Kernel Traffic #255 For 31 Mar 2004

I'm happy to announce the initial public release of Strace Test. I believe Strace Test is the most aggressive general purpose kernel tester available. Strace Test generally crashes my system within 5 minutes (2.6.1-rc2).

Strace Test uses a modified version of strace 4.5.1. Instead of printing out information about system calls, the modified version calls the syscalls with improper values. A patch and a binary for i386 are included in the strace_test tar ball.

Strace Test uses LTP to generate real world syscalls. Just unpack ltp and type 'make -k'. You don't need to install the test if you don't want to.

The modifications make the test scripts go haywire. To keep the test on track we restart it every 10 seconds. The first script is run as root and it spawns off the test as a test user. Every 10 seconds root kills off all the test user's processes and restarts the test. The actual tests are run with user permissions.

Strace Test is available from: http://67.113.20.209/strace_test.tar.bz2

Test Instructions (for i386)
Create a test user
Download and untar ltp (ltp.sf.net)
Cd to ltp and `make -k`
Untar strace_test
cd strace_test && ./go_go.sh
Enter the path to ltp
Enter the test user

Since Nischal didn't specify which port on the sound card he was thinking about, I started to think that a SPDIF digital output / input, would give much better results:

6 channels, 16 bit, 48KHz = 4.6Mbit/s

I don't know enough about the standard digital format and the maximum bandwidth that we could get from a sound card, but in theory we probably could connect the digital input on a sound card to digital output on another card (and vice-versa) and map the whole thing as a network interface :)

Anyway, this is an extremely crazy, time wasting, "just do it if you have loads of time to throw out the window" kind of project, because the cost of ethernet NIC's is extremely low these days.

Here's a patch that adds the inter-partition Virtual Ethernet driver for newer IBM iSeries and pSeries systems:

The patch applies against 2.6.3-rc3...

Jeff, can you formally add this driver to 2.6?... The differences between this driver and the 2.4 driver that you accepted are fairly trivial (i.e. workqueues instead of tasklets)... The architectural additions that I was waiting for have been applied to the mainline tree...

now that Intel has finally come clean about their x86-64 implementation (see http://www.intel.com/technology/64bitextensions/index.htm?iid=techtrends+spotlight_64bit for full details), can somebody write up a list of differences? I know there are people who have had access to the Intel docs for a while now, and obviously Intel is too frigging proud to list the differences explicitly.

From what I can tell from a quick look, it looks like it is basically just the 3DNow vs SSE3 thing, but I assume there are other details too. Can people who have been involved with this make a quick list for the rest of us who only got to see the final details today?

(And I assume there's somebody with a few patches pending..)

From what I can see from these docs, Intel's "IA-32e" is very very close to the natural combination of P4 with AMD64. No hyperlink stuff, but otherwise the same. The local APIC and performance counters should be exactly as in P4 :-)

What about naming? IA-64 is taken, AMD64 is too specific, Intel's "IA-32e" sounds too vague, and I find x86-64 / x86_64 difficult to type. "x64" perhaps?

x86-64 it is. Maybe you can remap one of your function keys to send the sequence ;)

This whole "ia32" crap has always been ridiculous - nobody has _ever_ called an x86 anything but x86, and Intel is just making it worse by adding random illogical letters to the end.

In contrast, x86-64 tells you _exactly_ what it's all about, and is what the kernel has always called the architecture anyway.

No. The filesystem policy _tends_ to be that dashes and spaces are turned into underscores when used as filenames. Don't ask me why (well, the space part is obvious, since real spaces tend to be a pain to use on the command line, but don't ask me why people tend to conver a dash to an underscore).

So the real name is (and has always been, as far as I can tell) x86-64.

Actually, I'm a bit disgusted at Intel for not even _mentioning_ AMD in their documentation or their releases, so I'd almost be inclined to rename the thing as "AMD64" just to give credit where credit is due. However, it's just not worth the pain and confusion.

Any Intel people on this list: tell your managers to be f*cking ashamed of themselves. Just because Intel didn't care about their customers and has been playing with some other 64-bit architecture that nobody wanted to use is no excuse for not giving credit to AMD for what they did with x86-64.

(I'm really happy Intel finally got with the program, but it's pretty petty to not even mention AMD in the documentation and try to make it look like it was all their idea).

As far as I know, the real reason for the underscore in x86_64 in Linux is that autoconf/configure hate dashes in arch names, because of this notation:

x86_64-gnu-linux-pc

If a dash were used, the string would be unparseable without prior knowledge of all arch names.

