| Message ID | 20201016200226.23994-1-ceggers@arri.de |
|---|---|
| Headers | show |
| Series | net: dsa: move skb reallocation to dsa_slave_xmit | expand |
On Fri, Oct 16, 2020 at 10:02:24PM +0200, Christian Eggers wrote: > Ensure that the skb is not cloned and has enough tail room for the tail > tag. This code will be removed from the drivers in the next commits. > > Signed-off-by: Christian Eggers <ceggers@arri.de> > --- Does 1588 work for you using this change, or you haven't finished implementing it yet? If you haven't, I would suggest finishing that part first. The post-reallocation skb looks nothing like the one before. Before: skb len=68 headroom=2 headlen=68 tailroom=186 mac=(2,14) net=(16,-1) trans=-1 shinfo(txflags=1 nr_frags=0 gso(size=0 type=0 segs=0)) csum(0x0 ip_summed=0 complete_sw=0 valid=0 level=0) hash(0x9d6927ec sw=1 l4=0) proto=0x88f7 pkttype=0 iif=0 dev name=swp2 feat=0x0x0002000000005020 sk family=17 type=3 proto=0 After: skb len=68 headroom=2 headlen=68 tailroom=186 mac=(2,16) net=(18,-17) trans=1 shinfo(txflags=0 nr_frags=0 gso(size=0 type=0 segs=0)) csum(0x0 ip_summed=0 complete_sw=0 valid=0 level=0) hash(0x0 sw=0 l4=0) proto=0x0000 pkttype=0 iif=0 Notice how you've changed shinfo(txflags), among other things. Which proves that you can't just copy&paste whatever you found in tag_trailer.c. I am not yet sure whether there is any helper that can be used instead of this crazy open-coding. Right now, not having tested anything yet, my candidates of choice would be pskb_expand_head or __pskb_pull_tail. You should probably also try to cater here for the potential reallocation done in the skb_cow_head() of non-tail taggers. Which would lean the balance towards pskb_expand_head(), I believe. Also, if the result is going to be longer than ~20 lines of code, I strongly suggest moving the reallocation to a separate function so you don't clutter dsa_slave_xmit. Also, please don't redeclare struct sk_buff *nskb, you don't need to.
On 10/16/2020 1:02 PM, Christian Eggers wrote: [snip] > On Friday, 16 October 2020, 20:03:11 CEST, Jakub Kicinski wrote: >> FWIW if you want to avoid the reallocs you may want to set >> needed_tailroom on the netdev. > I haven't looked for this yet. If this can really solve the tagging AND > padding problem, I would like to do this in a follow up patch. The comment in netdevice.h says: * @needed_headroom: Extra headroom the hardware may need, but not in all * cases can this be guaranteed * @needed_tailroom: Extra tailroom the hardware may need, but not in all * cases can this be guaranteed. Some cases also use * LL_MAX_HEADER instead to allocate the skb and while I have never seen a reallocation occur while pushing a descriptor status block in front of a frame on transmit after setting the correct needed_headroom, it was not exercised in a very complicated way either, just TCP or UDP over IPv4 or IPv6. This makes me think that the comment is cautionary about more complicated transmit scenarios with stacked devices, tunneling etc. > > Wishing a nice weekend for netdev. Likewise!
Hi Vladimir, On Saturday, 17 October 2020, 02:48:16 CEST, Vladimir Oltean wrote: > On Fri, Oct 16, 2020 at 10:02:24PM +0200, Christian Eggers wrote: > > Ensure that the skb is not cloned and has enough tail room for the tail > > tag. This code will be removed from the drivers in the next commits. > > > > Signed-off-by: Christian Eggers <ceggers@arri.de> > > --- > > Does 1588 work for you using this change, or you haven't finished > implementing it yet? If you haven't, I would suggest finishing that > part first. Yes it does. Just after finishing this topic, I would to sent the patches for PTP. Maybe I'll do it in parallel, anything but the combination of L2/E2E/SLOB seems to work. > The post-reallocation skb looks nothing like the one before. > > Before: > skb len=68 headroom=2 headlen=68 tailroom=186 > mac=(2,14) net=(16,-1) trans=-1 > shinfo(txflags=1 nr_frags=0 gso(size=0 type=0 segs=0)) > csum(0x0 ip_summed=0 complete_sw=0 valid=0 level=0) > hash(0x9d6927ec sw=1 l4=0) proto=0x88f7 pkttype=0 iif=0 > dev name=swp2 feat=0x0x0002000000005020 > sk family=17 type=3 proto=0 > > After: > skb len=68 headroom=2 headlen=68 tailroom=186 > mac=(2,16) net=(18,-17) trans=1 > shinfo(txflags=0 nr_frags=0 gso(size=0 type=0 segs=0)) > csum(0x0 ip_summed=0 complete_sw=0 valid=0 level=0) > hash(0x0 sw=0 l4=0) proto=0x0000 pkttype=0 iif=0 > > Notice how you've changed shinfo(txflags), among other things. I get a similar output when placing the two