The branch main has been updated by kbowling: URL: https://cgit.FreeBSD.org/src/commit/?id=bc5e7b0cbbb555ffebc7d73b273c421f9ee24c23
commit bc5e7b0cbbb555ffebc7d73b273c421f9ee24c23 Author: Kevin Bowling <[email protected]> AuthorDate: 2026-07-25 23:48:15 +0000 Commit: Kevin Bowling <[email protected]> CommitDate: 2026-07-26 01:00:02 +0000 e1000: make AIM counter sampling coherent Sample free-running counters by delta instead of clearing them from the interrupt filter, which can race their producers. Publish byte and packet counts together at the TX and RX doorbells so each sample is coherent. Aggregate every TX ring assigned to the interrupt vector so unequal RX and TX queue counts are safe. Count RX bytes only after a frame is accepted. MFC after: 1 week --- sys/dev/e1000/em_txrx.c | 8 +++- sys/dev/e1000/if_em.c | 97 +++++++++++++++++++++++++++++++++++------------- sys/dev/e1000/if_em.h | 46 +++++++++++++++++++++-- sys/dev/e1000/igb_txrx.c | 4 +- 4 files changed, 123 insertions(+), 32 deletions(-) diff --git a/sys/dev/e1000/em_txrx.c b/sys/dev/e1000/em_txrx.c index f2b48d6518f7..6ac41816b043 100644 --- a/sys/dev/e1000/em_txrx.c +++ b/sys/dev/e1000/em_txrx.c @@ -476,6 +476,8 @@ em_isc_txd_flush(void *arg, uint16_t txqid, qidx_t pidx) struct tx_ring *txr = &que->txr; E1000_WRITE_REG(&sc->hw, E1000_TDT(txr->me), pidx); + if (sc->hw.mac.type >= e1000_82540) + em_aim_publish(txr); } static int @@ -601,6 +603,8 @@ em_isc_rxd_flush(void *arg, uint16_t rxqid, uint8_t flid __unused, struct rx_ring *rxr = &que->rxr; E1000_WRITE_REG(&sc->hw, E1000_RDT(rxr->me), pidx); + if (sc->hw.mac.type >= e1000_82540) + em_aim_publish_rx(rxr); } static int @@ -679,7 +683,6 @@ lem_isc_rxd_pkt_get(void *arg, if_rxd_info_t ri) len = le16toh(rxd->length); ri->iri_len += len; - rxr->rx_bytes += len; eop = (status & E1000_RXD_STAT_EOP) != 0; @@ -701,6 +704,7 @@ lem_isc_rxd_pkt_get(void *arg, if_rxd_info_t ri) i++; } while (!eop); + rxr->rx_bytes += ri->iri_len; rxr->rx_packets++; if (scctx->isc_capenable & IFCAP_RXCSUM) @@ -745,7 +749,6 @@ em_isc_rxd_pkt_get(void *arg, if_rxd_info_t ri) len = le16toh(rxd->wb.upper.length); ri->iri_len += len; - rxr->rx_bytes += len; eop = (staterr & E1000_RXD_STAT_EOP) != 0; @@ -766,6 +769,7 @@ em_isc_rxd_pkt_get(void *arg, if_rxd_info_t ri) i++; } while (!eop); + rxr->rx_bytes += ri->iri_len; rxr->rx_packets++; if (scctx->isc_capenable & IFCAP_RXCSUM) diff --git a/sys/dev/e1000/if_em.c b/sys/dev/e1000/if_em.c index 1e54d5ea998d..3fa5a7ed9a5b 100644 --- a/sys/dev/e1000/if_em.c +++ b/sys/dev/e1000/if_em.c @@ -462,7 +462,7 @@ static int em_get_rs(SYSCTL_HANDLER_ARGS); static void em_print_debug_info(struct e1000_softc *); static int em_is_valid_ether_addr(u8 *); static void em_newitr(struct e1000_softc *, struct em_rx_queue *, - struct tx_ring *, struct rx_ring *); + struct rx_ring *); static bool em_automask_tso(if_ctx_t); static int em_sysctl_tso_tcp_flags_mask(SYSCTL_HANDLER_ARGS); static int em_sysctl_int_delay(SYSCTL_HANDLER_ARGS); @@ -1641,6 +1641,47 @@ em_if_init(if_ctx_t ctx) } } +/* + * RX publishes its byte and packet counters as one snapshot when iflib + * returns descriptors to hardware. This also covers watchdog-driven RX + * processing, which can run while the interrupt vector is unmasked. + */ +static __inline void +em_aim_rx_delta(struct rx_ring *rxr, u32 *bytes, u32 *packets) +{ + uint64_t snapshot; + u32 now_bytes, now_packets; + + snapshot = atomic_load_acq_64(&rxr->rx_aim_snapshot); + now_bytes = snapshot >> 32; + now_packets = (u32)snapshot; + *bytes = now_bytes - rxr->rx_bytes_last; + *packets = now_packets - rxr->rx_packets_last; + rxr->rx_bytes_last = now_bytes; + rxr->rx_packets_last = now_packets; +} + +/* + * TX publishes its byte and packet counters as one snapshot at the doorbell, + * because encapsulation can overlap the interrupt filter. The two halves + * remain independent free running u32 counters, so their deltas are correct + * across wrap. + */ +static __inline void +em_aim_tx_delta(struct tx_ring *txr, u32 *bytes, u32 *packets) +{ + uint64_t snapshot; + u32 now_bytes, now_packets; + + snapshot = atomic_load_acq_64(&txr->tx_aim_snapshot); + now_bytes = snapshot >> 32; + now_packets = (u32)snapshot; + *bytes = now_bytes - txr->tx_bytes_last; + *packets = now_packets - txr->tx_packets_last; + txr->tx_bytes_last = now_bytes; + txr->tx_packets_last = now_packets; +} + enum itr_latency_target { itr_latency_disabled = 0, itr_latency_lowest = 1, @@ -1654,16 +1695,32 @@ enum itr_latency_target { *********************************************************************/ static void em_newitr(struct e1000_softc *sc, struct em_rx_queue *que, - struct tx_ring *txr, struct rx_ring *rxr) + struct rx_ring *rxr) { struct e1000_hw *hw = &sc->hw; - unsigned long bytes, bytes_per_packet, packets; - unsigned long rxbytes, rxpackets, txbytes, txpackets; + struct em_tx_queue *tx_que; + u32 bytes, bytes_per_packet, packets; + u32 ringbytes, ringpackets, rxbytes, rxpackets, txbytes, txpackets; u32 newitr; u8 nextlatency; + int i; - rxbytes = atomic_load_long(&rxr->rx_bytes); - txbytes = atomic_load_long(&txr->tx_bytes); + em_aim_rx_delta(rxr, &rxbytes, &rxpackets); + + /* + * A vector can service more than one TX ring when iflib is configured + * with unequal RX and TX queue counts. Sample every ring routed to + * this vector rather than treating the vector as a TX queue index. + */ + txbytes = txpackets = 0; + for (i = 0; i < sc->tx_num_queues; i++) { + tx_que = &sc->tx_queues[i]; + if (tx_que->msix != que->msix) + continue; + em_aim_tx_delta(&tx_que->txr, &ringbytes, &ringpackets); + txbytes += ringbytes; + txpackets += ringpackets; + } /* Idle, do nothing */ if (txbytes == 0 && rxbytes == 0) @@ -1686,15 +1743,13 @@ em_newitr(struct e1000_softc *sc, struct em_rx_queue *que, goto em_set_next_itr; } - bytes = bytes_per_packet = 0; + bytes = bytes_per_packet = packets = 0; /* Get largest values from the associated tx and rx ring */ - txpackets = atomic_load_long(&txr->tx_packets); if (txpackets != 0) { bytes = txbytes; bytes_per_packet = txbytes / txpackets; packets = txpackets; } - rxpackets = atomic_load_long(&rxr->rx_packets); if (rxpackets != 0) { bytes = lmax(bytes, rxbytes); bytes_per_packet = @@ -1816,7 +1871,6 @@ em_intr(void *arg) struct e1000_softc *sc = arg; struct e1000_hw *hw = &sc->hw; struct em_rx_queue *que = &sc->rx_queues[0]; - struct tx_ring *txr = &sc->tx_queues[0].txr; struct rx_ring *rxr = &que->rxr; if_ctx_t ctx = sc->ctx; u32 reg_icr; @@ -1855,13 +1909,7 @@ em_intr(void *arg) sc->rx_overruns++; if (hw->mac.type >= e1000_82540) - em_newitr(sc, que, txr, rxr); - - /* Reset state */ - txr->tx_bytes = 0; - txr->tx_packets = 0; - rxr->rx_bytes = 0; - rxr->rx_packets = 0; + em_newitr(sc, que, rxr); return (FILTER_SCHEDULE_THREAD); } @@ -1916,18 +1964,11 @@ em_msix_que(void *arg) { struct em_rx_queue *que = arg; struct e1000_softc *sc = que->sc; - struct tx_ring *txr = &sc->tx_queues[que->msix].txr; struct rx_ring *rxr = &que->rxr; ++que->irqs; - em_newitr(sc, que, txr, rxr); - - /* Reset state */ - txr->tx_bytes = 0; - txr->tx_packets = 0; - rxr->rx_bytes = 0; - rxr->rx_packets = 0; + em_newitr(sc, que, rxr); return (FILTER_SCHEDULE_THREAD); } @@ -3543,6 +3584,9 @@ em_if_tx_queues_alloc(if_ctx_t ctx, caddr_t *vaddrs, uint64_t *paddrs, /* Set up some basics */ struct tx_ring *txr = &que->txr; + KASSERT(__is_aligned(&txr->tx_aim_snapshot, sizeof(uint64_t)), + ("%s: misaligned TX AIM snapshot %p", __func__, + &txr->tx_aim_snapshot)); txr->sc = que->sc = sc; que->me = txr->me = i; @@ -3596,6 +3640,9 @@ em_if_rx_queues_alloc(if_ctx_t ctx, caddr_t *vaddrs, uint64_t *paddrs, for (i = 0, que = sc->rx_queues; i < nrxqsets; i++, que++) { /* Set up some basics */ struct rx_ring *rxr = &que->rxr; + KASSERT(__is_aligned(&rxr->rx_aim_snapshot, sizeof(uint64_t)), + ("%s: misaligned RX AIM snapshot %p", __func__, + &rxr->rx_aim_snapshot)); rxr->sc = que->sc = sc; rxr->que = que; que->me = rxr->me = i; diff --git a/sys/dev/e1000/if_em.h b/sys/dev/e1000/if_em.h index 56c28597fa22..007f5f42d5a7 100644 --- a/sys/dev/e1000/if_em.h +++ b/sys/dev/e1000/if_em.h @@ -408,8 +408,18 @@ struct tx_ring { /* Soft stats */ unsigned long tx_irq; - unsigned long tx_packets; - unsigned long tx_bytes; + + /* + * Free running AIM counters. The producer updates these while + * encapsulating packets, then publishes both together at the TX + * doorbell. The interrupt handler samples only the published value, + * so it cannot observe one counter without the other. + */ + u32 tx_packets; + u32 tx_bytes; + uint64_t tx_aim_snapshot __aligned(8); + u32 tx_packets_last; + u32 tx_bytes_last; /* Saved csum offloading context information */ int csum_flags; @@ -424,6 +434,15 @@ struct tx_ring { uint32_t csum_txd_lower; /* last field */ }; +static __inline void +em_aim_publish(struct tx_ring *txr) +{ + uint64_t snapshot; + + snapshot = ((uint64_t)txr->tx_bytes << 32) | txr->tx_packets; + atomic_store_rel_64(&txr->tx_aim_snapshot, snapshot); +} + /* * The Receive ring, one per rx queue */ @@ -443,13 +462,32 @@ struct rx_ring { /* Soft stats */ unsigned long rx_irq; unsigned long rx_discarded; - unsigned long rx_packets; - unsigned long rx_bytes; + + /* + * Free running AIM counters. RX publishes both together when iflib + * returns descriptors to hardware. The interrupt handler samples only + * the published value, so watchdog-driven RX processing cannot expose + * one counter without the other. + */ + u32 rx_packets; + u32 rx_bytes; + uint64_t rx_aim_snapshot __aligned(8); + u32 rx_packets_last; + u32 rx_bytes_last; /* Next requested ITR latency */ u8 rx_nextlatency; }; +static __inline void +em_aim_publish_rx(struct rx_ring *rxr) +{ + uint64_t snapshot; + + snapshot = ((uint64_t)rxr->rx_bytes << 32) | rxr->rx_packets; + atomic_store_rel_64(&rxr->rx_aim_snapshot, snapshot); +} + struct em_tx_queue { struct e1000_softc *sc; u32 msix; diff --git a/sys/dev/e1000/igb_txrx.c b/sys/dev/e1000/igb_txrx.c index fc690dcfc100..f6f462cfd577 100644 --- a/sys/dev/e1000/igb_txrx.c +++ b/sys/dev/e1000/igb_txrx.c @@ -304,6 +304,7 @@ igb_isc_txd_flush(void *arg, uint16_t txqid, qidx_t pidx) struct tx_ring *txr = &que->txr; E1000_WRITE_REG(&sc->hw, E1000_TDT(txr->me), pidx); + em_aim_publish(txr); } static int @@ -395,6 +396,7 @@ igb_isc_rxd_flush(void *arg, uint16_t rxqid, uint8_t flid __unused, struct rx_ring *rxr = &que->rxr; E1000_WRITE_REG(&sc->hw, E1000_RDT(rxr->me), pidx); + em_aim_publish_rx(rxr); } static int @@ -458,7 +460,6 @@ igb_isc_rxd_pkt_get(void *arg, if_rxd_info_t ri) le32toh(rxd->wb.lower.lo_dword.data) & IGB_PKTTYPE_MASK; ri->iri_len += len; - rxr->rx_bytes += len; rxd->wb.upper.status_error = 0; eop = ((staterr & E1000_RXD_STAT_EOP) == E1000_RXD_STAT_EOP); @@ -487,6 +488,7 @@ igb_isc_rxd_pkt_get(void *arg, if_rxd_info_t ri) i++; } while (!eop); + rxr->rx_bytes += ri->iri_len; rxr->rx_packets++; if ((scctx->isc_capenable & IFCAP_RXCSUM) != 0)
