user-mode-linux-user Mailing List for User-mode Linux kernel port (Page 792)

I wish that were the case. I tried allocation diferent memory amounts,
ranging from the default 16M to 64M.
No difference, once I was not running anything on the uml machine.

Still the same problem.

The packages arrive at the uml host (tcpdump on the uml machines
shows them, and the um_eth_net_util dump shows them too), but there
is no answer for 9 out of 10 packages (except for the first 6).

On Sun, Dec 03, 2000 at 12:43:32PM -0800, Mooneer Salem wrote:
> It could be possible that you have not allocated enough RAM
> to the UML process. By default, 16MB is allocated. I generally
> allocate 64MB at the very least for each UML, and leave 128MB of
> RAM for the host system.
>=20
> Even with that, on my system ICMP pings sometimes freeze up for
> a few seconds and then resume, and ping would show 0% loss. It
> might be helpful to change the wait period between pings.

--=20
 Rodrigo Barbosa (morcego)  - rod...@co...
 Conectiva R&D Team         - http://distro.conectiva.com.br
 "Quis custodiet custodes?" - http://www.conectiva.com

It could be possible that you have not allocated enough RAM
to the UML process. By default, 16MB is allocated. I generally
allocate 64MB at the very least for each UML, and leave 128MB of
RAM for the host system.

Even with that, on my system ICMP pings sometimes freeze up for
a few seconds and then resume, and ping would show 0% loss. It
might be helpful to change the wait period between pings.

-----Original Message-----
From: use...@li...
[mailto:use...@li...]On Behalf Of
Rodrigo Barbosa (aka morcego)
Sent: Sunday, December 03, 2000 9:18 AM
To: Use...@li...
Subject: Re: [uml-user] NFS hanging badly


Well, I managed to make uml's eth0 interface work. I also managed
(using proxyarp) to make it looks just like any other machine on our
network.

The problem that still persist (with umn and eth0 both), is the high
package loosing.

Take a look at it (10.0.23.129 is my uml box. Morcego is the host):

morcego [/root] # ping 10.0.23.129
PING 10.0.23.129 (10.0.23.129): 56 octets data
64 octets from 10.0.23.129: icmp_seq=0 ttl=255 time=6.8 ms
64 octets from 10.0.23.129: icmp_seq=1 ttl=255 time=6.2 ms
64 octets from 10.0.23.129: icmp_seq=2 ttl=255 time=5.9 ms
64 octets from 10.0.23.129: icmp_seq=3 ttl=255 time=6.1 ms
64 octets from 10.0.23.129: icmp_seq=4 ttl=255 time=6.0 ms
64 octets from 10.0.23.129: icmp_seq=5 ttl=255 time=6.6 ms
64 octets from 10.0.23.129: icmp_seq=10 ttl=255 time=6.0 ms
64 octets from 10.0.23.129: icmp_seq=20 ttl=255 time=6.1 ms
64 octets from 10.0.23.129: icmp_seq=30 ttl=255 time=1.8 ms

uml is running as root (# ./linux), and the um_eth_net_util is running with
'nice -n -1', having high priority over anything else (almost. Lets not
forget init :-)).

I simply can't figure what is going wrong.

Any clues ?