Well, the thing is, I _like_ a vendor-neutral name.

I think it's important to have multiple sources for a chip, and I think one of the problems with IA-64 was that it was a locked-in chip with patents and no serious competition internally (ignore the Intel mouthing about "open").

The x86 is so great partly because there's been real competition. So I think it's very important to x86-64 to have real competition to make sure nobody gets too dishonest.

So AMD64 is a bad name, partly for the same reason IA32 is a horrible name (and who have you ever heard use the IA32 name except for people who are paid to do so by Intel?)

What I found so irritating is that _hours_ after the Intel announcement, people were _still_ confused about whether the new intel chip was actually compatible with AMD's chips. Why the f*ck not just come out and say so, and talk about it? It took people actually reading the manuals (which didn't mention it either) to convince some people on the architecture newsgroups that yes, "ia32e" was really the same as "amd64" except in the small details that have always set Intel and AMD apart.

So I don't really want to change the name. "x86-64" is a good name. I just wish there was more honesty involved, and less friggin *POSTURING*.

Sorry for the miscommunication. The page http://www.intel.com/technology/64bitextensions/faq.htm says at the _bottom_ at least:

Q9: Is it possible to write software that will run on Intel's processors with 64-bit extension technology, and AMD's 64-bit capable processors?

A9: With both companies designing entirely different architectures, the question is whether the operating system and software ported to each processor will run on the other processor, and the answer is yes in most cases.

Other than the standard IA-32 differences (eg. HT, SSE3, Intel Enhanced SpeedStep, etc.), there are few differences between the implementations of IA-32e and AMD64. The software visible ones are:

Fast system calls:

Syscall/sysret is supported only in 64-bit mode (not in compatibility mode). Sysenter/Sysexit is supported in both 64-bit and compatible mode.

CPUID:

If you look at Table 2-8 of Volume 1, you will find IA-32e specific things, including, GenuineIntel, HT, SSE3, monitor/mwait, Intel Enhanced SpeedStep, and cmpxchg16b.

The function 8000_0001h doesn't duplicate standard-feature bits from function 1 in EDX. It sets only the new features that are implemented.

MSRs:

Not all MSRs are architectural, and IA-32e does not implement SYSCFG, TOP_MEM, TOP_MEM2, for example. MSR usage should be vendor specific and be guarded with CPUID.Model

Fast-FXSAVE/FXRSTOR:

IA-32e always saves all of the FP state on FXSAVE/FXRSTOR. Does not support FXSAVE/FXRSTOR with reduced FP state.

Microcode Update:

IA-32e supports microcode update as the 32-bit mode does, as you already found the discussions in the mailing list.

NX (No-Execute) bit:

Initial implementation will not support the NX bit.

BSF/BSR when source is 0 & operand size is 32:

In 64-bit mode, the processor sets ZF, and the upper 32 bits of the destination are undefined. Should always check the ZF or do not use 32-bit operand size.

Near branch with 66H prefix:

As documented in PRM the behavior is implementation specific and should avoid using 66H prefix on near branches.

Not supported in IA-32e

3DNow instructions (including prefecthw or prefetch with the opcode 0f 0d)

Ok, it's out.

There were some minimal changes relative to the last -rc4, mostly some configuration and build fixes, but a few important one-liners too.

Here goes release candidate 4, including a few small fixes.

If nothing bad shows up, this will become final.

ftp://ftp.kernel.org/pub/linux/kernel/people/akpm/patches/2.6/2.6.3/2.6.3-mm1/

• Added the dm-crypt driver: a crypto layer for device-mapper.

  People need to test and use this please. There is documentation at http://www.saout.de/misc/dm-crypt/.

  We should get this tested and merged up. We can then remove the nasty bio remapping code from the loop driver. This will remove the current ordering guarantees which the loop driver provides for journalled filesystems. ie: ext3 on cryptoloop will no longer be crash-proof.

  After that we should remove cryptoloop altogether.

  It's a bit late but cyptoloop hasn't been there for long anyway and it doesn't even work right with highmem systems (that part is fixed in -mm).

• Added the fbdev cursor API patch. Not sure what this does apart from preventing the rivafb driver from linking. I'll let others decide if this is progress.
• There's a patch here to consolidate the 32->64 compat code for the IPC syscalls. Needs testing on various 64-bit machines.
• Various random fixes to things.

