I'm looking at sanitizing the I_* flags and the current implementation of I_CREATING reads like a bug. It showed up in this patchset: https://lore.kernel.org/all/20180729220453.13431-2-viro@xxxxxxxxxxxxxxxxxx/ The relevant commit: commit c2b6d621c4ffe9936adf7a55c8b1c769672c306f Author: Al Viro <viro@xxxxxxxxxxxxxxxxxx> Date: Thu Jun 28 15:53:17 2018 -0400 new primitive: discard_new_inode() We don't want open-by-handle picking half-set-up in-core struct inode from e.g. mkdir() having failed halfway through. [snip] Solution: new flag (I_CREATING) set by insert_inode_locked() and removed by unlock_new_inode() and a new primitive (discard_new_inode()) to be used by such halfway-through-setup failure exits instead of unlock_new_inode() / iput() combinations. That primitive unlocks new inode, but leaves I_CREATING in place. iget_locked() treats finding an I_CREATING inode as failure (-ESTALE, once we sort out the error propagation). insert_inode_locked() treats the same as instant -EBUSY. ilookup() treats those as icache miss. So as far as I understand the intent was to make it so that discarded inodes can be tested for with: (inode->i_state & (I_NEW | I_CREATING) == I_CREATING) But that's not what the patch is doing. In insert_inode_locked() every inserted inode gets both flags: if (likely(!old)) { spin_lock(&inode->i_lock); inode->i_state |= I_NEW | I_CREATING; hlist_add_head_rcu(&inode->i_hash, head); spin_unlock(&inode->i_lock); spin_unlock(&inode_hash_lock); return 0; } This means another call for the same inode will find it and: if (unlikely(old->i_state & I_CREATING)) { spin_unlock(&old->i_lock); spin_unlock(&inode_hash_lock); return -EBUSY; } ... return with -EBUSY instead of waiting to check what will happen with it. The call to wait_on_inode() which can be found later: __iget(old); spin_unlock(&old->i_lock); spin_unlock(&inode_hash_lock); wait_on_inode(old); if (unlikely(!inode_unhashed(old))) { iput(old); return -EBUSY; } ... only ever gets to execute if the inode is fully constructed *or* it was added to the hash with a routine which does not set I_CREATING, which does not add up. So if I understand correctly what was the intended behavior, my counterproposal is to retire I_CREATING and instead add I_DISCARDED which would be set by discard_new_inode(). Then insert_inode_locked() and others can do things like: if (unlikely(old->i_state & I_DISCARDED)) { spin_unlock(&old->i_lock); spin_unlock(&inode_hash_lock); return -EBUSY; } [snip] wait_on_inode(old); if (unlikely(old->i_state & I_DISCARDED || !inode_unhashed(old))) { iput(old); return -EBUSY; } The flag thing aside, there is weird bug in this routine in inode traversal: hlist_for_each_entry(old, head, i_hash) { if (old->i_ino != ino) continue; if (old->i_sb != sb) continue; spin_lock(&old->i_lock); if (old->i_state & (I_FREEING|I_WILL_FREE)) { spin_unlock(&old->i_lock); continue; } break; } This will intentionally skip I_FREEING|I_WILL_FREE instead of waiting on it to leave the hash. I verified this is the only place in the file doing this, others call __wait_on_freeing_inode(). It reads like a bug to me because it allows the caller looking for this inode to insert their own copy and progress outside of the routine. The current inode_unhashed checks after wait_on_inode could be replaced with something like: bool inode_can_use() { if (inode->i_state & I_DISCARDED) return false; if (inode_unhashed(inode)) return false; return true; } or maybe wait_on_inode could grow a return value and do the work inside. I also noticed wait_on_inode is defined in include/linux/writeback.h, but the only users are fs/gfs2 and fs/inode.c, so it will probably want a different home. Comments? Below is a non-operational WIP diff to replace the flag, it does not add the aforementioned checks after wait_on_inode yet or do any cleanups: diff --git a/fs/dcache.c b/fs/dcache.c index 60046ae23d51..c1188ff2fbd1 100644 --- a/fs/dcache.c +++ b/fs/dcache.c @@ -1981,7 +1981,7 @@ void d_instantiate_new(struct dentry *entry, struct inode *inode) spin_lock(&inode->i_lock); __d_instantiate(entry, inode); WARN_ON(!