Re: [PATCH net] net: Remove redundant calls of sk_tx_queue_clear().

From:   Eric Dumazet <edumazet@google.com>
Date:   Wed, 27 Jan 2021 18:05:24 +0100

On Wed, Jan 27, 2021 at 5:52 PM Kuniyuki Iwashima [off-list ref] wrote:

quoted

From:   Eric Dumazet <edumazet@google.com>
Date:   Wed, 27 Jan 2021 15:54:32 +0100

quoted

On Wed, Jan 27, 2021 at 1:50 PM Kuniyuki Iwashima [off-list ref] wrote:

quoted

The commit 41b14fb8724d ("net: Do not clear the sock TX queue in
sk_set_socket()") removes sk_tx_queue_clear() from sk_set_socket() and adds
it instead in sk_alloc() and sk_clone_lock() to fix an issue introduced in
the commit e022f0b4a03f ("net: Introduce sk_tx_queue_mapping"). However,
the original commit had already put sk_tx_queue_clear() in sk_prot_alloc():
the callee of sk_alloc() and sk_clone_lock(). Thus sk_tx_queue_clear() is
called twice in each path currently.

Are you sure ?

I do not clearly see the sk_tx_queue_clear() call from the cloning part.

Please elaborate.

If sk is not NULL in sk_prot_alloc(), sk_tx_queue_clear() is called [1].
Also the callers of sk_prot_alloc() are only sk_alloc() and sk_clone_lock().
If they finally return not NULL pointer, sk_tx_queue_clear() is called in
each function [2][3].

In the cloning part, sock_copy() is called after sk_prot_alloc(), but
skc_tx_queue_mapping is defined between skc_dontcopy_begin and
skc_dontcopy_end in struct sock_common [4]. So, sock_copy() does not
overwrite skc_tx_queue_mapping, and thus we can initialize it in
sk_prot_alloc().

That is a lot of assumptions.

What guarantees do we have that skc_tx_queue_mapping will never be
moved out of this section ?
AFAIK it was there by accident, for cache locality reasons, that might
change in the future as we add more stuff in socket.

I feel this optimization is risky for future changes, for a code path
that is spending thousands of cycles anyway.

If someone try to move skc_tx_queue_mapping out of the section, should
they take care about where it is used ?

But I agree that we should not write error-prone code.

Currently, sk_tx_queue_clear() is the only initialization code in
sk_prot_alloc(). So, does it make sense to remove sk_tx_queue_clear() in
sk_prot_alloc() so that it does only allocation and other fields are
initialized in each caller ?

quoted

[1] sk_prot_alloc
https://github.com/torvalds/linux/blob/master/net/core/sock.c#L1693

[2] sk_alloc
https://github.com/torvalds/linux/blob/master/net/core/sock.c#L1762

[3] sk_clone_lock
https://github.com/torvalds/linux/blob/master/net/core/sock.c#L1986

[4] struct sock_common
https://github.com/torvalds/linux/blob/master/include/net/sock.h#L218-L240

quoted

In any case, this seems to be a candidate for net-next, this is not
fixing a bug,
this would be an optimization at most, and potentially adding a bug.

So if you resend this patch, you can mention the old commit in the changelog,
but do not add a dubious Fixes: tag

I see.

I will remove the tag and resend this as a net-next candidate.

Thank you,
Kuniyuki

quoted

quoted

This patch removes the redundant calls of sk_tx_queue_clear() in sk_alloc()
and sk_clone_lock().

Fixes: 41b14fb8724d ("net: Do not clear the sock TX queue in sk_set_socket()")
CC: Tariq Toukan <redacted>
CC: Boris Pismenny <redacted>
Signed-off-by: Kuniyuki Iwashima <redacted>
Reviewed-by: Amit Shah <redacted>
---
 net/core/sock.c | 2 --
 1 file changed, 2 deletions(-)

diff --git a/net/core/sock.c b/net/core/sock.c
index bbcd4b97eddd..5c665ee14159 100644
--- a/net/core/sock.c
+++ b/net/core/sock.c

@@ -1759,7 +1759,6 @@ struct sock *sk_alloc(struct net *net, int family, gfp_t priority,
                cgroup_sk_alloc(&sk->sk_cgrp_data);
                sock_update_classid(&sk->sk_cgrp_data);
                sock_update_netprioidx(&sk->sk_cgrp_data);
-               sk_tx_queue_clear(sk);
        }

        return sk;

@@ -1983,7 +1982,6 @@ struct sock *sk_clone_lock(const struct sock *sk, const gfp_t priority)
                 */
                sk_refcnt_debug_inc(newsk);
                sk_set_socket(newsk, NULL);
-               sk_tx_queue_clear(newsk);
                RCU_INIT_POINTER(newsk->sk_wq, NULL);

                if (newsk->sk_prot->sockets_allocated)
--

2.17.2 (Apple Git-113)