dirty-rate calc

calc-dirty-rate 整体流程

qmp_calc_dirty_rate
  => qemu_thread_create(,MIGRATION_THREAD_DIRTY_RATE, get_dirtyrate_thread,,)
     (get_dirtyrate_thread)
     => dirtyrate_set_state(,,DIRTY_RATE_STATUS_MEASURING)
     => calculate_dirtyrate()
        => switch(mode)
           => DIRTY_RATE_MEASURE_MODE_DIRTY_BITMAP
              => calculate_dirtyrate_dirty_bitmap
           => DIRTY_RATE_MEASURE_MODE_DIRTY_RING
              => calculate_dirtyrate_dirty_ring
           => other:
              => calculate_dirtyrate_sample_vm
     => dirtyrate_set_state(,,DIRTY_RATE_STATUS_MEASURED)

dirty_bitmap

calculate_dirtyrate_dirty_bitmap
## may execute log_start(), may be not
=> memory_global_dirty_log_start(GLOBAL_DIRTY_DIRTY_RATE, )
=> record_dirtypages_bitmap(&dirty_pages ,true)
   => dirty_pages->start_pages = total_dirty_pages;
## get start end
=> start_time = get_clock_get_ms(QEMU_CLOCK_REALTIME)
=> global_dirty_log_sync(GLOBAL_DIRTY_DIRTY_RATE, true);
## WAIT... unit calc_time_ms + start_time expired
=> DirtyStat.calc_time_ms = dirty_stat_wait(config.calc_time_ms, start_time);
## record timeend dirtypages
=> record_dirtypages_bitmap(&dirty_pages, false);
   => dirty_pages->end_pages = total_dirty_pages;
=> do_calculate_dirtyrate
   => (end_pages-start_pages) * 1000 / calc_time_ms

dirty_ring

calculate_dirtyrate_dirty_ring
=> global_dirty_log_change(GLOBAL_DIRTY_DIRTY_RATE, true);
=> start_time = qemu_clock_get_ms(QEMU_CLOCK_HOST) / 1000;
=> vcpu_calculate_dirtyrate
   => vcpu_dirty_stat_collect(records, true);
      => foreach VCPU
         => record_dirtypages(,true)
            => dirty_pages[cpu->cpu_index].start_pages = cpu->dirty_pages

   ## WAIT... unit calc_time_ms + start_time expired
   => duration = dirty_stat_wait(calc_time_ms, init_time_ms);

   => global_dirty_log_sync(flag, one_shot);
   => vcpu_dirty_stat_collect
      => foreach VCPU
         => record_dirtypages(, false)
             => dirty_pages[cpu->cpu_index].start_pages = cpu->dirty_pages
   => foreach VCPU
      ## 计算每一个vcpu的脏页速率
      => do_calculate_dirtyrate

对于dirty_bitmap和dirty_ring两者统计dirty page number依据不一样:

• dirty_bitmap: total_dirty_pages
• dirty_ring: cpu->dirty_pages

query_migrate.ram.dirty-pages-rate

qmp_query_migrate
  => fill_source_migration_info
     => case MIGRATION_STATUS_ACTIVE
        => populate_time_info
        => populate_ram_info
           ...
           => info->ram->dirty_pages_rate = stat64_get(&mig_stats.dirty_pages_rate);
           ...
        => migration_populate_vfio_info

依据mig_stats.dirty_pages_rate

我们来看几个流程:

sync log

global_dirty_log_sync
  => memory_global_dirty_log_sync
     => memory_region_sync_dirty_bitmap
        => foreach memorylistener
           => if listener->log_sync
              => view = address_space_get_flatview()
                 => foreach flat range   ## foreach mr
                     => listener->log_sync() ## dirty bitmap
                        (kvm_log_sync)
           => if listener->log_sync_global
              => listener->log_sync_global() ## dirty ring
                 (kvm_log_sync_global)
     => if one_shot
        => memory_global_dirty_log_stop(flag)

kvm_log_sync:

kvm_log_sync
=> kvm_physical_sync_dirty_bitmap
   => foreach KVMSlot
      => kvm_slot_get_dirty_log()
         => kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d);
      => kvm_slot_sync_dirty_pages()
         => cpu_physical_memory_set_dirty_lebitmap()

kvm_log_sync_global:

kvm_log_sync_global
=> kvm_dirty_ring_flush()
   => kvm_cpu_synchronize_kick_all
      => kvm_dirty_ring_reap
         => kvm_dirty_ring_reap_locked
            => if (cpu)
               => total = kvm_dirty_ring_reap_one(s, cpu);
                  => foreach dirty_gfns[]
                     => kvm_dirty_ring_mark_page
                     => dirty_gfn_set_collected(cur)
                  ## !!!!!! 更新dirty_pages
                  => cpu->dirty_pages += count
            => else foreach_cpu
               => total += kvm_dirty_ring_reap_one(s, cpu);
            => ret = kvm_vm_ioctl(s, KVM_RESET_DIRTY_RINGS);
=> foreach KVMSlot
   => kvm_slot_sync_dirty_pages
      => cpu_physical_memory_set_dirty_lebitmap
   ## 这个指使用了dirty_ring_with_bitmap, 并且是last stage,
   ## 这种情况就说明，dirty_bitmap涉及的memory 并不多，不需要iter
   => if (s->kvm_dirty_ring_with_bitmap && last_stage &&
             kvm_slot_get_dirty_log(s, mem))
      => kvm_slot_sync_dirty_pages(mem);
   => kvm_slot_reset_dirty_pages

两者foreach KVMSlot的方式不同，但是个人感觉，应该最终结果是一样的. 并且最终都调用到了cpu_physical_memory_set_dirty_lebitmap.

kvm_log_sync和kvm_log_sync_global不同的是，kvm_log_sync需要调用 get_dirty_log来获取当前的dirty_bitmap, 而dirty_ring则调用 kvm_dirty_ring_reap_locked来完成”sync”动作，并更新cpu->dirty_pages