Linux-Development-Sys Digest #110, Volume #8 Tue, 29 Aug 00 20:13:12 EDT
Contents:
Re: Resolution of select timeout (Rick Ellis)
New nVidia drivers on 2.4 kernel ([EMAIL PROTECTED])
Re: New nVidia drivers on 2.4 kernel (Peter Mardahl)
Re: (YASP) Seagate Barracuda (ST318416W) Incompatibility (Dave Platt)
RAID-1 with NBD (Christian Fischbach)
Re: Linux memory perfomance is horrible (bill davidsen)
Re: Linux memory perfomance is horrible (bill davidsen)
how to handle local modules? (Richard Bonomo)
Re: Q: SMP question (bill davidsen)
USENIX Operating Systems Symposium (OSDI 2000) (Hali McGrath)
Re: RAID-1 with NBD (bill davidsen)
LARGE_FILE oddities ("Lawrence K. Chen, P.Eng.")
Re: (Q) Capturing everything at boot time (The Ghost In The Machine)
Re: checking properties of incoming connection (The Ghost In The Machine)
Re: EACCES on bind (The Ghost In The Machine)
----------------------------------------------------------------------------
From: [EMAIL PROTECTED] (Rick Ellis)
Subject: Re: Resolution of select timeout
Date: 29 Aug 2000 18:10:51 GMT
In article <[EMAIL PROTECTED]>,
Peter Mueller <[EMAIL PROTECTED]> wrote:
>I use select to get notefied of some timeouts. I found that the resolution of
>the timeout time is 10 milli seconds. Is my measurement true?
That makes sense with default 100 Hz clock.
--
http://www.fnet.net/~ellis/photo/linux.html
------------------------------
From: [EMAIL PROTECTED]
Subject: New nVidia drivers on 2.4 kernel
Date: Tue, 29 Aug 2000 18:53:32 GMT
Anybody managed to get these working yet? Looks
like some of the paging calls need to be changed.
Just thought I would see if anyone else has done
it before I give it a shot.
Steve White
[EMAIL PROTECTED]
He whose ranks are united in purpose will be
victorious.
- Sun Tzu, Art of War
Sent via Deja.com http://www.deja.com/
Before you buy.
------------------------------
From: [EMAIL PROTECTED] (Peter Mardahl)
Subject: Re: New nVidia drivers on 2.4 kernel
Date: 29 Aug 2000 20:07:21 GMT
In article <8oh0r2$u0n$[EMAIL PROTECTED]>, <[EMAIL PROTECTED]> wrote:
>
>
> Anybody managed to get these working yet? Looks
>like some of the paging calls need to be changed.
>
> Just thought I would see if anyone else has done
>it before I give it a shot.
Trying to use these drivers with my 2.4-test7 kernel failed.
Machine locked up.
PeterM
------------------------------
From: [EMAIL PROTECTED] (Dave Platt)
Crossposted-To: comp.os.linux.hardware
Subject: Re: (YASP) Seagate Barracuda (ST318416W) Incompatibility
Date: Tue, 29 Aug 2000 20:55:26 GMT
In article <[EMAIL PROTECTED]>,
D. Stimits <[EMAIL PROTECTED]> wrote:
>> I am running into what sounds like the same problem with a
>> Seagate Barracuda and a several year old AHA 2940UW. I'm quite
>> certain termination is correct. I've tried this drive alone
>> (on the wide bus, still have a CD-ROM and DAT on the narrow
>> portion of the bus) with termination enabled and with two other
>> drives with termination supplied by another drive. (Before
>> trying to add the drive, I had a working system with three IBM
>> Ultrastor UW drives runnning the 2.2.12 Kernel with no apparent
>> problems. Until one of the drives failed, that is.)
>>
>> I was able to format the drive under Windows NT and copy files
>> to it, so it seems that the hardware works. There must be
>> something different about how the Linux driver handles the
>> drive that triggers this problem.
>>
>> The first error message I see before the endless loop of SCSI
>> bus resets starts has text something along the line of 'Testing
>> Unit Ready'
>>
>> There seem to be others experiencing the same problem, so it
>> does not appear to be a sample defect with the drive.
>I have heard of some particular drives being a problem, or at least with
>certain mixes. I don't know which drives those are though, so I couldn't
>say if the Seagate Barracuda itself has known compatibility problems. It
>is possible though. Someone else with a Seagate Barracuda might mention
>if it is a problem or not.
