New mailing list for 2.6 Medley RAID (Silicon Image 3112 etc.) BIOS RAID development
Hi,
I've started a mailing list for those interested in the development of a Linux driver supporting Medley RAID (the BIOS/software RAID used by the Silicon Image 3112 and some CMD chipsets).
My main goal is that I want autodetection for these drives to work in 2.6 (like the ataraid framework in 2.4), but also to leverage the code in the existing MD and DM drivers as far as possilble. This currently works in the current 2.4 kernels, thanks to Arjan van de Ven's ataraid framework, which my Medley driver for 2.4 uses.
The reason I insist on autodetection is that I think it's important that if the BIOS will reckognise the drive without additional intervention, so will Linux. This will make the entry route for newbies much simpler.
If you are interested in taking part in the discussions/development of the 2.6 driver for Medley RAID (and other BIOS "wholedisk" RAID arrays, such as Highpoint), I would ask you to join this list.
On the new list we'll also welcome support questions for the 2.4 Medley driver and other related issues. I expect this to be a low volume list.
You can find more information about how to subscribe at
http://lists.infowares.com/cgi-bin/listinfo/medley
More informateion about the 2.4 driver, and the patch source, is at
http://www.infowares.com/linux/
// Thomas
Quoted 5 lines · On Monday February 9, [email protected] wrote:
Hi,
I've started a mailing list for those interested in the development of a Linux driver supporting Medley RAID (the BIOS/software RAID used by the Silicon Image 3112 and some CMD chipsets).
Why a new list? Why not just start a discussion on linux-raid (or linux-kernel) and if there is enough interest/traffic to annoy the spectators, then fork off a new list.
NeilBrown
Quoted 3 lines · Neil Brown
Why a new list? Why not just start a discussion on linux-raid (or linux-kernel) and if there is enough interest/traffic to annoy the spectators, then fork off a new list.
There already is, in my personal mailbox :) Hence a list for this topic.
// Thomas
Quoted 4 lines · Thomas Horsten
The reason I insist on autodetection is that I think it's important that if the BIOS will reckognise the drive without additional intervention, so will Linux. This will make the entry route for newbies much simpler.
do you call running devicemapper tools from the initrd autodetection ?
Quoted 5 lines · Arjan van de Ven
Quoted 3 lines
The reason I insist on autodetection is that I think it's important that if the BIOS will reckognise the drive without additional intervention, so will Linux. This will make the entry route for newbies much simpler.
do you call running devicemapper tools from the initrd autodetection ?
Probably not. I am working with several ways of doing it, and that's why I wanted to have a discussion about this.
Ideally I'd want something like the MD autodetect code, so that the whole thing can be set up by the kernel at boot-time if the necessary drivers are compiled in (by reading the Medley superblock the same way it's done for 0xfe partitions). And if the drivers are modules, then use the drive mapper tools from userspace to set it up.
This would be the most flexible solution, since it'd allow you to boot without an initrd if you wanted that, or to use the initrd and map the drives manually if you preferred that solution.
Having autodetection at kernel level would make it possible to boot from a kernel on a floppy disk without initrd support, and in general make a system easier to set up.
But the reason I wanted this discussion is to figure out the best way to go about it, and if there are some good arguments against autodetecting in the kernel I'll listen to them.
// Thomas
Quoted 8 lines · Thomas Horsten
Quoted 5 lines · Arjan van de Ven
Quoted 3 lines
The reason I insist on autodetection is that I think it's important that if the BIOS will reckognise the drive without additional intervention, so will Linux. This will make the entry route for newbies much simpler.
do you call running devicemapper tools from the initrd autodetection ?
Quoted 7 lines
Probably not. I am working with several ways of doing it, and that's why I wanted to have a discussion about this.
Ideally I'd want something like the MD autodetect code, so that the whole thing can be set up by the kernel at boot-time if the necessary drivers are compiled in (by reading the Medley superblock the same way it's done for
0xfepartitions).
I (and I suspect a lot of other folks) rather get rid of such autodetect and move it to userspace. Either via initrd or initramfs.
Quoted 3 lines
Having autodetection at kernel level would make it possible to boot from a kernel on a floppy disk without initrd support, and in general make a system easier to set up.
initrd/initramfs is increasingly becoming mandatory sort of, and it's actually easy if not even default to set up. (Eg on Fedora / Red Hat even just typing make install will auto-create this for you)
Quoted 3 lines
But the reason I wanted this discussion is to figure out the best way to go about it, and if there are some good arguments against autodetecting in the kernel I'll listen to them.
It doesn't really belong there.
Quoted 8 lines · Arjan van de Ven
Quoted 4 lines · Thomas Horsten
Ideally I'd want something like the MD autodetect code, so that the whole thing can be set up by the kernel at boot-time if the necessary drivers are compiled in (by reading the Medley superblock the same way it's done for
0xfepartitions).I (and I suspect a lot of other folks) rather get rid of such autodetect and move it to userspace. Either via initrd or initramfs.
