lkml.org 
[lkml]   [2000]   [Sep]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: networking todo, was Re: Linux-2.4.0-test9-pre2
On Tue, Sep 19, 2000 at 06:13:38PM -0700, David S. Miller wrote:
> Date: Wed, 20 Sep 2000 03:14:10 +0200
> From: "Andi Kleen" <ak@suse.de>
>
> The ipid handling is still fishy, it will break when you talk to
> more destinations than the inetpeer cache can take (I fixed it in
> my local tree with the appended patch)
>
> I don't like this change, please work with Alexey on this one.

What do you dislike ? Stateless ipid looks much nicer to me than stateful
one (and the "security" lost by not refetching the secret is minimal,
especially given the state of the random device on machines without diskio
and keyboard input)

> Receiver side SMP reordering is still there, but I'm not sure if it is
> fixable (but it'll surely hit people that cannot use Linux senders, I
> just see the reports)
>
> Reordering is a non-issue for ipv4/ipv6, all of the massive TCP input
> rewrite was strictly about dealing with this. No SMP reordering
> should cause any bogus fast retransmits etc. for example.

You mean TCP output rewrite ? The fix was strictly sender side only
(I first thought one of the ack changes alexey did was an attempt at a
hackish receiver side solution, but he told me that was false)

> What is the problem? Are you referering to the LAPB/X25 stuff?

When you have a non linux 2.4 sender you lose.

> The TCP connect running out of ports problem is still there (I
> fixed it locally, but the changes are probably far too extensive
> for 2.4.x now)
>
> I think people can change their ip_local_port_range and I really
> consider this a non-issue, at least, it's not a 2.4.x show stopper.

I agree.

>
> The TCP ISN computation is not SMP safe.
>
> "Big deal". So how much effort will you go to to get how much more
> protection and how much of it do we really need? How less secure are
> our TCP ISN's because of this? Not a show-stopper.
>
> I think you're reading too much tcp-impl :-)

No, it has nothing to do with tcp-impl ;), I discovered it while playing
with the ipids much earlier. With some bad luck could could get very similar
ISNs, which is probably bad (an easy fix is to do the secret hash with a
stack copy instead of the static)

(but it is not a show stopper I agree)


>
> include/linux/snmp.h is probably still not properly aligned for all
> cache line sizes.
>
> Read, and reread what Alexey tries to tell you over and over about
> this. It is not meant to be cache line aligned, it is meant to be a
> power of two and nothing more.

It has to be at least cache line aligned, otherwise the arrays wouldn't
make any sense at all, the power of two would just be a (imho misguided but
anyways) optimization.

When I count correctly sizeof(udp_mib) is 16 bytes currently (lots of
false sharing on 32/64byte cacheline sized architectures) and linux_mib
is 288 bytes currently (does not make any sense at all)


> UDP recvmsg error handling for csum errors is bogus (fix pending)
>
> Ok, send me a patch so I can see what the problem is.

Appended.

>
> I also would like to remove the inetpeer cache code before 2.4.0, now
> that nobody managed to salvage tw recycling and ipid can do without it.
>
> I would prefer not to, it seems to be too potentially destabilizing
> for questionable and unknown gain (ie. we don't know if tw recycling
> can be salvaged, so lets not take any changes).

The "unknown gain" is just removing a lot of complexity => speed and less
bugs.

The only advantage I know left is saving pmtus for a bit longer, but
I doubt it is worth the complexity.

Do you have any specific plan for salvaging tw recycling ? Just keeping
such an intrusive piece of code for benchmarks around looks wasteful
to me.

-Andi


Index: net/ipv4/udp.c
===================================================================
RCS file: /cvs/linux/net/ipv4/udp.c,v
retrieving revision 1.85
diff -u -u -d -r1.85 udp.c
--- net/ipv4/udp.c 2000/08/09 11:59:04 1.85
+++ net/ipv4/udp.c 2000/09/17 11:55:29
@@ -493,8 +493,6 @@
if (usin->sin_family != AF_INET) {
if (usin->sin_family != AF_UNSPEC)
return -EINVAL;
- if (net_ratelimit())
- printk("Remind Kuznetsov, he has to repair %s eventually\n", current->comm);
}

ufh.daddr = usin->sin_addr.s_addr;
@@ -678,6 +676,8 @@
if (flags & MSG_ERRQUEUE)
return ip_recv_error(sk, msg, len);

+
+ retry:
/*
* From here the generic datagram does a lot of the work. Come
* the finished NET3, it will do _ALL_ the work!
@@ -733,26 +733,21 @@
csum_copy_err:
UDP_INC_STATS_BH(UdpInErrors);

- /* Clear queue. */
- if (flags&MSG_PEEK) {
- int clear = 0;
+ if (flags&(MSG_PEEK|MSG_DONTWAIT)) {
+ struct sk_buff *skb2;
+
spin_lock_irq(&sk->receive_queue.lock);
- if (skb == skb_peek(&sk->receive_queue)) {
+ skb2 = skb_peek(&sk->receive_queue);
+ if ((flags & MSG_PEEK) && skb == skb2) {
__skb_unlink(skb, &sk->receive_queue);
- clear = 1;
}
spin_unlock_irq(&sk->receive_queue.lock);
- if (clear)
- kfree_skb(skb);
- }
-
- skb_free_datagram(sk, skb);
-
- /*
- * Error for blocking case is chosen to masquerade
- * as some normal condition.
- */
- return (flags&MSG_DONTWAIT) ? -EAGAIN : -EHOSTUNREACH;
+ skb_free_datagram(sk, skb);
+ if ((flags & MSG_DONTWAIT) && !skb2)
+ return -EAGAIN;
+ } else
+ skb_free_datagram(sk, skb);
+ goto retry;
}

int udp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/
\
 
 \ /
  Last update: 2005-03-22 12:39    [W:0.232 / U:0.208 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site