I've used some of the older (4-gig) Barracuda models with Linux (DTC
SCSI controller based on the NCR 53C810 chip) without difficulty.
You might want to check the drive configuration, and see if the drive
is jumpered or optioned in a way which would cause it to _initiate_
the SCSI option-negotiation for wide and fast transfers. I believe I
recall that certain Linux SCSI-adapter drivers have difficulty with
this - their firmware isn't set up to handle device-initiated
negotiations.
If your drives are configured/jumpered in this way, try disabling
drive-initiated option negotiation. The Linux kernel drivers should
initiate the appropriate fast/wide negotiation when they're good and
ready to do so.
I know that the IBM Ultrastar SCSI drives come with the
drive-initiated negotiation feature disabled by default... possibly
some Seagate drives have a different default.
--
Dave Platt [EMAIL PROTECTED]
Visit the Jade Warrior home page: http://www.radagast.org/jade-warrior/
I do _not_ wish to receive unsolicited commercial email, and I will
boycott any company which has the gall to send me such ads!
------------------------------
From: Christian Fischbach <[EMAIL PROTECTED]>
Subject: RAID-1 with NBD
Date: Tue, 29 Aug 2000 23:27:09 +0200
Hello!
I like to build a very redundant system with two exactly the same computer
as a File-Server for Windows (Samba).
The idea is, that if one system hangs, the other takes immediately its
job. I think, that it is not to difficult for the slave system to know,
when it has to do the job for the master. My problem is the
synchronisation of the two harddrives. Would it be able to use
Software-RAID Level 1 with one local (IDE) HD and one remote Network Block
Device? Are this two kernel-drive stable enough?
If anyone has an idea or a better solution, I would be very happy for an
answer.
Thank you,
Chris
------------------------------
From: [EMAIL PROTECTED] (bill davidsen)
Crossposted-To: comp.unix.bsd.freebsd.misc
Subject: Re: Linux memory perfomance is horrible
Date: 29 Aug 2000 21:35:24 GMT
In article <[EMAIL PROTECTED]>,
Kaz Kylheku <[EMAIL PROTECTED]> wrote:
| On Mon, 31 Jul 2000 16:54:20 GMT, Szabolcs Csetey <[EMAIL PROTECTED]> wrote:
| >BUFSIZ defined in stdio.h. FreeBSD and Solaris use 1024 while
| >Linux 8192. The memory tests are done with BUFSIZ chunks so
| >Linux should do 8x more work during the same time.
|
| What kind of a moron would design a benchmark that depends on
| implementation-defined constants like BUFSIZ?
More to the point, would use it to determine what you are trying to
do, but no when reporting what you have done?
It's a bad benchmark, what can you say? I think moron is a poor choice
of terms, most of us make mistakes from time to time.
--
bill davidsen <[EMAIL PROTECTED]> CTO, TMR Associates, Inc
Make the rules? I don't make the rules. I don't even FOLLOW the rules!
------------------------------
From: [EMAIL PROTECTED] (bill davidsen)
Subject: Re: Linux memory perfomance is horrible
Date: 29 Aug 2000 21:39:25 GMT
In article <[EMAIL PROTECTED]>,
Tim Moore <[EMAIL PROTECTED]> wrote:
| Here's some numbers I got by altering BUFSIZ. Which one reflects the
| actual memory throughput?
They all do, performance vs. BUFSIZE is a function or curve, not a
single number.
You are learning that not only is hard to write a good benchmark, it's
even harder to determine what it means.
--
bill davidsen <[EMAIL PROTECTED]> CTO, TMR Associates, Inc
Make the rules? I don't make the rules. I don't even FOLLOW the rules!
------------------------------
From: Richard Bonomo <[EMAIL PROTECTED]>
Subject: how to handle local modules?
Date: Tue, 29 Aug 2000 16:24:05 -0500
Here are a couple of things which I have managed
to miss thus far:
I have written and tested a character device driver
which I load manually and for which I manually create
the entry in /dev. This is running on an x86 platfrom
under RedHat Linux 6.1 (kernel v. 2.2.12-20).
I would like to have this load at boot time for the
convenience of the people who will be using it (or
at least have kmod load it dynamically).
1. What is the best place to put the compiled
module binary and associated header file? Is there
a customary place for such "locally generated" drivers?
2. As the major number is assigned dynamically, is
there a standard initialization or module-insertion
protocol which will transparently ensure the existence
of or create the necessary device name with appropriate
major number?
