2022-02-09 10:58:46

by Ido Schimmel

[permalink] [raw]
Subject: Re: [PATCH v7 net-next 2/2] net: drop_monitor: support drop reason

On Tue, Feb 08, 2022 at 03:28:36PM +0800, [email protected] wrote:
> From: Menglong Dong <[email protected]>
>
> In the commit c504e5c2f964 ("net: skb: introduce kfree_skb_reason()")
> drop reason is introduced to the tracepoint of kfree_skb. Therefore,
> drop_monitor is able to report the drop reason to users by netlink.
>
> The drop reasons are reported as string to users, which is exactly
> the same as what we do when reporting it to ftrace.
>
> Signed-off-by: Menglong Dong <[email protected]>
> ---
> v7:
> - take the size of NET_DM_ATTR_REASON into accounting in
> net_dm_packet_report_size()
> - let compiler define the size of drop_reasons
>
> v6:
> - check the range of drop reason in net_dm_packet_report_fill()
>
> v5:
> - check if drop reason larger than SKB_DROP_REASON_MAX
>
> v4:
> - report drop reasons as string
>
> v3:
> - referring to cb->reason and cb->pc directly in
> net_dm_packet_report_fill()
>
> v2:
> - get a pointer to struct net_dm_skb_cb instead of local var for
> each field
> ---
> include/uapi/linux/net_dropmon.h | 1 +
> net/core/drop_monitor.c | 34 ++++++++++++++++++++++++++++----
> 2 files changed, 31 insertions(+), 4 deletions(-)
>
> diff --git a/include/uapi/linux/net_dropmon.h b/include/uapi/linux/net_dropmon.h
> index 66048cc5d7b3..1bbea8f0681e 100644
> --- a/include/uapi/linux/net_dropmon.h
> +++ b/include/uapi/linux/net_dropmon.h
> @@ -93,6 +93,7 @@ enum net_dm_attr {
> NET_DM_ATTR_SW_DROPS, /* flag */
> NET_DM_ATTR_HW_DROPS, /* flag */
> NET_DM_ATTR_FLOW_ACTION_COOKIE, /* binary */
> + NET_DM_ATTR_REASON, /* string */
>
> __NET_DM_ATTR_MAX,
> NET_DM_ATTR_MAX = __NET_DM_ATTR_MAX - 1
> diff --git a/net/core/drop_monitor.c b/net/core/drop_monitor.c
> index 7b288a121a41..28c55d605566 100644
> --- a/net/core/drop_monitor.c
> +++ b/net/core/drop_monitor.c
> @@ -48,6 +48,19 @@
> static int trace_state = TRACE_OFF;
> static bool monitor_hw;
>
> +#undef EM
> +#undef EMe
> +
> +#define EM(a, b) [a] = #b,
> +#define EMe(a, b) [a] = #b
> +
> +/* drop_reasons is used to translate 'enum skb_drop_reason' to string,
> + * which is reported to user space.
> + */
> +static const char * const drop_reasons[] = {
> + TRACE_SKB_DROP_REASON
> +};
> +
> /* net_dm_mutex
> *
> * An overall lock guarding every operation coming from userspace.
> @@ -126,6 +139,7 @@ struct net_dm_skb_cb {
> struct devlink_trap_metadata *hw_metadata;
> void *pc;
> };
> + enum skb_drop_reason reason;
> };
>
> #define NET_DM_SKB_CB(__skb) ((struct net_dm_skb_cb *)&((__skb)->cb[0]))
> @@ -498,6 +512,7 @@ static void net_dm_packet_trace_kfree_skb_hit(void *ignore,
> {
> ktime_t tstamp = ktime_get_real();
> struct per_cpu_dm_data *data;
> + struct net_dm_skb_cb *cb;
> struct sk_buff *nskb;
> unsigned long flags;
>
> @@ -508,7 +523,9 @@ static void net_dm_packet_trace_kfree_skb_hit(void *ignore,
> if (!nskb)
> return;
>
> - NET_DM_SKB_CB(nskb)->pc = location;
> + cb = NET_DM_SKB_CB(nskb);
> + cb->reason = reason;
> + cb->pc = location;
> /* Override the timestamp because we care about the time when the
> * packet was dropped.
> */
> @@ -574,6 +591,8 @@ static size_t net_dm_packet_report_size(size_t payload_len)
> nla_total_size(sizeof(u32)) +
> /* NET_DM_ATTR_PROTO */
> nla_total_size(sizeof(u16)) +
> + /* NET_DM_ATTR_REASON */
> + nla_total_size(SKB_DR_MAX_LEN + 1) +

Nothing ensures that the reason is not longer than this length and
nothing ensures that this assumption remains valid as more reasons are
added.

I think "SKB_DR_MAX_LEN" can be removed completely. Pass "reason" to
this function and do "strlen(drop_reasons[reason]) + 1". Any reason it
can't work?

> /* NET_DM_ATTR_PAYLOAD */
> nla_total_size(payload_len);
> }
> @@ -606,8 +625,9 @@ static int net_dm_packet_report_in_port_put(struct sk_buff *msg, int ifindex,
> static int net_dm_packet_report_fill(struct sk_buff *msg, struct sk_buff *skb,
> size_t payload_len)
> {
> - u64 pc = (u64)(uintptr_t) NET_DM_SKB_CB(skb)->pc;
> + struct net_dm_skb_cb *cb = NET_DM_SKB_CB(skb);
> char buf[NET_DM_MAX_SYMBOL_LEN];
> + unsigned int reason;
> struct nlattr *attr;
> void *hdr;
> int rc;
> @@ -620,10 +640,16 @@ static int net_dm_packet_report_fill(struct sk_buff *msg, struct sk_buff *skb,
> if (nla_put_u16(msg, NET_DM_ATTR_ORIGIN, NET_DM_ORIGIN_SW))
> goto nla_put_failure;
>
> - if (nla_put_u64_64bit(msg, NET_DM_ATTR_PC, pc, NET_DM_ATTR_PAD))
> + if (nla_put_u64_64bit(msg, NET_DM_ATTR_PC, (u64)(uintptr_t)cb->pc,
> + NET_DM_ATTR_PAD))
> + goto nla_put_failure;
> +
> + reason = (unsigned int)cb->reason;
> + if (reason < SKB_DROP_REASON_MAX &&

In which cases can this happen? Might be better to perform this
validation in net_dm_packet_trace_kfree_skb_hit() and set "cb->reason"
to "SKB_DROP_REASON_NOT_SPECIFIED" in this case. That way we don't need
to perform the validation in later code paths

> + nla_put_string(msg, NET_DM_ATTR_REASON, drop_reasons[reason]))
> goto nla_put_failure;
>
> - snprintf(buf, sizeof(buf), "%pS", NET_DM_SKB_CB(skb)->pc);
> + snprintf(buf, sizeof(buf), "%pS", cb->pc);
> if (nla_put_string(msg, NET_DM_ATTR_SYMBOL, buf))
> goto nla_put_failure;
>
> --
> 2.34.1
>