On Tue 14-06-22 11:51:35, Linus Torvalds wrote: > On Tue, Jun 14, 2022 at 11:20 AM Thomas Backlund <tmb@xxxxxx> wrote: > > > > I "think" this is the suggested fix: > > > > https://git.kernel.org/pub/scm/linux/kernel/git/jack/linux-fs.git/commit/?h=for_next&id=46b6418e26c7c26f98ff9c2c2310bce5ae2aa4dd > > Ugh, this is just too ugly for words. > > That's not a fix. That's a "hide the problem" patch. I agree it is papering over the real problem. I consider that a stopgap solution so that machines can boot until we find a cleaner solution. > Now, admittedly clearly the "hide the problem" code already existed, > and was just moved earlier, but I really think this whole "we're > calling __mark_inode_dirty() on an inode that isn't even *initialized* > yet" is a much deeper issue, and shouldn't have some hacky work-around > in __mark_inode_dirty() that just happens to make it work. > > I don't mind that patch per se - moving the code is fine. > > But I *do* mind the patch when the reason is to hide that wrong > ordering of operations. > > Now, maybe a proper fix might be to say that new_inode_pseudo() should > always initialize i_state to I_DIRTY_ALL or something like that. The > comment already says that they cannot participate in writeback, so > maybe they should be disabled that way (ie a pseudo inode is always > dirty and marking it dirty does nothing). Sadly it is not so simple. Firstly, new_inode_pseudo() gets used for all inodes (through new_inode()), secondly, tmpfs allocates fully standard inodes through new_inode() as any other filesystem. We could check writeback capabilities of the sb->s_bdi in new_inode_pseudo() but that would not work for inodes that will become block device inodes because blockdev_superblock has noop_backing_dev_info so we'd have to specialcase that. Overall it looks a bit hairy to my taste. > And then you get rid of the noop_backing_dev_info entirely. And this would be even more difficult because there are other places that expect there's *some* bdi associated with each sb. > Or just make sure that noop_backing_dev_info is fully initialized > before it's used. > > Because I think the real problem here is that things have a pointer to > an uninitialized backing_dev_info. I fully agree with this. IMHO we need to initialize noop_backing_dev_info earlier but early init is not exactly my comfort zone so I have to verify whether various stuff in cgwb_bdi_init() is safe to call so early... Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