Thank you.
Rich B.
please reply to [EMAIL PROTECTED]
--
************************************************
Richard Bonomo
University of Wisconsin
************************************************
------------------------------
From: [EMAIL PROTECTED] (bill davidsen)
Subject: Re: Q: SMP question
Date: 29 Aug 2000 21:48:46 GMT
In article <[EMAIL PROTECTED]>,
Holger Eitzenberger <[EMAIL PROTECTED]> wrote:
| me and a colleague were discussing how SMP is implemented on Linux.
| I know that in true SMP you don't have the guarantee on which CPU a
| process will run the next time the process gets processor time. Is
| this different on Linux? My coleague is pretty sure that a process on
| Linux always runs on the same CPU.
Linux has a preference for running a process on the same CPU, which
means that a task will wait longer for dispatch if it can get the same
processor. By setting this higher the probability of getting the same
CPU can be set higher.
Note that it's a waste of time to set it higher, since you lose more
real time waiting than you gain by having the same CPU, and because the
longer the CPU runs in another task the lower the chance that anything
useful will be in the cache.
With 2.4 the gains are even smaller I believe, because the FPU
registers are always saved, instead of saved only when the "next" task
wanted to use them. Or so the list of changes reads.
If you really want to ensure running the same CPU build a uniprocessor
kernel.
--
bill davidsen <[EMAIL PROTECTED]> CTO, TMR Associates, Inc
Make the rules? I don't make the rules. I don't even FOLLOW the rules!
------------------------------
From: Hali McGrath <[EMAIL PROTECTED]>
Crossposted-To:
comp.arch,comp.object,comp.os.inferno,comp.os.linux.development.apps,comp.os.ms-windows.misc,comp.os.security,comp.protocols.nfs,comp.realtime,comp.security.misc
Subject: USENIX Operating Systems Symposium (OSDI 2000)
Reply-To: [EMAIL PROTECTED]
Date: Tue, 29 Aug 2000 15:16:12 -0700
4th USENIX Symposium on Operating Systems Design
and Implementation (OSDI 2000)
October 23-25, 2000
Paradise Point Resort
San Diego, California, USA
OSDI 2000 captures exciting, innovative work in the systems software
area, taking a broad view of what the area encompasses. The program
presents important results in a wide range of areas including new
storage devices, fault tolerance, software architectures for Internet
services, Internet content distribution, applying language techniques
to systems software, scheduling for variable-speed processors, storage
management, and distributed secure filesystems. OSDI 2000 also offers
more opportunities to learn about the state of the art in systems
software by connecting you with the researchers pursing promising
research ideas in their formative stages.
THREE DAYS OF REFEREED PAPERS AND INVITED SPEAKERS
================================================================
* A program of creative, highly original work opening up new
areas--not merely incremental results building on prior work.
* Diverse topic areas such as new storage devices, fault tolerance,
software architectures for Internet services, Internet content
distribution, applying language techniques systems software,
scheduling for variable-speed processors, storage management, and
distributed secure filesystems.
* Numerous opportunities to learn about the state of the art in
systems software through short Work-In-Progress presentations and an
evening Poster session where you'll get a chance to interact with
researchers pursing promising research ideas in their formative
stages.
* Keynote Danny Lewin, Chief Technical Officer of Akamai Corporation,
a leading Internet content delivery company, speaking on "Systems
Issues in Global Internet Content Delivery".
WIESS 2000 -- CO-LOCATED ONE DAY WORKSHOP
================================================================
The First Workshop on Industrial Experiences with Systems Software
(WIESS 2000), October 22, 2000
http://www.usenix.org/events/osdi2000/wiess2000
Co-sponsored by USENIX, IEEE TCOS, and ACM SIGOPS
"From the Trenches"
The First Workshop on Industrial Experiences with Systems Software
The First Workshop on Industrial Experiences with Systems Software
(WIESS 2000) is the ideal opportunity for OS professionals working "in
the trenches" to share their knowledge and connect with like-minded
peers. WIESS was conceived and co-sponsored by USENIX, ACM SIGOPS, and
IEEE-CS TCOS. It is designed to complement the SOSP and OSDI
conferences by focusing on industry results of immediate benefit and
use rather than long-term research. The workshop will be informal in
style, featuring short papers, invited talks, and posters that will
allow the participants to focus on the practical aspects and current
problems in the field.
================================================================
Register by October 2, 2000 and save on hotel and conference rates!
