vhost-user: questions regarding migration

vhost-user: questions regarding migration

[Anton Kuchin]

[-- Attachment #1: Type: text/plain, Size: 1030 bytes --]

Hi,

I'm implementing migration support in vhost-user backend and have a
couple of questions:

1. How master can be sure that logging was started?

We expect that right after set_fatures command with VHOST_F_LOG_ALL flag
all memory modifications will be tracked in log, but slave can need a
little time to process this command so there is a chance that some
requests can be untracked. Is there a way to ensure all requests are
logged or determine the moment since when tracking starts and master can
start migrating memory?

2. Why do we need separate log_addr for vring and how can it be not
covered by mem table?

As far as I understand slave receives used address in set_vring_addr
command and to map it correctly we do need valid entry in memory table.
So this field looks redundant to me. Am I missing something?

BTW the word "log_guest_addr" is mentioned only once in the document and
in "vring address description" payload it is just called "log",
shouldn't we should change this names to match?


[-- Attachment #2: pEpkey.asc --]
[-- Type: application/pgp-keys, Size: 2505 bytes --]

Re: vhost-user: questions regarding migration

[Dr. David Alan Gilbert]

* Anton Kuchin (antonkuchin@yandex-team.ru) wrote:
> Hi,
> 
> I'm implementing migration support in vhost-user backend and have a
> couple of questions:

Hi Anton,
  I'ts a while since I've looked at this code, but from memory:

> 1. How master can be sure that logging was started?
> 
> We expect that right after set_fatures command with VHOST_F_LOG_ALL flag
> all memory modifications will be tracked in log, but slave can need a
> little time to process this command so there is a chance that some
> requests can be untracked. Is there a way to ensure all requests are
> logged or determine the moment since when tracking starts and master can
> start migrating memory?

I think it happens after the set_mem_table command; I think if the
command had the 'NEED_REPLY' flag set then the client responds to the
qemu to say the command is done.

> 2. Why do we need separate log_addr for vring and how can it be not
> covered by mem table?

Hmm I don't think I understand that question; isn't the vring covered by
the memtable?

> As far as I understand slave receives used address in set_vring_addr
> command and to map it correctly we do need valid entry in memory table.
> So this field looks redundant to me. Am I missing something?
> 
> BTW the word "log_guest_addr" is mentioned only once in the document and
> in "vring address description" payload it is just called "log",
> shouldn't we should change this names to match?

Let's cc in Marc-Andre.

Dave



-- 
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK