Messages in this thread | | | Date | Fri, 23 Feb 2018 02:08:16 +0000 | From | Al Viro <> | Subject | Re: [PATCH v2 3/6] fs/dcache: Avoid the try_lock loop in d_delete() |
| |
On Fri, Feb 23, 2018 at 12:50:22AM +0100, John Ogness wrote:
> The trylock loop can be avoided with functionality similar to > lock_parent(). The fast path tries the trylock first, which is likely > to succeed. In the contended case it attempts locking in the correct > order. This requires to drop dentry->d_lock first, which allows > another task to free d_inode.
Wait a minute. _What_ allows another task to free ->d_inode on a dentry we are holding a reference to? Any place like that is a serious bug - after all, what's to prevent the same place doing that to dentry of an opened file, with obvious ugly results.
That's the whole reason why d_delete() is *NOT* making dentry negative when refcount is greater than 1 (i.e. when somebody else is holding a reference).
Rules for ->d_inode:
* initially NULL.
* only changes under ->d_lock
* __dentry_kill() makes it NULL after dentry has been + marked dead + evicted from all lists except possibly shrink one. with ->d_lock held through all of that. The only thing that can be done by anybody else with the ones stuck on shrink list is actually freeing them.
Note that once __dentry_kill() is called, that's it - dentry is ours, for all practical purposes. There'd better be no other references to that sucker and we make sure that no new ones will arise.
* prior to the call of __dentry_kill() any would-be changer of ->d_inode must be holding a reference to dentry.
* changes from non-NULL to NULL are possible only when there's nobody else holding references.
Changes from NULL to non-NULL _are_ possible (caller must be holding a reference, but that's it). However, feeding a negative dentry to your dentry_lock_inode() is an instant oops - it won't live to the point where you would recheck ->d_inode for changes.
So if you see any place where positive could be changed to negative under us, we do have a problem. Big one.
Refcount can change once we drop ->d_lock, but it can't get to zero - our reference is still with us.
Note that ->d_parent *CAN* change, no matter how many references are held. That's what rcu games in lock_parent() are about - dentry can be moved and ex-parent could've been freed if that was the last reference.
| |