Re: [PATCH] [openvswitch v4] openvswitch: Add support to count upcall packets
From: wangchuanlei
Date: Wed Nov 23 2022 - 21:24:34 EST
Hi,
Thank you for review! I will give a new verson of patch based on your comments,
and i give a explanation on every comments from you, please see below!
Best reagrds!
wangchuanlei
From: Alexander Lobakin [mailto:alexandr.lobakin@xxxxxxxxx]
To: wangchuanlei@xxxxxxxxxx
> From: wangchuanlei <wangchuanlei@xxxxxxxxxx>
> Date: Wed, 23 Nov 2022 04:18:43 -0500
> Add support to count upall packets, when kmod of openvswitch upcall to
> userspace , here count the number of packets for upcall succeed and
> failed, which is a better way to see how many packets upcalled to
> userspace(ovs-vswitchd) on every interfaces.
>
> Here optimize the function used by comments of v3.
>
> Changes since v3:
> - use nested NLA_NESTED attribute in netlink message
>
> Changes since v2:
> - add count of upcall failed packets
>
> Changes since v1:
> - add count of upcall succeed packets
>
> Signed-off-by: wangchuanlei <wangchuanlei@xxxxxxxxxx>
> ---
> include/uapi/linux/openvswitch.h | 19 ++++++++++++
> net/openvswitch/datapath.c | 52 ++++++++++++++++++++++++++++++++
> net/openvswitch/datapath.h | 12 ++++++++
> net/openvswitch/vport.c | 48 +++++++++++++++++++++++++++++
> net/openvswitch/vport.h | 6 ++++
> 5 files changed, 137 insertions(+)
>
> diff --git a/include/uapi/linux/openvswitch.h
> b/include/uapi/linux/openvswitch.h
> index 94066f87e9ee..fa13bce15fae 100644
> --- a/include/uapi/linux/openvswitch.h
> +++ b/include/uapi/linux/openvswitch.h
> @@ -126,6 +126,11 @@ struct ovs_vport_stats {
> __u64 tx_dropped; /* no space available in linux */
> };
>
> +struct ovs_vport_upcall_stats {
> + uint64_t upcall_success; /* total packets upcalls succeed */
> + uint64_t upcall_fail; /* total packets upcalls failed */
Please no uint64_t int the UAPI headers. __u64 as above. --Yes !
> +};
> +
> /* Allow last Netlink attribute to be unaligned */
> #define OVS_DP_F_UNALIGNED (1 << 0)
>
> @@ -277,11 +282,25 @@ enum ovs_vport_attr {
> OVS_VPORT_ATTR_PAD,
> OVS_VPORT_ATTR_IFINDEX,
> OVS_VPORT_ATTR_NETNSID,
> + OVS_VPORT_ATTR_UPCALL_STATS, /* struct ovs_vport_upcall_stats */
> __OVS_VPORT_ATTR_MAX
> };
>
> #define OVS_VPORT_ATTR_MAX (__OVS_VPORT_ATTR_MAX - 1)
>
> +/**
> + * enum ovs_vport_upcall_attr - attributes for %OVS_VPORT_UPCALL*
> +commands
> + * @OVS_VPORT_UPCALL_SUCCESS: 64-bit upcall success packets.
> + * @OVS_VPORT_UPCALL_FAIL: 64-bit upcall fail packets.
> + */
> +enum ovs_vport_upcall_attr {
> + OVS_VPORT_UPCALL_SUCCESS, /* 64-bit upcall success packets */
> + OVS_VPORT_UPCALL_FAIL, /* 64-bit upcall fail packets */
> + __OVS_VPORT_UPCALL_MAX
> +};
> +
> +#define OVS_VPORT_UPCALL_MAX (__OVS_VPORT_UPCALL_MAX-1)
Spaces around arithm operator ('-'). --Yes !
> +
> enum {
> OVS_VXLAN_EXT_UNSPEC,
> OVS_VXLAN_EXT_GBP, /* Flag or __u32 */
> diff --git a/net/openvswitch/datapath.c b/net/openvswitch/datapath.c
> index c8a9075ddd0a..5254c51cfa60 100644
> --- a/net/openvswitch/datapath.c
> +++ b/net/openvswitch/datapath.c
> @@ -209,6 +209,25 @@ static struct vport *new_vport(const struct vport_parms *parms)
> return vport;
> }
>
> +static void ovs_vport_upcalls(struct sk_buff *skb,
> + const struct dp_upcall_info *upcall_info,
> + bool upcall_success)
^^^^^^^^^^^^^^^^^^^
Just `bool success`? It's clear that is's about upcalls, I don't see a need to repeat it in every argument's name.
--Yes !
