Commits
Vladimir Oltean committed f0d8d28ed41
Revert "arm64: dts: LS1028a-rdb: use Ethernet PHY interrupt"
This reverts commit 841edb98671cfc4d6f010393ac429c78082ec4bd.
There are 2 separate issues with interrupts on the LS1028A-RDB board:
1. The GPIO1_DAT25 interrupt line is shared, so there is a real risk of
race conditions if used in edge-triggered mode, as we currently do.
This can be illustrated in the following setup:
- Take 2 LS1028A-RDB boards
- Connect swp0 to swp0, swp1 to swp1, swp2 to swp2
- Plug/unplug the power to board 2, 10 times in a row. This will make
the PHYs lose link simultaneously.
- Notice that at one point, the net devices on board 1 remain in a
state where not all the links are down (visible in "ip link"):
5: swp0: <BROADCAST,MULTICAST,UP> mtu 1468 qdisc pfifo_fast master br0 state UP mode DEFAULT group default qlen 1000
link/ether be:97:36:d3:3d:70 brd ff:ff:ff:ff:ff:ff
6: swp1: <BROADCAST,MULTICAST,UP> mtu 1468 qdisc pfifo_fast master br0 state UP mode DEFAULT group default qlen 1000
link/ether be:97:36:d3:3d:71 brd ff:ff:ff:ff:ff:ff
7: swp2: <NO-CARRIER,BROADCAST,MULTICAST,UP> mtu 1468 qdisc pfifo_fast master br0 state DOWN mode DEFAULT group default qlen 1000
link/ether be:97:36:d3:3d:72 brd ff:ff:ff:ff:ff:ff
This cannot be solved by making the interrupts level-triggered,
because the gpio-mpc8xxx controller only supports generating
edge-triggered interrupts. So the effective reality is that we
cannot not use shared interrupts connected to the gpio1
interrupt-parent.
2. The uBUS1 and uBUS2 slots that share this interrupt line with the
Ethernet PHYs are not pulled up by default, they are left floating on
current revisions of the LS1028A-RDB boards. So sufficient electrical
noise on these lines will make the CPLD think there's an interrupt
request, so it asserts the GPIO1_DAT25 signal and leaves it asserted.
This means that the PHYs on those boards will never have link when
used in interrupt mode, because their IRQ will be masked by the uBUS
line that is erroneously kept asserted. In poll mode this issue does
not occur.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>