fix: txring_put() refuses over-MTU packets instead of truncating them (#1108)

[fklassen]

Fixes #1108.

txring_put() used to copy in only the first max_len bytes of an over-MTU packet and queue that truncated result as though it were the whole packet. The kernel has no way to know it's short, so the corrupted frame reached the wire - while send_packets.c's caller-side check (comparing the requested length against txring_put()'s truncated return value) counted it a failure. "0 successful, N failed" described the bookkeeping, not what the interface actually transmitted - /sys/class/net/<iface>/statistics/tx_packets still advanced.

The fix

Chose the first of the two options the issue laid out: refuse outright rather than count truncated-but-sent as success. Moved the size check to the front of txring_put(), before it waits for a ring slot, and changed it to return -1 with errno set to EMSGSIZE instead of clamping and continuing. Nothing is written into the ring and no frame the kernel could transmit ever gets queued, so "failed" now means what it says.

sendpacket()'s existing generic error handling (src/common/sendpacket.c) already treats EMSGSIZE like any other hard failure - sets a descriptive message, counts it, doesn't retry (retries are reserved for EAGAIN/ENOBUFS, which would be pointless here: a packet that doesn't fit this MTU never will) - so no caller-side changes were needed.

Fragmenting instead of refusing is the /* TODO Fragment packet */ this replaces, and is still not attempted here - refusing is the smaller, no-worse-than-before change; actually fragmenting is a separate feature, not a bug fix.

Verification

Reproduced the issue's own repro (MTU 1280, test.pcap up to 1514 bytes) before and after:

                         before        after
Truncated packets:          32            0
Successful packets:          0            3

Packets that fit the MTU are no longer stuck in the ring behind ones that don't, so "Successful packets" now reports the packets that actually went out.

• test/unit/test_txring.c's oversized-packet case updated to assert the refusal (-1/EMSGSIZE, frame left TP_STATUS_AVAILABLE) instead of the old clamped-length/partial-copy behavior: 164/164 assertions.
• sudo make tcpreplay (full replay group): 19/19.
• CI's exact asan job config (ASan+UBSan, alignment included since [Bug] misaligned struct access is systemic in the packet parsers (15 sites) #1104): 78/78, 0 sanitizer reports.
• Plain non-sanitizer build and CMake: both 78/78.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NBmWiWg46r8BLbdwozKo6v