> +{
> + if (upcall_info->cmd == OVS_PACKET_CMD_MISS ||
> + upcall_info->cmd == OVS_PACKET_CMD_ACTION) {
if (cmd != MISS && cmd != ACTION)
return;
Saves 1 indent level. --you are right!
> + const struct vport *p = OVS_CB(skb)->input_vport;
> + struct vport_upcall_stats_percpu *vport_stats;
> +
> + vport_stats = this_cpu_ptr(p->vport_upcall_stats_percpu);
Why make a separate structure? You can just expand dp_stats_percpu, this function would then be just a couple lines in ovs_dp_upcall().
-- emm, beacause of this statistics based on vport, so new structure should insert to "struct vport"
> + u64_stats_update_begin(&vport_stats->syncp);
> + if (upcall_success)
> + u64_stats_inc(&vport_stats->n_upcall_success);
> + else
> + u64_stats_inc(&vport_stats->n_upcall_fail);
> + u64_stats_update_end(&vport_stats->syncp);
> + }
> +}
> +
> void ovs_dp_detach_port(struct vport *p) {
> ASSERT_OVSL();
> @@ -216,6 +235,9 @@ void ovs_dp_detach_port(struct vport *p)
> /* First drop references to device. */
> hlist_del_rcu(&p->dp_hash_node);
>
> + /* Free percpu memory */
> + free_percpu(p->vport_upcall_stats_percpu);
> +
> /* Then destroy it. */
> ovs_vport_del(p);
> }
> @@ -305,6 +327,8 @@ int ovs_dp_upcall(struct datapath *dp, struct sk_buff *skb,
> err = queue_userspace_packet(dp, skb, key, upcall_info, cutlen);
> else
> err = queue_gso_packets(dp, skb, key, upcall_info, cutlen);
> +
> + ovs_vport_upcalls(skb, upcall_info, !err);
> if (err)
> goto err;
Also, as you may see, your ::upcall_fail counter will be always exactly the same as stats->n_lost. So there's no point introducing a new one.
However, you can expand the structure dp_stats_percpu and add a new field there which would store the number of successfull upcalls.
...but I don't see a reason for this to be honest. From my PoV, it's better to count the number of successfully processed packets at the end of queue_userspace_packet() right before the 'out:'
label[0]. But please make sure then you don't duplicate some other counter (I'm not deep into OvS, so can't say for sure if there's anything similar to what you want).
--in ovs , as stats->n_lost only count the sum of packets of all ports, not on individal port , so expand the structure dp_stats_percpu may be not suitable
--and count upcall failed packets is useful beacuse no all of upcall packets are successfully sent。
>
> @@ -1825,6 +1849,13 @@ static int ovs_dp_cmd_new(struct sk_buff *skb, struct genl_info *info)
> goto err_destroy_portids;
> }
>
> + vport->vport_upcall_stats_percpu =
This can be at least twice shorter, e.g. 'upcall_stats'. Don't try to describe every detail in symbol names.
--yes!
> + netdev_alloc_pcpu_stats(struct vport_upcall_stats_percpu);
> + if (!vport->vport_upcall_stats_percpu) {
> + err = -ENOMEM;
> + goto err_destroy_upcall_stats;
I know you followed the previous label logics, but you actually aren't destroying the stats under this label. Here you should have `goto err_destroy_portids` as that's what you're actually doing on that error path.
--here is just keep format of code, and has no influence on function
> + }
> +
> err = ovs_dp_cmd_fill_info(dp, reply, info->snd_portid,
> info->snd_seq, 0, OVS_DP_CMD_NEW);
> BUG_ON(err < 0);
[...]
> @@ -2278,6 +2321,14 @@ static int ovs_vport_cmd_new(struct sk_buff *skb, struct genl_info *info)
> goto exit_unlock_free;
> }
>
> + vport->vport_upcall_stats_percpu =
> + netdev_alloc_pcpu_stats(struct vport_upcall_stats_percpu);
> +
> + if (!vport->vport_upcall_stats_percpu) {
> + err = -ENOMEM;
> + goto exit_unlock_free;
> + }
Why do you allocate them twice?
-- here is in different code segment on in vport_cmd_new , the other is in dp_cmd_new, they are has no collisions
> +
> err = ovs_vport_cmd_fill_info(vport, reply, genl_info_net(info),
> info->snd_portid, info->snd_seq, 0,
> OVS_VPORT_CMD_NEW, GFP_KERNEL);
[...]
> @@ -50,6 +50,18 @@ struct dp_stats_percpu {
> struct u64_stats_sync syncp;
> };
>
> +/**
> + * struct vport_upcall_stats_percpu - per-cpu packet upcall
> +statistics for
> + * a given vport.
