https://bugzilla.kernel.org/show_bug.cgi?id=201631 --- Comment #54 from Jan Kara (jack@xxxxxxx) --- Thanks for the results Erhard. Finally something to go on in the debug results. The pages have been normally faulted in for writing and PTE got set at time 1036580120591 from wp_page_reuse(). But then, although we've called clear_page_dirty_for_io() which calls page_mkclean() (at time 1036587452118), we've never reached the debug point in page_mkclean_one() where PTE write bit gets cleared. So it looks that for some reason page_mkclean() doesn't do what it should on your machines. I was considering a race with munmap() but page_mkclean() uses mapping->i_mmap tree to find the VMA and we remove vmas from that tree only after unmap_page_range() is complete so that does not look possible. Also it is very strange because this is common code for all architectures but you seem to be the only one hitting the problem so far. Aneesh, any idea? Anyway, I've added more debugging to page_mkclean() to see which paths it is taking and maybe that'll tell us something. -- You are receiving this mail because: You are watching the assignee of the bug.
