed-by: Andrew Lunn <andrew@lunn.ch>
+Reviewed-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Tested-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Signed-off-by: Paolo Abeni <pabeni@redhat.com>
+---
+ drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c    | 5 +----
+ drivers/net/ethernet/freescale/dpaa2/dpaa2-mac.c    | 8 ++++++++
+ drivers/net/ethernet/freescale/dpaa2/dpaa2-switch.c | 5 +----
+ 3 files changed, 10 insertions(+), 8 deletions(-)
+
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
+@@ -2082,10 +2082,8 @@ static int dpaa2_eth_open(struct net_dev
+               goto enable_err;
+       }
+-      if (dpaa2_eth_is_type_phy(priv)) {
++      if (dpaa2_eth_is_type_phy(priv))
+               dpaa2_mac_start(priv->mac);
+-              phylink_start(priv->mac->phylink);
+-      }
+       return 0;
+@@ -2159,7 +2157,6 @@ static int dpaa2_eth_stop(struct net_dev
+       int retries = 10;
+       if (dpaa2_eth_is_type_phy(priv)) {
+-              phylink_stop(priv->mac->phylink);
+               dpaa2_mac_stop(priv->mac);
+       } else {
+               netif_tx_stop_all_queues(net_dev);
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-mac.c
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-mac.c
+@@ -336,12 +336,20 @@ static void dpaa2_mac_set_supported_inte
+ void dpaa2_mac_start(struct dpaa2_mac *mac)
+ {
++      ASSERT_RTNL();
++
+       if (mac->serdes_phy)
+               phy_power_on(mac->serdes_phy);
++
++      phylink_start(mac->phylink);
+ }
+ void dpaa2_mac_stop(struct dpaa2_mac *mac)
+ {
++      ASSERT_RTNL();
++
++      phylink_stop(mac->phylink);
++
+       if (mac->serdes_phy)
+               phy_power_off(mac->serdes_phy);
+ }
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-switch.c
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-switch.c
+@@ -703,10 +703,8 @@ static int dpaa2_switch_port_open(struct
+       dpaa2_switch_enable_ctrl_if_napi(ethsw);
+-      if (dpaa2_switch_port_is_type_phy(port_priv)) {
++      if (dpaa2_switch_port_is_type_phy(port_priv))
+               dpaa2_mac_start(port_priv->mac);
+-              phylink_start(port_priv->mac->phylink);
+-      }
+       return 0;
+ }
+@@ -718,7 +716,6 @@ static int dpaa2_switch_port_stop(struct
+       int err;
+       if (dpaa2_switch_port_is_type_phy(port_priv)) {
+-              phylink_stop(port_priv->mac->phylink);
+               dpaa2_mac_stop(port_priv->mac);
+       } else {
+               netif_tx_stop_all_queues(netdev);
diff --git a/target/linux/armvirt/patches-6.1/701-v6.2-0004-net-dpaa2-mac-remove-defensive-check-in-dpaa2_mac_di.patch b/target/linux/armvirt/patches-6.1/701-v6.2-0004-net-dpaa2-mac-remove-defensive-check-in-dpaa2_mac_di.patch
new file mode 100644 (file)
index 0000000..c302835
--- /dev/null
@@ -0,0 +1,50 @@
+From 095ef388f714d622aa503fcccf20dc4095b72762 Mon Sep 17 00:00:00 2001
+From: Vladimir Oltean <vladimir.oltean@nxp.com>
+Date: Tue, 29 Nov 2022 16:12:13 +0200
+Subject: [PATCH 06/14] net: dpaa2-mac: remove defensive check in
+ dpaa2_mac_disconnect()
+
+dpaa2_mac_disconnect() will only be called with a NULL mac->phylink if
+dpaa2_mac_connect() failed, or was never called.
+
+The callers are these:
+
+dpaa2_eth_disconnect_mac():
+
+       if (dpaa2_eth_is_type_phy(priv))
+               dpaa2_mac_disconnect(priv->mac);
+
+dpaa2_switch_port_disconnect_mac():
+
+       if (dpaa2_switch_port_is_type_phy(port_priv))
+               dpaa2_mac_disconnect(port_priv->mac);
+
+priv->mac can be NULL, but in that case, dpaa2_eth_is_type_phy() returns
+false, and dpaa2_mac_disconnect() is never called. Similar for
+dpaa2-switch.
+
+When priv->mac is non-NULL, it means that dpaa2_mac_connect() returned
+zero (success), and therefore, priv->mac->phylink is also a valid
+pointer.
+
+Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
+Reviewed-by: Andrew Lunn <andrew@lunn.ch>
+Reviewed-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Tested-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Signed-off-by: Paolo Abeni <pabeni@redhat.com>
+---
+ drivers/net/ethernet/freescale/dpaa2/dpaa2-mac.c | 3 ---
+ 1 file changed, 3 deletions(-)
+
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-mac.c
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-mac.c
+@@ -446,9 +446,6 @@ err_pcs_destroy:
+ void dpaa2_mac_disconnect(struct dpaa2_mac *mac)
+ {
+-      if (!mac->phylink)
+-              return;
+-
+       phylink_disconnect_phy(mac->phylink);
+       phylink_destroy(mac->phylink);
+       dpaa2_pcs_destroy(mac);
diff --git a/target/linux/armvirt/patches-6.1/701-v6.2-0005-net-dpaa2-eth-assign-priv-mac-after-dpaa2_mac_connec.patch b/target/linux/armvirt/patches-6.1/701-v6.2-0005-net-dpaa2-eth-assign-priv-mac-after-dpaa2_mac_connec.patch
new file mode 100644 (file)
index 0000000..8c7ed6c
--- /dev/null
@@ -0,0 +1,101 @@
+From 06efc9b8a1360cad83cae6e71558e5458cc1fbf3 Mon Sep 17 00:00:00 2001
+From: Vladimir Oltean <vladimir.oltean@nxp.com>
+Date: Tue, 29 Nov 2022 16:12:14 +0200
+Subject: [PATCH 07/14] net: dpaa2-eth: assign priv->mac after
+ dpaa2_mac_connect() call
+
+There are 2 requirements for correct code:
+
+- Any time the driver accesses the priv->mac pointer at runtime, it
+  either holds NULL to indicate a DPNI-DPNI connection (or unconnected
+  DPNI), or a struct dpaa2_mac whose phylink instance was fully
+  initialized (created and connected to the PHY). No changes are made to
+  priv->mac while it is being used. Currently, rtnl_lock() watches over
+  the call to dpaa2_eth_connect_mac(), so it serves the purpose of
+  serializing this with all readers of priv->mac.
