Hello Peter,
On Mon, 3 Mar 2025 at 20:20, Peter Xu <[email protected]> wrote:
> We need the header.
* We need a section type, which is sent by qemu_savevm_command_send()
as 'QEMU_VM_COMMAND'.
> Maybe the easiest as of now is one more hook like
> qemu_savevm_state_complete_precopy_prepare(), and only use it in RAM as of
> now.
* What will this helper do?
> > * But earlier we discussed 'flush and sync' is enough for that, no?
>
> Yes it's ok I think, but this patch didn't do that.
>
> + multifd_send_flush();
> + multifd_send_sync_main(MULTIFD_SYNC_LOCAL);
> + qemu_savevm_send_multifd_recv_sync(s->to_dst_file);
>
> I don't think it sent RAM_SAVE_FLAG_MULTIFD_FLUSH. IIUC you need the
> complete multifd_ram_flush_and_sync(), and the new message not needed.
* If we look at multifd_ram_flush_and_sync(), it does:
1. multifd_send() <= this patch does it via
multifd_send_flush()
2. multifd_send_sync_main() <= this patch also calls it above
3. send RAM_SAVE_FLAG_MULTIFD_FLUSH <= this patch sends
MIG_CMD_MULTIFD_RECV_SYNC
* What is missing?
> Instead of I prepare the patch and whole commit message, please take your
> time and think about it, justify it, and if you also think it works put
> explanation into commit message and then we can go with it.
* The commit message does explain about flush and sync and how the
migration command helps. What else do we need to add?
> > * And multifd_recv_sync_main() function on the destination blocks the
> > 'main' thread until all multfd_recv_threads (mig/dst/recv_x) have
> > exited, only then it proceeds to accept the incoming new postcopy
> > connection.
>
> I don't think it makes sure threads have exited,
* 'multifd_recv_sync_main()' blocks the main thread on
'multifd_recv_state->sem_sync' semaphore. It is increased when
multifd_recv threads exit due to the shutdown message. ie. the 'main'
thread unblocks when all 'mig/dst/recv_x' threads have exited.
Thank you.
---
- Prasad
