The branch main has been updated by markj: URL: https://cgit.FreeBSD.org/src/commit/?id=669cd0d90ee7b5bed794ded5fa00b6be854a6598
commit 669cd0d90ee7b5bed794ded5fa00b6be854a6598 Author: Andrew Griffiths <[email protected]> AuthorDate: 2026-09-29 20:11:31 +0000 Commit: Mark Johnston <[email protected]> CommitDate: 2026-09-29 20:14:12 +0000 rge: Preserve replacement mbufs after defragmentation m_defrag() has pointer-return ownership semantics: success frees the original chain and returns a replacement, while failure returns NULL and leaves the original owned by the caller. rge_encap() compared that pointer as an integer status and could return failure after success, causing rge_tx_task() to free its stale old head. Pass the mbuf by reference, retain the replacement returned by m_defrag(), and publish it to the caller before retrying DMA mapping. A later mapping failure is then cleaned up through the current chain, and a successful transmission uses that same chain for BPF and TX ownership. An unprivileged process can reach the EFBIG branch in mapped-sendfile mode. Signed-off-by: Andrew Griffiths <[email protected]> Reviewed by: adrian, markj MFC after: 1 week Differential Revision: https://reviews.freebsd.org/D60142 --- sys/dev/rge/if_rge.c | 26 ++++++++++++++++---------- 1 file changed, 16 insertions(+), 10 deletions(-) diff --git a/sys/dev/rge/if_rge.c b/sys/dev/rge/if_rge.c index 7c11da5c3f53..89287450a01d 100644 --- a/sys/dev/rge/if_rge.c +++ b/sys/dev/rge/if_rge.c @@ -766,18 +766,20 @@ rge_tx_list_sync(struct rge_softc *sc, struct rge_queues *q, * * @param sc driver softc * @param q TX queue ring - * @param m mbuf to enqueue + * @param mp mbuf to enqueue, updated if defragmentation replaces it * @returns if the mbuf is enqueued, it's consumed here and the number of * TX descriptors used is returned; if there's no space then 0 is - * returned; if the mbuf couldn't be defragged and the caller - * should free it then -1 is returned. + * returned; if the mbuf couldn't be mapped and the caller should + * free it then -1 is returned. */ static int -rge_encap(struct rge_softc *sc, struct rge_queues *q, struct mbuf *m, int idx) +rge_encap(struct rge_softc *sc, struct rge_queues *q, struct mbuf **mp, + int idx) { struct rge_tx_desc *d = NULL; struct rge_txq *txq; bus_dmamap_t txmap; + struct mbuf *m; uint32_t cmdsts, cflags = 0; int cur, error, i; bus_dma_segment_t seg[RGE_TX_NSEGS]; @@ -785,6 +787,7 @@ rge_encap(struct rge_softc *sc, struct rge_queues *q, struct mbuf *m, int idx) RGE_ASSERT_LOCKED(sc); + m = *mp; txq = &q->q_tx.rge_txq[idx]; txmap = txq->txq_dmamap; @@ -800,10 +803,13 @@ rge_encap(struct rge_softc *sc, struct rge_queues *q, struct mbuf *m, int idx) case EFBIG: /* mbuf chain is too fragmented */ sc->sc_drv_stats.tx_encap_refrag_cnt++; nsegs = RGE_TX_NSEGS; - if (m_defrag(m, M_NOWAIT) == 0 && - bus_dmamap_load_mbuf_sg(sc->sc_dmat_tx_buf, txmap, m, - seg, &nsegs, BUS_DMA_NOWAIT) == 0) - break; + m = m_defrag(m, M_NOWAIT); + if (m != NULL) { + *mp = m; + if (bus_dmamap_load_mbuf_sg(sc->sc_dmat_tx_buf, txmap, + m, seg, &nsegs, BUS_DMA_NOWAIT) == 0) + break; + } /* FALLTHROUGH */ default: sc->sc_drv_stats.tx_encap_err_toofrag++; @@ -2498,8 +2504,8 @@ rge_tx_task(void *arg, int npending) if (m == NULL) break; - /* Attempt to encap */ - used = rge_encap(sc, q, m, idx); + /* Attempt to encap, m might change due to defrag */ + used = rge_encap(sc, q, &m, idx); if (used < 0) { if_inc_counter(sc->sc_ifp, IFCOUNTER_OQDROPS, 1); m_freem(m);
