Messages in this thread | | | Date | Sun, 20 Feb 2000 13:24:30 -0800 | From | "David S. Miller" <> | Subject | Re: IP changes in 2.3.4x make things wierd? |
| |
Date: Sun, 20 Feb 2000 21:09:28 +0000 (GMT) From: Alan Cox <alan@lxorguk.ukuu.org.uk>
Ok so how does a non setuid application receive the ICMP errors it needs to make intelligent decisions about load balancing UDP queries in 2.3.x ? If you can't answer that problem cleanly then your change is broken.
...
IP_RECVERR is one option.
Not reporting the error is even more stupid than reporting it. Its always been configurable before. Still it may explain why 2.3.x now hangs during boot starting userspace apps here.
Accidently reporting the error to the next application which happens to unluckily get that same port is even more dangerous.
I quote below a mail from Alexey from us discussing just this very issue, and which made me agree with him that the new behavior is what we want. Alan, you cannot propose going back to the old behavior without proposing a solution to this problem.
===============================================================
Let's wait a bit with this problem. I had very good reasons to make this and want to observe for reaction a bit more before starting some ado. I know that Alan advocated this Linux specific feature for years, I advocated it too and even more aggressively. But at time when Linux (and me) entered to mature state and started to serve thousands of processes and connections, I started to suspect something wrong. When IP_RECVERR was done, it were purely theoretical doubts.
Now I have no doubts that it is wrong.
The last drop was the accident with CommuniGatePRO. It has high preformance DNS resolver, using not-connected sockets concurrently and at some level of load it fails to resolve names even for local 100% reliable DNS services. I ridiculed its vendor for long time, but had to explain, why only Linux fails yet. The source appered clear, sockets get ICMPs from slow remote servers, which match new innocent sockets. It was the _end_, I shot the feature the same day when found that it is real.
We have to shot it as soon as possible simply because it does not match modern situation. The problem was not seen when nobody really used Linux as loaded server. Actually, exactly with nis it should occur very frequently.
Alexey
================================================================
Later, David S. Miller davem@redhat.com
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
| |