Hi Folks
Does anybody remember the history for the follow and can explain why the code is as it is?
The soft interface transmit function, batadv_interface_tx() calls batadv_skb_set_priority(skb, 0) to set the skb->priority based on the TOS bits or 801.p.
If the packet needs to be fragmented because of MTU issues, batadv_frag_create() is used to create the fragments. It overwrites the skb->priority in the original skb to TC_PRIO_CONTROL, and leaves the fragment skb with the default priority.
This seems a bit odd to me. I would of expected the priority to of been copied from the original into the fragment.
Thanks Andrew
Hi Andrew,
On Tuesday 12 April 2016 22:42:59 Andrew Lunn wrote:
Hi Folks
Does anybody remember the history for the follow and can explain why the code is as it is?
The soft interface transmit function, batadv_interface_tx() calls batadv_skb_set_priority(skb, 0) to set the skb->priority based on the TOS bits or 801.p.
If the packet needs to be fragmented because of MTU issues, batadv_frag_create() is used to create the fragments. It overwrites the skb->priority in the original skb to TC_PRIO_CONTROL, and leaves the fragment skb with the default priority.
This seems a bit odd to me. I would of expected the priority to of been copied from the original into the fragment.
I think this part could be improved. Right now, if I remember correctly, we set TC_PRIO_CONTROL by default and set another priority if we can parse the header (batadv_skb_set_priority()).
There are two cases:
1.) On the original sender, both fragments could adopt the priority as you suggest. The code probably doesn't take care of that yet, so that could be fixed.
2.) On routers on the way, the priority could only be set based on the first fragment, since the second fragment will not have a valid header to parse. And unless we remember the priority from the first fragment, we have no way to know to which priority we should set the second fragment.
I believe case 1 can be fixed easily, for case 2 I don't have an idea right now. :)
Cheers, Simon
On Wed, Apr 13, 2016 at 01:11:12PM +0200, Simon Wunderlich wrote:
Hi Andrew,
On Tuesday 12 April 2016 22:42:59 Andrew Lunn wrote:
Hi Folks
Does anybody remember the history for the follow and can explain why the code is as it is?
The soft interface transmit function, batadv_interface_tx() calls batadv_skb_set_priority(skb, 0) to set the skb->priority based on the TOS bits or 801.p.
If the packet needs to be fragmented because of MTU issues, batadv_frag_create() is used to create the fragments. It overwrites the skb->priority in the original skb to TC_PRIO_CONTROL, and leaves the fragment skb with the default priority.
This seems a bit odd to me. I would of expected the priority to of been copied from the original into the fragment.
Hi Simon
I think this part could be improved. Right now, if I remember correctly, we set TC_PRIO_CONTROL by default and set another priority if we can parse the header (batadv_skb_set_priority()).
Yes, that is what i thought the intention was. The implementation is a bit different.
There are two cases:
1.) On the original sender, both fragments could adopt the priority as you suggest. The code probably doesn't take care of that yet, so that could be fixed.
I will cook up a patch for this.
2.) On routers on the way, the priority could only be set based on the first fragment, since the second fragment will not have a valid header to parse. And unless we remember the priority from the first fragment, we have no way to know to which priority we should set the second fragment.
I don't think remembering works. It looks like it fragments from the tail towards the head. So we are not going to receive the IP header until we get the last fragment.
I believe case 1 can be fixed easily, for case 2 I don't have an idea right now. :)
There is one reasonable option i can think of. batadv_skb_set_priority() extracts three bits of priority information, either from the TOS bits, or the 801.q header.
The fragment header is:
struct batadv_frag_packet { u8 packet_type; u8 version; /* batman version field */ u8 ttl; #if defined(__BIG_ENDIAN_BITFIELD) u8 no:4; u8 reserved:4; #elif defined(__LITTLE_ENDIAN_BITFIELD) u8 reserved:4; u8 no:4; #else #error "unknown bitfield endianness" #endif u8 dest[ETH_ALEN]; u8 orig[ETH_ALEN]; __be16 seqno; __be16 total_size; };
Place the priority information into 3 of the 4 reserved bits. The receiver can then set the skb->priority of the fragment before passing it to the hard interface.
Andrew
Hi Andrew,
On Wednesday 13 April 2016 14:14:16 Andrew Lunn wrote:
[...]
There are two cases:
1.) On the original sender, both fragments could adopt the priority as you suggest. The code probably doesn't take care of that yet, so that could be fixed.
I will cook up a patch for this.
Excellent!
2.) On routers on the way, the priority could only be set based on the first fragment, since the second fragment will not have a valid header to parse. And unless we remember the priority from the first fragment, we have no way to know to which priority we should set the second fragment.
I don't think remembering works. It looks like it fragments from the tail towards the head. So we are not going to receive the IP header until we get the last fragment.
Ah, you are deeper into that. But my hunch was also that it would be messy, since we can't guarantee the order of the fragments anyway.
I believe case 1 can be fixed easily, for case 2 I don't have an idea right now. :)
There is one reasonable option i can think of. batadv_skb_set_priority() extracts three bits of priority information, either from the TOS bits, or the 801.q header.
The fragment header is:
struct batadv_frag_packet { u8 packet_type; u8 version; /* batman version field */ u8 ttl; #if defined(__BIG_ENDIAN_BITFIELD) u8 no:4; u8 reserved:4; #elif defined(__LITTLE_ENDIAN_BITFIELD) u8 reserved:4; u8 no:4; #else #error "unknown bitfield endianness" #endif u8 dest[ETH_ALEN]; u8 orig[ETH_ALEN]; __be16 seqno; __be16 total_size; };
Place the priority information into 3 of the 4 reserved bits. The receiver can then set the skb->priority of the fragment before passing it to the hard interface.
Ah, yes that sounds like an excellent idea! I like that. The default should be zero right now, which would also fit. :)
Thanks! Simon
b.a.t.m.a.n@lists.open-mesh.org