The question is, where do we draw the line between kernel and userspace setup of devices. For example, why is the frame buffer device detected in the kernel, it could just as well be done in the initrd.
My gut feeling is that if it is provided by the BIOS and reliable autodetection is possible, it should be autodetected. Why require the user to discover and supply information that the kernel could easily and reliably find out by itself? Besides, if there is no autodetection, the drive will come up with an invalid partition table (since the partition tables for these arrays are stored as the first block of the first disk in the array). If there is an extended partition, the kernel may try to read past the end of the disk, causing a lot of spurious error messages that confuse the user.
If the user tries to mount any of these partitions, or edit the partition table, they will probably corrupt the disk.
What I have in mind currently is a solution that uses the md and dm frameworks. Basically an option that adds support for reading and parsing the Medley superblock, and create an md raiddev based on it (using the standard RAID0/RAID1 personalities). Then a dm drive needs to be registered, to support the partition table on the whole disk array.
For the autodetection to work, when compiled into the kernel I would put a call into gendisk.c, just before it calls add_disk to parse the partition table. That way, if the Medley superblock is detected it will not try to parse the partition table.
The only thing I don't like about this solution is that we are relying on both the md and dm framework. I can't see an easy way around this, since we need to parse the partition table on the RAID. If we just create a separate md device for each partition, the user won't be able to change the partition table.
Quoted 7 lines
Quoted 3 lines
Having autodetection at kernel level would make it possible to boot from a kernel on a floppy disk without initrd support, and in general make a system easier to set up.
initrd/initramfs is increasingly becoming mandatory sort of, and it's actually easy if not even default to set up. (Eg on Fedora / Red Hat even just typing make install will auto-create this for you)
It's sort of mandatory for generic systems like distro installer disks, livecd's etc. But I've never needed to use it on any of my own systems after I compile a cutom kernel. I guess a lot of people are like me, and prefer to keep it simple with the essential drivers to get my specific system running compiled in, and the rest as modules.
Quoted 5 lines
Quoted 3 lines
But the reason I wanted this discussion is to figure out the best way to go about it, and if there are some good arguments against autodetecting in the kernel I'll listen to them.
It doesn't really belong there.
But where to draw the line?
I think if it can be implemented cleanly (as cleanly as the current md detection at least), and if it can be nearly 100% reliable, it doesn't add much complexity to the kernel and removes a lot from userspace.
// Thomas
Quoted 4 lines · Thomas Horsten
My gut feeling is that if it is provided by the BIOS and reliable autodetection is possible, it should be autodetected. Why require the user to discover and supply information that the kernel could easily and reliably find out by itself? Besides, if there is no autodetection, the
The autodetection will occur, reliably and without additional user information, from initramfs.
As Arjan said, we are moving this type of stuff out of the kernel.
Jeff
Quoted 6 lines · Jeff Garzik
Quoted 4 lines · Thomas Horsten
My gut feeling is that if it is provided by the BIOS and reliable autodetection is possible, it should be autodetected. Why require the user to discover and supply information that the kernel could easily and reliably find out by itself? Besides, if there is no autodetection, the
Quoted 6 lines
The autodetection will occur, reliably and without additional user information, from initramfs.
As Arjan said, we are moving this type of stuff out of the kernel.
Jeff
Is there any existing code using initramfs? So far all i have seen is very incomplete. I would love to move to initramfs and get rid of mkinitrd, but too many pieces seem to be still missing...
Regards,
Dominik Kubla
Signature · 2 lines
The more cordial the buyer's secretary, the greater the odds that the competition already has the order.
Quoted 1 line · Arjan van de Ven
It doesn't really belong there.
I didn't really know about how initramfs was supposed to work, now I've read up on it and I agree that it is the best solution.
I'll do the ataraid detection userside, but I am still not sure whether some additions to the device mapper might be needed.
For Medley RAID it's pretty straightforward to map striped sets, but how to deal with ataraids like Highpoint where they use non-standard zone models for the RAID0 sets?
// Thomas
Quoted 5 lines · On Monday February 9, [email protected] wrote:
The only thing I don't like about this solution is that we are relying on both the md and dm framework. I can't see an easy way around this, since we need to parse the partition table on the RAID. If we just create a separate md device for each partition, the user won't be able to change the partition table.
2.6.3-rc1-mm1 supports partitioning of md arrays though a new major number. I'm hopeful that this will get into 2.6.4 at least.
NeilBrown
Quoted 4 lines · On Monday February 9, [email protected] wrote:
For Medley RAID it's pretty straightforward to map striped sets, but how to deal with ataraids like Highpoint where they use non-standard zone models for the RAID0 sets?
Can you point me at documentation of code that gives the details of this "non-standard zone model" ??
NeilBrown