[PATCH 0/3] migration dirty bitmap support ARMv7

[PATCH 0/3] migration dirty bitmap support ARMv7

[Mario Smarduch]

The patch set supports migration dirty bitmap support implementation for
arm-kvm.  Spliting of pmd's to pte's as suggested is implemented on demand
when migration is started. 

I tested it on 4-way SMP ARMv7, with SMP guests.
2GB VMs with dirty shared memory segments upto 1.8 GB 
and relatively fast update rates 16Mb/5mS. 

Next course of action would be rmap support which 
scales much better on bigger systems. Although one
think that confused me, x86 migrations were sometimes
10 to 15 times slower, I think it must be something 
wrong with my configuration.


Mario Smarduch (3):
  headers for migration dirtybitmap support
  initial write protect of VM address space and on dirty log read
  hooks to interface with QEMU for initial write protect, dirty log read

 arch/arm/include/asm/kvm_host.h |    9 +++
 arch/arm/kvm/arm.c              |   62 ++++++++++++++-
 arch/arm/kvm/mmu.c              |  158 ++++++++++++++++++++++++++++++++++++++-
 3 files changed, 226 insertions(+), 3 deletions(-)
-- 
1.7.9.5

Re: [PATCH 0/3] migration dirty bitmap support ARMv7

[Marc Zyngier]

Mario,

On 15/04/14 02:24, Mario Smarduch wrote:
> 
> The patch set supports migration dirty bitmap support implementation for
> arm-kvm.  Spliting of pmd's to pte's as suggested is implemented on demand
> when migration is started. 
> 
> I tested it on 4-way SMP ARMv7, with SMP guests.
> 2GB VMs with dirty shared memory segments upto 1.8 GB 
> and relatively fast update rates 16Mb/5mS. 
> 
> Next course of action would be rmap support which 
> scales much better on bigger systems. Although one
> think that confused me, x86 migrations were sometimes
> 10 to 15 times slower, I think it must be something 
> wrong with my configuration.
> 
> 
> Mario Smarduch (3):
>   headers for migration dirtybitmap support
>   initial write protect of VM address space and on dirty log read
>   hooks to interface with QEMU for initial write protect, dirty log read
> 
>  arch/arm/include/asm/kvm_host.h |    9 +++
>  arch/arm/kvm/arm.c              |   62 ++++++++++++++-
>  arch/arm/kvm/mmu.c              |  158 ++++++++++++++++++++++++++++++++++++++-
>  3 files changed, 226 insertions(+), 3 deletions(-)
> 

Overall, I think this patch series requires some work, the most glaring
issue being the total lack of documentation (even by my own standards).
Also, it is fairly obvious that ARMv8 should be done at the same time
(as all that code is shared, and will avoid API abuse...), and that a
fair amount of the code is actually generic across architecture, and
should be shared.

I'm very much looking forward to the next iteration of this series.

Thanks,

	M.
-- 
Jazz is not dead. It just smells funny...

RE: [PATCH 0/3] migration dirty bitmap support ARMv7

[Mario Smarduch]

Hi Marc,
    Thanks for the feedback, very valuable going forward.  I'll clean the patch up and 
repost with much more documentation. One key error you pointed out,
and I overlooked  but knew about it, is to error out from get dirty log if whole 
VM can't be  WPed, at that point QEMU will abort.

- Mario

-----Original Message-----
From: Marc Zyngier [mailto:marc.zyngier@arm.com] 
Sent: Tuesday, April 15, 2014 2:13 AM
To: Mario Smarduch
Cc: kvmarm@lists.cs.columbia.edu; christoffer.dall@linaro.org; 이정석; 정성진; kvm@vger.kernel.org
Subject: Re: [PATCH 0/3] migration dirty bitmap support ARMv7

Mario,

On 15/04/14 02:24, Mario Smarduch wrote:
> 
> The patch set supports migration dirty bitmap support implementation 
> for arm-kvm.  Spliting of pmd's to pte's as suggested is implemented 
> on demand when migration is started.
> 
> I tested it on 4-way SMP ARMv7, with SMP guests.
> 2GB VMs with dirty shared memory segments upto 1.8 GB and relatively 
> fast update rates 16Mb/5mS.
> 
> Next course of action would be rmap support which scales much better 
> on bigger systems. Although one think that confused me, x86 migrations 
> were sometimes
> 10 to 15 times slower, I think it must be something wrong with my 
> configuration.
> 
> 
> Mario Smarduch (3):
>   headers for migration dirtybitmap support
>   initial write protect of VM address space and on dirty log read
>   hooks to interface with QEMU for initial write protect, dirty log 
> read
> 
>  arch/arm/include/asm/kvm_host.h |    9 +++
>  arch/arm/kvm/arm.c              |   62 ++++++++++++++-
>  arch/arm/kvm/mmu.c              |  158 ++++++++++++++++++++++++++++++++++++++-
>  3 files changed, 226 insertions(+), 3 deletions(-)
> 

Overall, I think this patch series requires some work, the most glaring issue being the total lack of documentation (even by my own standards).
Also, it is fairly obvious that ARMv8 should be done at the same time (as all that code is shared, and will avoid API abuse...), and that a fair amount of the code is actually generic across architecture, and should be shared.

I'm very much looking forward to the next iteration of this series.

Thanks,

	M.
--
Jazz is not dead. It just smells funny...