| Message ID | 20201031004918.463475-1-xie.he.0141@gmail.com |
|---|---|
| Headers | show |
| Series | net: hdlc_fr: Improve fr_rx and add support for any Ethertype | expand |
On Fri, Oct 30, 2020 at 8:49 PM Xie He <xie.he.0141@gmail.com> wrote: > > 1. > Change the skb->len check from "<= 4" to "< 4". > At first we only need to ensure a 4-byte header is present. We indeed > normally need the 5th byte, too, but it'd be more logical and cleaner > to check its existence when we actually need it. > > 2. > Add an fh->ea2 check to the initial checks in fr_rx. fh->ea2 == 1 means > the second address byte is the final address byte. We only support the > case where the address length is 2 bytes. If the address length is not > 2 bytes, the control field and the protocol field would not be the 3rd > and 4th byte as we assume. (Say it is 3 bytes, then the control field > and the protocol field would be the 4th and 5th byte instead.) > > Cc: Willem de Bruijn <willemdebruijn.kernel@gmail.com> > Cc: Krzysztof Halasa <khc@pm.waw.pl> > Signed-off-by: Xie He <xie.he.0141@gmail.com> Acked-by: Willem de Bruijn <willemb@google.com>
On Fri, Oct 30, 2020 at 8:50 PM Xie He <xie.he.0141@gmail.com> wrote: > > When the fr_rx function drops a received frame (because the protocol type > is not supported, or because the PVC virtual device that corresponds to > the DLCI number and the protocol type doesn't exist), the function frees > the skb and returns. > > The code for freeing the skb and returning is repeated several times, this > patch uses "goto rx_drop" to replace them so that the code looks cleaner. > > Cc: Willem de Bruijn <willemdebruijn.kernel@gmail.com> > Cc: Krzysztof Halasa <khc@pm.waw.pl> > Signed-off-by: Xie He <xie.he.0141@gmail.com> > --- > drivers/net/wan/hdlc_fr.c | 15 ++++++--------- > 1 file changed, 6 insertions(+), 9 deletions(-) > > diff --git a/drivers/net/wan/hdlc_fr.c b/drivers/net/wan/hdlc_fr.c > index 409e5a7ad8e2..4db0e01b96a9 100644 > --- a/drivers/net/wan/hdlc_fr.c > +++ b/drivers/net/wan/hdlc_fr.c > @@ -904,8 +904,7 @@ static int fr_rx(struct sk_buff *skb) > netdev_info(frad, "No PVC for received frame's DLCI %d\n", > dlci); > #endif > - dev_kfree_skb_any(skb); > - return NET_RX_DROP; > + goto rx_drop; > } > > if (pvc->state.fecn != fh->fecn) { > @@ -963,14 +962,12 @@ static int fr_rx(struct sk_buff *skb) > default: > netdev_info(frad, "Unsupported protocol, OUI=%x PID=%x\n", > oui, pid); > - dev_kfree_skb_any(skb); > - return NET_RX_DROP; > + goto rx_drop; > } > } else { > netdev_info(frad, "Unsupported protocol, NLPID=%x length=%i\n", > data[3], skb->len); > - dev_kfree_skb_any(skb); > - return NET_RX_DROP; > + goto rx_drop; > } > > if (dev) { > @@ -982,12 +979,12 @@ static int fr_rx(struct sk_buff *skb) > netif_rx(skb); > return NET_RX_SUCCESS; > } else { > - dev_kfree_skb_any(skb); > - return NET_RX_DROP; > + goto rx_drop; > } > > - rx_error: > +rx_error: > frad->stats.rx_errors++; /* Mark error */ > +rx_drop: > dev_kfree_skb_any(skb); > return NET_RX_DROP; I meant that I don't think errors should be double counted in rx_error and rx_drop. It is fine to count drops as either. Especially without that, I'm not sure this and the follow-on patch add much value. Minor code cleanups complicate backports of fixes.
