On Wed, May 06, 2020 at 12:16:13PM -0700, Eric Biggers wrote: > On Wed, May 06, 2020 at 02:47:19PM +0800, Gao Xiang wrote: > > On Wed, May 06, 2020 at 09:58:22AM +0800, Gao Xiang wrote: > > > On Tue, May 05, 2020 at 06:24:28PM -0700, Eric Biggers wrote: > > > > On Wed, May 06, 2020 at 08:14:07AM +0800, Gao Xiang wrote: > > > > > > > > > > > > Actually, I think this is wrong because the fsync can be done via a > > > > > > file > > > > > > descriptor that was opened to a now-deleted link to the file. > > > > > > > > > > I'm still confused about this... > > > > > > > > > > I don't know what's wrong with this version from my limited knowledge? > > > > > inode itself is locked when fsyncing, so > > > > > > > > > > if the fsync inode->i_nlink == 1, this inode has only one hard link > > > > > (not deleted yet) and should belong to a single directory; and > > > > > > > > > > the only one parent directory would not go away (not deleted as > > > > > well) > > > > > since there are some dirents in it (not empty). > > > > > > > > > > Could kindly explain more so I would learn more about this scenario? > > > > > Thanks a lot! > > > > > > > > i_nlink == 1 just means that there is one non-deleted link. There can > > > > be links > > > > that have since been deleted, and file descriptors can still be open to > > > > them. > > > > > > Thanks for your inspiration. You are right, thanks. > > > > > > Correct my words... I didn't check f2fs code just now, it seems f2fs > > > doesn't > > > take inode_lock as some other fs like __generic_file_fsync or ubifs_fsync. > > > > > > And i_sem locks nlink / try_to_fix_pino similarly in some extent. It seems > > > no race by using d_find_alias here. Thanks again. > > > > > > > (think more little bit just now...) > > > > Thread 1: Thread 2 (fsync): > > vfs_unlink try_to_fix_pino > > f2fs_unlink > > f2fs_delete_entry > > f2fs_drop_nlink (i_sem, inode->i_nlink = 1) > > > > (... but this dentry still hashed) i_sem, check > > inode->i_nlink = 1 > > i_sem d_find_alias > > > > d_delete > > > > I'm not sure if fsync could still use some wrong alias by chance.. > > completely untested, maybe just noise... > > > > Right, good observation. My patch makes it better, but it's still broken. > > I don't know how to fix it. If we see i_nlink == 1 and multiple hashed > dentries, there doesn't appear to be a way to distingush which one corresponds > to the remaining link on-disk (if any; it may not even be in the dcache), and > which correspond to links that vfs_unlink() has deleted from disk but hasn't > yet > done d_delete() on. > > One idea would be choose one, then take inode_lock_shared(dir) and do > __f2fs_find_entry() to check if the dentry is really still on-disk. That's > heavyweight and error-prone though, and the locking could cause problems. > > I'm wondering though, does f2fs really need try_to_fix_pino() at all, and did > it > ever really work? It never actually updates the f2fs_inode::i_name to match > the > new directory. So independently of this bug with deleted links, I don't see > how > it can possibly work as intended.
Part of my humble opinion would be "update pino in rename/unlink/link... such ops instead of in fsync" (maybe it makes better sense of locking)... But actually I'm not a f2fs folk now, just curious about what the original patch resolved with these new extra igrab/iput (as I said before, I could not find some clue previously). Thanks, Gao Xiang > > - Eric