On Sun, Dec 03, 2000 at 03:18:05PM -0200, Rodrigo Barbosa (aka morcego) wro=
te:
> Well, I managed to make uml's eth0 interface work. I also managed
> (using proxyarp) to make it looks just like any other machine on our
> network.
>=20
> The problem that still persist (with umn and eth0 both), is the high
> package loosing.
>=20
> Take a look at it (10.0.23.129 is my uml box. Morcego is the host):
>=20
> morcego [/root] # ping 10.0.23.129
> PING 10.0.23.129 (10.0.23.129): 56 octets data
> 64 octets from 10.0.23.129: icmp_seq=3D0 ttl=3D255 time=3D6.8 ms
> 64 octets from 10.0.23.129: icmp_seq=3D1 ttl=3D255 time=3D6.2 ms
> 64 octets from 10.0.23.129: icmp_seq=3D2 ttl=3D255 time=3D5.9 ms
> 64 octets from 10.0.23.129: icmp_seq=3D3 ttl=3D255 time=3D6.1 ms
> 64 octets from 10.0.23.129: icmp_seq=3D4 ttl=3D255 time=3D6.0 ms
> 64 octets from 10.0.23.129: icmp_seq=3D5 ttl=3D255 time=3D6.6 ms
> 64 octets from 10.0.23.129: icmp_seq=3D10 ttl=3D255 time=3D6.0 ms
> 64 octets from 10.0.23.129: icmp_seq=3D20 ttl=3D255 time=3D6.1 ms
> 64 octets from 10.0.23.129: icmp_seq=3D30 ttl=3D255 time=3D1.8 ms
>=20
> uml is running as root (# ./linux), and the um_eth_net_util is running wi=
th
> 'nice -n -1', having high priority over anything else (almost. Lets not
> forget init :-)).
>=20
> I simply can't figure what is going wrong.
>=20
> Any clues ?

I have some more info about what is going on. Maybe this will help.
Look at the tcpdump of the virtual eth0 of the uml host:

16:34:06.821653 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:06.821653 eth0 > 10.0.23.129 > 10.0.17.78: icmp: echo reply (DF)
16:34:07.820296 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:07.820296 eth0 > 10.0.23.129 > 10.0.17.78: icmp: echo reply (DF)
16:34:08.820524 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:08.820524 eth0 > 10.0.23.129 > 10.0.17.78: icmp: echo reply (DF)
16:34:09.820282 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:09.820282 eth0 > 10.0.23.129 > 10.0.17.78: icmp: echo reply (DF)
16:34:10.820282 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:10.820282 eth0 > 10.0.23.129 > 10.0.17.78: icmp: echo reply (DF)
16:34:11.820283 eth0 < arp who-has 10.0.23.129 tell 10.0.17.78
16:34:11.820283 eth0 > arp reply 10.0.23.129 (10:0:23:12:9:0) is-at 10:0:23=
:12:9:0 (0:c0:df:1:c7:2)
16:34:11.820283 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:11.820283 eth0 > 10.0.23.129 > 10.0.17.78: icmp: echo reply (DF)
16:34:12.820283 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:13.820283 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:14.820283 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:15.820282 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:16.820282 eth0 > arp who-has 10.0.17.78 tell 10.0.23.129 (10:0:23:12:=
9:0)
16:34:16.820282 eth0 < arp reply 10.0.17.78 is-at 0:c0:df:1:c7:2 (10:0:23:1=
2:9:0)
16:34:16.820282 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:17.820397 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:17.820397 eth0 > 10.0.23.129 > 10.0.17.78: icmp: echo reply (DF)
16:34:18.820283 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:19.820283 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:20.820283 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:21.820283 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:22.820284 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:23.820283 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:24.820285 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:25.820292 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:26.820306 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:27.820377 eth0 < 10.0.17.78 > 10.0.23.129: icmp: echo request
16:34:27.820377 eth0 > 10.0.23.129 > 10.0.17.78: icmp: echo reply (DF)


As we can see, we have the first 6 ECHO_REQ that get replies. After that,
they don't (sometimes they do, as seen on the previous c&p section).

We can see that the packages are ariving at the uml machine, which, as far =
as
I can gather, rules out problems with the host and with the um_eth_net_util
program.

I took a look at /proc/sys/net/ipv4/icmp*, and all values are
properly set.

One thing to notice is that everytime I change the value of the
forward flag, either on the /proc/sys/net/ipv4/ip_forward or the
/proc/sys/net/ipv4/conf/eth0/forwarding file, 6 packages get responses
without a problem. I must actualy change the value. Only writing the same
value twice to the file doesn't work.

Looks like kernel problems to me. I just don't know if it's 2.4 problems,
or problems on the uml virtual eth device.

Hope this helps.

