net: dsa: tag_mtk: add padding for tx packets
Subsystems (5)
PATCH v2 1/1 net: dsa: tag_mtk: add padding for tx packets
Padding for transmitted packets needs to account for the special tag. With not enough padding, garbage bytes are inserted by the switch at the end of small packets.
--- net/dsa/tag_mtk.c | 8 ++++++++ 1 file changed, 8 insertions(+)
net/dsa/tag_mtk.c +8 -0
@@ -25,6 +25,14 @@ static struct sk_buff *mtk_tag_xmit(struct sk_buff *skb,u8 xmit_tpid;u8 *mtk_tag; +/* The Ethernet switch we are interfaced with needs packets to be at +* least 64 bytes (including FCS) otherwise their padding might be +* corrupted. With tags enabled, we need to make sure that packets are +* at least 68 bytes (including FCS and tag). +*/ +if (__skb_put_padto (skb,ETH_ZLEN +MTK_HDR_LEN , false)) +return NULL ; +/* Build the special tag after the MAC Source Address. If VLAN header * is present, it's required that VLAN header and special tag is * being combined. Only in this way we can allow the switch can parse
Signature · 1 line
2.36.1
Vladimir Oltean
maintainerQuoted 3 lines · Felix Fietkau
Padding for transmitted packets needs to account for the special tag. With not enough padding, garbage bytes are inserted by the switch at the end of small packets.
I don't think padding bytes are guaranteed to be zeroes. Aren't they discarded? What is the issue?
Quoted 29 lines
Fixes 5cd8985a1909 ("net-next: dsa: add Mediatek tag RX/TX handler")Signed-off-by Felix Fietkau <[email protected]>--- net/dsa/tag_mtk.c | 8 ++++++++ 1 file changed, 8 insertions(+)net/dsa/tag_mtk.c +8
-0@@ -25,6 +25,14 @@ static struct sk_buff *mtk_tag_xmit(struct sk_buff *skb,u8 xmit_tpid;u8 *mtk_tag; +/* The Ethernet switch we are interfaced with needs packets to be at +* least 64 bytes (including FCS) otherwise their padding might be +* corrupted. With tags enabled, we need to make sure that packets are +* at least 68 bytes (including FCS and tag). +*/ +if (__skb_put_padto (skb,ETH_ZLEN +MTK_HDR_LEN , false)) +return NULL ; +/* Build the special tag after the MAC Source Address. If VLAN header * is present, it's required that VLAN header and special tag is * being combined. Only in this way we can allow the switch can parse Signature · 1 line
2.36.1
Quoted 7 lines · Vladimir Oltean
Quoted 3 lines · Felix Fietkau
Padding for transmitted packets needs to account for the special tag. With not enough padding, garbage bytes are inserted by the switch at the end of small packets.
I don't think padding bytes are guaranteed to be zeroes. Aren't they discarded? What is the issue?
With the broken padding, ARP requests are silently discarded on the receiver side in my test. Adding the padding explicitly fixes the issue.
- Felix
_______________________________________________
Linux-mediatek mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-mediatek
Vladimir Oltean
maintainerQuoted 7 lines · Felix Fietkau
Quoted 5 lines · Vladimir Oltean
Quoted 3 lines · Felix Fietkau
Padding for transmitted packets needs to account for the special tag. With not enough padding, garbage bytes are inserted by the switch at the end of small packets.
Quoted 6 lines
Quoted 2 lines
I don't think padding bytes are guaranteed to be zeroes. Aren't they discarded? What is the issue?
With the broken padding, ARP requests are silently discarded on the receiver side in my test. Adding the padding explicitly fixes the issue.
- Felix
Ok, I'm not going to complain too much about the patch, but I'm still curious where are the so-called "broken" packets discarded. I think the receiving MAC should be passing up to software a buffer without the extra padding beyond the L2 payload length (at least that's the behavior I'm familiar with).