+
+- dpaa2_mac_connect() should run unlocked, because inside it are 2
+  phylink calls with incompatible locking requirements: phylink_create()
+  requires that the rtnl_mutex isn't held, and phylink_fwnode_phy_connect()
+  requires that the rtnl_mutex is held. The only way to solve those
+  contradictory requirements is to let dpaa2_mac_connect() take
+  rtnl_lock() when it needs to.
+
+To solve both requirements, we need to identify the writer side of the
+priv->mac pointer, which can be wrapped in a mutex private to the driver
+in a future patch. The dpaa2_mac_connect() cannot be part of the writer
+side critical section, because of an AB/BA deadlock with rtnl_lock().
+
+So the strategy needs to be that where we prepare the DPMAC by calling
+dpaa2_mac_connect(), and only make priv->mac point to it once it's fully
+prepared. This ensures that the writer side critical section has the
+absolute minimum surface it can.
+
+The reverse strategy is adopted in the dpaa2_eth_disconnect_mac() code
+path. This makes sure that priv->mac is NULL when we start tearing down
+the DPMAC that we disconnected from, and concurrent code will simply not
+see it.
+
+No locking changes in this patch (concurrent code is still blocked by
+the rtnl_mutex).
+
+Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
+Reviewed-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Tested-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Signed-off-by: Paolo Abeni <pabeni@redhat.com>
+---
+ .../net/ethernet/freescale/dpaa2/dpaa2-eth.c  | 21 +++++++++++--------
+ 1 file changed, 12 insertions(+), 9 deletions(-)
+
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
+@@ -4443,9 +4443,8 @@ static int dpaa2_eth_connect_mac(struct
+       err = dpaa2_mac_open(mac);
+       if (err)
+               goto err_free_mac;
+-      priv->mac = mac;
+-      if (dpaa2_eth_is_type_phy(priv)) {
++      if (dpaa2_mac_is_type_phy(mac)) {
+               err = dpaa2_mac_connect(mac);
+               if (err && err != -EPROBE_DEFER)
+                       netdev_err(priv->net_dev, "Error connecting to the MAC endpoint: %pe",
+@@ -4454,11 +4453,12 @@ static int dpaa2_eth_connect_mac(struct
+                       goto err_close_mac;
+       }
++      priv->mac = mac;
++
+       return 0;
+ err_close_mac:
+       dpaa2_mac_close(mac);
+-      priv->mac = NULL;
+ err_free_mac:
+       kfree(mac);
+       return err;
+@@ -4466,15 +4466,18 @@ err_free_mac:
+ static void dpaa2_eth_disconnect_mac(struct dpaa2_eth_priv *priv)
+ {
+-      if (dpaa2_eth_is_type_phy(priv))
+-              dpaa2_mac_disconnect(priv->mac);
++      struct dpaa2_mac *mac = priv->mac;
++
++      priv->mac = NULL;
+-      if (!dpaa2_eth_has_mac(priv))
++      if (!mac)
+               return;
+-      dpaa2_mac_close(priv->mac);
+-      kfree(priv->mac);
+-      priv->mac = NULL;
++      if (dpaa2_mac_is_type_phy(mac))
++              dpaa2_mac_disconnect(mac);
++
++      dpaa2_mac_close(mac);
++      kfree(mac);
+ }
+ static irqreturn_t dpni_irq0_handler_thread(int irq_num, void *arg)
diff --git a/target/linux/armvirt/patches-6.1/701-v6.2-0006-net-dpaa2-switch-assign-port_priv-mac-after-dpaa2_ma.patch b/target/linux/armvirt/patches-6.1/701-v6.2-0006-net-dpaa2-switch-assign-port_priv-mac-after-dpaa2_ma.patch
new file mode 100644 (file)
index 0000000..e636549
--- /dev/null
@@ -0,0 +1,73 @@
+From a5e7f7e277bd4403c45c1c7922d56d0eb08dbc7c Mon Sep 17 00:00:00 2001
+From: Vladimir Oltean <vladimir.oltean@nxp.com>
+Date: Tue, 29 Nov 2022 16:12:15 +0200
+Subject: [PATCH 08/14] net: dpaa2-switch: assign port_priv->mac after
+ dpaa2_mac_connect() call
+
+The dpaa2-switch has the exact same locking requirements when connected
+to a DPMAC, so it needs port_priv->mac to always point either to NULL,
+or to a DPMAC with a fully initialized phylink instance.
+
+Make the same preparatory change in the dpaa2-switch driver as in the
+dpaa2-eth one.
+
+Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
+Reviewed-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Tested-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Signed-off-by: Paolo Abeni <pabeni@redhat.com>
+---
+ .../ethernet/freescale/dpaa2/dpaa2-switch.c   | 21 +++++++++++--------
+ 1 file changed, 12 insertions(+), 9 deletions(-)
+
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-switch.c
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-switch.c
+@@ -1450,9 +1450,8 @@ static int dpaa2_switch_port_connect_mac
+       err = dpaa2_mac_open(mac);
+       if (err)
+               goto err_free_mac;
+-      port_priv->mac = mac;
+-      if (dpaa2_switch_port_is_type_phy(port_priv)) {
++      if (dpaa2_mac_is_type_phy(mac)) {
+               err = dpaa2_mac_connect(mac);
+               if (err) {
+                       netdev_err(port_priv->netdev,
+@@ -1462,11 +1461,12 @@ static int dpaa2_switch_port_connect_mac
+               }
+       }
++      port_priv->mac = mac;
++
+       return 0;
+ err_close_mac:
+       dpaa2_mac_close(mac);
+-      port_priv->mac = NULL;
+ err_free_mac:
+       kfree(mac);
+       return err;
+@@ -1474,15 +1474,18 @@ err_free_mac:
+ static void dpaa2_switch_port_disconnect_mac(struct ethsw_port_priv *port_priv)
+ {
+-      if (dpaa2_switch_port_is_type_phy(port_priv))
+-              dpaa2_mac_disconnect(port_priv->mac);
++      struct dpaa2_mac *mac = port_priv->mac;
++
++      port_priv->mac = NULL;
+-      if (!dpaa2_switch_port_has_mac(port_priv))
++      if (!mac)
+               return;
+-      dpaa2_mac_close(port_priv->mac);
+-      kfree(port_priv->mac);
+-      port_priv->mac = NULL;
++      if (dpaa2_mac_is_type_phy(mac))
++              dpaa2_mac_disconnect(mac);
++
++      dpaa2_mac_close(mac);
++      kfree(mac);
+ }
+ static irqreturn_t dpaa2_switch_irq0_handler_thread(int irq_num, void *arg)
diff --git a/target/linux/armvirt/patches-6.1/701-v6.2-0007-net-dpaa2-publish-MAC-stringset-to-ethtool-S-even-if.patch b/target/linux/armvirt/patches-6.1/701-v6.2-0007-net-dpaa2-publish-MAC-stringset-to-ethtool-S-even-if.patch
new file mode 100644 (file)
index 0000000..c790ba1
--- /dev/null
@@ -0,0 +1,111 @@
+From ce44b6ed9ee65efa9b3025552c513842eabcab88 Mon Sep 17 00:00:00 2001
+From: Vladimir Oltean <vladimir.oltean@nxp.com>
+Date: Tue, 29 Nov 2022 16:12:16 +0200
+Subject: [PATCH 09/14] net: dpaa2: publish MAC stringset to ethtool -S even if
+ MAC is missing
+
+DPNIs and DPSW objects can connect and disconnect at runtime from DPMAC
+objects on the same fsl-mc bus. The DPMAC object also holds "ethtool -S"
+unstructured counters. Those counters are only shown for the entity
+owning the netdev (DPNI, DPSW) if it's connected to a DPMAC.