[]s




--=20
 Rodrigo Barbosa (morcego)  - rod...@co...
 Conectiva R&D Team         - http://distro.conectiva.com.br
 "Quis custodiet custodes?" - http://www.conectiva.com

Well, I managed to make uml's eth0 interface work. I also managed
(using proxyarp) to make it looks just like any other machine on our
network.

The problem that still persist (with umn and eth0 both), is the high
package loosing.

Take a look at it (10.0.23.129 is my uml box. Morcego is the host):

morcego [/root] # ping 10.0.23.129
PING 10.0.23.129 (10.0.23.129): 56 octets data
64 octets from 10.0.23.129: icmp_seq=3D0 ttl=3D255 time=3D6.8 ms
64 octets from 10.0.23.129: icmp_seq=3D1 ttl=3D255 time=3D6.2 ms
64 octets from 10.0.23.129: icmp_seq=3D2 ttl=3D255 time=3D5.9 ms
64 octets from 10.0.23.129: icmp_seq=3D3 ttl=3D255 time=3D6.1 ms
64 octets from 10.0.23.129: icmp_seq=3D4 ttl=3D255 time=3D6.0 ms
64 octets from 10.0.23.129: icmp_seq=3D5 ttl=3D255 time=3D6.6 ms
64 octets from 10.0.23.129: icmp_seq=3D10 ttl=3D255 time=3D6.0 ms
64 octets from 10.0.23.129: icmp_seq=3D20 ttl=3D255 time=3D6.1 ms
64 octets from 10.0.23.129: icmp_seq=3D30 ttl=3D255 time=3D1.8 ms

uml is running as root (# ./linux), and the um_eth_net_util is running with
'nice -n -1', having high priority over anything else (almost. Lets not
forget init :-)).

I simply can't figure what is going wrong.

Any clues ?

[]s

--=20
 Rodrigo Barbosa (morcego)  - rod...@co...
 Conectiva R&D Team         - http://distro.conectiva.com.br
 "Quis custodiet custodes?" - http://www.conectiva.com

mo...@ea... said:
> Just in case anyone else would like to know how I got UML to see the
> outside world, I'm posting the commands I used below

Thanks.

I added a link from my networking page (http://user-mode-linux.sourceforge.net/
networking.html) to that post in the archives.

				Jeff

Hi,

I've encontered strange behaviour from various applications when I SSH in
to a uml process.  This is with 2.2.40-test11 with the um patches from
CVS, hosted on 2.2.17, using the small debian root fs and the SLIP-based
virtual network driver.

When I run ircII, BitchX or lynx after a while the SSH session hangs and I
can't see anything change.  I can see SSH packets go back and forth
between my box and the uml process on another box, but the screen never
updates and killing the application or the shell has no effect; I have to
kill the sshd process associated with the connection in order to regain
control.  No messages in any logfiles or on the console relating to the
crash.

The problem is most reproducible with ircII as it always hangs like this
when it's finished logging into a server.  I had to take lynx to several
pages in order to have the same effect.  I am unable to reproduce these
problems when I run the same software on the uml console.

What should I do?

-don

Donald Gordon | wellington, new zealand | All opinions above are those of my
donald(@)gordon(0)co(0)nz | icq14995769 | deceased pet goldfish, "Fishy II".

I finally got it working for a few minutes to see if I can actually
access the outside world. Aside from a slight slowdown sometimes
(possibly due to the extra routing involved or from um_eth_net_util)
it worked great. I no longer had to use Squid to access the Internet
from UML.

Just in case anyone else would like to know how I got UML to see the
outside world, I'm posting the commands I used below (216.74.72.8 is
an IP on the host, and 216.74.72.9 is the UML IP). First I opened
3 ssh windows at once going to the same host. I login as root to all
three of them. In the first one, I booted UML. The second one ran
"um_eth_net_util tap0 100", and the third was the window I did all
the "From Host" commands in. I made sure UML and um_eth_net_util
started successfully, then I executed each of the commands under
"From UML" one at a time (inside the window running UML).

