Messages in this thread | | | From | Klaus Kudielka <> | Subject | [PATCH net-next v4 0/4] net: dsa: mv88e6xxx: accelerate C45 scan | Date | Wed, 15 Mar 2023 17:38:42 +0100 |
| |
Starting with commit 1a136ca2e089 ("net: mdio: scan bus based on bus capabilities for C22 and C45"), mdiobus_scan_bus_c45() is being called on buses with MDIOBUS_NO_CAP. On a Turris Omnia (Armada 385, 88E6176 switch), this causes a significant increase of boot time, from 1.6 seconds, to 6.3 seconds. The boot time stated here is until start of /init.
Further testing revealed that the C45 scan is indeed expensive (around 2.7 seconds, due to a huge number of bus transactions), and called twice.
Two things were suggested: (1) to move the expensive call of mv88e6xxx_mdios_register() from mv88e6xxx_probe() to mv88e6xxx_setup(). (2) to mask apparently non-existing phys during probing.
Before that: Patch #1 prepares the driver to handle the movement of mv88e6xxx_mdios_register() to mv88e6xxx_setup() for cross-chip DSA trees. Patch #2 is preparatory code movement, without functional change.
With those changes, boot time on the Turris Omnia is back to normal.
Link: https://lore.kernel.org/lkml/449bde236c08d5ab5e54abd73b645d8b29955894.camel@gmail.com/
v2: Add cover letter Extend the cleanup in mv88e6xxx_setup() to remove the mdio bus on failure Add patch "mask apparently non-existing phys during probing"
v3: Add patch "don't dispose of Global2 IRQ mappings from mdiobus code"
v4: Resubmit, this time hopefully with all subject prefixes correct
Klaus Kudielka (3): net: dsa: mv88e6xxx: re-order functions net: dsa: mv88e6xxx: move call to mv88e6xxx_mdios_register() net: dsa: mv88e6xxx: mask apparently non-existing phys during probing
Vladimir Oltean (1): net: dsa: mv88e6xxx: don't dispose of Global2 IRQ mappings from mdiobus code
drivers/net/dsa/mv88e6xxx/chip.c | 381 ++++++++++++++-------------- drivers/net/dsa/mv88e6xxx/global2.c | 20 +- 2 files changed, 196 insertions(+), 205 deletions(-)
-- 2.39.2
| |