Quoted 13 lines · Vladimir Oltean
Quoted 12 lines · Felix Fietkau
Quoted 7 lines · Vladimir Oltean
Quoted 3 lines · Felix Fietkau
Padding for transmitted packets needs to account for the special tag. With not enough padding, garbage bytes are inserted by the switch at the end of small packets.
I don't think padding bytes are guaranteed to be zeroes. Aren't they discarded? What is the issue?
With the broken padding, ARP requests are silently discarded on the receiver side in my test. Adding the padding explicitly fixes the issue.
Quoted 7 lines
Quoted 1 line
- Felix
Ok, I'm not going to complain too much about the patch, but I'm still curious where are the so-called "broken" packets discarded. I think the receiving MAC should be passing up to software a buffer without the extra padding beyond the L2 payload length (at least that's the behavior I'm familiar with).
I don't know where exactly these packets are discarded.
After digging through the devices I used during the tests, I just found some leftover pcap files that show the differences in the received packets. Since the packets are bigger after my patch, I can't rule out that packet size instead of the padding may have made a difference here in getting the ARP requests accepted by the receiver.
I've extracted the ARP requests and you can find them here:
http://nbd.name/arp-broken.pcap
http://nbd.name/arp-working.pcap
- Felix
Vladimir Oltean
maintainerQuoted 30 lines · Felix Fietkau
Quoted 13 lines · Vladimir Oltean
Quoted 12 lines · Felix Fietkau
Quoted 5 lines · Vladimir Oltean
Quoted 4 lines · Felix Fietkau
Padding for transmitted packets needs to account for the special tag. With not enough padding, garbage bytes are inserted by the switch at the end of small packets.
I don't think padding bytes are guaranteed to be zeroes. Aren'tthey
Quoted 1 line
discarded? What is the issue?
With the broken padding, ARP requests are silently discarded on the receiver side in my test. Adding the padding explicitly fixes the issue.
Quoted 7 lines
Quoted 1 line
- Felix
Ok, I'm not going to complain too much about the patch, but I'm still curious where are the so-called "broken" packets discarded. I think the receiving MAC should be passing up to software a buffer without the extra padding beyond the L2 payload length (at least that's the behavior I'm familiar with).
I don't know where exactly these packets are discarded.
After digging through the devices I used during the tests, I just found some leftover pcap files that show the differences in the received packets. Since the packets are bigger after my patch, I can't rule out that packet size instead of the padding may have made a difference here in getting the ARP requests accepted by the receiver.
Quoted 2 lines
I've extracted the ARP requests and you can find them here:
http://nbd.name/arp-broken.pcap
arp-broken.pcap was collected at the receiver MAC, right? So the packets actually exited the switch?
Quoted 3 lines
http://nbd.name/arp-working.pcap
- Felix
It sounds as if this is masking a problem on the receiver end, because not only does my enetc port receive the packet, it also replies to the ARP request.
pc # sudo tcpreplay -i eth1 arp-broken.pcap
root@debian:~# ip addr add 192.168.42.1/24 dev eno0 root@debian:~# tcpdump -i eno0 -e -n --no-promiscuous-mode arp tcpdump: verbose output suppressed, use -v[v]... for full protocol decode listening on eno0, link-type EN10MB (Ethernet), snapshot length 262144 bytes 22:18:58.846753 f4:d4:88:5e:6f:d2 > ff:ff:ff:ff:ff:ff, ethertype ARP (0x0806), length 60: Request who-has 192.168.42.1 tell 192.168.42.173, length 46 22:18:58.846806 00:04:9f:05:f4:ab > f4:d4:88:5e:6f:d2, ethertype ARP (0x0806), length 42: Reply 192.168.42.1 is-at 00:04:9f:05:f4:ab, length 28 ^C 2 packets captured 2 packets received by filter 0 packets dropped by kernel
What MAC/driver has trouble with these packets? Is there anything wrong in ethtool stats? Do they even reach software? You can also use "dropwatch -l kas" for some hints if they do.