AFAIK the two big security weaknesses in most version of cryptoloop are: (note that some versions have it fixed with various hacks)

• Weak IV
• Extremly bad key management

The first can be addressed in an crypto API module e.g. with an hashed IV, but it needs a stable IV format from dm-crypto (that is one of the things I asked for)

The second one is more a user space problem. However to solve it you need metadata. It would be much nicer if it was possible to store it on the block device directly (with a special losetup flag for compatibility). Otherwise you get into nasty chicken and egg problems with fully encrypted systems.

Supporting metadata can be quite simple - e.g. a standard header on the first blocks that has a length and a number of records with unique IDs. And the kernel driver would need to skip over these headers.

What definition of "stable kernel" do you use which includes removal of features which were reasons to migrate to 2.6 from 2.4? This change would mean having to add dm to the kernel which otherwise doesn't use it, carry dm utilities on the system whcih are otherwise unneeded, and train people to use and not use dm.

I expect major things to change in a development series, but less major things than this have been pushed to 2.7, why is this being forced in?

Why should we retain crypto capabilities which have widely understood vulnerabilities?

We mainly want to remove the bio remapping stuff from the loop driver because it's horrid and deadlocks under heavy memory pressure. Maybe we can leave crytoloop there with big "kindergarten crypto - do not use" labels all over it.

the point is that the new dmcrypto is only in -mm and cryptoloop is in both trees, those of us developing applications based on cryptoloop don't have a mainline kernel to even start testing dmcrypto against in the 2.6 series, so it is more a political issue than a technical issue which makes the removal of a feature like this from the 2.6 series a bad thing... (In my humble never contributed to the kernel opinion)

Technically speaking there is no doubt that you are correct to want to remove cryptoloop... but if people are depending on that support to stay in a stable kernel and are developing based on it and don't have the time to learn dm or dmcrypto and redesign whatever may need redesigning to use it, it strikes me as rude to pull that support.

This is actually an argument for removing cryptolooop. People are developing against a crypto infrastructure which has well-known weaknesses.

Pulling it out is an excellent way of communicating this fact. Right now, we're just deluding people.

Please apply this patch . They provide the bare minimum read-only support for ufs2 variant (from FreeBSD 5.x ) of the UFS filesystem .

The patch for 2.6.3 is here :
http://ufs-linux.sourceforge.net/ufs2/2.6.3/ufs2-read-only-p1.txt
http://ufs-linux.sourceforge.net/ufs2/2.6.3/ufs2-read-only-p2.txt

ooh, I see you have a mkfs.ufs there. Does it support UFS1 as well?

Does current UFS support little-endian machines? If so, has this code been tested on a little-endian host? The code _looks_ OK, but one does need to test...

Has the patched filesystem been regression tested against a UFS1 filesystem?

I plan to make rather important changes to the sysfs interface of I2C chip drivers in Linux 2.6. The topic has already been discussed on the lm_sensors mailing list, bug Greg KH suggested that I should explain my intentions here too, so here I am.

THE PROBLEM

The sysfs interface to I2C chips as it exists in Linux 2.6 isn't that bad. It has a number of benefits over the old procfs interface we used during the times of 2.2 and 2.4 kernels. The main change (apart from the pseudo-filesystem change itself) is that we opted for a one-value-per-file policy, with the obvious goal to make this interface more chip-independent. We also decided that all values, which are mainly fixed point values, would all use the same number of decimal places, again to make it less chip-dependent.

A few changes were made recently to enhance the chip-independency, such as differentiating between minimum temperatures and hysteresis temperatures.

The problem is that there are still several files that need you to know additional information about the chip to properly use them. Examples of that:

• Hysteresis temperature values can apply to several limits. For most chips it applies to the max temperature, but for some it applies to the critical temperature. One chip (LM80) even has the two of them, so its sysfs interface doesn't comply with the defined standard.
• Alarms are represented as bit vectors. The meaning of each bit is chipset-dependent.

There is a second, less important problem. The file naming scheme was not chosen in accordance with what was done in previous kernels and libsensors, the sensors information access library. When we would refer to "temp1_max" before, the corresponding sysfs file is now "temp_max1". I say it is less important because it doesn't affect the chip-independency of the current interface. Still it requires additional code in libsensors (not much, admittedly), makes "ls" group the sysfs files in an irrelevant order, and is likely to make sysfs the future interface "discovery" code a bit more complex.

THE REASONS FOR CHANGE

Although the problem detailed above probably makes it clear why I want to alter the sysfs interface, I think some of you will be interested in more details on why I want them to occur, and more specifically why I want the changes to be made as soon as possible, although in a stable series.