Next I switched over to the third window which I purposely left idle.
I entered each of the commands under "From Host", one at a time. After
this was successful, I went to the UML window and made sure I could ping
the host. Once I made sure I was able to access the host, I made sure
I could access the outside world by pinging the router. Once sasitified,
I considered myself fully hooked to the Internet from UML***.

(*** optionally you can execute the following commands in the third window
 after getting networking operational:

 ipchains -P forward DENY
 ipchains -A forward -s 216.74.72.9 -j ACCEPT -i eth0
 ipchains -A forward -d 216.74.72.9 -j ACCEPT -i tap0

 This will result in better security by limiting who can route through your
 system. Note: you may have to use different commands if you are running a
 2.4.x kernel on the host machine.)

Commands:

From Host:
ifconfig tap0 216.74.72.8
echo 1 >/proc/sys/net/ipv4/ip_forward
arp -i eth0 -s 216.74.72.9 -D tap0 pub
route add -host 216.74.72.9 dev tap0

From UML:
ifconfig eth0 hw ether 0:0:10:0:0:1
ifconfig eth0 216.74.72.9 netmask 255.255.255.0
route add -host 216.74.72.8 dev eth0
route add -net default gw 216.74.72.8 dev eth0

-------------------------------------------------------

Now since I got it up and running on the Internet is it possible
to use IP aliasing on the UML side so that I can attach more
than one IP to the virtual eth0? Would it involve going back
to the host and adding additional ARP entries?

-----Original Message-----
From: William Stearns [mailto:wst...@po...]
Sent: Thursday, November 30, 2000 8:10 PM
To: Mooneer Salem
Cc: ML-uml-user; William Stearns
Subject: Re: [uml-user] Making services in UML available to the outside
world


Good afternoon, Mooner,

On Thu, 30 Nov 2000, Mooneer Salem wrote:

> I've succeeded in getting UML to boot within my Redhat Linux 6.1
> box (running kernel 2.2.17, UML was running 2.4.0) I also got
> virtual ethernet to work on tap0 (I can contact the host and the
> host can contact UML, but UML cannot see the outside world without
> a proxy.
>
> The host has an IP of 216.74.72.5, which is accessable to the rest of
> the world. I know it's possible to set up IP Masquerating, but I would
> like people to be able to connect to service(s) I enable within UML.
> All I would need to do is set tap0 to 216.74.72.6 and eth0 in UML to
> 216.74.72.7 plus adding some ipchains rules correct? Or does it involve
> much more? (I already tried 'um_eth_net_util eth0 100', but that puts
> the Ethernet card in promiscuous mode, which overloads UML with network
> traffic.)

	- Set up uml with the following network configuration:
(netmask=255.255.255.255, ip addresses as you specified above, use the
host's tap0 IP as uml's default gateway).
	- Turn on ip forwarding in the kernel (if you have
an /etc/sysctl.conf, make sure it has "net.ipv4.ip_forward = 1", otherwise
add "echo 1 >/proc/sys/net/ipv4/ip_forward" to the end of
/etc/rc.d/rc.local or /etc/rc.d/init.d/network)
	- Tell the host to stand in for the uml on the ethernet network;
this is called proxyarp.  Here's an example, assuming that eth0's mac
address is 00:10:5A:CC:97:BF (you can find this in "ifconfig eth0"):
arp -i eth0 -s 216.74.72.7 00:10:5A:CC:97:BF pub
	Likewise, add this to /etc/rc.d/init.d/network or
/etc/rc.d/rc.local
	See ftp://mason.stearns.org/pub/doc/proxyarp-howto for a writeup I
did on proxyarp.
	One more note; you can conserve one more address by using
216.74.72.5 as the IP address for both eth0 and tap0.
	Cheers,
	- Bill

Good afternoon, Mooner,

On Thu, 30 Nov 2000, Mooneer Salem wrote:

> I've succeeded in getting UML to boot within my Redhat Linux 6.1
> box (running kernel 2.2.17, UML was running 2.4.0) I also got
> virtual ethernet to work on tap0 (I can contact the host and the
> host can contact UML, but UML cannot see the outside world without
> a proxy.
>
> The host has an IP of 216.74.72.5, which is accessable to the rest of
> the world. I know it's possible to set up IP Masquerating, but I would
> like people to be able to connect to service(s) I enable within UML.
> All I would need to do is set tap0 to 216.74.72.6 and eth0 in UML to
> 216.74.72.7 plus adding some ipchains rules correct? Or does it involve
> much more? (I already tried 'um_eth_net_util eth0 100', but that puts
> the Ethernet card in promiscuous mode, which overloads UML with network
> traffic.)