On Sat, Oct 31, 2020 at 7:33 AM Willem de Bruijn <willemdebruijn.kernel@gmail.com> wrote: > > > - rx_error: > > +rx_error: > > frad->stats.rx_errors++; /* Mark error */ > > +rx_drop: > > dev_kfree_skb_any(skb); > > return NET_RX_DROP; > > I meant that I don't think errors should be double counted in rx_error > and rx_drop. It is fine to count drops as either. OK. Can we do that in another patch? Because I feel this would make the code a little bit more complex. Let's keep this patch as only a simple clean-up. > Especially without that, I'm not sure this and the follow-on patch add > much value. Minor code cleanups complicate backports of fixes. To me this is necessary, because I feel hard to do any development on un-cleaned-up code. I really don't know how to add my code without these clean-ups, and even if I managed to do that, I would not be happy with my code.
On Sat, Oct 31, 2020 at 8:18 AM Xie He <xie.he.0141@gmail.com> wrote: > > > Especially without that, I'm not sure this and the follow-on patch add > > much value. Minor code cleanups complicate backports of fixes. > > To me this is necessary, because I feel hard to do any development on > un-cleaned-up code. I really don't know how to add my code without > these clean-ups, and even if I managed to do that, I would not be > happy with my code. And always keeping the user interface and even the code unchanged contradicts my motivation of contributing to the Linux kernel. All my contributions are motivated by the hope to clean things up. I'm not an actual user of any of the code I contribute. If we adhere to the philosophy of not doing any clean-ups, my contributions would be meaningless.
On Fri, Oct 30, 2020 at 5:49 PM Xie He <xie.he.0141@gmail.com> wrote: > > Add an fh->ea2 check to the initial checks in fr_rx. fh->ea2 == 1 means > the second address byte is the final address byte. We only support the > case where the address length is 2 bytes. If the address length is not > 2 bytes, the control field and the protocol field would not be the 3rd > and 4th byte as we assume. (Say it is 3 bytes, then the control field > and the protocol field would be the 4th and 5th byte instead.) No, I don't think adding this explanation (about why address lengths other than 2 are not supported) is necessary. The code of this driver completely doesn't support address lengths other than 2 for many reasons, not only the reason above. The code for parsing and generating the address field (q922_to_dlci and dlci_to_q922) supports only 2-byte address fields. Also, the structure of the sending code of this driver can only generate headers with a 2-byte address field. Frame Relay networks usually have a fixed address length, that means a certain network usually has only 1 fixed address length. If this driver sends frames with 2-byte address fields, it shouldn't expect it would receive frames with 3-byte or 4-byte address fields. There are too many reasons, and I think just saying we only support 2-byte address fields is completely enough. Explaining the reason as above seems weird and inadequate to me, and also unnecessary to me. Therefore, I will resend this patch series with this explanation (about why address lengths other than 2 are not supported) removed.
On Sat, Oct 31, 2020 at 12:02 PM Xie He <xie.he.0141@gmail.com> wrote: > > On Sat, Oct 31, 2020 at 8:18 AM Xie He <xie.he.0141@gmail.com> wrote: > > > > > Especially without that, I'm not sure this and the follow-on patch add > > > much value. Minor code cleanups complicate backports of fixes. > > > > To me this is necessary, because I feel hard to do any development on > > un-cleaned-up code. I really don't know how to add my code without > > these clean-ups, and even if I managed to do that, I would not be > > happy with my code. That is the reality of working in this space, I think. I have frequently restructured code, fixed a bug and then worked backwards to create a *minimal* bugfix that applies to the current code as well as older stable branches. Obviously this is more of a concern for stable fixes than for new code. But we have to keep in mind that every code churn will make future bug fixes harder to roll out to users. That is not to say that churn should be avoided, just that we need to balance a change's benefit against this cost. > And always keeping the user interface and even the code unchanged > contradicts my motivation of contributing to the Linux kernel. All my > contributions are motivated by the hope to clean things up. I'm not an > actual user of any of the code I contribute. If we adhere to the > philosophy of not doing any clean-ups, my contributions would be > meaningless. There are cleanups and cleanups. Dead code removal and deduplication of open coded logic, for instance, are very valuable. As is, for instance, your work in making sense of hard_header_len. Returning code in branches vs an error jump label seems more of a personal preference, and to me does not pass the benefit/cost threshold. FWIW, there is lots of code that I would jump at the opportunity to restructure. Starting with skb_segment, probably. Obviously, all this is just one opinion on the topic.