My ultimate goal is to make it possible to write a completely chip-independent sensors library. The way things were done so far, adding support for a new hardware monitoring chip always requires to update the library and user-space programs (such as "sensors"). At the moment, 75% of the libsensors code is chipset specific, and almost 90% of the "sensors" user-space tool code is. This demonstrates how much code we could save, would we have a truly chip-independant sysfs interface. This also means that the library and user-space tools wouldn't have to be updated with each new chip driver (with the exception of the sensors.conf configuration file).

Since only 30% of the sensors chip drivers have been ported to Linux 2.6 yet, I think that if we are to change the interface, we better do it before porting the other 70% (well, obviously less since some of them will probably never be ported).

Also, were we to wait for Linux 2.7 before starting these changes, as some of you might suggest, this would mean that the new sensors library would be either usable only there (which basically means maybe 3 years or so before a monitoring application can rely on this library only), or it would have to support both the 2.6 and 2.7 sysfs standards. This is likely to make the code be much more complex, for basically no benefit.

In fact, I don't feel bad to change the standard now because it is hardly established yet. It isn't as if I were planing to change a long-established standard. With each new Linux 2.6 release since .0, we have been updating the standard and modified the drivers to comply with it, so that we had to release a new libsensors each time. One more change won't hurt much more, especially if this is this ensures this is the last one. I also insist on the fact that, since the sysfs interface is not chip-independent yet, applications are discouraged to access the sysfs interface directly, and should rely on libsensors instead. Since we keep the library in sync with the Linux 2.6 kernel (and will continue to do so), such applications are not affected by the changes (except that libsensors has to be updated), even the relatively important one I propose to make here.

THE PLAN

I propose a three-step plan.

1. Change the base scheme (e.g. temp_min1 -> temp1_min). This is the more important change (in the sense it affects all drivers and the libsensors library) and correspond to the second problem listed above.

   Benefits of the new naming scheme are:

   • Back to the original (libsensors) names, more natural to the sensors developers. Allows some simplifications in the current libsensors. Represents the relations between the different values in a more understandable way IMHO. References in the sensors.conf configuration file will match the file names.
   • Better ordering by "ls" and the such. Groups name by sensor instead of limits.
   • Probably easier "browsing" code later. For example, if you want to treat all files related to the first temperature sensor, looking for files beginning with "temp1_" should be easier than looking for files beginning with "temp_" and ending with "1" not preceeded by another digit. Likewise, getting the sub-item is easy in the "temp1_min" scheme (access string+strlen(prefix)) and you're done) while it looks more tricky with the "temp_min1" scheme. Of course, all this depends on how you will be handling the whole thing in your application, and which language you use, but to make it short, it really looks like the "temp1_min" scheme can be handled with basic C functions while the "temp_min1" scheme would be more easily handled with regular expressions.

2. Change the hysteresis names (temp1_hyst -> temp1_max_hyst). Only some drivers are impacted. Changes required to the library as well.

   Benefits:

   Increase the chip-independency by making it clear which limits the hysteresis do apply to. The lm80 driver *needs* this since it has two hysteresis values per temperature channel.

3. Add splitted alarm files. This doesn't break the interface (these are new files), but on the other hand needs that we think about it a bit more so that our choices are extendable and correct for all known drivers.

I would like to make steps 1 and 2 occur as soon as in Linux 2.6.4 (i.e. immediately), because they break the interface. Step 3 is required for a fully chip-independent interface (and future library), but doesn't break the interface and cannot be used by the current libsensors, so we can give ourself some delay.

I will of course make the changes everywhere at once (all drivers and library, documentation) so that applications using libsensors won't be affected at all.

For those interested, the original thread on the lm_sensors mailing-list is available here:
http://archives.andrew.net.au/lm-sensors/msg06206.html
And an older one on the same topic:
http://archives.andrew.net.au/lm-sensors/msg05028.html

I have no idea if it is still in development. To be clear, it doesn't allow you to simply replace a kernel, but to add a second one, and possibly, to start transferring over tasks to it.

You can do much the same thing with user mode linux also. Again, not a kernel replacement in that sense, but something similar, sort of.

