IBMs LVM ?
Hi!
I heard rumors about IBM porting their LVM code from AIX to Linux. I guess the current LVM code is not from IBM ?
Will it be replaced with the one from IBM ?
Comments ?
Signature · 4 lines
David Balazic -------------- "Be excellent to each other." - Bill S. Preston, Esq., & "Ted" Theodore Logan - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Quoted 3 lines · In article <[email protected]> you wrote:
Hi!
I heard rumors about IBM porting their LVM code from AIX to Linux.
IBM has an OpenSource volume manager called evms, and although it does support AIX Volumes is has it's root in IBM's OS/2 volume management system. (I believe someone at IBM thinks of PCs when hearing Linux so all their ports start from OS/2...)
Quoted 1 line
I guess the current LVM code is not from IBM ?
The current code is from Heinz Maulshagen and now matained by Sistina, the company he works for.
Quoted 1 line
Will it be replaced with the one from IBM ?
ALthough the current LVM has its's issues I hope not so - the current EVMS is the best example on hhow to not write kernel subsystems.
Christoph
Signature · 1 line
Whip me. Beat me. Make me maintain AIX.
Quoted 6 lines · Christoph Hellwig
Quoted 1 line
I heard rumors about IBM porting their LVM code from AIX to Linux.
IBM has an OpenSource volume manager called evms, and although it does support AIX Volumes is has it's root in IBM's OS/2 volume management system. (I believe someone at IBM thinks of PCs when hearing Linux so all their ports start from OS/2...)
Actually, this isn't quite correct. Yes, the EVMS folks started with the LVM design from OS/2 (which was based on AIX LVM, but not compatible). However, IBM DID GPL the AIX 4.3 LVM code, and there IS support for AIX VGs under EVMS.
Quoted 4 lines
Quoted 1 line
Will it be replaced with the one from IBM ?
ALthough the current LVM has its's issues I hope not so - the current EVMS is the best example on hhow to not write kernel subsystems.
Do you have a specific complaint? The current LVM kernel code isn't all that robust either (and the user-space code is not very pretty).
I DO think that EVMS will replace the current LVM code in the kernel. Why?
- It already has 100% compatibility with the current LVM on-disk layout.
- It already has 100% compatibility with AIX LVM code.
- It is already possible to do LVM autoconfig of volumes within the kernel without needing an initrd to activate the volumes.
- It removes knowlege of partitions out of the disk drivers and makes them a simple form of LV.
- It is already possible to use partial Linux-LVM volumes (i.e. use LVs that are 100% available in a VG, even if a disk is bad/missing). You can't do that with the current LVM (sometimes you can't even do it if ALL of the disks are there).
- The AIX LVM on-disk metadata layout is FAR more robust than the current LVM metadata layout. While this is supposed to be fixed for Linux-LVM at some point, the previous Linux-LVM changes have been TERRIBLE with respect to compatibility, while EVMS is DESIGNED to support multiple on-disk metadata layouts.
- It is possible (not done yet) to add things like NT Logical Disk Manager (NT LVM-stuff) into EVMS.
- It is possible (not done yet) to add HPUX LVM/veritas/OSF volume stuff.
- It is possible to merge MD RAID support into EVMS also, to support all of the different RAID layouts with common code*.
All in all, instead of having things like striping/RAID-1/concatenation/ RAID-5 reimplemented in each subsystem (e.g. LDM/MD/veritas/etc) we can take the current MD code and make it a layer in EVMS which can handle all of these cases.
(*) While the exact on-disk layout of each RAID implementation may vary slightly (block sizes, stripe widths, etc), I'm guessing that the majority of it is common enough that we can re-use the existing MD code (parity calculation, resync, etc) to handle most kinds of RAID-0/1/5 setups.
