On Wed, 11 Jan 2012 12:23:11 +0400 Konstantin Khlebnikov <khlebnikov@xxxxxxxxxx> wrote: > KAMEZAWA Hiroyuki wrote: > > On Fri, 06 Jan 2012 21:38:56 +0400 > > Konstantin Khlebnikov<khlebnikov@xxxxxxxxxx> wrote: > > > >> Memory migration fill pte with migration entry and it didn't update rss counters. > >> Then it replace migration entry with new page (or old one if migration was failed). > >> But between this two passes this pte can be unmaped, or task can fork child and > >> it will get copy of this migration entry. Nobody account this into rss counters. > >> > >> This patch properly adjust rss counters for migration entries in zap_pte_range() > >> and copy_one_pte(). Thus we avoid extra atomic operations on migration fast-path. > >> > >> Signed-off-by: Konstantin Khlebnikov<khlebnikov@xxxxxxxxxx> > > > > It's better to show wheter this is a bug-fix or not in changelog. > > > > IIUC, the bug-fix is the 1st harf of this patch + patch [2/3]. > > Your new bug-check code is in patch[1/3] and 2nd half of this patch. > > > > No, there only one new bug-check in 1st patch, this is non-fatal warning. > I didn't hide this check under CONFIG_VM_DEBUG because it rather small and > rss counters covers whole page-table management, this is very good invariant. > Currently I can trigger this warning only on this rare race -- extremely loaded > memory compaction catches this every several seconds. > > 1/3 bug-check > 2/3 fix preparation > 3/3 bugfix in two places: > do rss++ in copy_one_pte() > do rss-- in zap_pte_range() > Hmm, ok, I read wrong. So, I think you should post the patch with [BUGFIX] and report 'what happens' and 'what is the bug' , 'what you fixed' explicitly. As... == This patch series fixes per-mm rss counter accounting bug. When pages are heavily migrated, the rss counters will go wrong by fork() and unmap() because they ignores migration_pte_entries. This rarelly happens but will make rss counter incorrect. This seires of patches will fix the issue by adding proper accounting of migration_pte_entries in unmap() and fork(). This series includes bug check code, too. == If [BUGFIX], people will have more interests. Anyway, thank you for bugfix. -Kame -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>