(inode->i_state & I_NEW)); - inode->i_state &= ~I_NEW & ~I_CREATING; + inode->i_state &= ~I_NEW; /* * Pairs with the barrier in prepare_to_wait_event() to make sure * ___wait_var_event() either sees the bit cleared or diff --git a/fs/inode.c b/fs/inode.c index 95fada5c45ea..9b6d5a644cf5 100644 --- a/fs/inode.c +++ b/fs/inode.c @@ -1029,7 +1029,7 @@ static struct inode *find_inode(struct super_block *sb, __wait_on_freeing_inode(inode, is_inode_hash_locked); goto repeat; } - if (unlikely(inode->i_state & I_CREATING)) { + if (unlikely(inode->i_state & I_DISCARDED)) { spin_unlock(&inode->i_lock); rcu_read_unlock(); return ERR_PTR(-ESTALE); @@ -1070,7 +1070,7 @@ static struct inode *find_inode_fast(struct super_block *sb, __wait_on_freeing_inode(inode, is_inode_hash_locked); goto repeat; } - if (unlikely(inode->i_state & I_CREATING)) { + if (unlikely(inode->i_state & I_DISCARDED)) { spin_unlock(&inode->i_lock); rcu_read_unlock(); return ERR_PTR(-ESTALE); @@ -1181,7 +1181,7 @@ void unlock_new_inode(struct inode *inode) lockdep_annotate_inode_mutex_key(inode); spin_lock(&inode->i_lock); WARN_ON(!(inode->i_state & I_NEW)); - inode->i_state &= ~I_NEW & ~I_CREATING; + inode->i_state &= ~I_NEW; /* * Pairs with the barrier in prepare_to_wait_event() to make sure * ___wait_var_event() either sees the bit cleared or @@ -1195,10 +1195,14 @@ EXPORT_SYMBOL(unlock_new_inode); void discard_new_inode(struct inode *inode) { + u32 state; lockdep_annotate_inode_mutex_key(inode); spin_lock(&inode->i_lock); - WARN_ON(!(inode->i_state & I_NEW)); - inode->i_state &= ~I_NEW; + state = inode->i_state; + WARN_ON(!(state & I_NEW)); + state &= ~I_NEW; + state |= I_DISCARDED; + WRITE_ONCE(inode->i_state, state); /* * Pairs with the barrier in prepare_to_wait_event() to make sure * ___wait_var_event() either sees the bit cleared or @@ -1783,6 +1787,7 @@ int insert_inode_locked(struct inode *inode) while (1) { struct inode *old = NULL; spin_lock(&inode_hash_lock); +repeat: hlist_for_each_entry(old, head, i_hash) { if (old->i_ino != ino) continue; @@ -1790,20 +1795,21 @@ int insert_inode_locked(struct inode *inode) continue; spin_lock(&old->i_lock); if (old->i_state & (I_FREEING|I_WILL_FREE)) { - spin_unlock(&old->i_lock); - continue; + __wait_on_freeing_inode(inode, true); + old = NULL; + goto repeat; } break; } if (likely(!old)) { spin_lock(&inode->i_lock); - inode->i_state |= I_NEW | I_CREATING; + inode->i_state |= I_NEW; hlist_add_head_rcu(&inode->i_hash, head); spin_unlock(&inode->i_lock); spin_unlock(&inode_hash_lock); return 0; } - if (unlikely(old->i_state & I_CREATING)) { + if (unlikely(old->i_state & I_DISCARDED)) { spin_unlock(&old->i_lock); spin_unlock(&inode_hash_lock); return -EBUSY; @@ -1826,7 +1832,6 @@ int insert_inode_locked4(struct inode *inode, unsigned long hashval, { struct inode *old; - inode->i_state |= I_CREATING; old = inode_insert5(inode, hashval, test, NULL, data); if (old != inode) { diff --git a/include/linux/fs.h b/include/linux/fs.h index 601d036a6c78..1928ffc55f09 100644 --- a/include/linux/fs.h +++ b/include/linux/fs.h @@ -2559,7 +2559,7 @@ static inline void kiocb_clone(struct kiocb *kiocb, struct kiocb *kiocb_src, * I_OVL_INUSE Used by overlayfs to get exclusive ownership on upper * and work dirs among overlayfs mounts. * - * I_CREATING New object's inode in the middle of setting up. + * I_DISCARDED Inode creation failed. * * I_DONTCACHE Evict inode as soon as it is not used anymore. * @@ -2595,7 +2595,7 @@ static inline void kiocb_clone(struct kiocb *kiocb, struct kiocb *kiocb_src, #define I_DIRTY_TIME (1 << 11) #define I_WB_SWITCH (1 << 12) #define I_OVL_INUSE (1 << 13) -#define I_CREATING (1 << 14) +#define I_DISCARDED (1 << 14) #define I_DONTCACHE (1 << 15) #define I_SYNC_QUEUED (1 << 16) #define I_PINNING_NETFS_WB (1 << 17) diff --git a/include/trace/events/writeback.h b/include/trace/events/writeback.h index 1e23919c0da9..cda77ca84bad 100644 --- a/include/trace/events/writeback.h +++ b/include/trace/events/writeback.h @@ -24,7 +24,7 @@ {I_LINKABLE, "I_LINKABLE"}, \ {I_WB_SWITCH, "I_WB_SWITCH"}, \ {I_OVL_INUSE, "I_OVL_INUSE"}, \ - {I_CREATING, "I_CREATING"}, \ + {I_DISCARDED, "I_DISCARDED"}, \ {I_DONTCACHE, "I_DONTCACHE"}, \ {I_SYNC_QUEUED, "I_SYNC_QUEUED"}, \ {I_PINNING_NETFS_WB, "I_PINNING_NETFS_WB"}, \
