> >> > >> Refer to function generic_file_read_iter(), in direct IO path, > >> file_accessed() is done before IO submission, so I think ovl_file_accessed() > >> should be done here no matter completion/error or IO is queued. > > > > Mmm, it doesn't matter much if atime is updated before or after, > > but ovl_file_accessed() does not only update atime, it also copies > > ctime which could have been modified as a result of the io, so > > I think it is safer to put it in the cleanup hook. > > > > Can you give a more detailed description that a read op will modify > ctime as a result of the io? > > I found that it will trigger BUG_ON(irqs_disabled()) while > calling ovl_file_accessed() on async IO return path. The calltrace > is pasted below: > > ovl_file_accessed > -> touch_atime > -> ovl_update_time > -> generic_update_time > -> __mark_inode_dirty > -> ext4_dirty_inode > -> __ext4_get_inode_loc > -> __find_get_block > -> lookup_bh_lru > -> check_irqs_on > > So I need more detail to find how to fix this issue. > It's not important. Please ignore. I din't know there was an issue with placing touch_atime() in completion context. Thanks, Amir.
