Re: [RFC net-next 1/6] net: add kfree_skb_for_sk function

[Yan Zhai]

On Fri, May 31, 2024 at 12:32 PM Eric Dumazet <edumazet@xxxxxxxxxx> wrote:
>
> On Fri, May 31, 2024 at 6:58 PM Yan Zhai <yan@xxxxxxxxxxxxxx> wrote:
> >
> > Hi Eric,
> >
> >  Thanks for the feedback.
> >
> > On Fri, May 31, 2024 at 1:51 AM Eric Dumazet <edumazet@xxxxxxxxxx> wrote:
> > >
> > > On Thu, May 30, 2024 at 11:46 PM Yan Zhai <yan@xxxxxxxxxxxxxx> wrote:
> > > >
> > > > Implement a new kfree_skb_for_sk to replace kfree_skb_reason on a few
> > > > local receive path. The function accepts an extra receiving socket
> > > > argument, which will be set in skb->cb for kfree_skb/consume_skb
> > > > tracepoint consumption. With this extra bit of information, it will be
> > > > easier to attribute dropped packets to netns/containers and
> > > > sockets/services for performance and error monitoring purposes.
> > >
> > > This is a lot of code churn...
> > >
> > > I have to ask : Why not simply adding an sk parameter to an existing
> > > trace point ?
> > >
> > Modifying a signature of the current tracepoint seems like a breaking
> > change, that's why I was saving the context inside skb->cb, hoping to
> > not impact any existing programs watching this tracepoint. But
> > thinking it twice, it might not cause a problem if the signature
> > becomes:
> >
> >  trace_kfree_skb(const struct sk_buff *skb, void *location, enum
> > skb_drop_reason reason, const struct sock *sk)
> >
> > As return values are usually not a thing for tracepoints, it is
> > probably still compatible. The cons is that the last "sk" still breaks
> > the integrity of naming. How about making a "kfree_skb_context"
> > internal struct and putting it as the last argument to "hide" the
> > naming confusion?
> >
> > > If this not possible, I would rather add new tracepoints, adding new classes,
> > > because it will ease your debugging :
> > >
> > > When looking for TCP drops, simply use a tcp_event_sk_skb_reason instance,
> > > and voila, no distractions caused by RAW/ICMP/ICMPv6/af_packet drops.
> > >
> > > DECLARE_EVENT_CLASS(tcp_event_sk_skb_reason,
> > >
> > >      TP_PROTO(const struct sock *sk, const struct sk_buff *skb, enum
> > > skb_drop_reason reason),
> > > ...
> > > );
> >
> > The alternative of adding another tracepoint could indeed work, we had
> > a few cases like that in the past, e.g.
> >
> > https://lore.kernel.org/lkml/20230711043453.64095-1-ivan@xxxxxxxxxxxxxx/
> > https://lore.kernel.org/netdev/20230707043923.35578-1-ivan@xxxxxxxxxxxxxx/
> >
> > But it does feel like a whack-a-mole thing. The problems are solvable
> > if we extend the kfree_skb tracepoint, so I would prefer to not add a
> > new tracepoint.
>
> Solvable with many future merge conflicts for stable teams.
>
I don't quite follow it. I think this specific commit using skb->cb is
unnecessary so I am going to re-work it. As you initially mentioned,
maybe I should just extend kfree_skb tracepoint. I saw a similar
change dd1b527831a3("net: add location to trace_consume_skb()"), is it
something I might follow, or do you specifically mean changes like
this can annoy stable teams?

>
> >
> > >
> > > Also, the name ( kfree_skb_for_sk) and order of parameters is confusing.
> > >
> > > I always prefer this kind of ordering/names :
> > >
> > > void sk_skb_reason_drop( [struct net *net ] // not relevant here, but
> > > to expand the rationale
> > >               struct sock *sk, struct sk_buff *skb, enum skb_drop_reason reason)
> > >
> > > Looking at the name, we immediately see the parameter order.
> > >
> > > The consume one (no @reason there) would be called
> > >
> > > void sk_skb_consume(struct sock *sk, struct sk_buff *skb);
> >
> > I was intending to keep the "kfree_skb" prefix initially since it
> > would appear less surprising to kernel developers who used kfree_skb
> > and kfree_skb_reason. But your points do make good sense. How about
> > "kfree_sk_skb_reason" and "consume_sk_skb" here?
> >
>
> IMO kfree_skb() and consume_skb() were a wrong choice. We have to live
> with them.
>
> It should have been skb_free(), skb_consume(), skb_alloc(),
> to be consistent.
>
> Following (partial) list was much better:
>
> skb_add_rx_frag_netmem, skb_coalesce_rx_frag, skb_pp_cow_data,
> skb_cow_data_for_xdp,
> skb_dump, skb_tx_error, skb_morph, skb_zerocopy_iter_stream, skb_copy_ubufs,
> skb_clone, skb_headers_offset_update, skb_copy_header, skb_copy,
> skb_realloc_headroom, skb_expand_head, skb_copy_expand, skb_put,
> skb_push, skb_pull, skb_pull_data, skb_trim, skb_copy_bits,
> skb_splice_bits, skb_send_sock_locked, skb_store_bits,
> skb_checksum, skb_copy_and_csum_bits, skb_zerocopy_headlen,
> skb_zerocopy, skb_copy_and_csum_dev, skb_dequeue,
> skb_dequeue_tail, skb_queue_purge_reason, skb_errqueue_purge,
> skb_queue_head, skb_queue_tail, skb_unlink, skb_append,
> skb_split, skb_prepare_seq_read, skb_seq_read, skb_abort_seq_read,
> skb_find_text, skb_append_pagefrags, skb_pull_rcsum, skb_segment_list,
> skb_segment, skb_to_sgvec, skb_to_sgvec_nomark, skb_cow_data, skb_clone_sk,
> skb_complete_tx_timestamp, skb_tstamp_tx, skb_complete_wifi_ack,
> skb_partial_csum_set, skb_checksum_setup, skb_checksum_trimmed,
> skb_try_coalesce, skb_scrub_packet, skb_vlan_untag, skb_ensure_writable,
> skb_ensure_writable_head_tail, skb_vlan_pop, skb_vlan_push, skb_eth_pop,
> skb_eth_push, skb_mpls_push, skb_mpls_pop, skb_mpls_update_lse,
> skb_mpls_dec_ttl, skb_condense, skb_ext_add, skb_splice_from_iter
>
> (just to make my point very very clear)
>
> Instead we have a myriad of functions with illogical parameter
> ordering vs their names.
>
> I see no reason to add more confusion for new helpers.

ACK. Thanks for clarifying.

Yan