	- Set up uml with the following network configuration:
(netmask=255.255.255.255, ip addresses as you specified above, use the
host's tap0 IP as uml's default gateway).
	- Turn on ip forwarding in the kernel (if you have
an /etc/sysctl.conf, make sure it has "net.ipv4.ip_forward = 1", otherwise
add "echo 1 >/proc/sys/net/ipv4/ip_forward" to the end of
/etc/rc.d/rc.local or /etc/rc.d/init.d/network)
	- Tell the host to stand in for the uml on the ethernet network;
this is called proxyarp.  Here's an example, assuming that eth0's mac
address is 00:10:5A:CC:97:BF (you can find this in "ifconfig eth0"):
arp -i eth0 -s 216.74.72.7 00:10:5A:CC:97:BF pub
	Likewise, add this to /etc/rc.d/init.d/network or
/etc/rc.d/rc.local
	See ftp://mason.stearns.org/pub/doc/proxyarp-howto for a writeup I
did on proxyarp.
	One more note; you can conserve one more address by using
216.74.72.5 as the IP address for both eth0 and tap0.
	Cheers,
	- Bill

---------------------------------------------------------------------------
	"Love is a snowmobile racing across the tundra and then suddenly it
flips over, pinning you underneath.  At night, the ice weasels come."
	-- Matt Groening
(Courtesy of Steve Dodd <di...@lo...>)
--------------------------------------------------------------------------
William Stearns (wst...@po...).  Mason, Buildkernel, named2hosts,
and ipfwadm2ipchains are at:                http://www.pobox.com/~wstearns
LinuxMonth; articles for Linux Enthusiasts! http://www.linuxmonth.com
--------------------------------------------------------------------------

I've succeeded in getting UML to boot within my Redhat Linux 6.1
box (running kernel 2.2.17, UML was running 2.4.0) I also got
virtual ethernet to work on tap0 (I can contact the host and the
host can contact UML, but UML cannot see the outside world without
a proxy.

The host has an IP of 216.74.72.5, which is accessable to the rest of
the world. I know it's possible to set up IP Masquerating, but I would
like people to be able to connect to service(s) I enable within UML. 
All I would need to do is set tap0 to 216.74.72.6 and eth0 in UML to
216.74.72.7 plus adding some ipchains rules correct? Or does it involve
much more? (I already tried 'um_eth_net_util eth0 100', but that puts
the Ethernet card in promiscuous mode, which overloads UML with network
traffic.)

--
Mooneer Salem
Sysadmin, Ultraspeed UK (http://www.ultraspeed.co.uk/)
GPLTrans (http://www.translator.cx/)
IRC: irc.slashnet.org/mind21_98
Personal Home Page (http://msalem.translator.cx/)

phi...@in... said:
>  I'm currently stuck in vread reading 0 bytes because vmlist is null.
> Here's the backtrace:
> #0  vread (buf=0x50738000 "", addr=0x0, count=4096) at vmalloc.c:269
> #1  0x10058c51 in read_kmem (file=0x500f5300,  

I don't see that this is my problem.  vmlist seems to be exclusively used for 
vmalloc/vfree and it will be NULL if there haven't been any vmallocs.  And 
under UML, vmalloc is pretty much only used for swapping and modules.  So, I 
not surprised that it's NULL.

				Jeff

Well, it's my first attempt with UML, so maybe I'm missing something
important.
Pointers for other documentation (RTFM) that may help me are welcome.