I've released the 018 version of udev. It can be found at:
kernel.org/pub/linux/utils/kernel/hotplug/udev-018.tar.gz

rpms built against Red Hat FC2-test1 are available at:
kernel.org/pub/linux/utils/kernel/hotplug/udev-018-1.i386.rpm
with the source rpm at:
kernel.org/pub/linux/utils/kernel/hotplug/udev-018-1.src.rpm

udev allows users to have a dynamic /dev and provides the ability to have persistent device names. It uses sysfs and /sbin/hotplug and runs entirely in userspace. It requires a 2.6 kernel with CONFIG_HOTPLUG enabled to run. Please see the udev FAQ for any questions about it:
kernel.org/pub/linux/utils/kernel/hotplug/udev-FAQ

For any udev vs devfs questions anyone might have, please see:
kernel.org/pub/linux/utils/kernel/hotplug/udev_vs_devfs

Major changes from the 017 version:

• We now handle the ability to generate all partitions for a device to allow removable devices to work in a sane manner. This has been requested by a lot of people.
• there's a new %s{} modifier available for people to use
• the SYSFS_ style rule has changed to SYSFS{}. The old style is still supported for now, but you have been warned
• %c1 style modifiers has been changed to %c{1}. Again, the old style format still works.
• scsi_id is built by default now in the rpm and is available in the pre-built rpm package. This should get it a wider range of testing.
• lots of bug fixes and other cleanups.

In all, there is nothing hugely major in this release, but any current users of udev will want this version for all of the bugfixes if for nothing else.

Thanks to everyone who has send me patches for this release, a full list of everyone, and their changes is below.

udev development is done in a BitKeeper repository located at:
bk://linuxusb.bkbits.net/udev

Daily snapshots of udev from the BitKeeper tree can be found at:
http://www.codemonkey.org.uk/projects/bitkeeper/udev/
If anyone ever wants a tarball of the current bk tree, just email me.

As of this release, I've been running with udev managing my /dev for me exclusively on my main email and development machine. This is a major milestone for udev and it proves that it is a viable solution.

I'd like to say thanks to everyone who has made this possible to do:

• Pat Mochel for creating sysfs and listening to my crazy ideas about how we could create a userspace devfs all those years ago.
• Dan Stekloff for prodding me to actually implement this crazy idea and who came up with a solid initial design document, without which this project would have never left the dream stage.
• Kay Sievers for almost single-handedly taking over the whole udev TODO list and converting udev from a small "proof of concept" toy into a powerful and useful tool.
• Pat Mansfield for creating the scsi_id tool and enabling udev to call external programs, which instantly made udev a real tool in the fine Unix tradition.
• All of the Gentoo developers who integrated udev into their distro and showed me that it can actually run a machine.
• everyone who has sent in udev patches, bug reports, and feature requests. Without these udev would only work for me, and not the rest of the world. A community is very important.
• the distros for picking up udev without me having to beg :)
• everyone else who I know I've forgotten...

udev development isn't done, but for anyone who has not checked it out yet, I suggest you do so. I'll post a small HOWTO that shows how to configure udev to manage your /dev without any problems or legacy issues (the 2.4 kernel will still work just fine on the same box.)

If anyone has any suggestions for things that are lacking in udev, please let me and the linux-hotplug-devel mailing list. This especially goes for any distro developers who are trying to integrate it into their systems.

I'm running udev for some time now on my main development machine and it works (almost) great. Thanks Greg and all the others who made it possible!

But I've a little issue left. My parallel port doesn't show up in /udev. I guess it's because of missing sysfs support? I'm running linux-2.6.3.

I did not find an entry for the parallel port in sysfs. If I create the device node manually I can access lp.

Here's a small document that I've added to the udev tarball that explains how I managed to get udev to manage my /dev tree on a Red Hat Fedora based machine. All you Gentoo developers can just laugh as it's already integrated into your distro.

Any users of other distros, feel free to send me updates to this to show how to do it for yours. Any distro maintainers, feel free to just integrate udev into your system so this kind of tweaking isn't necessary :)

HOWTO use udev to manage /dev

This document describes one way to get udev working on a Fedora-development machine to manage /dev. This procedure may be used to get udev to manage /dev on other distros, if you modify some of the steps.

This will only work if you use a 2.6 based kernel, preferably the most recent one. This does not prevent your machine from using a 2.4 kernel, if you boot into one, udev will not run and your old /dev will be present with no changes needed.

NOTE NOTE NOTE NOTE NOTE NOTE NOTE

This is completely unsupported. Attempting to do this may cause your machine to be unable to boot properly. USE AT YOUR OWN RISK. Always have a rescue disk or CD handy to allow you to fix up any errors that may occur.

