Re: [pve-devel] [PATCH storage 1/2] zfs: only use cache when listing images locally

[Fiona Ebner <f.ebner@proxmox.com>]

Am 16.11.22 um 12:18 schrieb Fabian Grünbichler:
> On November 7, 2022 12:00 pm, Fiona Ebner wrote:
>> The plugin for remote ZFS storages currently also uses the same
>> list_images() as the plugin for local ZFS storages. The issue with
>> this is that there is only one cache which does not remember the
>> target host where the information originated.
>>
>> Simply restrict the cache to be used for the local ZFS plugin only. An
>> alternative solution would be to use a cache for each target host, but
>> that seems a bit more involved and could still be added in the future.
> 
> wouldn't it be sufficient to just do
> 
> $cache->{zfs}->{$storeid}
> 
> when filling/querying the cache, and combining that with *always* listing only
> the storage-relevant pool?

Yes, should work. I'll send a v2 with that.

> 
> the only case where we actually benefit from listing *all* zfs volumes/datasets
> is when
> - there are multiple storages configured referencing overlapping parts of the
> ZFS hierarchy
> - vdisk_list is called with a volume_list with multiple such storages being part
> of the set, or with $vmid but no $storeid (rescan, or purging unreferenced guest
> disks on guest removal)

The cache is already useful if there two ZFS storages, nothing as fancy
as the above needed ;) Then for rescan and others which iterate all
storages, only one zfs list call is issued, rather than one for each ZFS
storage.

> 
> in practice, it likely doesn't make much difference since ZFS should cache the
> metadata for the overlapping parts in memory anyway (given that we'd then call
> 'zfs list' in a loop with different starting points).
> 
> whereas, for most regular cases listing happens without a cache anyway (or with
> a cache, but only a single storage involved), so there is no benefit in querying
> volumes belonging to other storages since we are not interested in them anyway.
> 

Yes, I'd also guess that in practice the benefit of the current list-all
cache is rather limited.

> sidenote: it seems like vdisk_list's volume_list is not used anywhere as parameter?
> 

Seems to be only used in the ZFS tests in pve-storage ;)