On 2015-09-22 11:39, Hugo Mills wrote:
> On Tue, Sep 22, 2015 at 10:54:45AM -0400, Austin S Hemmelgarn wrote:
>> On 2015-09-22 10:36, Hugo Mills wrote:
>>> On Tue, Sep 22, 2015 at 04:23:33PM +0200, David Sterba wrote:
>>>> On Tue, Sep 22, 2015 at 01:41:31PM +0000, Hugo Mills wrote:
>>>>> On Tue, Sep 22, 2015 at 03:36:43PM +0200, Holger Hoffstätte wrote:
>>>>>> On 09/22/15 14:59, Jeff Mahoney wrote:
>>>>>> (snip)
>>>>>>> So if they way we want to prevent the loss of raid type info is by
>>>>>>> maintaining the last block group allocated with that raid type, fine,
>>>>>>> but that's a separate discussion.  Personally, I think keeping 1GB
>>>>>>
>>>>>> At this point I'm much more surprised to learn that the RAID type can
>>>>>> apparently get "lost" in the first place, and is not persisted
>>>>>> separately. I mean..wat?
>>>>>
>>>>>     It's always been like that, unfortunately.
>>>>>
>>>>>     The code tries to use the RAID type that's already present to work
>>>>> out what the next allocation should be. If there aren't any chunks in
>>>>> the FS, the configuration is lost, because it's not stored anywhere
>>>>> else. It's one of the things that tripped me up badly when I was
>>>>> failing to rewrite the chunk allocator last year.
>>>>
>>>> Yeah, right now there's no persistent default for the allocator. I'm
>>>> still hoping that the object properties will magically solve that.
>>>
>>>     There's no obvious place that filesystem-wide properties can be
>>> stored, though. There's a userspace tool to manipulate the few current
>>> FS-wide properties, but that's all special-cased to use the
>>> "historical" ioctls for those properties, with no generalisation of a
>>> property store, or even (IIRC) any external API for them.
>>>
>>>     We're nominally using xattrs in the btrfs: namespace on directories
>>> and files, and presumably on the top directory of a subvolume for
>>> subvol-wide properties, but it's not clear where the FS-wide values
>>> should go: in the top directory of subvolid=5 would be confusing,
>>> because then you couldn't separate the properties for *that subvol*
>> >from the ones for the whole FS (say, the default replication policy,
>>> where you might want the top subvol to have different properties from
>>> everything else).
>> Possibly do special names for the defaults and store them there?  In
>> general, I personally see little value in having some special
>> 'default' properties however.
>
>     That would work.
>
>> The way I would expect things to work is that a new subvolume
>> inherits it's properties from it's parent (if it's a snapshot),
>
>     Definitely this.
>
>> or
>> from the next higher subvolume it's nested in.
>
>     I don't think I like this. I'm not quite sure why, though, at the
> moment.
>
>     It definitely makes the process at the start of allocating a new
> block group much more complex: you have to walk back up through an
> arbitrary depth of nested subvols to find the one that's actually got
> a replication policy record in it. (Because after this feature is
> brought in, there will be lots of filesystems without per-subvol
> replication policies in them, and we have to have some way of dealing
> with those as well).
ro-compat flag perhaps?
>
>     With an FS default policy, you only need check the current subvol,
> and then fall back to the FS default if that's not found.
>
>     These things are, I think, likely to be lightly used: I would be
> reasonably surprised to find more than two or possibly three storage
> policies in use on any given system with a sane sysadmin.
>
>     I'm actually not sure what the interactions of multiple storage
> policies are going to be like. It's entirely possible, particularly
> with some of the more exotic (but useful) suggestions I've thought of,
> that the behaviour of the FS is dependent on the order in which the
> block groups are allocated. (i.e. "20 GiB to subvol-A, then 20 GiB to
> subvol-B" results in different behaviour than "1 GiB to subvol-A then
> 1 GiB to subvol-B and repeat"). I tried some simple Monte-Carlo
> simulations, but I didn't get any concrete results out of it before
> the end of the train journey. :)
Yeah, I could easily see that getting complicated when you add in the 
(hopefully soon) possibility of n-copy replication.
>
>>   This would obviate
>> the need for some special 'default' properties, and would be
>> relatively intuitive behavior for a significant majority of people.
>
>     Of course, you shouldn't be nesting subvolumes anyway. It makes
> it much harder to manage them.
That depends though, I only ever do single nesting (ie, a subvolume in a 
subvolume), and I use it to exclude stuff from getting saved in 
snapshots (mostly stuff like clones of public git trees, or other stuff 
that's easy to reproduce without a backup).  Beyond that though, there 
are other inherent issues of course.