On Mon, May 23, 2022 at 11:56:14AM +0100, Dr. David Alan Gilbert wrote: > * Peter Xu (pet...@redhat.com) wrote: > > With preemption mode on, when we see a postcopy request that was requesting > > for exactly the page that we have preempted before (so we've partially sent > > the page already via PRECOPY channel and it got preempted by another > > postcopy request), currently we drop the request so that after all the > > other postcopy requests are serviced then we'll go back to precopy stream > > and start to handle that. > > > > We dropped the request because we can't send it via postcopy channel since > > the precopy channel already contains partial of the data, and we can only > > send a huge page via one channel as a whole. We can't split a huge page > > into two channels. > > > > That's a very corner case and that works, but there's a change on the order > > of postcopy requests that we handle since we're postponing this (unlucky) > > postcopy request to be later than the other queued postcopy requests. The > > problem is there's a possibility that when the guest was very busy, the > > postcopy queue can be always non-empty, it means this dropped request will > > never be handled until the end of postcopy migration. So, there's a chance > > that there's one dest QEMU vcpu thread waiting for a page fault for an > > extremely long time just because it's unluckily accessing the specific page > > that was preempted before. > > > > The worst case time it needs can be as long as the whole postcopy migration > > procedure. It's extremely unlikely to happen, but when it happens it's not > > good. > > > > The root cause of this problem is because we treat pss->postcopy_requested > > variable as with two meanings bound together, as the variable shows: > > > > 1. Whether this page request is urgent, and, > > 2. Which channel we should use for this page request. > > > > With the old code, when we set postcopy_requested it means either both (1) > > and (2) are true, or both (1) and (2) are false. We can never have (1) > > and (2) to have different values. > > > > However it doesn't necessarily need to be like that. It's very legal that > > there's one request that has (1) very high urgency, but (2) we'd like to > > use the precopy channel. Just like the corner case we were discussing > > above. > > > > To differenciate the two meanings better, introduce a new field called > > postcopy_target_channel, showing which channel we should use for this page > > request, so as to cover the old meaning (2) only. Then we leave the > > postcopy_requested variable to stand only for meaning (1), which is the > > urgency of this page request. > > > > With this change, we can easily boost priority of a preempted precopy page > > as long as we know that page is also requested as a postcopy page. So with > > the new approach in get_queued_page() instead of dropping that request, we > > send it right away with the precopy channel so we get back the ordering of > > the page faults just like how they're requested on dest. > > > > Alongside, I touched up find_dirty_block() to only set the postcopy fields > > in the pss section if we're going through a postcopy migration. That's a > > very light optimization and shouldn't affect much. > > > > Reported-by: manish.mis...@nutanix.com > > Signed-off-by: Peter Xu <pet...@redhat.com> > > So I think this is OK; getting a bit complicated!
Yes it is. I added some more comment, hopefully it'll help a little bit. > > Reviewed-by: Dr. David Alan Gilbert <dgilb...@redhat.com> Thanks! > > static bool find_dirty_block(RAMState *rs, PageSearchStatus *pss, bool > > *again) > > { > > - /* This is not a postcopy requested page */ > > - pss->postcopy_requested = false; > > + if (migration_in_postcopy()) { > > + /* > > + * This is not a postcopy requested page, mark it "not urgent", and > > + * use precopy channel to send it. > > + */ > > + pss->postcopy_requested = false; > > + pss->postcopy_target_channel = RAM_CHANNEL_PRECOPY; > > + } > > Do you need the 'if' here? Hmm good question.. precopy should always have these two fields cleared anyway so I wanted to avoid setting them every time, but I just noticed that pss is not initialized at all when used.. static int ram_find_and_save_block(RAMState *rs) { PageSearchStatus pss; ... } So either we'd reset pss explicitly on these fields, or simpler - let me drop the if.. Thanks, -- Peter Xu