On Sat, 31 Oct 2020 15:47:28 -0400 Willem de Bruijn wrote: > On Sat, Oct 31, 2020 at 12:02 PM Xie He <xie.he.0141@gmail.com> wrote: > > On Sat, Oct 31, 2020 at 8:18 AM Xie He <xie.he.0141@gmail.com> wrote: > > > > Especially without that, I'm not sure this and the follow-on patch add > > > > much value. Minor code cleanups complicate backports of fixes. > > > > > > To me this is necessary, because I feel hard to do any development on > > > un-cleaned-up code. I really don't know how to add my code without > > > these clean-ups, and even if I managed to do that, I would not be > > > happy with my code. > > That is the reality of working in this space, I think. I have > frequently restructured code, fixed a bug and then worked backwards to > create a *minimal* bugfix that applies to the current code as well as > older stable branches. > > Obviously this is more of a concern for stable fixes than for new > code. But we have to keep in mind that every code churn will make > future bug fixes harder to roll out to users. That is not to say that > churn should be avoided, just that we need to balance a change's > benefit against this cost. > > > And always keeping the user interface and even the code unchanged > > contradicts my motivation of contributing to the Linux kernel. All my > > contributions are motivated by the hope to clean things up. I'm not an > > actual user of any of the code I contribute. If we adhere to the > > philosophy of not doing any clean-ups, my contributions would be > > meaningless. > > There are cleanups and cleanups. Dead code removal and deduplication > of open coded logic, for instance, are very valuable. As is, for > instance, your work in making sense of hard_header_len. Or removing the buggy uses of IFF_TX_SKB_SHARING, for that matter (which at this point I agree we should just remove from ether_setup, and let people who care re-enable it). > Returning code in branches vs an error jump label seems more of a > personal preference, and to me does not pass the benefit/cost threshold. I must agree.
On Sat, Oct 31, 2020 at 12:48 PM Willem de Bruijn <willemdebruijn.kernel@gmail.com> wrote: > > Returning code in branches vs an error jump label seems more of a > personal preference, and to me does not pass the benefit/cost threshold. This patch is necessary for the 2nd and 5th patch in this series, because the 2nd and 5th patch would add a lot of places where we need to "error out" and drop the frame. Without this patch, the 2nd and 5th patch would add a lot of useless code. The 2nd patch is also necessary for the 5th patch, because otherwise I would not know how to produce the 5th patch. The logic is so convoluted for me. And it seems to me that the simplest way for me would make all code to follow the logic of eth_type_trans. The patch series was actually a single patch previously: http://patchwork.ozlabs.org/project/netdev/patch/20201017051951.363514-1-xie.he.0141@gmail.com/ I splitted it to make changes I do clearer. But really these patches should be as a whole. It's really hard for me to do the 5th patch without the 1st and 2nd patch.
The main purpose of this series is the last patch. The previous 4 patches are just code clean-ups so that the last patch will not make the code too messy. The patches must be applied in sequence. The receiving code of this driver doesn't support arbitrary Ethertype values. It only recognizes a few known Ethertypes when receiving and drops skbs with other Ethertypes. However, the standard document RFC 2427 allows Frame Relay to support any Ethertype values. This series adds support for this. Change from v5: Small fix to the commit messages. Change from v4: Drop the change related to stats.rx_dropped from the 1st patch. Switch the 3rd and 4th patch. Improve the commit message of the 4th patch by stating why only a 2-byte address field is accepted. Change from v3: Split the last patch into 2 patches. Improve the commit message of the 1st patch to explicitly state that the stats.rx_dropped count is also increased after "goto rx_error". Change from v2: Small fix to the commit messages. Change from v1: Small fix to the commit messages. Xie He (5): net: hdlc_fr: Simpify fr_rx by using "goto rx_drop" to drop frames net: hdlc_fr: Change the use of "dev" in fr_rx to make the code cleaner net: hdlc_fr: Do skb_reset_mac_header for skbs received on normal PVC devices net: hdlc_fr: Improve the initial checks when we receive an skb net: hdlc_fr: Add support for any Ethertype drivers/net/wan/hdlc_fr.c | 118 +++++++++++++++++++++++--------------- 1 file changed, 72 insertions(+), 46 deletions(-)