On Sun, Mar 13, 2022 at 3:23 PM Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote: > > On Sun, 13 Mar 2022 15:09:27 +0900 Ryusuke Konishi <konishi.ryusuke@xxxxxxxxx> wrote: > > > Hi Matthew, and Andrew, > > > > On Sat, Mar 12, 2022 at 7:56 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote: > > > > > > On Fri, Mar 11, 2022 at 08:43:57PM +0100, David Hildenbrand wrote: > > > > Hi, > > > > > > > > playing with swapfiles on random file systems, I stumbled over the > > > > following nilfs issue (and reproduced it on latest greatest > > > > linux/master -- v5.17-rc7+). I did not try finding out when this > > > > was introduced and I did not run into this issue on other file > > > > systems I tried. > > > > > > It's a known bug in NILFS, and I think yours is the fifth report > > > of it dating back eight months. > > > > The root cause of this issue is that NILFS uses two page caches > > per inode, one for data blocks and another for b-tree node blocks. > > > > Even though __folio_end_writeback(), __folio_start_writeback(), and > > __folio_mark_dirty() acquire lock for mapping->i_pages, > > inode_to_wb(inode) inside them performs lockdep test for the former one > > (i.e. inode->i_mapping->i_pages.xa_lock). > > > > So, mark_buffer_dirty(), end_page_writeback(), and set_page_writeback() > > for pages in the latter NILFS specific page cache hit the LOCKDEP warning. > > > > I tried to find a way to resolve this, but have no good idea so far. > > If things are set up appropriately, inode_to_wb() should be able to > test inode->i_mapping->host->i_mapping->i_pages.xa_lock and get the > desired result. > > At least, that's the case with blockdevs. I don't know if nilfs2 sets > things up that way. Yes, that is the problem of the current implementation of nilfs2. nilfs2 sets the page cache for data to inode->i_mapping but the page cache for b-tree nodes is not set to inode->i_mapping because both are kept in one nilfs_inode together. To satisfy relation inode->i_mapping->host == inode for both page caches, I need to change nilfs_inode to assign inode struct for both. Well, I'll consider if such modification is possible as a solution. Regards, Ryusuke Konishi
