Index: if_mcx.c
===================================================================
RCS file: /cvs/src/sys/dev/pci/if_mcx.c,v
retrieving revision 1.91
diff -u -p -r1.91 if_mcx.c
--- if_mcx.c	27 Dec 2020 01:00:25 -0000	1.91
+++ if_mcx.c	27 Dec 2020 06:37:04 -0000
@@ -2220,6 +2221,7 @@ struct mcx_dmamem {
 #define MCX_DMA_MAP(_mxm)	((_mxm)->mxm_map)
 #define MCX_DMA_DVA(_mxm)	((_mxm)->mxm_map->dm_segs[0].ds_addr)
 #define MCX_DMA_KVA(_mxm)	((void *)(_mxm)->mxm_kva)
+#define MCX_DMA_OFF(_mxm, _off)	((void *)((_mxm)->mxm_kva + (_off)))
 #define MCX_DMA_LEN(_mxm)	((_mxm)->mxm_size)
 
 struct mcx_hwmem {
@@ -2243,7 +2245,7 @@ struct mcx_eq {
 struct mcx_cq {
 	int			 cq_n;
 	struct mcx_dmamem	 cq_mem;
-	uint32_t		*cq_doorbell;
+	bus_addr_t		 cq_doorbell;
 	uint32_t		 cq_cons;
 	uint32_t		 cq_count;
 };
@@ -4477,6 +4480,8 @@ mcx_create_cq(struct mcx_softc *sc, stru
 	uint64_t *pas;
 	int insize, npages, paslen, i, token;
 
+	cq->cq_doorbell = MCX_CQ_DOORBELL_BASE + (MCX_CQ_DOORBELL_STRIDE * db);
+
 	npages = howmany((1 << MCX_LOG_CQ_SIZE) * sizeof(struct mcx_cq_entry),
 	    MCX_PAGE_SIZE);
 	paslen = npages * sizeof(*pas);
@@ -4516,8 +4521,7 @@ mcx_create_cq(struct mcx_softc *sc, stru
 	    (MCX_CQ_MOD_PERIOD << MCX_CQ_CTX_PERIOD_SHIFT) |
 	    MCX_CQ_MOD_COUNTER);
 	mbin->cmd_cq_ctx.cq_doorbell = htobe64(
-	    MCX_DMA_DVA(&sc->sc_doorbell_mem) +
-	    MCX_CQ_DOORBELL_BASE + (MCX_CQ_DOORBELL_STRIDE * db));
+	    MCX_DMA_DVA(&sc->sc_doorbell_mem) + cq->cq_doorbell);
 
 	bus_dmamap_sync(sc->sc_dmat, MCX_DMA_MAP(&cq->cq_mem),
 	    0, MCX_DMA_LEN(&cq->cq_mem), BUS_DMASYNC_PREREAD);
@@ -4546,11 +4550,13 @@ mcx_create_cq(struct mcx_softc *sc, stru
 	cq->cq_n = mcx_get_id(out->cmd_cqn);
 	cq->cq_cons = 0;
 	cq->cq_count = 0;
-	cq->cq_doorbell = MCX_DMA_KVA(&sc->sc_doorbell_mem) +
-	    MCX_CQ_DOORBELL_BASE + (MCX_CQ_DOORBELL_STRIDE * db);
 
 	mcx_dmamem_free(sc, &mxm);
 
+	bus_dmamap_sync(sc->sc_dmat, MCX_DMA_MAP(&sc->sc_doorbell_mem),
+	    cq->cq_doorbell, sizeof(struct mcx_cq_doorbell),
+	    BUS_DMASYNC_PREWRITE);
+
 	mcx_arm_cq(sc, cq, uar);
 
 	return (0);
@@ -4600,6 +4606,10 @@ mcx_destroy_cq(struct mcx_softc *sc, str
 		return -1;
 	}
 
+	bus_dmamap_sync(sc->sc_dmat, MCX_DMA_MAP(&sc->sc_doorbell_mem),
+	    cq->cq_doorbell, sizeof(struct mcx_cq_doorbell),
+	    BUS_DMASYNC_POSTWRITE);
+
 	bus_dmamap_sync(sc->sc_dmat, MCX_DMA_MAP(&cq->cq_mem),
 	    0, MCX_DMA_LEN(&cq->cq_mem), BUS_DMASYNC_POSTREAD);
 	mcx_dmamem_free(sc, &cq->cq_mem);
@@ -6791,24 +6801,32 @@ mcx_next_cq_entry(struct mcx_softc *sc, 
 static void
 mcx_arm_cq(struct mcx_softc *sc, struct mcx_cq *cq, int uar)
 {
+	struct mcx_cq_doorbell *db;
 	bus_size_t offset;
 	uint32_t val;
 	uint64_t uval;
 
-	offset = (MCX_PAGE_SIZE * uar);
 	val = ((cq->cq_count) & 3) << MCX_CQ_DOORBELL_ARM_CMD_SN_SHIFT;
 	val |= (cq->cq_cons & MCX_CQ_DOORBELL_ARM_CI_MASK);
 
-	cq->cq_doorbell[0] = htobe32(cq->cq_cons & MCX_CQ_DOORBELL_ARM_CI_MASK);
-	cq->cq_doorbell[1] = htobe32(val);
+	db = MCX_DMA_OFF(&sc->sc_doorbell_mem, cq->cq_doorbell);
+
+	bus_dmamap_sync(sc->sc_dmat, MCX_DMA_MAP(&sc->sc_doorbell_mem),
+	    cq->cq_doorbell, sizeof(*db), BUS_DMASYNC_POSTWRITE);
 
-	uval = val;
-	uval <<= 32;
+	htobem32(&db->db_update_ci, cq->cq_cons & MCX_CQ_DOORBELL_ARM_CI_MASK);
+	htobem32(&db->db_arm_ci, val);
+
+	bus_dmamap_sync(sc->sc_dmat, MCX_DMA_MAP(&sc->sc_doorbell_mem),
+	    cq->cq_doorbell, sizeof(*db), BUS_DMASYNC_PREWRITE);
+
+	offset = (MCX_PAGE_SIZE * uar) + MCX_UAR_CQ_DOORBELL;
+
+	uval = (uint64_t)val << 32;
 	uval |= cq->cq_n;
-	bus_space_write_raw_8(sc->sc_memt, sc->sc_memh,
-	    offset + MCX_UAR_CQ_DOORBELL, htobe64(uval));
-	mcx_bar(sc, offset + MCX_UAR_CQ_DOORBELL, sizeof(uint64_t),
-	    BUS_SPACE_BARRIER_WRITE);
+
+	bus_space_write_raw_8(sc->sc_memt, sc->sc_memh, offset, htobe64(uval));
+	mcx_bar(sc, offset, sizeof(uval), BUS_SPACE_BARRIER_WRITE);
 }
 
 void