+
+The ethtool stringset code path is split into multiple callbacks, but
+currently, connecting and disconnecting the DPMAC takes the rtnl_lock().
+This blocks the entire ethtool code path from running, see
+ethnl_default_doit() -> rtnl_lock() -> ops->prepare_data() ->
+strset_prepare_data().
+
+This is going to be a problem if we are going to no longer require
+rtnl_lock() when connecting/disconnecting the DPMAC, because the DPMAC
+could appear between ops->get_sset_count() and ops->get_strings().
+If it appears out of the blue, we will provide a stringset into an array
+that was dimensioned thinking the DPMAC wouldn't be there => array
+accessed out of bounds.
+
+There isn't really a good way to work around that, and I don't want to
+put too much pressure on the ethtool framework by playing locking games.
+Just make the DPMAC counters be always available. They'll be zeroes if
+the DPNI or DPSW isn't connected to a DPMAC.
+
+Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
+Reviewed-by: Andrew Lunn <andrew@lunn.ch>
+Reviewed-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Tested-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Signed-off-by: Paolo Abeni <pabeni@redhat.com>
+---
+ drivers/net/ethernet/freescale/dpaa2/dpaa2-ethtool.c | 12 +++---------
+ .../ethernet/freescale/dpaa2/dpaa2-switch-ethtool.c  | 11 ++---------
+ 2 files changed, 5 insertions(+), 18 deletions(-)
+
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-ethtool.c
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-ethtool.c
+@@ -186,7 +186,6 @@ static int dpaa2_eth_set_pauseparam(stru
+ static void dpaa2_eth_get_strings(struct net_device *netdev, u32 stringset,
+                                 u8 *data)
+ {
+-      struct dpaa2_eth_priv *priv = netdev_priv(netdev);
+       u8 *p = data;
+       int i;
+@@ -200,22 +199,17 @@ static void dpaa2_eth_get_strings(struct
+                       strscpy(p, dpaa2_ethtool_extras[i], ETH_GSTRING_LEN);
+                       p += ETH_GSTRING_LEN;
+               }
+-              if (dpaa2_eth_has_mac(priv))
+-                      dpaa2_mac_get_strings(p);
++              dpaa2_mac_get_strings(p);
+               break;
+       }
+ }
+ static int dpaa2_eth_get_sset_count(struct net_device *net_dev, int sset)
+ {
+-      int num_ss_stats = DPAA2_ETH_NUM_STATS + DPAA2_ETH_NUM_EXTRA_STATS;
+-      struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
+-
+       switch (sset) {
+       case ETH_SS_STATS: /* ethtool_get_stats(), ethtool_get_drvinfo() */
+-              if (dpaa2_eth_has_mac(priv))
+-                      num_ss_stats += dpaa2_mac_get_sset_count();
+-              return num_ss_stats;
++              return DPAA2_ETH_NUM_STATS + DPAA2_ETH_NUM_EXTRA_STATS +
++                     dpaa2_mac_get_sset_count();
+       default:
+               return -EOPNOTSUPP;
+       }
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-switch-ethtool.c
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-switch-ethtool.c
+@@ -145,14 +145,9 @@ dpaa2_switch_set_link_ksettings(struct n
+ static int
+ dpaa2_switch_ethtool_get_sset_count(struct net_device *netdev, int sset)
+ {
+-      struct ethsw_port_priv *port_priv = netdev_priv(netdev);
+-      int num_ss_stats = DPAA2_SWITCH_NUM_COUNTERS;
+-
+       switch (sset) {
+       case ETH_SS_STATS:
+-              if (port_priv->mac)
+-                      num_ss_stats += dpaa2_mac_get_sset_count();
+-              return num_ss_stats;
++              return DPAA2_SWITCH_NUM_COUNTERS + dpaa2_mac_get_sset_count();
+       default:
+               return -EOPNOTSUPP;
+       }
+@@ -161,7 +156,6 @@ dpaa2_switch_ethtool_get_sset_count(stru
+ static void dpaa2_switch_ethtool_get_strings(struct net_device *netdev,
+                                            u32 stringset, u8 *data)
+ {
+-      struct ethsw_port_priv *port_priv = netdev_priv(netdev);
+       u8 *p = data;
+       int i;
+@@ -172,8 +166,7 @@ static void dpaa2_switch_ethtool_get_str
+                              ETH_GSTRING_LEN);
+                       p += ETH_GSTRING_LEN;
+               }
+-              if (port_priv->mac)
+-                      dpaa2_mac_get_strings(p);
++              dpaa2_mac_get_strings(p);
+               break;
+       }
+ }
diff --git a/target/linux/armvirt/patches-6.1/701-v6.2-0008-net-dpaa2-switch-replace-direct-MAC-access-with-dpaa.patch b/target/linux/armvirt/patches-6.1/701-v6.2-0008-net-dpaa2-switch-replace-direct-MAC-access-with-dpaa.patch
new file mode 100644 (file)
index 0000000..0663bf6
--- /dev/null
@@ -0,0 +1,28 @@
+From c838d9fd7e6ba9ddd6006bf0a296396266e9f121 Mon Sep 17 00:00:00 2001
+From: Vladimir Oltean <vladimir.oltean@nxp.com>
+Date: Tue, 29 Nov 2022 16:12:17 +0200
+Subject: [PATCH 10/14] net: dpaa2-switch replace direct MAC access with
+ dpaa2_switch_port_has_mac()
+
+The helper function will gain a lockdep annotation in a future patch.
