Sorry about the long wait.
I think this is the right direction, I like the idea of having a thread
local map of the txqs. It will allow us to assign the txqs in a more
sophisticated way.
I see some unit tests failing with this patch applied. Could you try
running the testsuite with 'make check'?
The txq distribution code is run only when a pmd device is added.
If a "system" or a "dummy" device is added its txq will not be used
by any pmd thread (or to the main thread).
Two more comments below.
On 06/08/2015 15:32, "Ilya Maximets" <i.maxim...@samsung.com> wrote:
>Currently tx_qid is equal to pmd->core_id. This leads to unexpected
>behavior if pmd-cpu-mask different from '/(0*)(1|3|7)?(f*)/',
>e.g. if core_ids are not sequential, or doesn't start from 0, or both.
>
>Example:
> starting 2 pmd threads with 1 port, 2 rxqs per port,
> pmd-cpu-mask = 00000014 and let dev->real_n_txq = 2
>
> It that case pmd_1->tx_qid = 2, pmd_2->tx_qid = 4 and
> txq_needs_locking = true (if device hasn't ovs_numa_get_n_cores()+1
> queues).
>
> In that case, after truncating in netdev_dpdk_send__():
> 'qid = qid % dev->real_n_txq;'
> pmd_1: qid = 2 % 2 = 0
> pmd_2: qid = 4 % 2 = 0
>
> So, both threads will call dpdk_queue_pkts() with same qid = 0.
> This is unexpected behavior if there is 2 tx queues in device.
> Queue #1 will not be used and both threads will lock queue #0
> on each send.
>
>Fix that by introducing per pmd thread hash map 'tx_queues', where will
>be stored all available tx queues for that pmd thread with
>port_no as a key(hash). All tx_qid-s will be unique per port and
>sequential to prevent described unexpected mapping after truncating.
>
>Implemented infrastructure can be used in the future to choose
>between all tx queues available for that pmd thread.
>
>Signed-off-by: Ilya Maximets <i.maxim...@samsung.com>
>---
> lib/dpif-netdev.c | 110
>++++++++++++++++++++++++++++++++++++++++++++----------
> 1 file changed, 91 insertions(+), 19 deletions(-)
>
>diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
>index c144352..309994c 100644
>--- a/lib/dpif-netdev.c
>+++ b/lib/dpif-netdev.c
>@@ -365,6 +365,13 @@ struct dp_netdev_pmd_cycles {
> atomic_ullong n[PMD_N_CYCLES];
> };
>
>+struct dp_netdev_pmd_txq {
>+ struct cmap_node node; /* In owning dp_netdev_pmd_thread's */
>+ /* 'tx_queues'. */
>+ struct dp_netdev_port *port;
>+ int tx_qid;
>+};
>+
> /* PMD: Poll modes drivers. PMD accesses devices via polling to
>eliminate
> * the performance overhead of interrupt processing. Therefore netdev
>can
> * not implement rx-wait for these devices. dpif-netdev needs to poll
>@@ -420,8 +427,8 @@ struct dp_netdev_pmd_thread {
> /* threads on same numa node. */
> unsigned core_id; /* CPU core id of this pmd thread. */
> int numa_id; /* numa node id of this pmd thread.
>*/
>- int tx_qid; /* Queue id used by this pmd thread
>to
>- * send packets on all netdevs */
>+ struct cmap tx_queues; /* Queue ids used by this pmd thread
>to
>+ * send packets to ports */
>
> /* Only a pmd thread can write on its own 'cycles' and 'stats'.
> * The main thread keeps 'stats_zero' and 'cycles_zero' as base
>@@ -2602,6 +2609,46 @@ dpif_netdev_wait(struct dpif *dpif)
> seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq);
> }
>
>+static void
>+pmd_flush_tx_queues(struct dp_netdev_pmd_thread *pmd)
>+{
>+ struct dp_netdev_pmd_txq *txq;
>+
>+ CMAP_FOR_EACH (txq, node, &pmd->tx_queues) {
>+ cmap_remove(&pmd->tx_queues, &txq->node,
>+ hash_port_no(txq->port->port_no));
>+ port_unref(txq->port);
>+ free(txq);
>+ }
>+}
The term "flush" used for txqs makes me think that the
packets in the device transmission queue will be sent.
