On Sun, Jul 11, 2021 at 11:27:13PM +0800, Hyman Huang wrote: > > IMHO we can directly do the calculation when synchronizing the dirty bits in > > below functions: > > > > cpu_physical_memory_set_dirty_range > > cpu_physical_memory_set_dirty_lebitmap > > cpu_physical_memory_sync_dirty_bitmap > > > > Maybe we can define a global statistics for that? > uhhh... Do you mean that we can reuse the DIRTY_MEMORY_MIGRATION dirty bits > to stat the new dirty pages number and just define the global var to count > the increased dirty pages during the calculation time?
I think I misguided you.. Sorry :) cpu_physical_memory_sync_dirty_bitmap() should not really be in the list above, as it's fetching the bitmap in ram_list.dirty_memory[DIRTY_MEMORY_MIGRATION]. If you see the other two functions, they all apply dirty bits upon the same bitmap (actually ram_list.dirty_memory[*] instead of migration-only). It's used by e.g. memory region log_sync() to deliver lower level dirty bits to upper, e.g., see kvm's log_sync[_global]() and kvm_slot_sync_dirty_pages(). Using cpu_physical_memory_sync_dirty_bitmap() is not a good idea to me (which I saw you used in your latest version), as it could affect migration. See its only caller now at ramblock_sync_dirty_bitmap(): when migration calls it, it'll start to count less than it should for rs->migration_dirty_pages. So what I wanted to suggest is we do some general counting in both cpu_physical_memory_set_dirty_range and cpu_physical_memory_set_dirty_lebitmap. Then to sync for dirty rate measuring, we use memory_global_dirty_log_sync(). That'll sync all dirty bits e.g. in kernel to ram_list.dirty_memory[*], along which we do the accounting. Would that work? -- Peter Xu