NOTE NOTE NOTE NOTE NOTE NOTE NOTE

• Build and install udev as specified in the README that comes with udev. I recommend using the following build options to get the smallest possible binaries:

  make USE_KLIBC=true USE_LOG=false DEBUG=false

• disable udev from the boot process by running:

  chkconfig udev off

  or

  chkconfig --del udev

  as root.

• place the start_udev script somewhere that is accessible by your initscripts. I placed it into /etc/rc.d with the following command:

  copy extras/start_udev /etc/rc.d/

• modify the rc.sysinit script to call the start_udev script as one of the first things that it does, but after /proc and /sys are mounted. I did this with the latest Fedora startup scripts with the patch at the end of this file.

• make sure the /etc/udev/udev.conf file lists the udev_root as /dev. It should contain the following line in order to work properly.

  udev_root="/dev/index.html"

• reboot into a 2.6 kernel and watch udev create all of the initial device nodes in /dev

If anyone has any problems with this, please let me, and the linux-hotplug-devel@lists.sourceforge.net mailing list know.

A big thanks go out to the Gentoo developers for showing me that this is possible to do.

ftp://ftp.kernel.org/pub/linux/kernel/people/akpm/patches/2.6/2.6.3/2.6.3-mm2/

• More parport fixes/cleanups from Al Viro
• Various patches were folded together to make them a bit more logical. -mm has less than 200 patches for the first time in a long time. Things are actually slowing down.
• Added an absolutely gargantuan MIPS update.
• More CPU scheduler changes.
• 2.6.3-mm2 builds and runs on x86 and ia64, and compiles on x86_64 and ppc64.

Here's another device mapper update, some of these are quite big patches, so I'll run through the list:

endio method
We've been using this code for many months (years?). Needed for the more complicated targets.

Remove the version-1 ioctl interface
This didn't get in last time I submitted it. Leave it out if you still disagree.

Audit for list_for_each_*entry*
Trivial, please merge

Queue limits
Please merge.

List targets ioctl
Adds a command that lets tools query the kernel to see what targets/versions are available.

Multipath target
People really want this, so I'm probably pushing it sooner than I'd like. It would be good if it got a wider audience in the -mm tree or as an experimental target in vanilla.

The Linux Stability page continues to post test result data for the post 2.6.0 kernels. http://www.osdl.org/projects/26lnxstblztn/results/

Below lists some recent changes to the page (in case you haven't visited in a while).

1. 2.6.x kernels tested upon release:
   -mm
   -rc
   -bk
2. Detaild Test Result Links show links to continual updated results to re-aim-7, tiobench, and iozone tests run in STP (1-way, 2-way, 4-way and 8-way, as appopriate)
3. New links to database performance reports in the Database Performance Reports section (replaces old database section)
4. New section added to link to various Lilnux kernel test report

If anyone knows of other useful test result information or information providing a current state of the Linux kernel that can be linked from this page, let me know and I'll add the link.

Attached to this email is a gzipped patch for the next revision of the Intel Software RAID driver "iswraid" (which is an ataraid subdriver) for the 2.4.x series kernels. The patch is taken against kernel 2.4.25 with Jeff Garzik's 2.4.25-libata1.patch already applied. (This driver is useless without libata.) My apologies about the size and the binary nature of this attachment.

Also attached is a very small patch that enables (partial?) Intel's ICH6R chipset's SATA support in the ata_piix driver. I am sure that Jeff can enable this properly (or perhaps already has), but for those who find the chipset unidentified, this may be a quick fix. ICH6R is then usable in both the IDE and RAID modes.

The driver is intended to be used with Intel's ICH5R/ICH6R chipsets. Configuration of RAID volumes is done from the Option ROM. This "iswraid" driver differs from the other ataraid sub-drivers in that it operates on SCSI block devices rather than the ATA/IDE ones. The "ata_piix" driver (part of libata1) presents the SATA disks connected to ICH5R/ICH6R chipsets as SCSI disks. The "iswraid" driver is considered experimental, use at your own risk.

Thanks to Boji Kannanthanam for answering my questions and to Jeff Garzik for providing a great low level driver to utilize.

Here's a list of pending issues with fbdev (either upstream or in the fbdev bk treee), I figured posting it here may help getting more people on those issues as my time is sparse and I suppose James too.