+Make sure to benefit from it.
+
+Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
+Reviewed-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Tested-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Signed-off-by: Paolo Abeni <pabeni@redhat.com>
+---
+ drivers/net/ethernet/freescale/dpaa2/dpaa2-switch-ethtool.c | 2 +-
+ 1 file changed, 1 insertion(+), 1 deletion(-)
+
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-switch-ethtool.c
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-switch-ethtool.c
+@@ -189,7 +189,7 @@ static void dpaa2_switch_ethtool_get_sta
+                                  dpaa2_switch_ethtool_counters[i].name, err);
+       }
+-      if (port_priv->mac)
++      if (dpaa2_switch_port_has_mac(port_priv))
+               dpaa2_mac_get_ethtool_stats(port_priv->mac, data + i);
+ }
diff --git a/target/linux/armvirt/patches-6.1/701-v6.2-0009-net-dpaa2-eth-connect-to-MAC-before-requesting-the-e.patch b/target/linux/armvirt/patches-6.1/701-v6.2-0009-net-dpaa2-eth-connect-to-MAC-before-requesting-the-e.patch
new file mode 100644 (file)
index 0000000..37831f2
--- /dev/null
@@ -0,0 +1,93 @@
+From e0ea63162cb5f1ca7f844d6ef5fc4079448ee2d5 Mon Sep 17 00:00:00 2001
+From: Vladimir Oltean <vladimir.oltean@nxp.com>
+Date: Tue, 29 Nov 2022 16:12:18 +0200
+Subject: [PATCH 11/14] net: dpaa2-eth: connect to MAC before requesting the
+ "endpoint changed" IRQ
+
+dpaa2_eth_connect_mac() is called both from dpaa2_eth_probe() and from
+dpni_irq0_handler_thread().
+
+It could happen that the DPNI gets connected to a DPMAC on the fsl-mc
+bus exactly during probe, as soon as the "endpoint change" interrupt is
+requested in dpaa2_eth_setup_irqs(). This will cause the
+dpni_irq0_handler_thread() to register a phylink instance for that DPMAC.
+
+Then, the probing function will also try to register a phylink instance
+for the same DPMAC, operation which should fail (and this will fail the
+probing of the driver).
+
+Reorder dpaa2_eth_setup_irqs() and dpaa2_eth_connect_mac(), such that
+dpni_irq0_handler_thread() never races with the DPMAC-related portion of
+the probing path.
+
+Also reorder dpaa2_eth_disconnect_mac() to be in the mirror position of
+dpaa2_eth_connect_mac() in the teardown path.
+
+Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
+Reviewed-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Tested-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Signed-off-by: Paolo Abeni <pabeni@redhat.com>
+---
+ .../net/ethernet/freescale/dpaa2/dpaa2-eth.c   | 18 +++++++++---------
+ 1 file changed, 9 insertions(+), 9 deletions(-)
+
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
+@@ -4710,6 +4710,10 @@ static int dpaa2_eth_probe(struct fsl_mc
+       }
+ #endif
++      err = dpaa2_eth_connect_mac(priv);
++      if (err)
++              goto err_connect_mac;
++
+       err = dpaa2_eth_setup_irqs(dpni_dev);
+       if (err) {
+               netdev_warn(net_dev, "Failed to set link interrupt, fall back to polling\n");
+@@ -4722,10 +4726,6 @@ static int dpaa2_eth_probe(struct fsl_mc
+               priv->do_link_poll = true;
+       }
+-      err = dpaa2_eth_connect_mac(priv);
+-      if (err)
+-              goto err_connect_mac;
+-
+       err = dpaa2_eth_dl_alloc(priv);
+       if (err)
+               goto err_dl_register;
+@@ -4759,13 +4759,13 @@ err_dl_port_add:
+ err_dl_trap_register:
+       dpaa2_eth_dl_free(priv);
+ err_dl_register:
+-      dpaa2_eth_disconnect_mac(priv);
+-err_connect_mac:
+       if (priv->do_link_poll)
+               kthread_stop(priv->poll_thread);
+       else
+               fsl_mc_free_irqs(dpni_dev);
+ err_poll_thread:
++      dpaa2_eth_disconnect_mac(priv);
++err_connect_mac:
+       dpaa2_eth_free_rings(priv);
+ err_alloc_rings:
+ err_csum:
+@@ -4813,9 +4813,6 @@ static int dpaa2_eth_remove(struct fsl_m
+ #endif
+       unregister_netdev(net_dev);
+-      rtnl_lock();
+-      dpaa2_eth_disconnect_mac(priv);
+-      rtnl_unlock();
+       dpaa2_eth_dl_port_del(priv);
+       dpaa2_eth_dl_traps_unregister(priv);
+@@ -4826,6 +4823,9 @@ static int dpaa2_eth_remove(struct fsl_m
+       else
+               fsl_mc_free_irqs(ls_dev);
++      rtnl_lock();
++      dpaa2_eth_disconnect_mac(priv);
++      rtnl_unlock();
+       dpaa2_eth_free_rings(priv);
+       free_percpu(priv->fd);
+       free_percpu(priv->sgt_cache);
diff --git a/target/linux/armvirt/patches-6.1/701-v6.2-0010-net-dpaa2-eth-serialize-changes-to-priv-mac-with-a-m.patch b/target/linux/armvirt/patches-6.1/701-v6.2-0010-net-dpaa2-eth-serialize-changes-to-priv-mac-with-a-m.patch
new file mode 100644 (file)
index 0000000..89f5841
--- /dev/null
@@ -0,0 +1,320 @@
+From 5e448a17dfa2e95166534df7f677a3694ef6187d Mon Sep 17 00:00:00 2001
+From: Vladimir Oltean <vladimir.oltean@nxp.com>
+Date: Tue, 29 Nov 2022 16:12:19 +0200
+Subject: [PATCH 12/14] net: dpaa2-eth: serialize changes to priv->mac with a
+ mutex
+
+The dpaa2 architecture permits dynamic connections between objects on
+the fsl-mc bus, specifically between a DPNI object (represented by a
+struct net_device) and a DPMAC object (represented by a struct phylink).
+
+The DPNI driver is notified when those connections are created/broken
+through the dpni_irq0_handler_thread() method. To ensure that ethtool
+operations, as well as netdev up/down operations serialize with the
+connection/disconnection of the DPNI with a DPMAC,
+dpni_irq0_handler_thread() takes the rtnl_lock() to block those other
+operations from taking place.
+
+There is code called by dpaa2_mac_connect() which wants to acquire the
+rtnl_mutex once again, see phylink_create() -> phylink_register_sfp() ->
+sfp_bus_add_upstream() -> rtnl_lock(). So the strategy doesn't quite
+work out, even though it's fairly simple.
