[PATCH net-next 1/3] macsec: remove superfluous function calls

[PATCH net-next 1/3] macsec: remove superfluous function calls

[Andreas Steinmetz]

Remove superfluous skb_share_check() and skb_unshare().
macsec_decrypt is only called by macsec_handle_frame which
already does a skb_unshare().

Signed-off-by: Andreas Steinmetz <ast@domdv.de>

--- a/drivers/net/macsec.c	2019-06-30 22:02:54.906908179 +0200
+++ b/drivers/net/macsec.c	2019-06-30 22:03:07.315785186 +0200
@@ -939,9 +939,6 @@
 	u16 icv_len = secy->icv_len;
 
 	macsec_skb_cb(skb)->valid = false;
-	skb = skb_share_check(skb, GFP_ATOMIC);
-	if (!skb)
-		return ERR_PTR(-ENOMEM);
 
 	ret = skb_cow_data(skb, 0, &trailer);
 	if (unlikely(ret < 0)) {
@@ -973,11 +970,6 @@
 
 		aead_request_set_crypt(req, sg, sg, len, iv);
 		aead_request_set_ad(req, macsec_hdr_len(macsec_skb_cb(skb)->has_sci));
-		skb = skb_unshare(skb, GFP_ATOMIC);
-		if (!skb) {
-			aead_request_free(req);
-			return ERR_PTR(-ENOMEM);
-		}
 	} else {
 		/* integrity only: all headers + data authenticated */
 		aead_request_set_crypt(req, sg, sg, icv_len, iv);

Re: [PATCH net-next 1/3] macsec: remove superfluous function calls

[Willem de Bruijn]

On Sun, Jun 30, 2019 at 4:48 PM Andreas Steinmetz <ast@domdv.de> wrote:
>
> Remove superfluous skb_share_check() and skb_unshare().
> macsec_decrypt is only called by macsec_handle_frame which
> already does a skb_unshare().

There is a subtle difference. skb_unshare() acts on cloned skbs, not
shared skbs.

It creates a private copy of data if clones may access it
concurrently, which clearly is needed when modifying for decryption.

At rx_handler, I don't think a shared skb happen (unlike clones, e.g.,
from packet sockets). But it is peculiar that most, if not all,
rx_handlers seem to test for it. That have started with the bridge
device:

commit 7b995651e373d6424f81db23f2ec503306dfd7f0
Author: Herbert Xu <herbert@gondor.apana.org.au>
Date:   Sun Oct 14 00:39:01 2007 -0700

    [BRIDGE]: Unshare skb upon entry

    Due to the special location of the bridging hook, it should never see a
    shared packet anyway (certainly not with any in-kernel code).  So it
    makes sense to unshare the skb there if necessary as that will greatly
    simplify the code below it (in particular, netfilter).

Anyway, remove the check only if certain.

>
> Signed-off-by: Andreas Steinmetz <ast@domdv.de>
>
> --- a/drivers/net/macsec.c      2019-06-30 22:02:54.906908179 +0200
> +++ b/drivers/net/macsec.c      2019-06-30 22:03:07.315785186 +0200
> @@ -939,9 +939,6 @@
>         u16 icv_len = secy->icv_len;
>
>         macsec_skb_cb(skb)->valid = false;
> -       skb = skb_share_check(skb, GFP_ATOMIC);
> -       if (!skb)
> -               return ERR_PTR(-ENOMEM);
>
>         ret = skb_cow_data(skb, 0, &trailer);
>         if (unlikely(ret < 0)) {
> @@ -973,11 +970,6 @@
>
>                 aead_request_set_crypt(req, sg, sg, len, iv);
>                 aead_request_set_ad(req, macsec_hdr_len(macsec_skb_cb(skb)->has_sci));
> -               skb = skb_unshare(skb, GFP_ATOMIC);
> -               if (!skb) {
> -                       aead_request_free(req);
> -                       return ERR_PTR(-ENOMEM);
> -               }
>         } else {
>                 /* integrity only: all headers + data authenticated */
>                 aead_request_set_crypt(req, sg, sg, icv_len, iv);
>

Re: [PATCH net-next 1/3] macsec: remove superfluous function calls

[Willem de Bruijn]

On Sun, Jun 30, 2019 at 9:57 PM Willem de Bruijn
<willemdebruijn.kernel@gmail.com> wrote:
>
> On Sun, Jun 30, 2019 at 4:48 PM Andreas Steinmetz <ast@domdv.de> wrote:
> >
> > Remove superfluous skb_share_check() and skb_unshare().
> > macsec_decrypt is only called by macsec_handle_frame which
> > already does a skb_unshare().
>
> There is a subtle difference. skb_unshare() acts on cloned skbs, not
> shared skbs.
>
> It creates a private copy of data if clones may access it
> concurrently, which clearly is needed when modifying for decryption.
>
> At rx_handler, I don't think a shared skb happen (unlike clones, e.g.,
> from packet sockets). But it is peculiar that most, if not all,
> rx_handlers seem to test for it. That have started with the bridge
> device:
>
> commit 7b995651e373d6424f81db23f2ec503306dfd7f0
> Author: Herbert Xu <herbert@gondor.apana.org.au>
> Date:   Sun Oct 14 00:39:01 2007 -0700
>
>     [BRIDGE]: Unshare skb upon entry
>
>     Due to the special location of the bridging hook, it should never see a
>     shared packet anyway (certainly not with any in-kernel code).  So it
>     makes sense to unshare the skb there if necessary as that will greatly
>     simplify the code below it (in particular, netfilter).
>
> Anyway, remove the check only if certain.

Did you see this point? I notice the v2 without this mentioned. I
don't think you can remove an skb_share_check solely on the basis of a
previous skb_unshare. I agree that here a shared skb is unlikely, but
that is for a very different, less obvious, reason.




>
> >
> > Signed-off-by: Andreas Steinmetz <ast@domdv.de>
> >
> > --- a/drivers/net/macsec.c      2019-06-30 22:02:54.906908179 +0200
> > +++ b/drivers/net/macsec.c      2019-06-30 22:03:07.315785186 +0200
> > @@ -939,9 +939,6 @@
> >         u16 icv_len = secy->icv_len;
> >
> >         macsec_skb_cb(skb)->valid = false;
> > -       skb = skb_share_check(skb, GFP_ATOMIC);
> > -       if (!skb)
> > -               return ERR_PTR(-ENOMEM);
> >
> >         ret = skb_cow_data(skb, 0, &trailer);
> >         if (unlikely(ret < 0)) {
> > @@ -973,11 +970,6 @@
> >
> >                 aead_request_set_crypt(req, sg, sg, len, iv);
> >                 aead_request_set_ad(req, macsec_hdr_len(macsec_skb_cb(skb)->has_sci));
> > -               skb = skb_unshare(skb, GFP_ATOMIC);
> > -               if (!skb) {
> > -                       aead_request_free(req);
> > -                       return ERR_PTR(-ENOMEM);
> > -               }
> >         } else {
> >                 /* integrity only: all headers + data authenticated */
> >                 aead_request_set_crypt(req, sg, sg, icv_len, iv);
> >