Cheers, Andreas
Signature · 3 lines
Andreas Dilger \ "If a man ate a pound of pasta and a pound of antipasto,
\ would they cancel out, leaving him still hungry?"
http://www-mddsp.enel.ucalgary.ca/People/adilger/ -- DogbertQuoted 5 lines · Andreas Dilger
Quoted 4 lines · Christoph Hellwig
Quoted 1 line
I heard rumors about IBM porting their LVM code from AIX to Linux.
IBM has an OpenSource volume manager called evms, and although it does support AIX Volumes is has it's root in IBM's OS/2
^^^^^^^^^^^^^^^^^^
Quoted 7 lines
Quoted 2 lines
volume management system. (I believe someone at IBM thinks of PCs when hearing Linux so all their ports start from OS/2...)
Actually, this isn't quite correct. Yes, the EVMS folks started with the LVM design from OS/2 (which was based on AIX LVM, but not compatible). However, IBM DID GPL the AIX 4.3 LVM code, and there IS support for AIX VGs under EVMS.
Where is the AIX LVM code?
Quoted 7 lines
Quoted 4 lines
Quoted 1 line
Will it be replaced with the one from IBM ?
ALthough the current LVM has its's issues I hope not so - the current EVMS is the best example on hhow to not write kernel subsystems.
Do you have a specific complaint? The current LVM kernel code isn't all that robust either (and the user-space code is not very pretty).
Please take a barf-bag, look at the code, look at the unneded abstrations, look at the uneeded interfaces, look at the code-duplication, look at the over-engineering.
IF you don't have enough read the archives of the mailinglist about god-mode that allows even root to do anything and we come back.
Quoted 3 lines
I DO think that EVMS will replace the current LVM code in the kernel. Why?
- It already has 100% compatibility with the current LVM on-disk layout.
With _which_ LVM on-disk layout (OOPS, this was a complaint to Sistina :))
Quoted 1 line
- It already has 100% compatibility with AIX LVM code.
The code is there, yes.
Quoted 2 lines
- It is already possible to do LVM autoconfig of volumes within the kernel without needing an initrd to activate the volumes.
I had that for LVM 0.8 as well, that's a year ago now.
Quoted 2 lines
- It removes knowlege of partitions out of the disk drivers and makes them a simple form of LV.
No - it duplicates the partitions scanning code. It does not remove the simple partitioning code (which, BTW is not part of the drivers at all under Linux), makeing it's copy more complicated than the original version.
Quoted 4 lines
- It is already possible to use partial Linux-LVM volumes (i.e. use LVs that are 100% available in a VG, even if a disk is bad/missing). You can't do that with the current LVM (sometimes you can't even do it if ALL of the disks are there).
Yupp, nice thing.
Quoted 5 lines
- The AIX LVM on-disk metadata layout is FAR more robust than the current LVM metadata layout. While this is supposed to be fixed for Linux-LVM at some point, the previous Linux-LVM changes have been TERRIBLE with respect to compatibility, while EVMS is DESIGNED to support multiple on-disk metadata layouts.
Yes, because it is a Meta-LVM, not an actualy inplementation. I _really_ want something like that in 2.5 - but not this horrible IBM implementation.
Quoted 3 lines
- It is possible (not done yet) to add things like NT Logical Disk Manager (NT LVM-stuff) into EVMS.
- It is possible (not done yet) to add HPUX LVM/veritas/OSF volume stuff.
That's the same argument as above.
Quoted 8 lines
- It is possible to merge MD RAID support into EVMS also, to support all of the different RAID layouts with common code*.
All in all, instead of having things like striping/RAID-1/concatenation/ RAID-5 reimplemented in each subsystem (e.g. LDM/MD/veritas/etc) we can take the current MD code and make it a layer in EVMS which can handle all of these cases.
Quoted 4 lines
(*) While the exact on-disk layout of each RAID implementation may vary slightly (block sizes, stripe widths, etc), I'm guessing that the majority of it is common enough that we can re-use the existing MD code (parity calculation, resync, etc) to handle most kinds of RAID-0/1/5 setups.
Nope - I'm currently working on implementing VxVM support, and I have to redo all the RAID stuff because it is so incomaptible.