skb_dump() calls in the current ksz_common_xmit() code: [ 5052.662168] old:skb len=58 headroom=2 headlen=58 tailroom=68 [ 5052.662168] mac=(2,14) net=(16,-1) trans=-1 [ 5052.662168] shinfo(txflags=1 nr_frags=0 gso(size=0 type=0 segs=0)) [ 5052.662168] csum(0x0 ip_summed=0 complete_sw=0 valid=0 level=0) [ 5052.662168] hash(0x0 sw=0 l4=0) proto=0x88f7 pkttype=0 iif=0 [ 5052.676360] old:dev name=lan0 feat=0x0x0002000000005220 [ 5052.679001] old:sk family=17 type=3 proto=0 [ 5052.681140] old:skb linear: 00000000: 01 1b 19 00 00 00 52 d9 a9 5d a1 40 88 f7 01 02 [ 5052.685236] old:skb linear: 00000010: 00 2c 00 00 00 00 00 00 00 00 00 00 00 00 00 00 [ 5052.689342] old:skb linear: 00000020: 00 00 52 d9 a9 ff fe 5d a1 40 00 01 00 00 01 7f [ 5052.693418] old:skb linear: 00000030: 00 00 00 00 00 00 00 00 00 00 [ 5052.696843] new:skb len=65 headroom=2 headlen=65 tailroom=61 [ 5052.696843] mac=(2,16) net=(18,-17) trans=1 [ 5052.696843] shinfo(txflags=0 nr_frags=0 gso(size=0 type=0 segs=0)) [ 5052.696843] csum(0x0 ip_summed=0 complete_sw=0 valid=0 level=0) [ 5052.696843] hash(0x0 sw=0 l4=0) proto=0x0000 pkttype=0 iif=0 [ 5052.711215] new:skb linear: 00000000: 01 1b 19 00 00 00 52 d9 a9 5d a1 40 88 f7 01 02 [ 5052.715305] new:skb linear: 00000010: 00 2c 00 00 00 00 00 00 00 00 00 00 00 00 00 00 [ 5052.719407] new:skb linear: 00000020: 00 00 52 d9 a9 ff fe 5d a1 40 00 01 00 00 01 7f [ 5052.723484] new:skb linear: 00000030: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 [ 5052.727587] new:skb linear: 00000040: 00 Note that whilst some skb members differ, the two hexdumps look correct. > Which proves that you can't just copy&paste whatever you found in > tag_trailer.c. I did. tag_trailer and tag_ksz are quite similar here, so I took a combination of them. > I am not yet sure whether there is any helper that can be used instead > of this crazy open-coding. Right now, not having tested anything yet, my > candidates of choice would be pskb_expand_head or __pskb_pull_tail. You > should probably also try to cater here for the potential reallocation > done in the skb_cow_head() of non-tail taggers. Which would lean the > balance towards pskb_expand_head(), I believe. The "open coding" is from the existing code (which doesn't say that it is correct). I will investigate why the copied skb is different and whether psk_expand_head can do better. I don't like to touch the non-tail taggers, this is too much out of the scope of my current work. > Also, if the result is going to be longer than ~20 lines of code, I > strongly suggest moving the reallocation to a separate function so you > don't clutter dsa_slave_xmit. As Florian requested I'll likely put the code into a separate function in slave.c and call it from the individual tail-taggers in order not to put extra conditionals in dsa_slave_xmit. regards Christian
On Sat, Oct 17, 2020 at 08:53:19PM +0200, Christian Eggers wrote: > > Does 1588 work for you using this change, or you haven't finished > > implementing it yet? If you haven't, I would suggest finishing that > > part first. > Yes it does. Just after finishing this topic, I would to sent the patches for > PTP. Maybe I'll do it in parallel, anything but the combination of L2/E2E/SLOB > seems to work. 2 aspects: - net-next is closed for this week and the next one, due to the merge window. You'll have to wait until it reopens. - Actually I was asking you this because sja1105 PTP no longer works after this change, due to the change of txflags. > I don't like to touch the non-tail taggers, this is too much out of the scope > of my current work. Do you want me to try and send a version using pskb_expand_head and you can test if it works for your tail-tagging switch? > > Also, if the result is going to be longer than ~20 lines of code, I > > strongly suggest moving the reallocation to a separate function so you > > don't clutter dsa_slave_xmit. > As Florian requested I'll likely put the code into a separate function in > slave.c and call it from the individual tail-taggers in order not to put > extra conditionals in dsa_slave_xmit. I think it would be best to use the unlikely(tail_tag) approach though. The reallocation function should still be in the common code path. Even for a non-1588 switch, there are other code paths that clone packets on TX. For example, the bridge does that, when flooding packets. Currently, DSA ensures that the header area is writable by calling skb_cow_head, as far as I can see. But the point is, maybe we can do TX reallocation centrally.
