Each completion ring entry has a valid bit to indicate that the entry contains a valid completion event. The driver's main poll loop __bnxt_poll_work() has the proper dma_rmb() to make sure the valid bit of the next entry has been checked before proceeding further. But when we call bnxt_rx_pkt() to process the RX event, the RX completion event consists of two completion entries and only the first entry has been checked to be valid. We need the same barrier after checking the next completion entry. Add missing dma_rmb() barriers in bnxt_rx_pkt() and other similar locations. Fixes: 67a95e2022c7 ("bnxt_en: Need memory barrier when processing the completion ring.") Reported-by: Lance Richardson Reviewed-by: Andy Gospodarek Reviewed-by: Lance Richardson Signed-off-by: Michael Chan --- drivers/net/ethernet/broadcom/bnxt/bnxt.c | 12 ++++++++++++ 1 file changed, 12 insertions(+) diff --git a/drivers/net/ethernet/broadcom/bnxt/bnxt.c b/drivers/net/ethernet/broadcom/bnxt/bnxt.c index 207ef7dc9bff..8a97640cdfe7 100644 --- a/drivers/net/ethernet/broadcom/bnxt/bnxt.c +++ b/drivers/net/ethernet/broadcom/bnxt/bnxt.c @@ -1788,6 +1788,10 @@ static int bnxt_rx_pkt(struct bnxt *bp, struct bnxt_cp_ring_info *cpr, if (!RX_CMP_VALID(rxcmp1, tmp_raw_cons)) return -EBUSY; + /* The valid test of the entry must be done first before + * reading any further. + */ + dma_rmb(); prod = rxr->rx_prod; if (cmp_type == CMP_TYPE_RX_L2_TPA_START_CMP) { @@ -2010,6 +2014,10 @@ static int bnxt_force_rx_discard(struct bnxt *bp, if (!RX_CMP_VALID(rxcmp1, tmp_raw_cons)) return -EBUSY; + /* The valid test of the entry must be done first before + * reading any further. + */ + dma_rmb(); cmp_type = RX_CMP_TYPE(rxcmp); if (cmp_type == CMP_TYPE_RX_L2_CMP) { rxcmp1->rx_cmp_cfa_code_errors_v2 |= @@ -2475,6 +2483,10 @@ static int bnxt_poll_nitroa0(struct napi_struct *napi, int budget) if (!TX_CMP_VALID(txcmp, raw_cons)) break; + /* The valid test of the entry must be done first before + * reading any further. + */ + dma_rmb(); if ((TX_CMP_TYPE(txcmp) & 0x30) == 0x10) { tmp_raw_cons = NEXT_RAW_CMP(raw_cons); cp_cons = RING_CMP(tmp_raw_cons); -- 2.18.1