Hi Vladimir,
Quoted 16 lines · Vladimir Oltean
It sounds as if this is masking a problem on the receiver end, because not only does my enetc port receive the packet, it also replies to the ARP request.
pc # sudo tcpreplay -i eth1 arp-broken.pcap
root@debian:~# ip addr add 192.168.42.1/24 dev eno0 root@debian:~# tcpdump -i eno0 -e -n --no-promiscuous-mode arp tcpdump: verbose output suppressed, use -v[v]... for full protocol decode listening on eno0, link-type EN10MB (Ethernet), snapshot length 262144 bytes 22:18:58.846753 f4:d4:88:5e:6f:d2 > ff:ff:ff:ff:ff:ff, ethertype ARP (0x0806), length 60: Request who-has 192.168.42.1 tell 192.168.42.173, length 46 22:18:58.846806 00:04:9f:05:f4:ab > f4:d4:88:5e:6f:d2, ethertype ARP (0x0806), length 42: Reply 192.168.42.1 is-at 00:04:9f:05:f4:ab, length 28 ^C 2 packets captured 2 packets received by filter 0 packets dropped by kernel
Quoted 3 lines
What MAC/driver has trouble with these packets? Is there anything wrong in ethtool stats? Do they even reach software? You can also use "dropwatch -l kas" for some hints if they do.
For some reason I can't reproduce the issue of ARPs not getting replies anymore.
The garbage data is still present in the ARP packets without my patch though. So regardless of whether ARP packets are processed correctly or if they just trip up on some receivers under specific conditions, I believe my patch is valid and should be applied.
Who knows, maybe the garbage padding even leaks some data from previous packets, or some other information from within the switch.
- Felix
Vladimir Oltean
maintainerQuoted 23 lines · Felix Fietkau
Hi Vladimir,
Quoted 16 lines · Vladimir Oltean
It sounds as if this is masking a problem on the receiver end, because not only does my enetc port receive the packet, it also replies to the ARP request.
pc # sudo tcpreplay -i eth1 arp-broken.pcap
root@debian:~# ip addr add 192.168.42.1/24 dev eno0 root@debian:~# tcpdump -i eno0 -e -n --no-promiscuous-mode arp tcpdump: verbose output suppressed, use -v[v]... for full protocol decode listening on eno0, link-type EN10MB (Ethernet), snapshot length 262144 bytes 22:18:58.846753 f4:d4:88:5e:6f:d2 > ff:ff:ff:ff:ff:ff, ethertype ARP (0x0806), length 60: Request who-has 192.168.42.1 tell 192.168.42.173, length 46 22:18:58.846806 00:04:9f:05:f4:ab > f4:d4:88:5e:6f:d2, ethertype ARP (0x0806), length 42: Reply 192.168.42.1 is-at 00:04:9f:05:f4:ab, length 28 ^C 2 packets captured 2 packets received by filter 0 packets dropped by kernelQuoted 3 lines
What MAC/driver has trouble with these packets? Is there anything wrong in ethtool stats? Do they even reach software? You can also use "dropwatch -l kas" for some hints if they do.
Quoted 6 lines
For some reason I can't reproduce the issue of ARPs not getting replies anymore.
The garbage data is still present in the ARP packets without my patch though. So regardless of whether ARP packets are processed correctly or if they just trip up on some receivers under specific conditions, I believe my patch is valid and should be applied.
I don't have a very strong opinion regarding whether to apply the patch or not. I think we've removed it from bug fix territory now, until proven otherwise. I do care about the justification (commit message, comments) being correct though. If you cannot reproduce now, someone one year from now surely cannot reproduce it either, and won't know why the code is there. FYI, the reason why you call __skb_put_padto() is not the reason why others call __skb_put_padto().
Quoted 2 lines
Who knows, maybe the garbage padding even leaks some data from previous packets, or some other information from within the switch.
I mean, the padding has to come from somewhere, no? Although I'd probably imagine non-scrubbed buffer cells rather than data structures...
