I'm putting this RFC out early so we can discuss the issues around the SYNC packet of the multifd protocol. There's a related series posted by Elena Ufimtseva:
https://lore.kernel.org/r/20230922065625.21848-1-elena.ufimts...@oracle.com My interest in this (aside from correctness) is that when doing the (upcoming) fixed-ram[1] migration, I would like to have multifd ignore the concept of packets altogether, since the file: migration is not synchronous. The main problem I hit is that multifd (ab)uses the knowledge that a sync packet is sent after a batch of pages and relies (perhaps inadvertently) on the last post to sem_sync to finish the migration. Which means that without the sync, the main thread just rushes and does cleanup while packets are still in flight. I have add another patch to this series that introduces a multifd-nopackets option (placeholder name), but it's probably too early to discuss that so I'm leaving it out. 1- https://lore.kernel.org/r/20230330180336.2791-1-faro...@suse.de Fabiano Rosas (3): migration/multifd: Move channels_ready semaphore migration/multifd: Decouple control flow from the SYNC packet migration/multifd: Extract sem_done waiting into a function migration/multifd.c | 97 +++++++++++++++++++++++++----------------- migration/multifd.h | 8 ++-- migration/trace-events | 2 +- 3 files changed, 63 insertions(+), 44 deletions(-) -- 2.35.3