>+
>+static void OVS_UNUSED
>+pmd_tx_queues_print(struct dp_netdev_pmd_thread *pmd)
>+{
>+ struct dp_netdev_pmd_txq *txq;
>+
>+ CMAP_FOR_EACH (txq, node, &pmd->tx_queues) {
>+ VLOG_INFO("Core_id: %d, Port: %s, tx_qid: %d\n",
>+ pmd->core_id, netdev_get_name(txq->port->netdev),
>+ txq->tx_qid);
>+ }
>+}
>+
>+static struct dp_netdev_pmd_txq *
>+pmd_lookup_txq(const struct dp_netdev_pmd_thread *pmd,
>+ const struct dp_netdev_port *port)
>+{
>+ struct dp_netdev_pmd_txq *txq;
>+
>+ CMAP_FOR_EACH_WITH_HASH (txq, node, hash_port_no(port->port_no),
>+ &pmd->tx_queues) {
>+ if (txq->port == port) {
>+ return txq;
>+ }
>+ }
>+ return NULL;
>+}
>+
> struct rxq_poll {
> struct dp_netdev_port *port;
> struct netdev_rxq *rx;
>@@ -2613,16 +2660,20 @@ pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
> {
> struct rxq_poll *poll_list = *ppoll_list;
> struct dp_netdev_port *port;
>- int n_pmds_on_numa, index, i;
>+ struct dp_netdev_pmd_txq *txq;
>+ int n_pmds_on_numa, rx_index, tx_index, i;
>
> /* Simple scheduler for netdev rx polling. */
>+ pmd_flush_tx_queues(pmd);
>+
> for (i = 0; i < poll_cnt; i++) {
> port_unref(poll_list[i].port);
> }
>
> poll_cnt = 0;
> n_pmds_on_numa = get_n_pmd_threads_on_numa(pmd->dp, pmd->numa_id);
>- index = 0;
>+ rx_index = 0;
>+ tx_index = 0;
>
> CMAP_FOR_EACH (port, node, &pmd->dp->ports) {
> /* Calls port_try_ref() to prevent the main thread
>@@ -2633,7 +2684,7 @@ pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
> int i;
>
> for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
>- if ((index % n_pmds_on_numa) == pmd->index) {
>+ if ((rx_index % n_pmds_on_numa) == pmd->index) {
> poll_list = xrealloc(poll_list,
> sizeof *poll_list * (poll_cnt +
>1));
>
>@@ -2642,7 +2693,20 @@ pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
> poll_list[poll_cnt].rx = port->rxq[i];
> poll_cnt++;
> }
>- index++;
>+ rx_index++;
>+ }
>+
>+ /* Last queue reserved for non pmd threads */
>+ for (i = 0; i < netdev_n_txq(port->netdev) - 1; i++) {
>+ if ((tx_index % n_pmds_on_numa) == pmd->index) {
>+ txq = xmalloc(sizeof *txq);
>+ port_ref(port);
>+ txq->port = port;
>+ txq->tx_qid = i;
>+ cmap_insert(&pmd->tx_queues, &txq->node,
>+ hash_port_no(port->port_no));
>+ }
>+ tx_index++;
> }
> }
> /* Unrefs the port_try_ref(). */
>@@ -2712,6 +2776,8 @@ reload:
> goto reload;
> }
>
>+ pmd_flush_tx_queues(pmd);
>+
> for (i = 0; i < poll_cnt; i++) {
> port_unref(poll_list[i].port);
> }
>@@ -2825,16 +2891,6 @@ dp_netdev_pmd_get_next(struct dp_netdev *dp,
>struct cmap_position *pos)
> return next;
> }
>
>-static int
>-core_id_to_qid(unsigned core_id)
>-{
>- if (core_id != NON_PMD_CORE_ID) {
>- return core_id;
>- } else {
>- return ovs_numa_get_n_cores();
>- }
>-}
>-