Let's see what others have to say. I've been wanting to make the policy of whether to call __skb_put_padto() standardized for all tagging protocol drivers (similar to what is done in dsa_realloc_skb() and below it). We pad for tail taggers, maybe we can always pad and this removes a conditional, and simplifies taggers. Side note, I already dislike that the comment in tag_brcm.c is out of sync with the code. It says that padding up to ETH_ZLEN is necessary, but proceeds to pad up until ETH_ZLEN + tag len, only to add the tag len once more below via skb_push(). It would be nice if we could use the simple eth_skb_pad().
But there will be a small performance degradation for small packets due to the memset in __skb_pad(), which I'm not sure is worth the change.
Quoted 31 lines · Vladimir Oltean
Quoted 23 lines · Felix Fietkau
Hi Vladimir,
Quoted 16 lines · Vladimir Oltean
It sounds as if this is masking a problem on the receiver end, because not only does my enetc port receive the packet, it also replies to the ARP request.
pc # sudo tcpreplay -i eth1 arp-broken.pcap
root@debian:~# ip addr add 192.168.42.1/24 dev eno0 root@debian:~# tcpdump -i eno0 -e -n --no-promiscuous-mode arp tcpdump: verbose output suppressed, use -v[v]... for full protocol decode listening on eno0, link-type EN10MB (Ethernet), snapshot length 262144 bytes 22:18:58.846753 f4:d4:88:5e:6f:d2 > ff:ff:ff:ff:ff:ff, ethertype ARP (0x0806), length 60: Request who-has 192.168.42.1 tell 192.168.42.173, length 46 22:18:58.846806 00:04:9f:05:f4:ab > f4:d4:88:5e:6f:d2, ethertype ARP (0x0806), length 42: Reply 192.168.42.1 is-at 00:04:9f:05:f4:ab, length 28 ^C 2 packets captured 2 packets received by filter 0 packets dropped by kernelQuoted 3 lines
What MAC/driver has trouble with these packets? Is there anything wrong in ethtool stats? Do they even reach software? You can also use "dropwatch -l kas" for some hints if they do.
Quoted 6 lines
For some reason I can't reproduce the issue of ARPs not getting replies anymore.
The garbage data is still present in the ARP packets without my patch though. So regardless of whether ARP packets are processed correctly or if they just trip up on some receivers under specific conditions, I believe my patch is valid and should be applied.
Quoted 2 lines
I don't have a very strong opinion regarding whether to apply the patch or not. I think we've removed it from bug fix territory now, until proven otherwise.
I strongly disagree. Without my fix we're relying on undefined behavior of the hardware, since the switch requires padding that accounts for the special tag.
Quoted 3 lines
I do care about the justification (commit message, comments) being correct though. If you cannot reproduce now, someone one year from now surely cannot reproduce it either, and won't know why the code is there.
I think there is some misunderstanding here. I absolutely can reproduce the corrupted padding reliably, and it matches what I put into commit message and comments.
The issue that I can't reproduce reliably at the moment (ARP reception failure) is something that I only pointed out in a reply to this thread. This is what prompted me to look into the padding issue in the first place, and it also matches reports about connectivity issues that I got from other people.
Quoted 2 lines
FYI, the reason why you call
__skb_put_padto()is not the reason why others call__skb_put_padto().
It matches the call in tag_brcm.c (because I copied it from there), it's just that the symptoms that I'm fixing are different (undefined behavior instead of hard packet drop in the switch logic).
Quoted 16 lines
Quoted 2 lines
Who knows, maybe the garbage padding even leaks some data from previous packets, or some other information from within the switch.
I mean, the padding has to come from somewhere, no? Although I'd probably imagine non-scrubbed buffer cells rather than data structures...
Let's see what others have to say. I've been wanting to make the policy of whether to call
__skb_put_padto()standardized for all tagging protocol drivers (similar to what is done indsa_realloc_skb()and below it). We pad for tail taggers, maybe we can always pad and this removes a conditional, and simplifies taggers. Side note, I already dislike that the comment intag_brcm.cis out of sync with the code. It says that padding up toETH_ZLENis necessary, but proceeds to pad up untilETH_ZLEN+ tag len, only to add the tag len once more below viaskb_push(). It would be nice if we could use the simpleeth_skb_pad().
