[rtems-libbsd commit] dpaa: Use if_transmit instead of legacy if_start
Sebastian Huber
sebh at rtems.org
Tue Jan 15 07:01:22 UTC 2019
Module: rtems-libbsd
Branch: master
Commit: 5aa6ee55fc75308bba7d6236e0d200dfa5bc8e56
Changeset: http://git.rtems.org/rtems-libbsd/commit/?id=5aa6ee55fc75308bba7d6236e0d200dfa5bc8e56
Author: Sebastian Huber <sebastian.huber at embedded-brains.de>
Date: Tue Jan 15 07:56:36 2019 +0100
dpaa: Use if_transmit instead of legacy if_start
This avoids a lock contention on the send queue.
---
.../net/ethernet/freescale/dpaa/if_fmanmac.c | 147 +++++++++------------
1 file changed, 65 insertions(+), 82 deletions(-)
diff --git a/rtemsbsd/sys/powerpc/drivers/net/ethernet/freescale/dpaa/if_fmanmac.c b/rtemsbsd/sys/powerpc/drivers/net/ethernet/freescale/dpaa/if_fmanmac.c
index 92ad140..de16762 100644
--- a/rtemsbsd/sys/powerpc/drivers/net/ethernet/freescale/dpaa/if_fmanmac.c
+++ b/rtemsbsd/sys/powerpc/drivers/net/ethernet/freescale/dpaa/if_fmanmac.c
@@ -113,108 +113,90 @@ fman_mac_enable_tx_csum(struct mbuf *m, struct qm_fd *fd,
fd->cmd |= FM_FD_CMD_RPD | FM_FD_CMD_DTC;
}
-static void
-fman_mac_txstart_locked(struct ifnet *ifp, struct fman_mac_softc *sc)
+static int
+fman_mac_tx(struct ifnet *ifp, struct mbuf *m)
{
+ struct fman_mac_softc *sc;
+ struct fman_mac_sgt *sgt;
+ struct mbuf *n;
+ struct qm_fd fd;
+ struct dpaa_priv *priv;
+ struct qman_fq *egress_fq;
+ int queue = 0;
+ size_t i;
+ int err;
- FMAN_MAC_ASSERT_LOCKED(sc);
-
- for (;;) {
- struct fman_mac_sgt *sgt;
- struct mbuf *m;
- struct mbuf *n;
- struct qm_fd fd;
- struct dpaa_priv *priv;
- struct qman_fq *egress_fq;
- int queue = 0;
- size_t i;
- int err;
-
- IFQ_DRV_DEQUEUE(&ifp->if_snd, m);
- if (m == NULL) {
- break;
- }
+ sc = ifp->if_softc;
- sgt = uma_zalloc(fman_mac_sgt_zone, M_NOWAIT);
- if (sgt == NULL) {
- if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
- m_freem(m);
- continue;
- }
+ sgt = uma_zalloc(fman_mac_sgt_zone, M_NOWAIT);
+ if (unlikely(sgt == NULL)) {
+ if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
+ m_freem(m);
+ return (ENOBUFS);
+ }
- qm_fd_clear_fd(&fd);
- qm_fd_set_sg(&fd, offsetof(struct fman_mac_sgt, sg), m->m_pkthdr.len);
- fd.bpid = FSL_DPAA_BPID_INV;
- fd.cmd |= cpu_to_be32(FM_FD_CMD_FCO);
- qm_fd_addr_set64(&fd, (uintptr_t)sgt);
- fman_mac_enable_tx_csum(m, &fd, &sgt->prs);
+ qm_fd_clear_fd(&fd);
+ qm_fd_set_sg(&fd, offsetof(struct fman_mac_sgt, sg), m->m_pkthdr.len);
+ fd.bpid = FSL_DPAA_BPID_INV;
+ fd.cmd |= cpu_to_be32(FM_FD_CMD_FCO);
+ qm_fd_addr_set64(&fd, (uintptr_t)sgt);
+ fman_mac_enable_tx_csum(m, &fd, &sgt->prs);
repeat_with_collapsed_mbuf_chain:
- i = 0;
- n = m;
+ i = 0;
+ n = m;
- while (n != NULL && i < DPAA_SGT_MAX_ENTRIES) {
