Based-on: <20220920223800.47467-1-pet...@redhat.com> [PATCH 0/5] migration: Bug fixes (prepare for preempt-full)
Tree is here: https://github.com/xzpeter/qemu/tree/preempt-full RFC: https://lore.kernel.org/qemu-devel/20220829165659.96046-1-pet...@redhat.com This patchset is the v1 formal version of preempt-full series. The RFC tag is removed as more tests was done, and all items mentioned in the TODO items previously in the RFC cover letters are implemented in this patchset. A few patches are added. Most of the patches are still the same as the RFC ones but may have some trivial touch ups here and there. E.g., comment touch-ups as Dave used to suggest on bitmap_mutex. If to have a look on the diff stat we'll see mostly "+" is the same as "-" this time though, because with the rp-return thread change we can logically drop a lot of complicated preempt logic previously maintained in migration thread: 3 files changed, 371 insertions(+), 399 deletions(-) Feel free to have a look at patch "migration: Remove old preempt code around state maintainance" where we dropped a lot of old code on preempt state maintainance of migration thread (but the major part of the preempt old code still needs to be there, e.g. channel managements) along with the break-huge parameter (as we never need to break-huge anymore.. because we already run in parallel). Comparing to the recently merged preempt mode I called it "preempt-full" because it threadifies the postcopy channels so now urgent pages can be fully handled separately outside of the ram save loop. The existing preempt code has reduced ramdom page req latency over 10Gbps network from ~12ms to ~500us which has already landed. This preempt-full series can further reduces that ~500us to ~230us per my initial test. More to share below. Note that no new capability is needed, IOW it's fully compatible with the existing preempt mode. So the naming is actually not important but just to identify the difference on the binaries. The logic of the series is simple: send urgent pages in rp-return thread rather than migration thread. It also means rp-thread will take over the ownership of the newly created preempt channel. It can slow down rp-return thread on receiving page requests, but I don't really see a major issue with it so far but only benefits. For detailed performance numbers, please refer to the rfc cover letter. Please have a look, thanks. Peter Xu (14): migration: Add postcopy_preempt_active() migration: Cleanup xbzrle zero page cache update logic migration: Trivial cleanup save_page_header() on same block check migration: Remove RAMState.f references in compression code migration: Yield bitmap_mutex properly when sending/sleeping migration: Use atomic ops properly for page accountings migration: Teach PSS about host page migration: Introduce pss_channel migration: Add pss_init() migration: Make PageSearchStatus part of RAMState migration: Move last_sent_block into PageSearchStatus migration: Send requested page directly in rp-return thread migration: Remove old preempt code around state maintainance migration: Drop rs->f migration/migration.c | 47 +-- migration/multifd.c | 2 +- migration/ram.c | 721 ++++++++++++++++++++---------------------- 3 files changed, 371 insertions(+), 399 deletions(-) -- 2.32.0