For Detailed conference information is online
http://www.usenix.org/events/osdi2000
================================================================
Sponsored by USENIX, the Advanced Computing Systems Association
Co-sponsored by IEEE TCOS and ACM SIGOPS
------------------------------
From: [EMAIL PROTECTED] (bill davidsen)
Subject: Re: RAID-1 with NBD
Date: 29 Aug 2000 22:23:08 GMT
In article <[EMAIL PROTECTED]>,
Christian Fischbach <[EMAIL PROTECTED]> wrote:
| I like to build a very redundant system with two exactly the same computer
| as a File-Server for Windows (Samba).
|
| The idea is, that if one system hangs, the other takes immediately its
| job. I think, that it is not to difficult for the slave system to know,
| when it has to do the job for the master. My problem is the
| synchronisation of the two harddrives. Would it be able to use
| Software-RAID Level 1 with one local (IDE) HD and one remote Network Block
| Device? Are this two kernel-drive stable enough?
As many people have found, it is VERY hard to know when the master is
hung, or down, or whatever, as opposed to just being a tad slow, or the
heartbeat link slow or failed.
The case where the backup decides to become master but the master
isn't in need of it is called the "palace revolution" by some people,
and in the case of IP address takeover can result in some very amusing
horrible failures. When the slave takes over the master absolutely must
be out of the picture.
| If anyone has an idea or a better solution, I would be very happy for an
| answer.
Not that doesn't involve a bunch of expensive commercial hardware,
which is designed for the job.
I really like your idea, if you have the ability to test it I would
love to hear what you find. But the slave has to take over the NBD as
well, which might add to the problem. Could it mount it as an NBD from
itself? Hum, have to look at the code.
Another possible solution would involve using a SCSI bus between the
systems (called multiple initiator) and allowing the master to write
drives physically on the slave. Again, to be totally secure would take
kernel mods.
If you want to snoop the net you could have the slave process the
write commands to keep its data up to date, but I think you would have
to do hacking on SAMBA to really make this work.
Hope you do this, in test you can start writing and shoot the master,
then see what the slave does. I would expect the read case to work, the
data is presumably current on the slave.
--
bill davidsen <[EMAIL PROTECTED]> CTO, TMR Associates, Inc
Make the rules? I don't make the rules. I don't even FOLLOW the rules!
------------------------------
From: "Lawrence K. Chen, P.Eng." <[EMAIL PROTECTED]>
Subject: LARGE_FILE oddities
Date: Tue, 29 Aug 2000 18:13:22 -0400
I was experimenting to see what kind of errors (errno) that I would get if I
tried to exceed 2GB using 32bit file I/O. Since we're planning to let our
betatest version go out where internally it could access 8GB files, but the
file API is only 32bit. (fcntl support for F_SETLK64/F_GETLK64 is missing,
and what is defined makes it broken....which prevents our app for working on
current releases of RedHat & SuSE).
The testing is on a clean install of RedHat 6.2 (comparison is against a
Solaris 7 box).
Deviations noted are:
open()'ng a file that is larger than 2GB should yield EOVERFLOW, but on Linux
it succeeds. Though other failures occur as expected as if the file was only
2GB.
lseek(fd,+ve,SEEK_END)
yields ECHILD on Linux (instead of EINVAL).
pread() and pwrite() do care where the file pointer is.
If I lseek() to 4K from the end of a 2GB file, and pread() or pwrite() 8K.
All operations fail with EOVERFLOW, even if the offset used by pread/pwrite
would be valid. lseek() to somewhere where the operation from that point
would work, the pread()/pwrite() works/fails as expected.
--
Snail: Lawrence K. Chen Email: [EMAIL PROTECTED]
4411 Camden Circle URL 1: http://www.bigfoot.com/~TheDreamer/
Dublin, OH 43016-3553 URL 2: http://members.xoom.com/TheDreamer/
Phone: 614-791-2130 Fax: 614-792-7544 ICQ: 5235156
------------------------------
From: [EMAIL PROTECTED] (The Ghost In The Machine)
Subject: Re: (Q) Capturing everything at boot time
Date: Tue, 29 Aug 2000 22:25:54 GMT
In comp.os.linux.development.system, Kevin Lacquement
<[EMAIL PROTECTED]>
wrote
on Thu, 20 Jul 2000 05:35:27 GMT
<[EMAIL PROTECTED]>:
>Timothy Murphy wrote:
>>
>> dmesg doesn't capture everything that flashes by at boot-time,
>> eg inconsistencies between modules.