+
+Create a different strategy, where all code paths in the dpaa2-eth
+driver access priv->mac only while they are holding priv->mac_lock.
+The phylink instance is not created or connected to the PHY under the
+priv->mac_lock, but only assigned to priv->mac then. This will eliminate
+the reliance on the rtnl_mutex.
+
+Add lockdep annotations and put comments where holding the lock is not
+necessary, and priv->mac can be dereferenced freely.
+
+Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
+Reviewed-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Tested-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Signed-off-by: Paolo Abeni <pabeni@redhat.com>
+---
+ .../net/ethernet/freescale/dpaa2/dpaa2-eth.c  | 43 ++++++++++++--
+ .../net/ethernet/freescale/dpaa2/dpaa2-eth.h  |  6 ++
+ .../ethernet/freescale/dpaa2/dpaa2-ethtool.c  | 58 +++++++++++++++----
+ 3 files changed, 91 insertions(+), 16 deletions(-)
+
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
+@@ -2020,8 +2020,11 @@ static int dpaa2_eth_link_state_update(s
+       /* When we manage the MAC/PHY using phylink there is no need
+        * to manually update the netif_carrier.
++       * We can avoid locking because we are called from the "link changed"
++       * IRQ handler, which is the same as the "endpoint changed" IRQ handler
++       * (the writer to priv->mac), so we cannot race with it.
+        */
+-      if (dpaa2_eth_is_type_phy(priv))
++      if (dpaa2_mac_is_type_phy(priv->mac))
+               goto out;
+       /* Chech link state; speed / duplex changes are not treated yet */
+@@ -2060,6 +2063,8 @@ static int dpaa2_eth_open(struct net_dev
+                          priv->dpbp_dev->obj_desc.id, priv->bpid);
+       }
++      mutex_lock(&priv->mac_lock);
++
+       if (!dpaa2_eth_is_type_phy(priv)) {
+               /* We'll only start the txqs when the link is actually ready;
+                * make sure we don't race against the link up notification,
+@@ -2078,6 +2083,7 @@ static int dpaa2_eth_open(struct net_dev
+       err = dpni_enable(priv->mc_io, 0, priv->mc_token);
+       if (err < 0) {
++              mutex_unlock(&priv->mac_lock);
+               netdev_err(net_dev, "dpni_enable() failed\n");
+               goto enable_err;
+       }
+@@ -2085,6 +2091,8 @@ static int dpaa2_eth_open(struct net_dev
+       if (dpaa2_eth_is_type_phy(priv))
+               dpaa2_mac_start(priv->mac);
++      mutex_unlock(&priv->mac_lock);
++
+       return 0;
+ enable_err:
+@@ -2156,6 +2164,8 @@ static int dpaa2_eth_stop(struct net_dev
+       int dpni_enabled = 0;
+       int retries = 10;
++      mutex_lock(&priv->mac_lock);
++
+       if (dpaa2_eth_is_type_phy(priv)) {
+               dpaa2_mac_stop(priv->mac);
+       } else {
+@@ -2163,6 +2173,8 @@ static int dpaa2_eth_stop(struct net_dev
+               netif_carrier_off(net_dev);
+       }
++      mutex_unlock(&priv->mac_lock);
++
+       /* On dpni_disable(), the MC firmware will:
+        * - stop MAC Rx and wait for all Rx frames to be enqueued to software
+        * - cut off WRIOP dequeues from egress FQs and wait until transmission
+@@ -2488,12 +2500,20 @@ static int dpaa2_eth_ts_ioctl(struct net
+ static int dpaa2_eth_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
+ {
+       struct dpaa2_eth_priv *priv = netdev_priv(dev);
++      int err;
+       if (cmd == SIOCSHWTSTAMP)
+               return dpaa2_eth_ts_ioctl(dev, rq, cmd);
+-      if (dpaa2_eth_is_type_phy(priv))
+-              return phylink_mii_ioctl(priv->mac->phylink, rq, cmd);
++      mutex_lock(&priv->mac_lock);
++
++      if (dpaa2_eth_is_type_phy(priv)) {
++              err = phylink_mii_ioctl(priv->mac->phylink, rq, cmd);
++              mutex_unlock(&priv->mac_lock);
++              return err;
++      }
++
++      mutex_unlock(&priv->mac_lock);
+       return -EOPNOTSUPP;
+ }
+@@ -4453,7 +4473,9 @@ static int dpaa2_eth_connect_mac(struct
+                       goto err_close_mac;
+       }
++      mutex_lock(&priv->mac_lock);
+       priv->mac = mac;
++      mutex_unlock(&priv->mac_lock);
+       return 0;
+@@ -4466,9 +4488,12 @@ err_free_mac:
+ static void dpaa2_eth_disconnect_mac(struct dpaa2_eth_priv *priv)
+ {
+-      struct dpaa2_mac *mac = priv->mac;
++      struct dpaa2_mac *mac;
++      mutex_lock(&priv->mac_lock);
++      mac = priv->mac;
+       priv->mac = NULL;
++      mutex_unlock(&priv->mac_lock);
+       if (!mac)
+               return;
+@@ -4487,6 +4512,7 @@ static irqreturn_t dpni_irq0_handler_thr
+       struct fsl_mc_device *dpni_dev = to_fsl_mc_device(dev);
+       struct net_device *net_dev = dev_get_drvdata(dev);
+       struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
++      bool had_mac;
+       int err;
+       err = dpni_get_irq_status(dpni_dev->mc_io, 0, dpni_dev->mc_handle,
+@@ -4504,7 +4530,12 @@ static irqreturn_t dpni_irq0_handler_thr
+               dpaa2_eth_update_tx_fqids(priv);
+               rtnl_lock();
+-              if (dpaa2_eth_has_mac(priv))
++              /* We can avoid locking because the "endpoint changed" IRQ
++               * handler is the only one who changes priv->mac at runtime,
++               * so we are not racing with anyone.