- int len = n->m_len;
+ while (n != NULL && i < DPAA_SGT_MAX_ENTRIES) {
+ int len = n->m_len;
- if (len > 0) {
- qm_sg_entry_set_len(&sgt->sg[i], len);
- sgt->sg[i].bpid = FSL_DPAA_BPID_INV;
- sgt->sg[i].offset = 0;
- qm_sg_entry_set64(&sgt->sg[i],
- mtod(n, uintptr_t));
- ++i;
- }
-
- n = n->m_next;
+ if (len > 0) {
+ qm_sg_entry_set_len(&sgt->sg[i], len);
+ sgt->sg[i].bpid = FSL_DPAA_BPID_INV;
+ sgt->sg[i].offset = 0;
+ qm_sg_entry_set64(&sgt->sg[i],
+ mtod(n, uintptr_t));
+ ++i;
}
- if (n != NULL && i == DPAA_SGT_MAX_ENTRIES) {
- struct mbuf *c;
+ n = n->m_next;
+ }
- c = m_collapse(m, M_NOWAIT, DPAA_SGT_MAX_ENTRIES);
- if (c == NULL) {
- if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
- m_freem(m);
- uma_zfree(fman_mac_sgt_zone, sgt);
- continue;
- }
+ if (unlikely(n != NULL && i == DPAA_SGT_MAX_ENTRIES)) {
+ struct mbuf *c;
- m = c;
- goto repeat_with_collapsed_mbuf_chain;
+ c = m_collapse(m, M_NOWAIT, DPAA_SGT_MAX_ENTRIES);
+ if (c == NULL) {
+ if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
+ m_freem(m);
+ uma_zfree(fman_mac_sgt_zone, sgt);
+ return (ENOBUFS);
}
- sgt->sg[i - 1].cfg |= cpu_to_be32(QM_SG_FIN);
- sgt->m = m;
- priv = netdev_priv(&sc->mac_dev.net_dev);
- egress_fq = priv->egress_fqs[queue];
- fd.cmd |= cpu_to_be32(qman_fq_fqid(priv->conf_fqs[queue]));
+ m = c;
+ goto repeat_with_collapsed_mbuf_chain;
+ }
- for (i = 0; i < DPAA_ENQUEUE_RETRIES; ++i) {
- err = qman_enqueue(egress_fq, &fd);
- if (err != -EBUSY) {
- break;
- }
- }
+ sgt->sg[i - 1].cfg |= cpu_to_be32(QM_SG_FIN);
+ sgt->m = m;
+ priv = netdev_priv(&sc->mac_dev.net_dev);
+ egress_fq = priv->egress_fqs[queue];
+ fd.cmd |= cpu_to_be32(qman_fq_fqid(priv->conf_fqs[queue]));
- if (unlikely(err < 0)) {
- if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
- m_freem(m);
- continue;
+ for (i = 0; i < DPAA_ENQUEUE_RETRIES; ++i) {
+ err = qman_enqueue(egress_fq, &fd);
+ if (likely(err != -EBUSY)) {
+ break;
}
}
-}
-static void
-fman_mac_txstart(struct ifnet *ifp)
-{
- struct fman_mac_softc *sc;
-
- sc = ifp->if_softc;
+ if (unlikely(err < 0)) {
+ if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
+ m_freem(m);
+ return (ENOBUFS);
+ }
- FMAN_MAC_LOCK(sc);
- fman_mac_txstart_locked(ifp, sc);
- FMAN_MAC_UNLOCK(sc);
+ return (0);
}
static void
@@ -447,7 +429,8 @@ fman_mac_dev_attach(device_t dev)
IFCAP_VLAN_MTU | IFCAP_JUMBO_MTU;
ifp->if_capenable = ifp->if_capabilities;
ifp->if_hwassist = FMAN_MAC_CSUM;
- ifp->if_start = fman_mac_txstart;
+ ifp->if_transmit = fman_mac_tx;
+ ifp->if_qflush = if_qflush;
ifp->if_ioctl = fman_mac_ioctl;
ifp->if_init = fman_mac_init;
IFQ_SET_MAXLEN(&ifp->if_snd, 128);
More information about the vc
mailing list