If you have a design to share the RAID engine for very different layouts:
nice - but I don'T see the relation to EVMS.
Christoph
Signature · 1 line
Whip me. Beat me. Make me maintain AIX.
Quoted 7 lines · Christoph Hellwig
Quoted 2 lines · Andreas Dilger
Do you have a specific complaint? The current LVM kernel code isn't all that robust either (and the user-space code is not very pretty).
Please take a barf-bag, look at the code, look at the unneded abstrations, look at the uneeded interfaces, look at the code-duplication, look at the over-engineering.
Well, you don't get anything for free. You need to have some abstraction in order to handle multiple LVM-type systems in a useful way, I don't see any way around that. As for the kernel code, I don't see TOO much abstraction.
- You need code to read each separate on-disk format (partition headers, LVM metadata, etc)
- You need code to combine each of these into some useful "volume"
- You need code to do mapping from input volume/block to output disk/block
- You need code to be able to modify a volume metadata.
You need code to do this for each different metadata format or partition type. Luckily (hopefully?) most people won't have systems with more than a couple different metadata formats at a time. HOWEVER, you NEED to be able to migrate from one format to another, so you will have periods when you have extra modules loaded during a migration.
Consider EVMS to be the VFS for block devices. The VFS abstracts a lot of common code out of the filesystems so that each individual one can be less complex. Yes, there is still duplication within the filesystems, but it would be much moreso if the VFS didn't exist. There is a lot of abstraction in the VFS, but it is NOT bloat, it is code consolidation. As enough filesystems need a specific feature, it is abstracted out of the multiple filesystems and put into the VFS for all to use.
Quoted 2 lines
IF you don't have enough read the archives of the mailinglist about god-mode that allows even root to do anything and we come back.
It is not clear if you are you for or against "god-mode"? Would you like it so that it is easy to shoot themselves in the head (or it is so complex/manual that it is hard not to), or rather everyone in straight-jackets in rubber rooms (ala Windows)? Some safety is needed, but there are also reasons to bypass that safety. I don't know what the current state of "god-mode" is in the EVMS code (if it is there, needed, whatever, so it is a moot point).
Quoted 4 lines
Quoted 2 lines
I DO think that EVMS will replace the current LVM code in the kernel. Why?
- It already has 100% compatibility with the current LVM on-disk layout.With _which_ LVM on-disk layout (OOPS, this was a complaint to Sistina :))
Exactly. It was me and the EVMS LVM-emulation author that pointed out to the Sistina folks that there was NO need to have an incompatible change to the on-disk layout. In fact, the EVMS code can handle BOTH formats, and did so before LVM did.
Quoted 4 lines
Quoted 2 lines
- It is already possible to do LVM autoconfig of volumes within the kernel without needing an initrd to activate the volumes.
I had that for LVM 0.8 as well, that's a year ago now.
But, strangely it didn't make it into the LVM codebase. Why is that? I'm not blaming you, of course. EVMS is at least a SourceForge project.
Quoted 6 lines
Quoted 2 lines
- It removes knowlege of partitions out of the disk drivers and makes them a simple form of LV.
No - it duplicates the partitions scanning code. It does not remove the simple partitioning code (which, BTW is not part of the drivers at all under Linux), makeing it's copy more complicated than the original version.
Well, there are obvious reasons for that. Some people won't want to have EVMS in the kernel (clearly you are one of them), so their patch can't just rip out the existing partition scanning code. Maybe there will be a config option to not compile it later, or there will be an "EVMS-lite" which handles basic partitions, or whatever.
In any case, I doubt an EVMS patch would be accepted if it removed that code to start with. HOWEVER, since EVMS can handle ALL disk devices, even just regular partitions, at some point it COULD be possible to get rid of the mess of different major/minor numbers for different disk types (hdX, sdX, cciss, rd, ida, etc) and assign all of them to EVMS. Since EVMS only needs a minor number for the end-result volume (which may represent many individual disks, and doesn't need a minor for unused partitions), we would likely not have any shortage of block major numbers. Consider 10*hdX*256 + 8*sdX*256 + MD*256 + 8*CCISS*256 + 8*DAC960*256 + 8*IDA*256... and it is a long time until you have 10k+ VISIBLE volumes on a single system.
