aboutsummaryrefslogtreecommitdiffstats
path: root/fs
diff options
context:
space:
mode:
authorLinus Torvalds <torvalds@linux-foundation.org>2013-09-03 08:08:21 -0700
committerLinus Torvalds <torvalds@linux-foundation.org>2013-09-03 08:08:21 -0700
commitfc6d0b037678f50014ef409c92c5bedc01208fcd (patch)
treeb7de25e97b03c31ea6c5f2540f641b0be2c91832 /fs
parent6e4664525b1db28f8c4e1130957f70a94c19213e (diff)
parentbc08b449ee14ace4d869adaa1bb35a44ce68d775 (diff)
downloadkernel_goldelico_gta04-fc6d0b037678f50014ef409c92c5bedc01208fcd.zip
kernel_goldelico_gta04-fc6d0b037678f50014ef409c92c5bedc01208fcd.tar.gz
kernel_goldelico_gta04-fc6d0b037678f50014ef409c92c5bedc01208fcd.tar.bz2
Merge branch 'lockref' (locked reference counts)
Merge lockref infrastructure code by me and Waiman Long. I already merged some of the preparatory patches that didn't actually do any semantic changes earlier, but this merges the actual _reason_ for those preparatory patches. The "lockref" structure is a combination "spinlock and reference count" that allows optimized reference count accesses. In particular, it guarantees that the reference count will be updated AS IF the spinlock was held, but using atomic accesses that cover both the reference count and the spinlock words, we can often do the update without actually having to take the lock. This allows us to avoid the nastiest cases of spinlock contention on large machines under heavy pathname lookup loads. When updating the dentry reference counts on a large system, we'll still end up with the cache line bouncing around, but that's much less noticeable than actually having to spin waiting for the lock. * lockref: lockref: implement lockless reference count updates using cmpxchg() lockref: uninline lockref helper functions vfs: reimplement d_rcu_to_refcount() using lockref_get_or_lock() vfs: use lockref_get_not_zero() for optimistic lockless dget_parent() lockref: add 'lockref_get_or_lock() helper
Diffstat (limited to 'fs')
-rw-r--r--fs/dcache.c17
-rw-r--r--fs/namei.c90
2 files changed, 80 insertions, 27 deletions
diff --git a/fs/dcache.c b/fs/dcache.c
index b949af8..96655f4 100644
--- a/fs/dcache.c
+++ b/fs/dcache.c
@@ -611,8 +611,23 @@ static inline void __dget(struct dentry *dentry)
struct dentry *dget_parent(struct dentry *dentry)
{
+ int gotref;
struct dentry *ret;
+ /*
+ * Do optimistic parent lookup without any
+ * locking.
+ */
+ rcu_read_lock();
+ ret = ACCESS_ONCE(dentry->d_parent);
+ gotref = lockref_get_not_zero(&ret->d_lockref);
+ rcu_read_unlock();
+ if (likely(gotref)) {
+ if (likely(ret == ACCESS_ONCE(dentry->d_parent)))
+ return ret;
+ dput(ret);
+ }
+
repeat:
/*
* Don't need rcu_dereference because we re-check it was correct under
@@ -1771,7 +1786,7 @@ static noinline enum slow_d_compare slow_dentry_cmp(
* without taking d_lock and checking d_seq sequence count against @seq
* returned here.
*
- * A refcount may be taken on the found dentry with the __d_rcu_to_refcount
+ * A refcount may be taken on the found dentry with the d_rcu_to_refcount
* function.
*
* Alternatively, __d_lookup_rcu may be called again to look up the child of
diff --git a/fs/namei.c b/fs/namei.c
index 7720fbd..2c30c84 100644
--- a/fs/namei.c
+++ b/fs/namei.c
@@ -494,6 +494,50 @@ static inline void unlock_rcu_walk(void)
br_read_unlock(&vfsmount_lock);
}
+/*
+ * When we move over from the RCU domain to properly refcounted
+ * long-lived dentries, we need to check the sequence numbers
+ * we got before lookup very carefully.
+ *
+ * We cannot blindly increment a dentry refcount - even if it
+ * is not locked - if it is zero, because it may have gone
+ * through the final d_kill() logic already.
+ *
+ * So for a zero refcount, we need to get the spinlock (which is
+ * safe even for a dead dentry because the de-allocation is
+ * RCU-delayed), and check the sequence count under the lock.