Quoted 2 lines
But there will be a small performance degradation for small packets due to the memset in
__skb_pad(), which I'm not sure is worth the change.
I guess we have different views on this. In my opinion, correctness matters more in this case than the tiny performance degradation.
- Felix
Vladimir Oltean
maintainerQuoted 76 lines · Felix Fietkau
Quoted 30 lines · Vladimir Oltean
Quoted 22 lines · Felix Fietkau
Hi Vladimir,
Quoted 4 lines · Vladimir Oltean
It sounds as if this is masking a problem on the receiver end, because not only does my enetc port receive the packet, it also replies to the ARP request.
Quoted 1 line
pc # sudo tcpreplay -i eth1 arp-broken.pcap
Quoted 11 lines
root@debian:~# ip addr add 192.168.42.1/24 dev eno0 root@debian:~# tcpdump -i eno0 -e -n --no-promiscuous-mode arp tcpdump: verbose output suppressed, use -v[v]... for full protocol decode listening on eno0, link-type EN10MB (Ethernet), snapshot length 262144 bytes 22:18:58.846753 f4:d4:88:5e:6f:d2 > ff:ff:ff:ff:ff:ff, ethertype ARP (0x0806), length 60: Request who-has 192.168.42.1 tell 192.168.42.173, length 46 22:18:58.846806 00:04:9f:05:f4:ab > f4:d4:88:5e:6f:d2, ethertype ARP (0x0806), length 42: Reply 192.168.42.1 is-at 00:04:9f:05:f4:ab, length 28 ^C 2 packets captured 2 packets received by filter 0 packets dropped by kernelQuoted 1 line
What MAC/driver has trouble with these packets? Is there
anything wrong
Quoted 2 lines
in ethtool stats? Do they even reach software? You can also use "dropwatch -l kas" for some hints if they do.
Quoted 6 lines
For some reason I can't reproduce the issue of ARPs not getting replies anymore.
The garbage data is still present in the ARP packets without my patch though. So regardless of whether ARP packets are processed correctly or if they just trip up on some receivers under specific conditions, I believe my patch is valid and should be applied.Quoted 2 lines
I don't have a very strong opinion regarding whether to apply the patch or not. I think we've removed it from bug fix territory now, until proven otherwise.
I strongly disagree. Without my fix we're relying on undefined behavior of the hardware, since the switch requires padding that accounts for the special tag.
Quoted 3 lines
I do care about the justification (commit message, comments) being correct though. If you cannot reproduce now, someone one year from now surely cannot reproduce it either, and won't know why the code is there.
I think there is some misunderstanding here. I absolutely can reproduce the corrupted padding reliably, and it matches what I put into commit message and comments.
The issue that I can't reproduce reliably at the moment (ARP reception failure) is something that I only pointed out in a reply to this thread. This is what prompted me to look into the padding issue in the first place, and it also matches reports about connectivity issues that I got from other people.
Quoted 2 lines
FYI, the reason why you call
__skb_put_padto()is not the reason why others call__skb_put_padto().It matches the call in
tag_brcm.c(because I copied it from there), it's just that the symptoms that I'm fixing are different (undefined behavior instead of hard packet drop in the switch logic).Quoted 16 lines
Quoted 2 lines
Who knows, maybe the garbage padding even leaks some data from previous packets, or some other information from within the switch.
I mean, the padding has to come from somewhere, no? Although I'd probably imagine non-scrubbed buffer cells rather than data structures...
Let's see what others have to say. I've been wanting to make the policy of whether to call
__skb_put_padto()standardized for all tagging protocol drivers (similar to what is done indsa_realloc_skb()and below it). We pad for tail taggers, maybe we can always pad and this removes a conditional, and simplifies taggers. Side note, I already dislike that the comment intag_brcm.cis out of sync with the code. It says that padding up toETH_ZLENis necessary, but proceeds to pad up untilETH_ZLEN+ tag len, only to add the tag len once more below viaskb_push(). It would be nice if we could use the simpleeth_skb_pad().
