Skip to contentSkip to conversation

r8169: rtl_xmit_pkt corrupts TX descriptor ring on zero-length mid-chain segment

DPDK/ethdev Bug 1890 [email protected] 1 participant 1 message Feb 19

lore
Share & tools
lore
#0 lore raw

http://bugs.dpdk.org/show_bug.cgi?id=1890

Bug ID: 1890
Summary: r8169: rtl_xmit_pkt corrupts TX descriptor ring on
zero-length mid-chain segment
Product: DPDK
Version: 25.11
Hardware: All
OS: All
Status: UNCONFIRMED
Severity: minor
Priority: Normal
Component: ethdev
Assignee: [email protected]
Reporter: [email protected]
Target Milestone: ---

This is not urgent bug, since giving a driver an mbuf like this is likely an application bug.

rtl_xmit_pkt() in drivers/net/r8169/r8169_rxtx.c breaks out of its segment loop when it encounters a zero-length segment in a multi-segment mbuf. If earlier segments have already been written to the descriptor ring (with FirstFrag set but no LastFrag), the ring is left in a corrupted state with an incomplete packet owned by the hardware.

In rtl_xmit_pkt() (r8169_rxtx.c, line ~1624), the per-segment loop breaks immediately when it encounters a segment with data_len == 0:

```c
    for (m_seg = tx_pkt; m_seg; m_seg = m_seg->next) {
        opts1 = opts[0];
        opts2 = opts[1];

        len = m_seg->data_len;

        if (len == 0)
            break;

        txd = &txq->hw_ring[tail];
        ...
        opts1 |= len;
        if (m_seg == tx_pkt)
            opts1 |= FirstFrag;
        if (!m_seg->next)
            opts1 |= LastFrag;
        ...
        txd->opts1 = rte_cpu_to_le_32(opts1);

        tail = (tail + 1) % nb_tx_desc;
        desc_count++;
        ...
    }

    txq->tx_tail += desc_count;
    txq->tx_free -= desc_count;
```

Consider a 3-segment mbuf where segment 2 (the middle one) has data_len == 0:

1. **Segment 1** (first): written to descriptor ring with FirstFrag set. DescOwn is set, so hardware considers it owned.
2. **Segment 2**: data_len == 0, loop breaks.
3. **Segment 3** (last): never processed — LastFrag is never written.

The result is:

- The descriptor ring contains a partial packet: one or more descriptors with DescOwn and FirstFrag but no corresponding LastFrag.
- txq->tx_tail and txq->tx_free are updated to reflect the partial write (desc_count > 0).
- When the doorbell is rung in rtl_xmit_pkts(), the hardware will attempt to process these descriptors. The behavior is undefined — the hardware may hang waiting for LastFrag, process garbage, or trigger a DMA error.
- The tx_clean path may also malfunction since the sw_ring entries and descriptor states are inconsistent.

This is a data path corruption bug that could cause the transmit queue to hang or produce undefined hardware behavior.

## Steps to Reproduce

1. Construct a multi-segment mbuf (nb_segs >= 2) where a non-first segment has data_len == 0.
2. Submit it via rte_eth_tx_burst() on an r8169 port.
3. The hardware receives an incomplete descriptor chain.

## Suggested Fix

The zero-length segment check should be handled before any descriptors are written, or alternatively, zero-length segments should be skipped with continue rather than aborting the entire packet.

**Option A — reject the packet before writing any descriptors** (simplest, pairs with Bug #1 fix):

Validate all segments up front in rtl_xmit_pkts() before calling rtl_xmit_pkt():

```c
        /* Validate all segments before committing to descriptor ring */
        bool valid = true;
        for (struct rte_mbuf *seg = tx_pkt; seg; seg = seg->next) {
            if (seg->data_len == 0) {
                valid = false;
                break;
            }
        }
        if (!valid) {
            rte_pktmbuf_free(tx_pkt);
            txq->sw_stats.tx_errors++;
            continue;
        }
```

**Option B — skip zero-length segments** (if they are considered legal):

```c
        if (len == 0)
            continue;  /* skip empty segments instead of aborting */
```

However, Option B still needs to ensure LastFrag is set on the actual last non-empty segment, which requires restructuring the LastFrag logic. Option A is safer and more straightforward.

Signature · 2 lines
You are receiving this mail because:
You are the assignee for the bug.