mirror of
https://github.com/openwrt/openwrt.git
synced 2026-10-13 21:58:06 +04:00
Gennaro Cimmino
3b2bc55dcb
The transmit path frees a TX skb only when its descriptor slot is reused, RTETH_TX_RING_SIZE transmissions later. Until then the skb stays charged to the sending socket's send buffer. With standard frames the pinned amount is negligible, but with jumbo frames a handful of completed transmissions is enough to exhaust the sender's budget: eight ~9k echo replies exceed the kernel ICMP socket limit of 2 * SKB_TRUESIZE(64 * 1024), after which the stack drops every further reply of any size (Icmp OutErrors) until unrelated transmissions happen to recycle the slots. Observed on the Hasivo S1100W-8XGT-SE (RTL9303) as a total ICMP blackout setting in after exactly eight jumbo echo replies, healed by any sixteen device-originated frames. Track monotonic send/clean counters in the driver-private ring info (their difference is the number of unreleased skbs, their low bits the ring slots), and release the contiguous completed run at the start of the RX poll, before the received packets are processed, so replies generated while handling the poll always find their send budget released. The walk runs under the TX queue lock the transmit path already holds, stops at the first descriptor the hardware still owns, and is skipped entirely - no lock taken, no access to the uncached descriptor memory - while nothing is pending. The transmit path reuses the same walk for the ring-wrap case instead of its old per-slot cleanup, and stopped or watchdog-frozen queues are left to the rteth_tx_timeout() recovery, which keeps releasing every buffer itself. Assisted-by: Claude:claude-fable-5 Signed-off-by: Gennaro Cimmino <gcimmino@rayonra.net> Link: https://github.com/openwrt/openwrt/pull/24406 Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>