• First one, I will deal with, just writing it for completeness: When switching from KD_GRAPHICS to KD_TEXT (either same console changing mode or switching to a different console), we need a callback so the fbdev has a chance to restore the accel engine setting (or even the whole mode, to be safe).
• Memory corruption problems. There is still at least one identified when using stty. Just use crazy values, like flipping rows & cols (like stty rows 132 cols 30) and usually, you'll see the box blow up very soon with heavy memory corruption.
• mach64 lockups on LT-G (I'll try on an LT-Pro soon) plus other mach64 bugs in the new version in the bk fbdev, I'll have a patch for some of the problems, but I didn't find a good explanation for the accel lockups yet
• Logo problems. When booting with a logo, then going to getty, the logo doesn't get erased until we actually switch to another console (or reset the console). At this point, using things like vi & scrolling up doesn't work properly. Actually, last time I tried, I had to switch back & forth twice before my console that had the logo got fully working with vi.
• Back buffer problem: maybe related to the logo ? After boot, doing shift-pageup to go back to the boot message, usually you get crap displayed at various places.
• On x86, various junk displayed when the fbdev takes over. Reported by radeonfb users, I couldn't test myself, I don't have an x86 with radeon at hand for the moment. Apparently, the takeover from vgacon doesn't properly "convert" the previous VGA text buffer content
• stty & mode picking. Currently, fbcon_resize() (called when stty is used to resize the console) will hack a "var" strcture by just putting new width/height in it and pass that to set_var. The way the various drivers react to that mostly broken "var" structure is rather random. We need to explicitely differenciate between a mode that is "complete" (like what fbset or X passes down the fbdev) or a mode that is just width/height and eventually a hint of frequency, like what fbcon passes in this case. I added FB_ACTIVATE_FIND for that purpose, but that needs better driver support to "pick" up a proper mode. The algorithm for that isn't trivial. Could be moved to common code.
• fbset doesn't resize the console. I consider that a regression from 2.4. I have some code based on the notification mecanism to address that, but it tends to trigger the same memory corruption problem as reported with stty & bogus coordinates. There is something hairy going on with console resizes. That code is a bit foreign to me though.

Ok, that's all that comes to my mind right now, help is welcome :)

The patch '[Bluetooth] Initial sysfs and device class support', is the usual misguided sysfs conversion attempt levaing oopsable races, so please back it out until the bluetooth folks have understood and cared for the lifetime rule issues of using the driver model.

Guys, if you do use the driver model you have to adhere to it's life time rules. And that's most importantely you _must_ free the device only in it's ->release routine.

And no, the code doesn't warn about the lack of a release method only that everyone adds an empty one, sigh..

Release 1.0.0 of Libsysfs is now available as part of the sysfsutils package at:

http://linux-diag.sourceforge.net

Libsysfs provides APIs to access the sysfs filesystem for device information.

Major changes in this release include:

• A comprehensive testsuite to exercise APIs exported by Libsysfs. An easy to use config file is provided so users can modify appropriately for the system under test.
• Improvements to "refresh" functions (udev requirement).
• dlist elements now stored in sorted order.
• A reworked systool for better output.
• A number of miscellaneous bugfixes.

udev has been shipping with more or less the latest code for quite a few releases now.

Thanks to everyone who have been using the library and providing valuable feedback.

I would like to introduce a new device driver for the I2C based Videotext/Teletext decoder SAA5246A from Philips. The interface is identical to the existing driver for the SAA5249 chip. User programs can access these devices via /dev/vtx0 ... /dev/vtx31.

Templete was drivers/media/video/saa5249.c, some of the SAA5246A functionality was taken from the original driver from Martin Buck (http://home.pages.de/~videotext/). In order to make the driver sources easier to understand I added lots of symbolic constants related to the data sheet of the SAA5246A from Philips.

The driver is available as a patch to the linux kernel version 2.6.3 from http://www.michaelgeng.de/linux/saa5246a-2.6.3.patch

I know that today's TV cards do no more have such a teletext processor chip on board, they use the CPU for decoding videotext pages. This is much faster if you want to extract all the pages from a TV station for reading like a journal. I also wanted to change to the vbi interface but after some trial I went back to decoding via SAA5246A because there were more errors in the pages decoded with the vbi interface than with the vtx interface using a SAA5246A. I don't know why. And have you tried to capture a special subpage from a page that contains 10 suppages or more? When it is finally transmitted after 10 Minutes you will have to wait for another 10 Minutes if your CPU missed it. According to my experience a teletext processor will more probably capture the page.

Of courde I have tested the driver both as a module and compiled into the kernel. It is licensed under the GPL.

new reiser4 snapshot against 2.6.3 kernel is available at

http://www.namesys.com/snapshots/2004.02.25

It contains bug fixes and stability improvements. On-disk format is compatible with the previous snapshot. Also,

• raid0 preliminary support, and
• improved loop back support

were added.