> /* Configures the 'pmd' based on the input argument. */
> static void
> dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct
>dp_netdev *dp,
>@@ -2843,7 +2899,6 @@ dp_netdev_configure_pmd(struct dp_netdev_pmd_thread
>*pmd, struct dp_netdev *dp,
> pmd->dp = dp;
> pmd->index = index;
> pmd->core_id = core_id;
>- pmd->tx_qid = core_id_to_qid(core_id);
> pmd->numa_id = numa_id;
>
> ovs_refcount_init(&pmd->ref_cnt);
>@@ -2854,9 +2909,20 @@ dp_netdev_configure_pmd(struct
>dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
> ovs_mutex_init(&pmd->flow_mutex);
> dpcls_init(&pmd->cls);
> cmap_init(&pmd->flow_table);
>- /* init the 'flow_cache' since there is no
>+ cmap_init(&pmd->tx_queues);
>+ /* init the 'flow_cache' and 'tx_queues' since there is no
> * actual thread created for NON_PMD_CORE_ID. */
> if (core_id == NON_PMD_CORE_ID) {
>+ struct dp_netdev_port *port;
>+ CMAP_FOR_EACH (port, node, &pmd->dp->ports) {
>+ if (port_try_ref(port)) {
>+ struct dp_netdev_pmd_txq * txq = xmalloc(sizeof *txq);
>+ txq->port = port;
>+ txq->tx_qid = ovs_numa_get_n_cores();
>+ cmap_insert(&pmd->tx_queues, &txq->node,
>+ hash_port_no(port->port_no));
>+ }
>+ }
> emc_cache_init(&pmd->flow_cache);
> }
> cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *,
>&pmd->node),
>@@ -2869,6 +2935,7 @@ dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread
>*pmd)
> dp_netdev_pmd_flow_flush(pmd);
> dpcls_destroy(&pmd->cls);
> cmap_destroy(&pmd->flow_table);
>+ cmap_destroy(&pmd->tx_queues);
> ovs_mutex_destroy(&pmd->flow_mutex);
> latch_destroy(&pmd->exit_latch);
> xpthread_cond_destroy(&pmd->cond);
>@@ -2885,6 +2952,7 @@ dp_netdev_del_pmd(struct dp_netdev_pmd_thread *pmd)
> * no actual thread uninit it for NON_PMD_CORE_ID. */
> if (pmd->core_id == NON_PMD_CORE_ID) {
> emc_cache_uninit(&pmd->flow_cache);
>+ pmd_flush_tx_queues(pmd);
> } else {
> latch_set(&pmd->exit_latch);
> dp_netdev_reload_pmd__(pmd);
>@@ -3480,13 +3548,17 @@ dp_execute_cb(void *aux_, struct dp_packet
>**packets, int cnt,
> struct dp_netdev *dp = pmd->dp;
> int type = nl_attr_type(a);
> struct dp_netdev_port *p;
>+ struct dp_netdev_pmd_txq *txq;
> int i;
>
> switch ((enum ovs_action_attr)type) {
> case OVS_ACTION_ATTR_OUTPUT:
> p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
> if (OVS_LIKELY(p)) {
>- netdev_send(p->netdev, pmd->tx_qid, packets, cnt, may_steal);
>+ txq = pmd_lookup_txq(pmd, p);
>+ if (OVS_UNLIKELY(!txq))
>+ break;
>+ netdev_send(p->netdev, txq->tx_qid, packets, cnt, may_steal);
I think that performing two lookups (for the port and for the txq) is more
expensive than necessary. One solution could be to include a pointer to
the
netdev in the 'txq'.
> return;
> }
> break;
>--
>2.1.4
>
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
