Both of these come out of the same investigation on a pair of ASMedia
ASM4242 hosts wired to each other, where every teardown of a host-to-host
DMA tunnel spends the full 500 ms pending-bit timeout on the host
interface adapter's hop and fails anyway.
1 gives the software connection manager the same contract the ICM one
already has: return the first error seen while deactivating hops,
instead of returning 0 unconditionally. Teardown still runs to
completion either way. Today a caller cannot tell the two connection
managers apart, and tbnet's "failed to disable DMA paths" warning can
only ever fire on ICM based hosts.
2 remembers on the adapter that its pending bit stayed set, and stops
re-learning that on every subsequent teardown. Cycling the interface
down and up 200 times across two such hosts produced 213 and 206 of
those timeouts before, and 1 each after; the run comes out 110
seconds shorter over its 200 cycles.
Patch 2 is the one I am least sure of. I do not have the USB4 Connection
Manager guide, so I cannot tell whether a protocol adapter's pending bit
is vendor defined the way its IFC, ISE and Path Credits Allocated fields
are - if it is, not waiting on it at all would be the right fix and this
is working around the symptom. Its own notes say what else I could not
show. I would rather send it with that written down than leave it out.
Note what the pair does together on this hardware: 2 leaves the first
failure to be reported and silences the rest, so the signal 1 adds fires
once per adapter here rather than on every teardown.
Two further patches from the same investigation touch
drivers/net/thunderbolt/ and went to netdev separately:
https://lore.kernel.org/netdev/20260809-b4-tbnet-hopid-v1-0-97aaf2aa0fc7@gmail.com/
They are not needed to apply these two, but they are what the runs above
were measured across - the second of them removes a warning splat that
fires throughout any prolonged run of link cycling.
Per-patch notes, including what I am not claiming, are below the cut in
each patch.
---
Fan Ye (2):
thunderbolt: Report DMA path teardown failures to the caller
thunderbolt: Stop waiting on a path pending bit that never clears
drivers/thunderbolt/path.c | 38 ++++++++++++++++++++++++++++++++------
drivers/thunderbolt/tb.c | 35 +++++++++++++++++++++++------------
drivers/thunderbolt/tb.h | 5 ++++-
drivers/thunderbolt/tunnel.c | 19 +++++++++++++++----
drivers/thunderbolt/tunnel.h | 2 +-
5 files changed, 75 insertions(+), 24 deletions(-)
---
base-commit: 06cf61899d6498b33e4b7c87d99d5bd471ccc375
change-id: 20260809-b4-tb-teardown-d26c32482c96
Best regards,
--
Fan Ye <fy15309206903@gmail.com>