Re: [Qemu-devel] [PATCH v1 1/2] reduce qemu's heap Rss size from 12252kB to 2752KB

[Paolo Bonzini <pbonzini@redhat.com>]

----- Original Message -----
> From: "Anthony Xu" <anthony.xu@intel.com>
> To: "Paolo Bonzini" <pbonzini@redhat.com>, "Yang Zhong" <yang.zhong@intel.com>, qemu-devel@nongnu.org
> Cc: "Chao P Peng" <chao.p.peng@intel.com>
> Sent: Friday, March 10, 2017 8:30:06 PM
> Subject: RE: [PATCH v1 1/2] reduce qemu's heap Rss size from 12252kB to 2752KB
> 
> > > Ideally, freeing QOM object should not require a global lock.
> > > If you see any other QOM requiring a global lock, please let us know, we
> > are willing to fix it.
> > 
> > All of them.  When unplugging a device, the device object will be freed
> > from an RCU callback.
> > 
> 
> Thanks for your reply,
> Some objects may not need lock at all to be freed,
> Some objects may just need a small lock to be freed,
> 
> Should we let RCU callbacks to decide which lock they need instead of
> enforcing the global lock in RCU thread?

Splitting the global lock and deciding which object requires which lock
is very, very hard.

The problem is that everything calling object_unref potentially needs the
global lock, and that includes the following call chain: object_unref
-> memory_region_unref -> flatview_destroy -> flatview_unref ->
address_space_update_topology -> memory_region_transaction_commit.

Paolo

> As for the device object, can we get the global lock inside the object
> free/destroy function for now?
> 
> If we can remove the global lock inside RCU thread, we can save 9MB heap
> memory, that's a lot!
> 
> Please share with us if you have other idea to do this.
> 
> Thanks
> Anthony
> 
> 
> 
>