Here is my setting.

Host is kernel 2.2.17+patchs (mostly .18pre).
UML Kernel is 2.4.0-test11-1um, custom build.
root_fs is Conectiva Linux 6.0.

Everything is working fine, except for 2 things:

When using the uml network interface, I can't do nfs over it.
I can mount and ls the mounted dir, but when I try to access(read) any
file, it hangs. Also, trying to umount the filesystem results in a hang.

Most notable is this behaviour (thendara =3D host, uml =3D uml):

thendara [/home/rodrigob] > ping 192.168.0.253
PING 192.168.0.253 (192.168.0.253): 56 octets data
64 octets from 192.168.0.253: icmp_seq=3D0 ttl=3D255 time=3D1.3 ms
64 octets from 192.168.0.253: icmp_seq=3D2 ttl=3D255 time=3D0.6 ms
64 octets from 192.168.0.253: icmp_seq=3D13 ttl=3D255 time=3D0.5 ms
64 octets from 192.168.0.253: icmp_seq=3D23 ttl=3D255 time=3D0.7 ms

[root@uml /root]# ping 192.168.0.253
PING 192.168.0.253 (192.168.0.253): 56 octets data
64 octets from 192.168.0.253: icmp_seq=3D0 ttl=3D255 time=3D2.4 ms
64 octets from 192.168.0.253: icmp_seq=3D1 ttl=3D255 time=3D1.5 ms
64 octets from 192.168.0.253: icmp_seq=3D2 ttl=3D255 time=3D1.6 ms
64 octets from 192.168.0.253: icmp_seq=3D3 ttl=3D255 time=3D1.5 ms
64 octets from 192.168.0.253: icmp_seq=3D4 ttl=3D255 time=3D1.5 ms
64 octets from 192.168.0.253: icmp_seq=3D5 ttl=3D255 time=3D1.5 ms
64 octets from 192.168.0.253: icmp_seq=3D6 ttl=3D255 time=3D1.5 ms
64 octets from 192.168.0.253: icmp_seq=3D7 ttl=3D255 time=3D1.6 ms
64 octets from 192.168.0.253: icmp_seq=3D8 ttl=3D255 time=3D1.5 ms
64 octets from 192.168.0.253: icmp_seq=3D9 ttl=3D255 time=3D1.5 ms

Ping from host to uml, 85% package loss. I guess that is what is making nfs
hang. Actualy, I'm pretty sure thats the cause but, then again, I may be
wrong.

About the uml eth0 interface, I have no clue how to use it. Maybe that is
the solution for my problem, but I found no documentation anywere.

As an extra doubt. When I start an umn interface, the host sistem gets
and slip interface with the 192.168.0.254 ip address. The uml IP is easy
enough to configure, but I would like to know if there is any=20
kernel paramter I can use at load time to specify the IP address
to be used on the host slip interface (changing the hardcoded value is
what I'm doing now).

Thank you very much.

Any further information (.config file, versions and so on) are avaliable
under request.

[]s

--=20
 Rodrigo Barbosa (morcego)  - rod...@co...
 Conectiva R&D Team         - http://distro.conectiva.com.br
 "Quis custodiet custodes?" - http://www.conectiva.com

phi...@in... said:
> #0  vread (buf=0x50738000 "", addr=0x0, count=4096) at vmalloc.c:269
> #1  0x10058c51 in read_kmem (file=0x500f5300,  

Yeah, some /dev and /proc entries probably need some arch support that I 
haven't implemented yet...