Quoted 9 lines
Quoted 5 lines
- The AIX LVM on-disk metadata layout is FAR more robust than the current LVM metadata layout. While this is supposed to be fixed for Linux-LVM at some point, the previous Linux-LVM changes have been TERRIBLE with respect to compatibility, while EVMS is DESIGNED to support multiple on-disk metadata layouts.
Yes, because it is a Meta-LVM, not an actualy inplementation. I _really_ want something like that in 2.5 - but not this horrible IBM implementation.
Maybe you can help them work on it? They have recently just redone a lot of the code, partly based on input from Andrew Clausen. I don't see any other similar projects out there, and I think it is a waste of effort to complain about work that is done rather than fix it. I tried for a long time (while it was still my job to do so) to fix the current Linux-LVM code, and I only had very minimal impact. Even so, Linux-LVM will never become a Meta-LVM without AT LEAST as much "mess" as EVMS, and probably will contain a lot more.
IBM is at least doing something about this by putting their code where their mouth is. Nothing is ever perfect. Consider MD RAID - it isn't even compatible between "stock" 2.2 and 2.2+Mingo-0.9RAID patches (which is one reason why Alan never accepted the 0.9 MD RAID into 2.2, despite the fact that ALL distros used the 0.9 MD RAID patches). With a framework like EVMS it is possible to have both of the on-disk formats supported, and migrate from one to the next (maybe even while the system is in use), and then unload the module that supports the old format.
Quoted 7 lines
Quoted 4 lines
(*) While the exact on-disk layout of each RAID implementation may vary slightly (block sizes, stripe widths, etc), I'm guessing that the majority of it is common enough that we can re-use the existing MD code (parity calculation, resync, etc) to handle most kinds of RAID-0/1/5 setups.
Nope - I'm currently working on implementing VxVM support, and I have to redo all the RAID stuff because it is so incomaptible.
The question is - will the VxVM support be yet ANOTHER separate code base in the kernel? You complain about code bloat in EVMS, but having LVM, MD, VxVM, NT LDM, etc. all separate is also code bloat, and each one needs separate admin tools. Remember - EVMS is also modular, so you only need the support for "LVM" types that you use. The real benefit, however, is that since it is integrated, you _could_ cleanly migrate from Linux-LVM and/or MD-RAID and/or NT-LDM to whatever you wanted (maybe even on a live system) if EVMS is handling all of the kernel interfaces for them. I don't think you could ever do that without full backup/restore on separate tools.
Quoted 2 lines
If you have a design to share the RAID engine for very different layouts:
nice - but I don'T see the relation to EVMS.
No, I can't say I do - it was purely speculation based on the concept that RAID-1 and RAID-5 are GENERALLY the same concept. Maybe the reason VxVM is so different is that it also incorporates "LVM" style function as well? Note that NT-LDM ALSO has RAID-0/1/5 modes (not yet supported by the LDM driver) which will eventually need support. Do you want the number of RAID implementations in the kernel to grow without bound, or rather try and consolidate them (where possible) into a generic interface like EVMS?
Again, if EVMS is the VFS of block devices, then we can consolidate RAID code into a single place. Yes, there may be RAID layouts that are too different from the common code, but hopefully we will avoid duplication for EACH new layout that we need to support.
Cheers, Andreas
Signature · 3 lines
Andreas Dilger \ "If a man ate a pound of pasta and a pound of antipasto,
\ would they cancel out, leaving him still hungry?"
http://www-mddsp.enel.ucalgary.ca/People/adilger/ -- DogbertQuoted 10 lines · Andreas Dilger
Well, you don't get anything for free. You need to have some abstraction in order to handle multiple LVM-type systems in a useful way, I don't see any way around that. As for the kernel code, I don't see TOO much abstraction.
