On Tue, Apr 27, 2021 at 01:16:44AM +0800, haosdent wrote: > > really should not assume ->d_inode stable > > Hi, Alexander, sorry to disturb you again. Today I try to check what > `dentry->d_inode` and `nd->link_inode` looks like when `dentry` is > already been killed in `__dentry_kill`. > > ``` > nd->last.name: net/sockstat, dentry->d_lockref.count: -128, > dentry->d_inode: (nil), nd->link_inode: 0xffffffffab299966 > nd->last.name: net/sockstat, dentry->d_lockref.count: -128, > dentry->d_inode: (nil), nd->link_inode: 0xffffffffab299966 > nd->last.name: net/sockstat, dentry->d_lockref.count: -128, > dentry->d_inode: (nil), nd->link_inode: 0xffffffffab299966 > ``` > > It looks like `dentry->d_inode` could be NULL while `nd->link_inode` > is always has value. > But this make me confuse, by right `nd->link_inode` is get from > `dentry->d_inode`, right? It's sampled from there, yes. And in RCU mode there's nothing to prevent a previously positive dentry from getting negative and/or killed. ->link_inode (used to - it's gone these days) go with ->seq, which had been sampled from dentry->d_seq before fetching ->d_inode and then verified to have ->d_seq remain unchanged. That gives you "dentry used to have this inode at the time it had this d_seq", and that's what gets used to validate the sucker when we switch to non-RCU mode (look at legitimize_links()). IOW, we know that * at some point during the pathwalk that sucker had this inode * the inode won't get freed until we drop out of RCU mode * if we need to go to non-RCU (and thus grab dentry references) while we still need that inode, we will verify that nothing has happened to that link (same ->d_seq, so it still refers to the same inode) and grab dentry reference, making sure it won't go away or become negative under us. Or we'll fail (in case something _has_ happened to dentry) and repeat the entire thing in non-RCU mode.
