Hi, Steve.

On 2021/8/5 4:50, Steven Sistare wrote:
> On 7/30/2021 9:10 AM, Zheng Chuan wrote:
>> Hi, Steve
>> I have saw the discussion about the fork+exec mode below:
>> https://www.mail-archive.com/qemu-devel@nongnu.org/msg815956.html
>>
>> And I am still very curious and I want to discuss about the possibility to 
>> support both fork+exec and exec in cpr framework.
>>
>> 1.Why
>> fork+exec could have some advantages and also drawbacks versus execvp() 
>> directly.
>> Advantages
>> i) fork+exec give the chance to fallback to original process even after we 
>> do exec which is important for workload seamless if any error happens.
>> ii) smaller downtime since we could remove the vm_start() downtime out of 
>> the frozen window.
>> Drawbacks
>> i)need more codes to handle including fork,address/ports conflict between 
>> parent and child.
>> ii)more complex life cycle management for the two processes.
>>
>> 2.How
>> The cpr framework is flexible and scalable, and maybe we can make use of 
>> most codes to support both execvp and fork+exec mode.
>> However, we may need to do more work compared to execvp method.
>> i) do fork mode in a thread like migration thread
>> ii) make parent and child talk to each other through socket or anonymous pipe
>> iii)make use of sharing mechanism of fds in cpr framework including memfd, 
>> vfio and devices fds
>> iv)deal with the conflict about the socket address and port like vnc (do by 
>> reuse port and pass the different args by cprexec)
>> v) do life cycle managements for two qemu processes and need parent exit and 
>> reconnection for the child at last by the management service
>>
>> Please tell me if I am missing something important:)
> 
> Hi Zheng, that list sounds about right.  However, additional kernel changes 
> would be needed to 
> change ownership of the vfio device descriptors. The vfio module records the 
> mm of the creating
> process, and does not allow other processes to unmap ranges.  Also, the 
> mm->locked_vm would 
> need to be transferred to the new process.
> 
Oh,That's right. Maybe we could update them along with VFIO_DMA_MAP_FLAG_VADDR.

> Fork+exec could be a new mode to the cprsave command.
>
Yeah, I would be very happy to see it in the future:)
> - Steve
> .
> 

-- 
Regards.
Chuan

Reply via email to