Quoted 5 lines
Quoted 2 lines
But there will be a small performance degradation for small packets due to the memset in
__skb_pad(), which I'm not sure is worth the change.I guess we have different views on this. In my opinion, correctness matters more in this case than the tiny performance degradation.
It seems that we are in disagreement about what it is that I am disputing. I am not disputing that the switch inserts non-zero padding octets, I am disputing the claim that this somehow violates any standard.
IEEE 802.3 clause 4.2.8 Frame transmission details this beyond any doubt. It says:
ComputePad appends an array of arbitrary bits to the MAC client data to pad the frame to the minimum frame size:
function ComputePad(var dataParam: DataValue): DataValue; begin
ComputePad := {Append an array of size padSize of arbitrary bits to the MAC client dataField}end; {ComputePad}
Clause 4.2.9 Frame reception then proceeds to say (too long to copy it, sorry) that the RemovePad function truncates the dataParam when possible (which has to do with whether the FCS is passed up to software or not) to the value represented by the lengthOrTypeParam (in octets), therefore *not* looking at the contents of the padding (other than to validate the FCS).
So, what correctness are we talking about?
_______________________________________________
Linux-mediatek mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-mediatek
Andrew Lunn
maintainerQuoted 9 lines
Let's see what others have to say. I've been wanting to make the policy of whether to call
__skb_put_padto()standardized for all tagging protocol drivers (similar to what is done indsa_realloc_skb()and below it). We pad for tail taggers, maybe we can always pad and this removes a conditional, and simplifies taggers. Side note, I already dislike that the comment intag_brcm.cis out of sync with the code. It says that padding up toETH_ZLENis necessary, but proceeds to pad up untilETH_ZLEN+ tag len, only to add the tag len once more below viaskb_push(). It would be nice if we could use the simpleeth_skb_pad().
There are some master devices which will perform padding on their own, in hardware. So for taggers which insert the header at the head, forcing such padding would be a waste of CPU time.
For tail taggers, padding short packets by default does however make sense. The master device is probably going to pad in the wrong way if it does padding.
Andrew
Quoted 10 lines · Andrew Lunn
Quoted 9 lines
Let's see what others have to say. I've been wanting to make the policy of whether to call
__skb_put_padto()standardized for all tagging protocol drivers (similar to what is done indsa_realloc_skb()and below it). We pad for tail taggers, maybe we can always pad and this removes a conditional, and simplifies taggers. Side note, I already dislike that the comment intag_brcm.cis out of sync with the code. It says that padding up toETH_ZLENis necessary, but proceeds to pad up untilETH_ZLEN+ tag len, only to add the tag len once more below viaskb_push(). It would be nice if we could use the simpleeth_skb_pad().
Quoted 7 lines
There are some master devices which will perform padding on their own, in hardware. So for taggers which insert the header at the head, forcing such padding would be a waste of CPU time.
For tail taggers, padding short packets by default does however make sense. The master device is probably going to pad in the wrong way if it does padding.
I just ran some more tests, here's what I found:
The switch automatically pads all forwarded packets to 64 bytes.
When packets are forwarded from one external port to another, the padding is all zero.
Only when packets are sent from a CPU port to an external port, the last 4 bytes contain garbage. The garbage bytes are different for every packet, and I can't tell if it's leaking contents of previous packets or what else is in there. Based on that, I'm pretty sure that the hardware simply has a quirk where it does not account for the special tag when generating its own padding internally.
I found that replacing my __skb_put_padto call with eth_skb_pad also works, so I'm going to send v3 with that and an updated comment.
- Felix
Andrew Lunn
maintainerHi Felix
Thanks for the additional testing.
Quoted 11 lines
I just ran some more tests, here's what I found:
The switch automatically pads all forwarded packets to 64 bytes. When packets are forwarded from one external port to another, the padding is all zero.