				Jeff

I did

  grep foo / -r

on my debian-small root_fs and I'm currently stuck in vread reading 0
bytes because vmlist is null.  Here's the backtrace:

#0  vread (buf=0x50738000 "", addr=0x0, count=4096) at vmalloc.c:269
#1  0x10058c51 in read_kmem (file=0x500f5300, 
    buf=0x4014f000 "#! /bin/sh -\n\nRCSID='$Id: MAKEDEV 2.3 1997/09/29 19:03:13 johnsonm Exp $'\n\n#---#---#---#---#---#---#---#---#---#---#---#---#---#---#---#---#---#---#---#\n# Customisation:\n#   The devices fall into vari"..., 
    count=1490944, ppos=0x500f5320) at mem.c:258
#2  0x1001f858 in sys_read (fd=3, 
    buf=0x4014f000 "#! /bin/sh -\n\nRCSID='$Id: MAKEDEV 2.3 1997/09/29 19:03:13 johnsonm Exp $'\n\n#---#---#---#---#---#---#---#---#---#---#---#---#---#---#---#---#---#---#---#\n# Customisation:\n#   The devices fall into vari"..., 
    count=1490944) at read_write.c:133
#3  0x1006b7db in execute_syscall (syscall=3, args=0x5029bce4)
    at syscall_kern.c:340
#4  0x1006bc4b in syscall_handler (unused=0) at syscall_user.c:113
#5  0x1006893d in fork_handler (sig=10) at process.c:96
#6  0x10073208 in __restore ()
    at ../sysdeps/unix/sysv/linux/i386/sigaction.c:127
(gdb) 

-- 
Daniel

phi...@in... said:
> Rik van Riel gave me this little gem for stress testing the buffer
> cache:
> 	usermode:~# dd if=/dev/ubd0 of=/dev/null
> After about 30 seconds of this we get a couple of:
> end_request: I/O error, dev 62:00 (User-mode block device), sector 204800
> end_request: I/O error, dev 62:00 (User-mode block device), sector 204802
> and can't break out of it. 

I found the bug.  The patch is below.

				Jeff

--- ../cvs/linux/arch/um/drivers/ubd.c  Tue Nov 28 11:41:48 2000
+++ arch/um/drivers/ubd.c       Tue Nov 28 13:16:30 2000
@@ -352,23 +352,27 @@
 static void do_one_request(void)
 {
        struct io_thread_req req;
-       int block, nsect, dev;
+       int block, nsect, dev, again;
 
        if(CURRENT->rq_status == RQ_INACTIVE) return;
-       if (DEVICE_INTR) return;
-       INIT_REQUEST;
-       block = CURRENT->sector;
-       nsect = CURRENT->current_nr_sectors;
-       dev = MINOR(CURRENT->rq_dev);
-       req.read = (CURRENT->cmd == READ);
-       req.fd = ubd_dev[dev].fd;
-       req.offset = block << 9;
-       req.length = nsect << 9;
-       req.buffer = CURRENT->buffer;
-       if((req.offset >= ubd_dev[dev].size) && (ubd_dev[dev].size != -1)){
-               end_request(0);
-               return;
-       }
+       if (DEVICE_INTR) return;
+       do {
+               again = 0;
+               INIT_REQUEST;
+               block = CURRENT->sector;
+               nsect = CURRENT->current_nr_sectors;
+               dev = MINOR(CURRENT->rq_dev);
+               req.read = (CURRENT->cmd == READ);
+               req.fd = ubd_dev[dev].fd;
+               req.offset = block << 9;
+               req.length = nsect << 9;
+               req.buffer = CURRENT->buffer;
+               if((req.offset >= ubd_dev[dev].size) && 
+                  (ubd_dev[dev].size != -1)){
+                       end_request(0);
+                       again = 1;
+               }
+       } while(again);
        if(thread_fds[0] == -1){
                do_io(&req);
                ubd_finish();

On Tue, 28 Nov 2000, Jeff Dike wrote:
> The latest CVS includes the "ubd=sync" option, which makes the block driver do 
> sychronous I/O.  When you put this switch on the command line, the driver puts 
> this out on the console:
> 	ubd : Synchronous mode
> 
> It also happens to be somewhat faster than doing asynchronous I/O.
> 
> This lets uml run dbench without messing up the filesystem, so this ought to 
> let you bang on it.
> 
> Also "debug=go" is in there.