- You need code to read each separate on-disk format (partition headers, LVM metadata, etc)
- You need code to combine each of these into some useful "volume"
- You need code to do mapping from input volume/block to output disk/block
- You need code to be able to modify a volume metadata.
Yupp.
Quoted 5 lines
You need code to do this for each different metadata format or partition type. Luckily (hopefully?) most people won't have systems with more than a couple different metadata formats at a time. HOWEVER, you NEED to be able to migrate from one format to another, so you will have periods when you have extra modules loaded during a migration.
No - there is no reason we need support for migration in our volume managment system. The reason to have a framework is excatly that we can have multiple different formats. If there is a tool to provide migration it's nice, but it shouldn't complicate the actual volume managment core.
Quoted 9 lines
Quoted 1 line
god-mode that allows even root to do anything and we come back.
It is not clear if you are you for or against "god-mode"? Would you like it so that it is easy to shoot themselves in the head (or it is so complex/manual that it is hard not to), or rather everyone in straight-jackets in rubber rooms (ala Windows)? Some safety is needed, but there are also reasons to bypass that safety. I don't know what the current state of "god-mode" is in the EVMS code (if it is there, needed, whatever, so it is a moot point).
In unix we have the useruser (or more precise the capabilities model), no need for a damn god mode.
Quoted 4 lines
Exactly. It was me and the EVMS LVM-emulation author that pointed out to the Sistina folks that there was NO need to have an incompatible change to the on-disk layout. In fact, the EVMS code can handle BOTH formats, and did so before LVM did.
Of course they don't, but Sistina doesn't like staying with one version more than a few month. See my compat code to support 0.8 volumes on 0.9 that got rejected.
Quoted 1 line
But, strangely it didn't make it into the LVM codebase. Why is that?
Why didn't all your nice stuff get merged? :P
Quoted 1 line
I'm not blaming you, of course. EVMS is at least a SourceForge project.
With the problem that the IBM guys ignore me since someone told them that I'm "not part of the OpenSource movement". I started sending fixes in the very beginning..
Quoted 5 lines
Well, there are obvious reasons for that. Some people won't want to have EVMS in the kernel (clearly you are one of them), so their patch can't just rip out the existing partition scanning code. Maybe there will be a config option to not compile it later, or there will be an "EVMS-lite" which handles basic partitions, or whatever.
The naming doesn't matter. (BTW, EVMS is a damn stupid name..). Either we do the - as you say - VFS for blockdevices or we don't - it doesn't make sense to do it conditional. Wether the complicated drivers are actually use, and wether people want to use the integrated userspace soloutions is a very different question.
Quoted 3 lines
In any case, I doubt an EVMS patch would be accepted if it removed that code to start with.
I think such an patch would be accepted much more likely. You know Linus (and we all :)) likes ripping out code.
If you come and say: this patch nukes all special cases for MD, LVM and partition handling, the code is now X lines less I bet he will like it.
Quoted 1 line
HOWEVER, since EVMS can handle ALL disk devices,
A few days I read that thread on the EVMS list about only detecting SCSI and IDE drives yet - I probably got that wrong?..
Quoted 8 lines
even just regular partitions, at some point it COULD be possible to get rid of the mess of different major/minor numbers for different disk types (hdX, sdX, cciss, rd, ida, etc) and assign all of them to EVMS. Since EVMS only needs a minor number for the end-result volume (which may represent many individual disks, and doesn't need a minor for unused partitions), we would likely not have any shortage of block major numbers. Consider 10*hdX*256 + 8*sdX*256 + MD*256 + 8*CCISS*256 + 8*DAC960*256 + 8*IDA*256... and it is a long time until you have 10k+ VISIBLE volumes on a single system.
Yes, that's a nice thing and in fact linux already posted his opinion about all disk drivers sharing the major numbers (for 2.5?).