Only when packets are sent from a CPU port to an external port, the last 4 bytes contain garbage. The garbage bytes are different for every packet, and I can't tell if it's leaking contents of previous packets or what else is in there.
Based on that, I'm pretty sure that the hardware simply has a quirk where it does not account for the special tag when generating its own padding internally.
This does not yet explain why your receiver is dropping the frame. As Vladimir pointed out, the contents of the pad should not matter.
Is it also getting the FCS wrong when it pads? That would cause the receiver to drop the frame.
Or do we have an issue in the receiver where it is looking at the contents of the pad?
Andrew
Hi Andrew,
Quoted 12 lines · Andrew Lunn
Quoted 11 lines
I just ran some more tests, here's what I found:
The switch automatically pads all forwarded packets to 64 bytes. When packets are forwarded from one external port to another, the padding is all zero.
Only when packets are sent from a CPU port to an external port, the last 4 bytes contain garbage. The garbage bytes are different for every packet, and I can't tell if it's leaking contents of previous packets or what else is in there.
Based on that, I'm pretty sure that the hardware simply has a quirk where it does not account for the special tag when generating its own padding internally.
Quoted 8 lines
This does not yet explain why your receiver is dropping the frame. As Vladimir pointed out, the contents of the pad should not matter.
Is it also getting the FCS wrong when it pads? That would cause the receiver to drop the frame.
Or do we have an issue in the receiver where it is looking at the contents of the pad?
On the devices that I used for testing before, FCS wasn't reported in my captures. Since I can't reproduce the issue of the receiver dropping frames anymore, I currently have no way of figuring out what went wrong.
When I was able to reproduce the issue, I'm sure that I switched between patched and unpatched builds a few times to make sure that my change actually made a difference, which it did.
I do agree that having the garbage bytes in there is technically compliant with the spec. On the other hand, based on my observations I believe that the hardware's behavior of filling the last 4 bytes with seemingly random values only in the case of small frames being sent with a CPU special tag is clearly not intentional nor by design.
The issue is also clearly limited to processing packets with the tag (which can only come from the CPU), so in my opinion the tag driver is the right place to deal with it.
So I guess it comes down to whether you guys think that this is worth fixing.
I still consider it worth fixing, because:
- It looks like a hardware bug to me with potentially unknown consequences.
- If it caused issues in my setup, it might do so in other people's setups as well.
- I can't rule out potential information leakage from those 4 bytes
I guess if you guys don't think the issue is worth the price of a very small performance hit from padding the packets, I will just have to keep this as an out-of-tree patch.
- Felix
Andrew Lunn
maintainerQuoted 3 lines
I guess if you guys don't think the issue is worth the price of a very small performance hit from padding the packets, I will just have to keep this as an out-of-tree patch.
I'm thinking it is worth fixing, but i also wonder if your receiver has a bug. And if it does have a bug, it will probably come back to bite you sometime in the future.
Andrew
Andrew Lunn
maintainerQuoted 7 lines
The garbage data is still present in the ARP packets without my patch though. So regardless of whether ARP packets are processed correctly or if they just trip up on some receivers under specific conditions, I believe my patch is valid and should be applied.
Who knows, maybe the garbage padding even leaks some data from previous packets, or some other information from within the switch.
I somewhat agree with Vladimir at the moment. We don't seem to fully understand why the change makes things work better. And without that understanding, it is hard to say if this is the correct fix or not.
If this fix is making the transmitter work around bugs in the receiver, we really should be fixing the receiver, otherwise the receiver is going to be broken in other scenarios involving padded short packets.
Andrew
Jakub Kicinski
maintainerQuoted 6 lines · On Tue, 10 May 2022 11:40:13 +0200 Felix Fietkau wrote:
Padding for transmitted packets needs to account for the special tag. With not enough padding, garbage bytes are inserted by the switch at the end of small packets.
Fixes 5cd8985a1909 ("net-next: dsa: add Mediatek tag RX/TX handler")Signed-off-by Felix Fietkau <[email protected]>
You should answer all questions and address or dispute all feedback.