Previously, exhausting hardware buffers would begin blocking the lwip
thread. This patch changes the emac layer to simply drop ethernet
frames, leaving recovery up to a higher level protocol.
This is consistent with the behaviour of the emac layer when unable
to allocate dynamic memory.
This was actually several bugs colluding together.
1. Confusion on the buffer-semaphore paradigm used led to misuse of the
tx semaphore and potential for odd behaviour.
2. Equality tests on tx_consume_index and tx_produce_index did not
handle overflow correctly. This would allow tx_consume_index to catch
up to tx_produce_index and trick the k64f_rx_reclaim function into
forgetting about a whole buffer of pbufs.
3. On top of all of that, the ENET_BUFFDESCRIPTOR_TX_READ_MASK was not
correctly read immediately after being set due to either a compiler
optimization or hardware delays. This caused k64f_low_level_output
to eagerly overrun existing buff-descriptors before they had been
completely sent. Adopting the counting-semaphore paradigm for 1 avoided
this concern.
As pointed out by @infinnovation, the overflow only occurs in the rare
case that the 120MHz CPU can actually generate packets faster than the
ENET hardware can transmit on a 100Mbps link.
Mahadevan Mahesh
Kevin Bracey
Bartek Szatkowski
Mika Leppänen
Christopher Haster