:-)

-- 
Daniel

The latest CVS includes the "ubd=sync" option, which makes the block driver do 
sychronous I/O.  When you put this switch on the command line, the driver puts 
this out on the console:
	ubd : Synchronous mode

It also happens to be somewhat faster than doing asynchronous I/O.

This lets uml run dbench without messing up the filesystem, so this ought to 
let you bang on it.

Also "debug=go" is in there.

				Jeff

phi...@in... said:
> Of course, it could also be a race in -test11... 

I don't think so.  The fact that the problem goes away when the driver is run 
synchronously points the finger at my code.

				Jeff

On Mon, 27 Nov 2000, Jeff Dike wrote:
> I finally got a clean run of dbench after I gave it a 500M filesystem of its 
> own.  The errors I was getting from it were caused by it running out of room.
> 
> The forced fsck afterwards found inconsistencies.  I ran it again with the 
> block driver doing synchronous I/O and the filesystem was clean afterwards.
> 
> So, it looks like there's a race or something in the block driver.

Of course, it could also be a race in -test11...

-- 
Daniel

I finally got a clean run of dbench after I gave it a 500M filesystem of its 
own.  The errors I was getting from it were caused by it running out of room.

The forced fsck afterwards found inconsistencies.  I ran it again with the 
block driver doing synchronous I/O and the filesystem was clean afterwards.

So, it looks like there's a race or something in the block driver.

				Jeff

phi...@in... said:
> Have you tried running dbench?

I just did.

> For me, this messes ext2 up seriously on uml (2.4.0-test10  host/
> guest). What happens with you? 

dbench gave me a flood of error messages.  I ^C'd it, halted, did e2fsck -f on 
the filesystem and that turned up a few inconsistencies.

It looks like my block driver needs some looking at.

				Jeff

ri...@co... said:
> This makes me curious ... how does UML decide which pages from an
> mmap()ed file to write back and which pages to just discard ? 

It doesn't.  That's the job of the generic kernel, not the port.  But Daniel 
clued me in that the port needs to set dirty bits so that the generic kernel 
can tell what it needs to write back.

				Jeff

Hi Jeff.

Have you tried running dbench?  It's a stress tester from Andrew
Tridgell of Samba fame:

  ftp://samba.org/pub/tridge/dbench/dbench-1.1.tar.gz

For me, this messes ext2 up seriously on uml (2.4.0-test10 
host/guest). What happens with you?

(install instructions: unzip; make; ignore the compile warning)

I haven't tried it on a real test10 system so it's not certain this is
a uml problem.

-- 
Daniel

On Tue, 21 Nov 2000, Jeff Dike wrote:
> phi...@in... said:
> > Well, I've got a solid day's vm hacking under uml, and I have to say
> > it works like a dream.  It's really nice that the whole vm
> > infrastructure is there. 
> 
> There's just one tiny little problem if you want to do VM
> hacking.  There's no equivalent of hardware access or dirty bits
> (I think i386 provides an access bit, not sure about dirty).
> 
> If you need them, I can put in some emulation.  I haven't done
> it because I haven't heard of anyone needing them.

This makes me curious ... how does UML decide which pages
from an mmap()ed file to write back and which pages to
just discard ?

regards,

Rik
--
Hollywood goes for world dumbination,
	Trailer at 11.

http://www.conectiva.com/		http://www.surriel.com/

Hi

yeah that i think would be a good thing
i keep forgetting the command args

but hey i am managing to install slackware
at the minute which is a pritty good thing :)
oh BTW i am using the slackware 7.1 root disks
not the ones from the web site

cya
	James

In local.user-mode-linux, you wrote:
>How about supporting --help:
>
>  linux --help
>
>-- 
>Daniel
>_______________________________________________
>User-mode-linux-user mailing list
>Use...@li...
>http://lists.sourceforge.net/mailman/listinfo/user-mode-linux-user
>


-- 
---------------------------------------------
Check Out: http://stev.org
E-Mail: mi...@st...
 12:20am  up 41 days, 12:16,  7 users,  load average: 1.14, 1.36, 0.97

How about supporting --help:

  linux --help

-- 
Daniel