++               */
++              had_mac = !!priv->mac;
++              if (had_mac)
+                       dpaa2_eth_disconnect_mac(priv);
+               else
+                       dpaa2_eth_connect_mac(priv);
+@@ -4605,6 +4636,8 @@ static int dpaa2_eth_probe(struct fsl_mc
+       priv = netdev_priv(net_dev);
+       priv->net_dev = net_dev;
++      mutex_init(&priv->mac_lock);
++
+       priv->iommu_domain = iommu_get_domain_for_dev(dev);
+       priv->tx_tstamp_type = HWTSTAMP_TX_OFF;
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h
+@@ -580,6 +580,8 @@ struct dpaa2_eth_priv {
+ #endif
+       struct dpaa2_mac *mac;
++      /* Serializes changes to priv->mac */
++      struct mutex            mac_lock;
+       struct workqueue_struct *dpaa2_ptp_wq;
+       struct work_struct      tx_onestep_tstamp;
+       struct sk_buff_head     tx_skbs;
+@@ -733,11 +735,15 @@ static inline unsigned int dpaa2_eth_rx_
+ static inline bool dpaa2_eth_is_type_phy(struct dpaa2_eth_priv *priv)
+ {
++      lockdep_assert_held(&priv->mac_lock);
++
+       return dpaa2_mac_is_type_phy(priv->mac);
+ }
+ static inline bool dpaa2_eth_has_mac(struct dpaa2_eth_priv *priv)
+ {
++      lockdep_assert_held(&priv->mac_lock);
++
+       return priv->mac ? true : false;
+ }
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-ethtool.c
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-ethtool.c
+@@ -86,11 +86,16 @@ static void dpaa2_eth_get_drvinfo(struct
+ static int dpaa2_eth_nway_reset(struct net_device *net_dev)
+ {
+       struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
++      int err = -EOPNOTSUPP;
++
++      mutex_lock(&priv->mac_lock);
+       if (dpaa2_eth_is_type_phy(priv))
+-              return phylink_ethtool_nway_reset(priv->mac->phylink);
++              err = phylink_ethtool_nway_reset(priv->mac->phylink);
++
++      mutex_unlock(&priv->mac_lock);
+-      return -EOPNOTSUPP;
++      return err;
+ }
+ static int
+@@ -98,10 +103,18 @@ dpaa2_eth_get_link_ksettings(struct net_
+                            struct ethtool_link_ksettings *link_settings)
+ {
+       struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
++      int err;
+-      if (dpaa2_eth_is_type_phy(priv))
+-              return phylink_ethtool_ksettings_get(priv->mac->phylink,
+-                                                   link_settings);
++      mutex_lock(&priv->mac_lock);
++
++      if (dpaa2_eth_is_type_phy(priv)) {
++              err = phylink_ethtool_ksettings_get(priv->mac->phylink,
++                                                  link_settings);
++              mutex_unlock(&priv->mac_lock);
++              return err;
++      }
++
++      mutex_unlock(&priv->mac_lock);
+       link_settings->base.autoneg = AUTONEG_DISABLE;
+       if (!(priv->link_state.options & DPNI_LINK_OPT_HALF_DUPLEX))
+@@ -116,11 +129,17 @@ dpaa2_eth_set_link_ksettings(struct net_
+                            const struct ethtool_link_ksettings *link_settings)
+ {
+       struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
++      int err = -EOPNOTSUPP;
+-      if (!dpaa2_eth_is_type_phy(priv))
+-              return -EOPNOTSUPP;
++      mutex_lock(&priv->mac_lock);
++
++      if (dpaa2_eth_is_type_phy(priv))
++              err = phylink_ethtool_ksettings_set(priv->mac->phylink,
++                                                  link_settings);
+-      return phylink_ethtool_ksettings_set(priv->mac->phylink, link_settings);
++      mutex_unlock(&priv->mac_lock);
++
++      return err;
+ }
+ static void dpaa2_eth_get_pauseparam(struct net_device *net_dev,
+@@ -129,11 +148,16 @@ static void dpaa2_eth_get_pauseparam(str
+       struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
+       u64 link_options = priv->link_state.options;
++      mutex_lock(&priv->mac_lock);
++
+       if (dpaa2_eth_is_type_phy(priv)) {
+               phylink_ethtool_get_pauseparam(priv->mac->phylink, pause);
++              mutex_unlock(&priv->mac_lock);
+               return;
+       }
++      mutex_unlock(&priv->mac_lock);
++
+       pause->rx_pause = dpaa2_eth_rx_pause_enabled(link_options);
+       pause->tx_pause = dpaa2_eth_tx_pause_enabled(link_options);
+       pause->autoneg = AUTONEG_DISABLE;
+@@ -152,9 +176,17 @@ static int dpaa2_eth_set_pauseparam(stru
+               return -EOPNOTSUPP;
+       }
+-      if (dpaa2_eth_is_type_phy(priv))
+-              return phylink_ethtool_set_pauseparam(priv->mac->phylink,
+-                                                    pause);
++      mutex_lock(&priv->mac_lock);
++
++      if (dpaa2_eth_is_type_phy(priv)) {
++              err = phylink_ethtool_set_pauseparam(priv->mac->phylink,
++                                                   pause);
++              mutex_unlock(&priv->mac_lock);
++              return err;
++      }
++
++      mutex_unlock(&priv->mac_lock);
++
+       if (pause->autoneg)
+               return -EOPNOTSUPP;
+@@ -309,8 +341,12 @@ static void dpaa2_eth_get_ethtool_stats(
+       }
+       *(data + i++) = buf_cnt;
++      mutex_lock(&priv->mac_lock);
++
+       if (dpaa2_eth_has_mac(priv))
+               dpaa2_mac_get_ethtool_stats(priv->mac, data + i);
++
++      mutex_unlock(&priv->mac_lock);
+ }
+ static int dpaa2_eth_prep_eth_rule(struct ethhdr *eth_value, struct ethhdr *eth_mask,
diff --git a/target/linux/armvirt/patches-6.1/701-v6.2-0011-net-dpaa2-switch-serialize-changes-to-priv-mac-with.patch b/target/linux/armvirt/patches-6.1/701-v6.2-0011-net-dpaa2-switch-serialize-changes-to-priv-mac-with.patch
new file mode 100644 (file)
index 0000000..7ea4465
--- /dev/null
@@ -0,0 +1,203 @@
+From 80d12452a5f160c39d63efc1be07df36f9d07133 Mon Sep 17 00:00:00 2001
+From: Vladimir Oltean <vladimir.oltean@nxp.com>
+Date: Tue, 29 Nov 2022 16:12:20 +0200
+Subject: [PATCH 13/14] net: dpaa2-switch: serialize changes to priv->mac with
+ a mutex
+
+The dpaa2-switch driver uses a DPMAC in the same way as the dpaa2-eth
+driver, so we need to duplicate the locking solution established by the
+previous change to the switch driver as well.
