From: Vitaliy Makkoveev <mvs@openbsd.org>
Subject: Re: 64-bit DMA
To: Mark Kettenis <mark.kettenis@xs4all.nl>
Cc: tech@openbsd.org
Date: Wed, 25 Feb 2026 13:34:13 +0000

On Wed, Feb 25, 2026 at 02:09:11PM +0100, Mark Kettenis wrote:
> > Date: Wed, 25 Feb 2026 12:50:37 +0000
> > From: Vitaliy Makkoveev <mvs@openbsd.org>
> > 
> > re(4) supports 64-bit DMA. Tested with
> > 
> > re0 at pci3 dev 0 function 0 "Realtek 8168" rev 0x11: RTL8168G/8111G
> > (0x4c00), msi
> 
> Not quite as easy I fear.  Older PCI variants apparently may have
> issues with dual address cycle (DAC) transactions.

I have and old Core2 motherboard, and if I'm not wrong it has re(4)
related NIC. Is it old enough? However is easy to separate PCIe devices
and enable 64-bit DMA only for them.

Index: sys/dev/ic/re.c
===================================================================
RCS file: /cvs/src/sys/dev/ic/re.c,v
diff -u -p -r1.220 re.c
--- sys/dev/ic/re.c	10 May 2025 11:08:26 -0000	1.220
+++ sys/dev/ic/re.c	25 Feb 2026 13:27:01 -0000
@@ -667,6 +667,8 @@ re_attach(struct rl_softc *sc, const cha
 	int		error = 0, i;
 	const struct re_revision *rr;
 	const char	*re_name = NULL;
+	int		bus_dma64 = sc->rl_flags |= RL_FLAG_PCIE ?
+			    BUS_DMA_64BIT : 0;
 
 	sc->sc_hwrev = CSR_READ_4(sc, RL_TXCFG) & RL_TXCFG_HWREV;
 
@@ -913,7 +915,7 @@ re_attach(struct rl_softc *sc, const cha
 	if ((error = bus_dmamem_alloc(sc->sc_dmat, RL_TX_LIST_SZ(sc),
 		    RL_RING_ALIGN, 0, &sc->rl_ldata.rl_tx_listseg, 1,
 		    &sc->rl_ldata.rl_tx_listnseg, BUS_DMA_NOWAIT |
-		    BUS_DMA_ZERO)) != 0) {
+		    BUS_DMA_ZERO | bus_dma64)) != 0) {
 		printf("%s: can't allocate tx listseg, error = %d\n",
 		    sc->sc_dev.dv_xname, error);
 		goto fail_0;
@@ -930,7 +932,7 @@ re_attach(struct rl_softc *sc, const cha
 	}
 
 	if ((error = bus_dmamap_create(sc->sc_dmat, RL_TX_LIST_SZ(sc), 1,
-		    RL_TX_LIST_SZ(sc), 0, 0,
+		    RL_TX_LIST_SZ(sc), 0, bus_dma64,
 		    &sc->rl_ldata.rl_tx_list_map)) != 0) {
 		printf("%s: can't create tx list map, error = %d\n",
 		    sc->sc_dev.dv_xname, error);
@@ -949,7 +951,7 @@ re_attach(struct rl_softc *sc, const cha
 	for (i = 0; i < sc->rl_ldata.rl_tx_desc_cnt; i++) {
 		error = bus_dmamap_create(sc->sc_dmat,
 		    RL_JUMBO_FRAMELEN, sc->rl_ldata.rl_tx_ndescs,
-		    RL_JUMBO_FRAMELEN, 0, 0,
+		    RL_JUMBO_FRAMELEN, 0, bus_dma64,
 		    &sc->rl_ldata.rl_txq[i].txq_dmamap);
 		if (error) {
 			printf("%s: can't create DMA map for TX\n",
@@ -962,7 +964,7 @@ re_attach(struct rl_softc *sc, const cha
 	if ((error = bus_dmamem_alloc(sc->sc_dmat, RL_RX_DMAMEM_SZ(sc),
 		    RL_RING_ALIGN, 0, &sc->rl_ldata.rl_rx_listseg, 1,
 		    &sc->rl_ldata.rl_rx_listnseg, BUS_DMA_NOWAIT |
-		    BUS_DMA_ZERO)) != 0) {
+		    BUS_DMA_ZERO | bus_dma64)) != 0) {
 		printf("%s: can't allocate rx listnseg, error = %d\n",
 		    sc->sc_dev.dv_xname, error);
 		goto fail_4;
@@ -980,7 +982,7 @@ re_attach(struct rl_softc *sc, const cha
 	}
 
 	if ((error = bus_dmamap_create(sc->sc_dmat, RL_RX_DMAMEM_SZ(sc), 1,
-		    RL_RX_DMAMEM_SZ(sc), 0, 0,
+		    RL_RX_DMAMEM_SZ(sc), 0, bus_dma64,
 		    &sc->rl_ldata.rl_rx_list_map)) != 0) {
 		printf("%s: can't create rx list map, error = %d\n",
 		    sc->sc_dev.dv_xname, error);
@@ -999,7 +1001,7 @@ re_attach(struct rl_softc *sc, const cha
 	for (i = 0; i < sc->rl_ldata.rl_rx_desc_cnt; i++) {
 		error = bus_dmamap_create(sc->sc_dmat,
 		    RL_FRAMELEN(sc->rl_max_mtu), 1,
-		    RL_FRAMELEN(sc->rl_max_mtu), 0, 0,
+		    RL_FRAMELEN(sc->rl_max_mtu), 0, bus_dma64,
 		    &sc->rl_ldata.rl_rxsoft[i].rxs_dmamap);
 		if (error) {
 			printf("%s: can't create DMA map for RX\n",