https://bugzilla.kernel.org/show_bug.cgi?id=201631 --- Comment #12 from Jan Kara (jack@xxxxxxx) --- Thanks for info Ben! First one slightly off-topic comment: Spuriously setting dirty bits in page tables is certainly not OK for file systems. This results in set_page_dirty() being called for a page when it gets unmapped and that means that file system is told there's new data in the page. But the file system is not prepared for that - it needs to prepare its structures for data modification (usually happens in ->page_mkwrite callback during page fault) like making sure there's enough free space, do COW on the page in case it was reflinked etc. As you say, this is likely not a problem of this particular reporter but I wanted to make sure the message gets across :). Now to this report: It's good to know that setting of PTE dirty bit should be atomic at least on G5 wrt concurrent invalidation and page write protection. But something seems to have happened with 4.20-rcs that dirty bits do get set as Erhard started to see these warnings. I haven't heard any report of this from x86, this seems to be ppc specific. Since it happens once in couple of weeks, bisection is not really an option. I guess, I'll write a debug patch to dump PTE and page state when such problematic situation occurs and maybe that'll tell us something more to debug this. -- You are receiving this mail because: You are watching the assignee of the bug.