> + * @n_upcall_success: Number of packets that upcall to userspace succeed.
> + * @n_upcall_fail: Number of packets that upcall to userspace failed.
> + */
> +struct vport_upcall_stats_percpu {
> + u64_stats_t n_upcall_success;
> + u64_stats_t n_upcall_fail;
> + struct u64_stats_sync syncp;
Nit: syncp would feel better at the start. You could then sort the structure by the field hit probability and reduce cache misses %)
--ok
> +};
> +
> /**
> * struct dp_nlsk_pids - array of netlink portids of for a datapath.
> * This is used when OVS_DP_F_DISPATCH_UPCALL_PER_CPU
> diff --git a/net/openvswitch/vport.c b/net/openvswitch/vport.c index
> 82a74f998966..a69c9356b57c 100644
> --- a/net/openvswitch/vport.c
> +++ b/net/openvswitch/vport.c
> @@ -284,6 +284,54 @@ void ovs_vport_get_stats(struct vport *vport, struct ovs_vport_stats *stats)
> stats->tx_packets = dev_stats->tx_packets; }
>
> +/**
> + * ovs_vport_get_upcall_stats - retrieve upcall stats
> + *
> + * @vport: vport from which to retrieve the stats
> + * @ovs_vport_upcall_stats: location to store stats
> + *
> + * Retrieves upcall stats for the given device.
> + *
> + * Must be called with ovs_mutex or rcu_read_lock.
> + */
> +void ovs_vport_get_upcall_stats(struct vport *vport, struct
> +ovs_vport_upcall_stats *stats) {
> + int i;
> +
> + stats->upcall_success = 0;
> + stats->upcall_fail = 0;
> +
> + for_each_possible_cpu(i) {
> + const struct vport_upcall_stats_percpu *percpu_upcall_stats;
You wouldn't need to linewrap the lines below if you didn't make its name so huge.
--you are right, i would change the name!
> + unsigned int start;
> +
> + percpu_upcall_stats = per_cpu_ptr(vport->vport_upcall_stats_percpu, i);
> + do {
> + start = u64_stats_fetch_begin(&percpu_upcall_stats->syncp);
> + stats->upcall_success +=
> + u64_stats_read(&percpu_upcall_stats->n_upcall_success);
> + stats->upcall_fail += u64_stats_read(&percpu_upcall_stats->n_upcall_fail);
> + } while (u64_stats_fetch_retry(&percpu_upcall_stats->syncp, start));
> + }
> +}
> +
> +int ovs_vport_put_upcall_stats(struct sk_buff *skb,
> + struct ovs_vport_upcall_stats *stats) {
> + if (nla_put_u64_64bit(skb, OVS_VPORT_UPCALL_SUCCESS, stats->upcall_success,
> + OVS_VPORT_ATTR_PAD))
> + goto nla_put_failure;
> +
> + if (nla_put_u64_64bit(skb, OVS_VPORT_UPCALL_FAIL, stats->upcall_fail,
> + OVS_VPORT_ATTR_PAD))
> + goto nla_put_failure;
> +
> + return 0;
> +
> +nla_put_failure:
> + return -EMSGSIZE;
goto with only one action makes no sense, just exit directly.
--you are right!
> +}
> +
> /**
> * ovs_vport_get_options - retrieve device options
> *
> diff --git a/net/openvswitch/vport.h b/net/openvswitch/vport.h index
> 7d276f60c000..02cf8c589588 100644
> --- a/net/openvswitch/vport.h
> +++ b/net/openvswitch/vport.h
> @@ -32,6 +32,11 @@ struct vport *ovs_vport_locate(const struct net
> *net, const char *name);
>
> void ovs_vport_get_stats(struct vport *, struct ovs_vport_stats *);
>
> +void ovs_vport_get_upcall_stats(struct vport *vport,
> + struct ovs_vport_upcall_stats *stats); int
> +ovs_vport_put_upcall_stats(struct sk_buff *skb,
> + struct ovs_vport_upcall_stats *stats);
> +
> int ovs_vport_set_options(struct vport *, struct nlattr *options);
> int ovs_vport_get_options(const struct vport *, struct sk_buff *);
>
> @@ -78,6 +83,7 @@ struct vport {
> struct hlist_node hash_node;
> struct hlist_node dp_hash_node;
> const struct vport_ops *ops;
> + struct vport_upcall_stats_percpu __percpu
> +*vport_upcall_stats_percpu;
Almost 80 columns in one field definition :D
--yes i would change the name !
>
> struct list_head detach_list;
> struct rcu_head rcu;
> --
> 2.27.0
[0] https://elixir.bootlin.com/linux/v6.1-rc6/source/net/openvswitch/datapath.c#L557
Thanks,
Olek