ceph: do not carry i_lock for readdir from dcache

We were taking dcache_lock inside of i_lock, which introduces a dependency
not found elsewhere in the kernel, complicationg the vfs locking
scalability work.  Since we don't actually need it here anyway, remove
it.

We only need i_lock to test for the I_COMPLETE flag, so be careful to do
so without dcache_lock held.

Signed-off-by: Sage Weil <sage@newdream.net>

2 files changed, 18 insertions, 26 deletions

diff --git a/fs/ceph/dir.c b/fs/ceph/dir.c
index 495825268cb9..e0a2dc6fcafc 100644
--- a/fs/ceph/dir.c
+++ b/fs/ceph/dir.c
@@ -95,10 +95,7 @@ static unsigned fpos_off(loff_t p)
  */
 static int __dcache_readdir(struct file *filp,
                             void *dirent, filldir_t filldir)
-                __releases(inode->i_lock)
-                __acquires(inode->i_lock)
 {
-        struct inode *inode = filp->f_dentry->d_inode;
         struct ceph_file_info *fi = filp->private_data;
         struct dentry *parent = filp->f_dentry;
         struct inode *dir = parent->d_inode;
@@ -154,7 +151,6 @@ more:
 
         atomic_inc(&dentry->d_count);
         spin_unlock(&dcache_lock);
-        spin_unlock(&inode->i_lock);
 
         dout(" %llu (%llu) dentry %p %.*s %p\n", di->offset, filp->f_pos,
              dentry, dentry->d_name.len, dentry->d_name.name, dentry->d_inode);
@@ -172,35 +168,30 @@ more:
                 } else {
                         dput(last);
                 }
-                last = NULL;
         }
-
-        spin_lock(&inode->i_lock);
-        spin_lock(&dcache_lock);
-
         last = dentry;
 
         if (err < 0)
-                goto out_unlock;
+                goto out;
 
-        p = p->prev;
         filp->f_pos++;
 
         /* make sure a dentry wasn't dropped while we didn't have dcache_lock */
-        if ((ceph_inode(dir)->i_ceph_flags & CEPH_I_COMPLETE))
-                goto more;
-        dout(" lost I_COMPLETE on %p; falling back to mds\n", dir);
-        err = -EAGAIN;
+        if (!ceph_i_test(dir, CEPH_I_COMPLETE)) {
+                dout(" lost I_COMPLETE on %p; falling back to mds\n", dir);
+                err = -EAGAIN;
+                goto out;
+        }
+
+        spin_lock(&dcache_lock);
+        p = p->prev;    /* advance to next dentry */
+        goto more;
 
 out_unlock:
         spin_unlock(&dcache_lock);
-
-        if (last) {
-                spin_unlock(&inode->i_lock);
+out:
+        if (last)
                 dput(last);
-                spin_lock(&inode->i_lock);
-        }
-
         return err;
 }
 
@@ -272,13 +263,13 @@ static int ceph_readdir(struct file *filp, void *dirent, filldir_t filldir)
             ceph_snap(inode) != CEPH_SNAPDIR &&
             (ci->i_ceph_flags & CEPH_I_COMPLETE) &&
             __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1)) {
+                spin_unlock(&inode->i_lock);
                 err = __dcache_readdir(filp, dirent, filldir);
-                if (err != -EAGAIN) {
-                        spin_unlock(&inode->i_lock);
+                if (err != -EAGAIN)
                         return err;
-                }
+        } else {
+                spin_unlock(&inode->i_lock);
         }
-        spin_unlock(&inode->i_lock);
         if (fi->dentry) {
                 err = note_last_dentry(fi, fi->dentry->d_name.name,
                                        fi->dentry->d_name.len);
diff --git a/fs/ceph/super.h b/fs/ceph/super.h
index e2e904442ce2..1886294e12f7 100644
--- a/fs/ceph/super.h
+++ b/fs/ceph/super.h
@@ -400,8 +400,9 @@ static inline bool ceph_i_test(struct inode *inode, unsigned mask)
         struct ceph_inode_info *ci = ceph_inode(inode);
         bool r;
 
-        smp_mb();
+        spin_lock(&inode->i_lock);
         r = (ci->i_ceph_flags & mask) == mask;
+        spin_unlock(&inode->i_lock);
         return r;
 }