>> Is there any simple way -- a kind of kernel "script" --
>> which will save everything that appears on the screen at boot-up?
>>
>> --
>
>cat /var/log/bootlog
Pedant point: /var/log/boot.log, at least on my RedHat system.
(Dunno if they patched it or not. Shouldn't be hard to find
regardless, either way.)
--
[EMAIL PROTECTED] -- insert random misquote here
------------------------------
From: [EMAIL PROTECTED] (The Ghost In The Machine)
Subject: Re: checking properties of incoming connection
Date: Tue, 29 Aug 2000 23:35:28 GMT
In comp.os.linux.development.system, Bhavin Shah
<[EMAIL PROTECTED]>
wrote
on Fri, 11 Aug 2000 10:44:13 -0700
<[EMAIL PROTECTED]>:
>Hi,
>
>Is there a way to check the properties (say ip addr)
>of an incoming connection from listen/accept?
>
>Thanks in advance.
>From the manpage:
int accept(int s, struct sockaddr *addr, int *addrlen);
Three guesses what the next two parameters are for, and the
first two don't count. :-)
Note, however, that *addrlen has to contain the size allocated
for *addr upon routine entry, which means usual code would look
something like:
struct sockaddr sin;
int sinl = sizeof(sin);
int endpt = accept(sock, &sin, &sinl);
and then sinl will contain the number of bytes in sin that
actually make sense. (If it's different than what you fed in,
you might have a problem, although Linux tries its best to
ensure that sockaddr, sockaddr_in, and sockaddr_un all
have the same size.)
You might want to look at gethostbyaddr() to complete
the solution to your problem; be warned, though, that
I suspect gethostbyaddr() is going to be slow if it has
to do a reverse DNS lookup.
'man 2 accept' and 'man 3 gethostbyaddr' for details.
>
>Bhavin
>
>
--
[EMAIL PROTECTED] -- insert random misquote here
------------------------------
From: [EMAIL PROTECTED] (The Ghost In The Machine)
Subject: Re: EACCES on bind
Date: Tue, 29 Aug 2000 23:56:56 GMT
In comp.os.linux.development.system, Bhavin Shah
<[EMAIL PROTECTED]>
wrote
on Mon, 14 Aug 2000 12:46:19 -0700
<[EMAIL PROTECTED]>:
>Hi,
>
>I'm trying to bind a socket to an address, but
>am getting an EACCES errno, which states that
>the address is protected and the user isn't a
>super user. I'm running the program as a user
>other than root, and gave him all priv's.
>In linuxconf, I granted super-user equilvalence,
>but I can only bind w/ root. How would I allow
>this other user to bind? Thanks.
>
>Bhavin
>
>
I'm not at all sure what it means to give a non-root user
all privileges -- most likely, it assigned him to various
groups, one of them being root. However, AFAIK, groups
aren't looked at when opening a socket. (I'm not that
familiar with linuxconf.)
You are basically running afoul of a protective system limit,
which disables binding any socket with a port # less than
1024 or 4096 -- I forget which.
One rather simple -- and very dangerous -- way is to grant
setuid privilege (chmod +s) to the program opening the socket,
after chowning it to root; the danger is that, if the program
is compromised by an enterprising hacker, instant root access.
Also, if the program forks off subshells or uses system(),
care must be taken to ensure the correct user ID is used;
one doesn't want to grant root access to subprograms that shouldn't
have it!
(And yes, this has happened, as far back as about 1982 or
thereabouts; the 'mail' program had a bug that gave anyone
using a shell escape ('!') superuser access! Needless to say, this
was fixed fairly quickly.)
A safer method is to rewrite, if possible, the two
communicating programs (you did not indicate what two programs
were communicating) so that they use non-privileged ports.
My $.02, admittedly.
--
[EMAIL PROTECTED] -- insert random misquote here
------------------------------
** FOR YOUR REFERENCE **
The service address, to which questions about the list itself and requests
to be added to or deleted from it should be directed, is:
Internet: [EMAIL PROTECTED]
You can send mail to the entire list (and comp.os.linux.development.system) via:
Internet: [EMAIL PROTECTED]
Linux may be obtained via one of these FTP sites:
ftp.funet.fi pub/Linux
tsx-11.mit.edu pub/linux
sunsite.unc.edu pub/Linux
End of Linux-Development-System Digest
******************************