+ *
+ * Once we have checked the sequence count, we know it is live,
+ * and since we hold the spinlock it cannot die from under us.
+ *
+ * In contrast, if the reference count wasn't zero, we can just
+ * increment the lockref without having to take the spinlock.
+ * Even if the sequence number ends up being stale, we haven't
+ * gone through the final dput() and killed the dentry yet.
+ */
+static inline int d_rcu_to_refcount(struct dentry *dentry, seqcount_t *validate, unsigned seq)
+{
+ int gotref;
+
+ gotref = lockref_get_or_lock(&dentry->d_lockref);
+
+ /* Does the sequence number still match? */
+ if (read_seqcount_retry(validate, seq)) {
+ if (gotref)
+ dput(dentry);
+ else
+ spin_unlock(&dentry->d_lock);
+ return -ECHILD;
+ }
+
+ /* Get the ref now, if we couldn't get it originally */
+ if (!gotref) {
+ dentry->d_lockref.count++;
+ spin_unlock(&dentry->d_lock);
+ }
+ return 0;
+}
+
/**
* unlazy_walk - try to switch to ref-walk mode.
* @nd: nameidata pathwalk data
@@ -518,29 +562,28 @@ static int unlazy_walk(struct nameidata *nd, struct dentry *dentry)
nd->root.dentry != fs->root.dentry)
goto err_root;
}
- spin_lock(&parent->d_lock);
+
+ /*
+ * For a negative lookup, the lookup sequence point is the parents
+ * sequence point, and it only needs to revalidate the parent dentry.
+ *
+ * For a positive lookup, we need to move both the parent and the
+ * dentry from the RCU domain to be properly refcounted. And the
+ * sequence number in the dentry validates *both* dentry counters,
+ * since we checked the sequence number of the parent after we got
+ * the child sequence number. So we know the parent must still
+ * be valid if the child sequence number is still valid.
+ */
if (!dentry) {
- if (!__d_rcu_to_refcount(parent, nd->seq))
- goto err_parent;
+ if (d_rcu_to_refcount(parent, &parent->d_seq, nd->seq) < 0)
+ goto err_root;
BUG_ON(nd->inode != parent->d_inode);
} else {
- if (dentry->d_parent != parent)
+ if (d_rcu_to_refcount(dentry, &dentry->d_seq, nd->seq) < 0)
+ goto err_root;
+ if (d_rcu_to_refcount(parent, &dentry->d_seq, nd->seq) < 0)
goto err_parent;
- spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
- if (!__d_rcu_to_refcount(dentry, nd->seq))
- goto err_child;
- /*
- * If the sequence check on the child dentry passed, then
- * the child has not been removed from its parent. This
- * means the parent dentry must be valid and able to take
- * a reference at this point.
- */
- BUG_ON(!IS_ROOT(dentry) && dentry->d_parent != parent);
- BUG_ON(!parent->d_lockref.count);
- parent->d_lockref.count++;
- spin_unlock(&dentry->d_lock);
}
- spin_unlock(&parent->d_lock);
if (want_root) {
path_get(&nd->root);
spin_unlock(&fs->lock);
@@ -551,10 +594,8 @@ static int unlazy_walk(struct nameidata *nd, struct dentry *dentry)
nd->flags &= ~LOOKUP_RCU;
return 0;
-err_child:
- spin_unlock(&dentry->d_lock);
err_parent:
- spin_unlock(&parent->d_lock);
+ dput(dentry);
err_root:
if (want_root)
spin_unlock(&fs->lock);
@@ -585,14 +626,11 @@ static int complete_walk(struct nameidata *nd)
nd->flags &= ~LOOKUP_RCU;
if (!(nd->flags & LOOKUP_ROOT))
nd->root.mnt = NULL;
- spin_lock(&dentry->d_lock);
- if (unlikely(!__d_rcu_to_refcount(dentry, nd->seq))) {
- spin_unlock(&dentry->d_lock);
+
+ if (d_rcu_to_refcount(dentry, &dentry->d_seq, nd->seq) < 0) {
unlock_rcu_walk();
return -ECHILD;
}
- BUG_ON(nd->inode != dentry->d_inode);
- spin_unlock(&dentry->d_lock);
mntget(nd->path.mnt);
unlock_rcu_walk();
}