Messages in this thread | | | From | David Howells <> | Subject | Re: [PATCH 3/6] vfs: Allow searching of the icache under RCU conditions [ver #2] | Date | Thu, 25 Apr 2019 16:45:27 +0100 |
| |
Al Viro <viro@zeniv.linux.org.uk> wrote:
> Hmm... Why do these stores to ->i_state need WRITE_ONCE, while an arseload > of similar in fs/fs-writeback.c does not?
Because what matters in find_inode_rcu() are the I_WILL_FREE and I_FREEING flags - and there's a gap during iput_final() where neither is set.
if (!drop) { inode->i_state |= I_WILL_FREE; spin_unlock(&inode->i_lock); write_inode_now(inode, 1); spin_lock(&inode->i_lock); WARN_ON(inode->i_state & I_NEW); inode->i_state &= ~I_WILL_FREE; ---> }
inode->i_state |= I_FREEING;
It's normally covered by i_lock, but it's a problem if anyone looks at the pair without taking i_lock.
Even flipping the order:
if (!drop) { inode->i_state |= I_WILL_FREE; spin_unlock(&inode->i_lock); write_inode_now(inode, 1); spin_lock(&inode->i_lock); WARN_ON(inode->i_state & I_NEW); inode->i_state |= I_FREEING; inode->i_state &= ~I_WILL_FREE; } else { inode->i_state |= I_FREEING; }
isn't a guarantee of the order in which the compiler will do things AIUI. Maybe I've been listening to Paul McKenney too much. So the WRITE_ONCE() should guarantee that both bits will change atomically.
Note that ocfs2_drop_inode() looks a tad suspicious:
int ocfs2_drop_inode(struct inode *inode) { struct ocfs2_inode_info *oi = OCFS2_I(inode);
trace_ocfs2_drop_inode((unsigned long long)oi->ip_blkno, inode->i_nlink, oi->ip_flags);
assert_spin_locked(&inode->i_lock); inode->i_state |= I_WILL_FREE; spin_unlock(&inode->i_lock); write_inode_now(inode, 1); spin_lock(&inode->i_lock); WARN_ON(inode->i_state & I_NEW); inode->i_state &= ~I_WILL_FREE;
return 1; }
David
| |