On Wed, Apr 20, 2022 at 12:05:24PM +0100, Daniel P. Berrangé wrote: > On Thu, Mar 31, 2022 at 11:08:48AM -0400, Peter Xu wrote: > > This patch enables postcopy-preempt feature. > > > > It contains two major changes to the migration logic: > > > > (1) Postcopy requests are now sent via a different socket from precopy > > background migration stream, so as to be isolated from very high page > > request delays. > > > > (2) For huge page enabled hosts: when there's postcopy requests, they can > > now > > intercept a partial sending of huge host pages on src QEMU. > > > > After this patch, we'll live migrate a VM with two channels for postcopy: > > (1) > > PRECOPY channel, which is the default channel that transfers background > > pages; > > and (2) POSTCOPY channel, which only transfers requested pages. > > > > There's no strict rule of which channel to use, e.g., if a requested page is > > already being transferred on precopy channel, then we will keep using the > > same > > precopy channel to transfer the page even if it's explicitly requested. In > > 99% > > of the cases we'll prioritize the channels so we send requested page via the > > postcopy channel as long as possible. > > > > On the source QEMU, when we found a postcopy request, we'll interrupt the > > PRECOPY channel sending process and quickly switch to the POSTCOPY channel. > > After we serviced all the high priority postcopy pages, we'll switch back to > > PRECOPY channel so that we'll continue to send the interrupted huge page > > again. > > There's no new thread introduced on src QEMU. > > Implicit in this approach is that the delay in sending postcopy > OOB pages is from the pending socket buffers the kernel already > has, and not any delay caused by the QEMU sending thread being > busy doing other stuff.
Yes. > > Is there any scenario in which the QEMU sending thread is stalled > in sendmsg() with a 1GB huge page waiting for the kernel to > get space in the socket outgoing buffer ? Another yes.. It doesn't necessarily to be during sending a 1GB huge page, the guest can be using small pages and IMHO we could get stuck at sendmsg() for a precopy small page while there's actually postcopy requests in the queue. We can't solve this as long as we keep using 1 single thread for sending page. This patchset doesn't solve this issue, yet. And it's actually the chunk discussed and mention in the cover letter too in the section "Avoid precopy write() blocks postcopy" as an TODO item. Logically in the future we could try to make two or more sender threads so postcopy pages can use a separate sender thread. Note that this change will _not_ require interface change either from qemu cmdline or on migration protocol, because this patchset should have handled all the migration protocol already even for that, but then if it'll work well we could get pure speed up on further shrinked latency when preempt mode enabled comparing to before. The other thing is I never measured such an effect, so I can't tell how would it perform at last. We need more work on top if we'd like to persue it, mostly on doing proper synchronizations on senders. Thanks, -- Peter Xu
