Paolo Bonzini <pbonz...@redhat.com> wrote: > On 13/02/2017 18:19, Juan Quintela wrote: >> case RAM_SAVE_FLAG_MULTIFD_PAGE: >> + fd_num = qemu_get_be16(f); >> + multifd_recv_page(host, fd_num); >> qemu_get_buffer(f, host, TARGET_PAGE_SIZE); >> break; > > Why do you need RAM_SAVE_FLAG_MULTIFD_PAGE? I understand the > orchestration of sent pages from a single thread, but could the receive > threads proceed independently, each reading its own socket? They do not > even need to tell the central thread "I'm done" (they can do so just by > exiting, and the central thread does qemu_thread_join when it sees the > marker for end of live data).
We can send multiple sends in one go, the whole idea was to send the pages "aligned", and being able to read also in place on destination. But this showed that we still have bottlenecks on the code that search for pages :-( Later, Juan.