In "standard" configuration reiser4 is stable except for a multiple umount/mount issue we are still analyzing. See the READ.ME file for not-well-supported features and options.

Hopefully our next snapshot will be ready for inclusion, and hopefully it is just a few days away.

Here goes -pre1.

It contains a big SCTP merge (to match 2.6 API), networking updates, network driver updates (including the addition of nVidia Force driver).

Also includes ACPI upstream merge, amongst others.

We've moved the openlogging server back to San Francisco and updated the DNS entries. BK users should be just fine but if anything appears strange let us know.

One side effect of this is we're making the web pages work again, should be going today or tomorrow.

Ok, as usual, there was a lot of stuff for the -rc1, but as seems to be more and more true it is mainly in the "periphery".

This is a big patch, but the bulk of it is a long-overdue MIPS update, and the (working) HFS/HFS+ filesystem update (total rewrite), and ISDN updates. And some large s390 driver updates too, for that matter.

So please keep bigger updates to yourself, and let's calm this down.

ftp://ftp.kernel.org/pub/linux/kernel/people/akpm/patches/2.6/2.6.4-rc1/2.6.4-rc1-mm1/

• Added the POSIX message queue implementation. We're still stitching together a decent description of all of this. Reference information is at

  http://www.mat.uni.torun.pl/~wrona/posix_ipc/

  and

  http://www.opengroup.org/onlinepubs/007904975/basedefs/mqueue.h.html

• A fair amount of work against the page reclaim code. Mainly reorganisation and simplification of various little glitches. This means that a few of the optimisations which were in 2.6.3-mm4 were broken and were dropped. But this is a better basis upon which to reintroduce them.

  Performance, however, seems similar to 2.6.3-mm4 in a few tests. Inter-zone balancing is much better than 2.6.3 but still could be improved a little.

  Slab reclaim balancing is improved, however with some (artificial) workloads slab is still being a problem because of tremendous internal fragmentation problems: 6% occupancy of the pages which are allocated to dcache, for example. More work is needed to account for this.

  I tested the swapout code with 7.2G of tmpfs pagecache on a 7G machine. The rotate_reclaimable_page() logic seems to work fine here - only 200M of memory was added to swapcache and was swapped out.

• Plus the usual various little fixes and cleanups.

• The contents of the broken-out/ directory are now available in the 2.6.4-rc1-mm1-broken-out.tar.gz file. This includes the `series' file which describes the patching order.

  People who are using patch-scripts can recreate the patching machinery by doing:

          cd /usr/src/linux
          tar xfz ~/2.6.4-rc1-mm1-broken-out.tar.gz
          mv broken-out/*.patch patches
          mv broken-out/series .
          for i in $(cat-series series)
          do
                  pcpatch $i
          done
          rmdir broken-out

But the glibc headers don't describe 2.6, do they? Don't work with 2.6? We went around with this for cdrecord unless I misread which headers are involved.

It's not reasonable to expect people to rebuild glibc for each kernel, even if it is "the right thing to do" in some purist sense. It would be better to have something like user header in the kernel for just the interface, and have the kernel include start by pulling in the user header and then adding things the user doesn't need.

Changes between kernel series are always unpleasant during the time when people have to boot back and forth, we should think about a better way to do this for 2.7.

We have two sets of kgdb patches as of now: [core-lite, i386-lite, 8250] and [core, i386, ppc, x86_64, eth]. First set of kgdb patches (lite) is fairly clean. Let's consider it to be a candicate for submission to mainline kernel.

I am freezing the lite patches wrt. feature updates. Only bug-fixes and code cleanups will be allowed in lite patches. You can make any feature enhancements to second set of patches.

I propose following schedule for pushing kgdb lite into mainline kernel: Take 1: 8th , Take 2: 15th, Take 3: 22nd, Take 4:29th. I'll download the kernel snapshot (http://www.kernel.org/pub/linux/kernel/v2.6/snapshots/) on these dates and submit a single patch for possible acceptance into mainline kenrel and feedback from community. Hopefully we'll succeed by end of this month.

Please checkin any fixes or cleanups by end of this week. I plan to add some documentation to core-lite.patch this week (will send it for review in a separate email)

ppc.patch removes arch/ppc/kernel/ppc-stub.c and adds a new file kgdb.c I think that has a greater rejection chance.

Let's not change the direction now. Some time ago there was another view that x86_64 would be easier. We have already had sufficient headache because of split -lite -heavy patches. Let's try to finish that asap.