+
+Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
+Reviewed-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Tested-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Signed-off-by: Paolo Abeni <pabeni@redhat.com>
+---
+ .../freescale/dpaa2/dpaa2-switch-ethtool.c    | 32 +++++++++++++++----
+ .../ethernet/freescale/dpaa2/dpaa2-switch.c   | 31 ++++++++++++++++--
+ .../ethernet/freescale/dpaa2/dpaa2-switch.h   |  2 ++
+ 3 files changed, 55 insertions(+), 10 deletions(-)
+
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-switch-ethtool.c
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-switch-ethtool.c
+@@ -60,11 +60,18 @@ dpaa2_switch_get_link_ksettings(struct n
+ {
+       struct ethsw_port_priv *port_priv = netdev_priv(netdev);
+       struct dpsw_link_state state = {0};
+-      int err = 0;
++      int err;
+-      if (dpaa2_switch_port_is_type_phy(port_priv))
+-              return phylink_ethtool_ksettings_get(port_priv->mac->phylink,
+-                                                   link_ksettings);
++      mutex_lock(&port_priv->mac_lock);
++
++      if (dpaa2_switch_port_is_type_phy(port_priv)) {
++              err = phylink_ethtool_ksettings_get(port_priv->mac->phylink,
++                                                  link_ksettings);
++              mutex_unlock(&port_priv->mac_lock);
++              return err;
++      }
++
++      mutex_unlock(&port_priv->mac_lock);
+       err = dpsw_if_get_link_state(port_priv->ethsw_data->mc_io, 0,
+                                    port_priv->ethsw_data->dpsw_handle,
+@@ -99,9 +106,16 @@ dpaa2_switch_set_link_ksettings(struct n
+       bool if_running;
+       int err = 0, ret;
+-      if (dpaa2_switch_port_is_type_phy(port_priv))
+-              return phylink_ethtool_ksettings_set(port_priv->mac->phylink,
+-                                                   link_ksettings);
++      mutex_lock(&port_priv->mac_lock);
++
++      if (dpaa2_switch_port_is_type_phy(port_priv)) {
++              err = phylink_ethtool_ksettings_set(port_priv->mac->phylink,
++                                                  link_ksettings);
++              mutex_unlock(&port_priv->mac_lock);
++              return err;
++      }
++
++      mutex_unlock(&port_priv->mac_lock);
+       /* Interface needs to be down to change link settings */
+       if_running = netif_running(netdev);
+@@ -189,8 +203,12 @@ static void dpaa2_switch_ethtool_get_sta
+                                  dpaa2_switch_ethtool_counters[i].name, err);
+       }
++      mutex_lock(&port_priv->mac_lock);
++
+       if (dpaa2_switch_port_has_mac(port_priv))
+               dpaa2_mac_get_ethtool_stats(port_priv->mac, data + i);
++
++      mutex_unlock(&port_priv->mac_lock);
+ }
+ const struct ethtool_ops dpaa2_switch_port_ethtool_ops = {
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-switch.c
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-switch.c
+@@ -603,8 +603,11 @@ static int dpaa2_switch_port_link_state_
+       /* When we manage the MAC/PHY using phylink there is no need
+        * to manually update the netif_carrier.
++       * We can avoid locking because we are called from the "link changed"
++       * IRQ handler, which is the same as the "endpoint changed" IRQ handler
++       * (the writer to port_priv->mac), so we cannot race with it.
+        */
+-      if (dpaa2_switch_port_is_type_phy(port_priv))
++      if (dpaa2_mac_is_type_phy(port_priv->mac))
+               return 0;
+       /* Interrupts are received even though no one issued an 'ifconfig up'
+@@ -684,6 +687,8 @@ static int dpaa2_switch_port_open(struct
+       struct ethsw_core *ethsw = port_priv->ethsw_data;
+       int err;
++      mutex_lock(&port_priv->mac_lock);
++
+       if (!dpaa2_switch_port_is_type_phy(port_priv)) {
+               /* Explicitly set carrier off, otherwise
+                * netif_carrier_ok() will return true and cause 'ip link show'
+@@ -697,6 +702,7 @@ static int dpaa2_switch_port_open(struct
+                            port_priv->ethsw_data->dpsw_handle,
+                            port_priv->idx);
+       if (err) {
++              mutex_unlock(&port_priv->mac_lock);
+               netdev_err(netdev, "dpsw_if_enable err %d\n", err);
+               return err;
+       }
+@@ -706,6 +712,8 @@ static int dpaa2_switch_port_open(struct
+       if (dpaa2_switch_port_is_type_phy(port_priv))
+               dpaa2_mac_start(port_priv->mac);
++      mutex_unlock(&port_priv->mac_lock);
++
+       return 0;
+ }
+@@ -715,6 +723,8 @@ static int dpaa2_switch_port_stop(struct
+       struct ethsw_core *ethsw = port_priv->ethsw_data;
+       int err;
++      mutex_lock(&port_priv->mac_lock);
++
+       if (dpaa2_switch_port_is_type_phy(port_priv)) {
+               dpaa2_mac_stop(port_priv->mac);
+       } else {
+@@ -722,6 +732,8 @@ static int dpaa2_switch_port_stop(struct
+               netif_carrier_off(netdev);
+       }
++      mutex_unlock(&port_priv->mac_lock);
++
+       err = dpsw_if_disable(port_priv->ethsw_data->mc_io, 0,
+                             port_priv->ethsw_data->dpsw_handle,
+                             port_priv->idx);
+@@ -1461,7 +1473,9 @@ static int dpaa2_switch_port_connect_mac
+               }
+       }
++      mutex_lock(&port_priv->mac_lock);
+       port_priv->mac = mac;
++      mutex_unlock(&port_priv->mac_lock);
+       return 0;
+@@ -1474,9 +1488,12 @@ err_free_mac:
+ static void dpaa2_switch_port_disconnect_mac(struct ethsw_port_priv *port_priv)
+ {
+-      struct dpaa2_mac *mac = port_priv->mac;
++      struct dpaa2_mac *mac;
++      mutex_lock(&port_priv->mac_lock);
++      mac = port_priv->mac;
+       port_priv->mac = NULL;
++      mutex_unlock(&port_priv->mac_lock);
+       if (!mac)
+               return;
+@@ -1495,6 +1512,7 @@ static irqreturn_t dpaa2_switch_irq0_han
+       struct ethsw_port_priv *port_priv;
+       u32 status = ~0;
+       int err, if_id;
++      bool had_mac;
+       err = dpsw_get_irq_status(ethsw->mc_io, 0, ethsw->dpsw_handle,
+                                 DPSW_IRQ_INDEX_IF, &status);
+@@ -1513,7 +1531,12 @@ static irqreturn_t dpaa2_switch_irq0_han
+       if (status & DPSW_IRQ_EVENT_ENDPOINT_CHANGED) {
+               rtnl_lock();
+-              if (dpaa2_switch_port_has_mac(port_priv))
++              /* We can avoid locking because the "endpoint changed" IRQ
++               * handler is the only one who changes priv->mac at runtime,
++               * so we are not racing with anyone.