But that just needs a small assignement layer (like sound_core), not something as complex as EVMS. Or a bugfree devfs.
Quoted 6 lines
Quoted 4 lines
Yes, because it is a Meta-LVM, not an actualy inplementation. I _really_ want something like that in 2.5 - but not this horrible IBM implementation.
Maybe you can help them work on it?
If they stop ignoring me - sure.
Quoted 8 lines
They have recently just redone a lot
of the code, partly based on input from Andrew Clausen. I don't see any other similar projects out there, and I think it is a waste of effort to complain about work that is done rather than fix it. I tried for a long time (while it was still my job to do so) to fix the current Linux-LVM code, and I only had very minimal impact. Even so, Linux-LVM will never become a Meta-LVM without AT LEAST as much "mess" as EVMS, and probably will contain a lot more.
Linux-LVM doesn't even try to be a "Meta-LVM".
Quoted 6 lines
IBM is at least doing something about this by putting their code where their mouth is. Nothing is ever perfect. Consider MD RAID - it isn't even compatible between "stock" 2.2 and 2.2+Mingo-0.9RAID patches (which is one reason why Alan never accepted the 0.9 MD RAID into 2.2, despite the fact that ALL distros used the 0.9 MD RAID patches).
It's a different issue. For 2.4 ti sould be very easy to support both, 2.2 is missing a generic block remapping interface.
Quoted 5 lines
Quoted 2 lines
Nope - I'm currently working on implementing VxVM support, and I have to redo all the RAID stuff because it is so incomaptible.
The question is - will the VxVM support be yet ANOTHER separate code base in the kernel?
Of course I plug into one of the many generic volume managment frameworks that are available - that's why we have this thread, heh?
Quoted 2 lines
You complain about code bloat in EVMS, but having LVM, MD, VxVM, NT LDM, etc. all separate is also code bloat,
I think al these are together smaller than EVMS.
At lest they don't have their own 200k linked list implementation :P
Quoted 6 lines
Quoted 2 lines
If you have a design to share the RAID engine for very different layouts:
nice - but I don'T see the relation to EVMS.No, I can't say I do - it was purely speculation based on the concept that RAID-1 and RAID-5 are GENERALLY the same concept. Maybe the reason VxVM is so different is that it also incorporates "LVM" style function as well?
Yes. It has a completly different layering. FreeBSD Vinum has a structure that is somewhat similar if you want to take a look at another opensource project.
Quoted 4 lines
Note that NT-LDM ALSO has RAID-0/1/5 modes (not yet supported by the LDM driver) which will eventually need support. Do you want the number of RAID implementations in the kernel to grow without bound, or rather try and consolidate them (where possible) into a generic interface like EVMS?
You forgot ATA fakeraid by arjan :)
Christoph
Signature · 1 line
Of course it doesn't work. We've performed a software upgrade.
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Christoph" == Christoph Hellwig <[email protected]> writes:
Christoph> On Tue, Sep 11, 2001 at 11:57:13AM -0600, Andreas Dilger
Quoted 3 lines · Christoph> wrote:
Quoted 3 lines
Do you have a specific complaint? The current LVM kernel code isn't all that robust either (and the user-space code is not very pretty).
Christoph> Please take a barf-bag, look at the code, look at the Christoph> unneded abstrations, look at the uneeded interfaces, look Christoph> at the code-duplication, look at the over-engineering.
At least the IBM guys seem interested in listening and taking code/suggestions from the community. There was a fairly productive BOF at OLS a couple of months ago. Thats a lot more than we can say about the current LVM.
Jes
Quoted 4 lines
At least the IBM guys seem interested in listening and taking code/suggestions from the community. There was a fairly productive BOF at OLS a couple of months ago. Thats a lot more than we can say about the current LVM.
I've been taking LVM patches into -ac via Joe and the impression I get is that they are working sanely and sensibly towards cleaning up the existing LVM code. It certainly isnt a trivial job to take any piece of code and make it excellent, and I am sure they would benefit from productive help and patches
Alan