Re: [PATCH 0/4] multifd: various fixes

[Fabiano Rosas]

Elena Ufimtseva <elena.ufimtseva@oracle.com> writes:

> Hello
>
> While working and testing various live migration scenarios,
> a few issues were found.
>
> This is my first patches in live migration and I will
> appreciate the suggestions from the community if these
> patches could be done differently.
>
> [PATCH 1/4] multifd: wait for channels_ready before sending sync
> I am not certain about this change since it seems that
> the sync flag could be the part of the packets with pages that are
> being sent out currently.
> But the traces show this is not always the case:
> multifd_send 230.873 pid=55477 id=0x0 packet_num=0x6f4 normal=0x40 flags=0x1 
> next_packet_size=0x40000
> multifd_send 14.718 pid=55477 id=0x1 packet_num=0x6f5 normal=0x0 flags=0x1 
> next_packet_size=0x80000
> If the sync packet is indeed can be a standalone one, then waiting for
> channels_ready before seem to be appropriate, but waisting iteration on
> sync only packet.

I haven't looked at this code for a while, so there's some context
switching to be made, but you're definitely on the right track here. I
actually have an unsent patch doing almost the same as your patch
1/4. I'll comment more there.

About the sync being standalone, I would expect that to always be the
case since we're incrementing packet_num at that point.

> [PATCH 4/4] is also relevant to 1/4, but fixes the over-accounting in
> case of sync only packet.
>
>
> Thank you in advance and looking forward for your feedback.
>
> Elena
>
> Elena Ufimtseva (4):
>   multifd: wait for channels_ready before sending sync
>   migration: check for rate_limit_max for RATE_LIMIT_DISABLED
>   multifd: fix counters in multifd_send_thread
>   multifd: reset next_packet_len after sending pages
>
>  migration/migration-stats.c |  8 ++++----
>  migration/multifd.c         | 11 ++++++-----
>  2 files changed, 10 insertions(+), 9 deletions(-)