On Saturday, 17 October 2020, 21:12:47 CEST, Vladimir Oltean wrote: > On Sat, Oct 17, 2020 at 08:53:19PM +0200, Christian Eggers wrote: > > > Does 1588 work for you using this change, or you haven't finished > > > implementing it yet? If you haven't, I would suggest finishing that > > > part first. > > > > Yes it does. Just after finishing this topic, I would to sent the patches > > for PTP. Maybe I'll do it in parallel, anything but the combination of > > L2/E2E/SLOB seems to work. > > 2 aspects: > - net-next is closed for this week and the next one, due to the merge > window. You'll have to wait until it reopens. The status page seems to be out of date: http://vger.kernel.org/~davem/net-next.html The FAQ says: "Do not send new net-next content to netdev...". So there is no possibility for code review, is it? > - Actually I was asking you this because sja1105 PTP no longer works > after this change, due to the change of txflags. The tail taggers seem to be immune against this change. > > I don't like to touch the non-tail taggers, this is too much out of the > > scope of my current work. > > Do you want me to try and send a version using pskb_expand_head and you > can test if it works for your tail-tagging switch? I already wanted to ask... My 2nd try (checking for !skb_cloned()) was already sufficient (for me). Hacking linux-net is very interesting, but I have many other items open... Testing would be no problem. > > > Also, if the result is going to be longer than ~20 lines of code, I > > > strongly suggest moving the reallocation to a separate function so you > > > don't clutter dsa_slave_xmit. > > > > As Florian requested I'll likely put the code into a separate function in > > slave.c and call it from the individual tail-taggers in order not to put > > extra conditionals in dsa_slave_xmit. > > I think it would be best to use the unlikely(tail_tag) approach though. > The reallocation function should still be in the common code path. Even > for a non-1588 switch, there are other code paths that clone packets on > TX. For example, the bridge does that, when flooding packets. You already mentioned that you don't want to pass cloned packets to the tag drivers xmit() functions. I've no experience with the problems caused by cloned packets, but would cloned packets work anyway? Or must cloned packets not be changed (e.g. by tail-tagging)? Is there any value in first cloning in dsa_skb_tx_timestamp() and then unsharing in dsa_slave_xmit a few lines later? The issue I currently have only affects a very minor number of packets (cloned AND < ETH_ZLEN AND CONFIG_SLOB), so only these packets would need a copying. > Currently, DSA ensures that the header area is writable by calling > skb_cow_head, as far as I can see. But the point is, maybe we can do TX > reallocation centrally. regards Christian
On Sat, Oct 17, 2020 at 10:56:24PM +0200, Christian Eggers wrote: > The status page seems to be out of date: > http://vger.kernel.org/~davem/net-next.html Yeah, it can do that sometimes. Extremely rarely, but it happens. But net-next is still closed, nonetheless. > The FAQ says: "Do not send new net-next content to netdev...". So there is no > possibility for code review, is it? You can always send patches as RFC (Request For Comments). In fact that's what I'm going to do right now. > > - Actually I was asking you this because sja1105 PTP no longer works > > after this change, due to the change of txflags. > The tail taggers seem to be immune against this change. How? > > Do you want me to try and send a version using pskb_expand_head and you > > can test if it works for your tail-tagging switch? > I already wanted to ask... My 2nd try (checking for !skb_cloned()) was already > sufficient (for me). Hacking linux-net is very interesting, but I have many > other items open... Testing would be no problem. Ok, incoming..... > > I think it would be best to use the unlikely(tail_tag) approach though. > > The reallocation function should still be in the common code path. Even > > for a non-1588 switch, there are other code paths that clone packets on > > TX. For example, the bridge does that, when flooding packets. > You already mentioned that you don't want to pass cloned packets to the tag > drivers xmit() functions. I've no experience with the problems caused by > cloned packets, but would cloned packets work anyway? Or must cloned packets > not be changed (e.g. by tail-tagging)? Is there any value in first cloning in > dsa_skb_tx_timestamp() and then unsharing in dsa_slave_xmit a few lines later? > The issue I currently have only affects a very minor number of packets (cloned > AND < ETH_ZLEN AND CONFIG_SLOB), so only these packets would need a copying. Yes, we need to clone and then unshare immediately afterwards because sja1105_xmit calls sja1105_defer_xmit, which schedules a workqueue. The sja1105 driver assumes that the skb has already been cloned by then. So basically, the sja1105 driver introduces a strict ordering requirement that dsa_skb_tx_timestamp needs to be first, then p->xmit second. So we necessarily must reallocate freshly cloned skbs, as things stand now. I'll think about avoiding that, but not now. We were always reallocating those frames before, using skb_cow_head. The only difference now is that the skb, as it is passed to the tagger's xmit() function, is directly writable. You'll see...