++               */
++              had_mac = !!port_priv->mac;
++              if (had_mac)
+                       dpaa2_switch_port_disconnect_mac(port_priv);
+               else
+                       dpaa2_switch_port_connect_mac(port_priv);
+@@ -3256,6 +3279,8 @@ static int dpaa2_switch_probe_port(struc
+       port_priv->netdev = port_netdev;
+       port_priv->ethsw_data = ethsw;
++      mutex_init(&port_priv->mac_lock);
++
+       port_priv->idx = port_idx;
+       port_priv->stp_state = BR_STATE_FORWARDING;
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-switch.h
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-switch.h
+@@ -161,6 +161,8 @@ struct ethsw_port_priv {
+       struct dpaa2_switch_filter_block *filter_block;
+       struct dpaa2_mac        *mac;
++      /* Protects against changes to port_priv->mac */
++      struct mutex            mac_lock;
+ };
+ /* Switch data */
diff --git a/target/linux/armvirt/patches-6.1/701-v6.2-0012-net-dpaa2-mac-move-rtnl_lock-only-around-phylink.patch b/target/linux/armvirt/patches-6.1/701-v6.2-0012-net-dpaa2-mac-move-rtnl_lock-only-around-phylink.patch
new file mode 100644 (file)
index 0000000..976c2a0
--- /dev/null
@@ -0,0 +1,113 @@
+From 4ea2faf5bb13d9ba9f07e996d495c4cbe34a4236 Mon Sep 17 00:00:00 2001
+From: Vladimir Oltean <vladimir.oltean@nxp.com>
+Date: Tue, 29 Nov 2022 16:12:21 +0200
+Subject: [PATCH 14/14] net: dpaa2-mac: move rtnl_lock() only around
+ phylink_{,dis}connect_phy()
+
+After the introduction of a private mac_lock that serializes access to
+priv->mac (and port_priv->mac in the switch), the only remaining purpose
+of rtnl_lock() is to satisfy the locking requirements of
+phylink_fwnode_phy_connect() and phylink_disconnect_phy().
+
+But the functions these live in, dpaa2_mac_connect() and
+dpaa2_mac_disconnect(), have contradictory locking requirements.
+While phylink_fwnode_phy_connect() wants rtnl_lock() to be held,
+phylink_create() wants it to not be held.
+
+Move the rtnl_lock() from top-level (in the dpaa2-eth and dpaa2-switch
+drivers) to only surround the phylink calls that require it, in the
+dpaa2-mac library code.
+
+This is possible because dpaa2_mac_connect() and dpaa2_mac_disconnect()
+run unlocked, and there isn't any danger of an AB/BA deadlock between
+the rtnl_mutex and other private locks.
+
+Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
+Reviewed-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Tested-by: Ioana Ciornei <ioana.ciornei@nxp.com>
+Signed-off-by: Paolo Abeni <pabeni@redhat.com>
+---
+ drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c    | 4 ----
+ drivers/net/ethernet/freescale/dpaa2/dpaa2-mac.c    | 5 +++++
+ drivers/net/ethernet/freescale/dpaa2/dpaa2-switch.c | 4 ----
+ 3 files changed, 5 insertions(+), 8 deletions(-)
+
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
+@@ -4529,7 +4529,6 @@ static irqreturn_t dpni_irq0_handler_thr
+               dpaa2_eth_set_mac_addr(netdev_priv(net_dev));
+               dpaa2_eth_update_tx_fqids(priv);
+-              rtnl_lock();
+               /* We can avoid locking because the "endpoint changed" IRQ
+                * handler is the only one who changes priv->mac at runtime,
+                * so we are not racing with anyone.
+@@ -4539,7 +4538,6 @@ static irqreturn_t dpni_irq0_handler_thr
+                       dpaa2_eth_disconnect_mac(priv);
+               else
+                       dpaa2_eth_connect_mac(priv);
+-              rtnl_unlock();
+       }
+       return IRQ_HANDLED;
+@@ -4856,9 +4854,7 @@ static int dpaa2_eth_remove(struct fsl_m
+       else
+               fsl_mc_free_irqs(ls_dev);
+-      rtnl_lock();
+       dpaa2_eth_disconnect_mac(priv);
+-      rtnl_unlock();
+       dpaa2_eth_free_rings(priv);
+       free_percpu(priv->fd);
+       free_percpu(priv->sgt_cache);
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-mac.c
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-mac.c
+@@ -428,7 +428,9 @@ int dpaa2_mac_connect(struct dpaa2_mac *
+       }
+       mac->phylink = phylink;
++      rtnl_lock();
+       err = phylink_fwnode_phy_connect(mac->phylink, dpmac_node, 0);
++      rtnl_unlock();
+       if (err) {
+               netdev_err(net_dev, "phylink_fwnode_phy_connect() = %d\n", err);
+               goto err_phylink_destroy;
+@@ -446,7 +448,10 @@ err_pcs_destroy:
+ void dpaa2_mac_disconnect(struct dpaa2_mac *mac)
+ {
++      rtnl_lock();
+       phylink_disconnect_phy(mac->phylink);
++      rtnl_unlock();
++
+       phylink_destroy(mac->phylink);
+       dpaa2_pcs_destroy(mac);
+       of_phy_put(mac->serdes_phy);
+--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-switch.c
++++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-switch.c
+@@ -1530,7 +1530,6 @@ static irqreturn_t dpaa2_switch_irq0_han
+       }
+       if (status & DPSW_IRQ_EVENT_ENDPOINT_CHANGED) {
+-              rtnl_lock();
+               /* We can avoid locking because the "endpoint changed" IRQ
+                * handler is the only one who changes priv->mac at runtime,
+                * so we are not racing with anyone.
+@@ -1540,7 +1539,6 @@ static irqreturn_t dpaa2_switch_irq0_han
+                       dpaa2_switch_port_disconnect_mac(port_priv);
+               else
+                       dpaa2_switch_port_connect_mac(port_priv);
+-              rtnl_unlock();
+       }
+ out:
+@@ -2958,9 +2956,7 @@ static void dpaa2_switch_remove_port(str
+ {
+       struct ethsw_port_priv *port_priv = ethsw->ports[port_idx];
+-      rtnl_lock();
+       dpaa2_switch_port_disconnect_mac(port_priv);
+-      rtnl_unlock();
+       free_netdev(port_priv->netdev);
+       ethsw->ports[port_idx] = NULL;
+ }