[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: Xenstore quota and driver domains
 
 
On 04.06.20 10:07, Paul Durrant wrote:
 
-----Original Message-----
From: Xen-devel <xen-devel-bounces@xxxxxxxxxxxxxxxxxxxx> On Behalf Of Jürgen 
Groß
Sent: 04 June 2020 06:03
To: xen-devel@xxxxxxxxxxxxxxxxxxxx
Subject: Xenstore quota and driver domains
A recent report on xen-users surfaced a problem we have with driver
domains in medium sized or large configuration: the driver domain can
easily hit the default Xenstore quota (in the report it was a driver
domain for disks which hit the quota when 15 domUs were active at the
same time).
 
 
Which quota is hit? Node or watch?
 
 
Node.
 
 
Setting the quota to a higher limit will work, of course, but this
enables "normal" domUs to use this quota, too, thus hogging lots of
Xenstore resources.
I believe the most sensible way to solve that would be to have a way
to set per-domain quota in Xenstore. As the original reporter already
raised concerns regarding rebooting the server for applying new global
quota, I believe new quota settings should be possible at runtime.
The question is how this interface should look like. Right now I could
think of two variants:
1. A new XS_SET_QUOTA and XS_GET_QUOTA wire command pair which can set
     and get the quota (both, default values for new domains and for any
     existing domain)
2. A new XS_CONTROL command for setting/getting quota (same scope as 1.)
Both variants have advantages and disadvantages:
1. Pros: - clear interface, easily usable
     Cons: - requires more code churn in tools (libxenstore, xl, libxl)
             for support in domain-config
           - any extension will again require similar churn
2. Pros: - easy extensible domain-config possible (via a single item,
             e.g. "xenstore-quota='watches=10 entries=2000'")
     Cons: - text parsing in Xenstore instead of xl/libxl
           - XS_CONTROL sub-options for quota will be ABI
Any thoughts?
 
Even though 1 requires more code churn, I still think it is the better way to 
go.
 
 
Noted.
Juergen
 
 
    
     |