Whamcloud - gitweb
LU-1994 llite: atomic_open support
[fs/lustre-release.git] / lustre / llite / dcache.c
index 88db91b..6ce32f7 100644 (file)
@@ -36,7 +36,6 @@
 
 #include <linux/fs.h>
 #include <linux/sched.h>
-#include <linux/smp_lock.h>
 #include <linux/quotaops.h>
 
 #define DEBUG_SUBSYSTEM S_LLITE
 
 #include "llite_internal.h"
 
+static void free_dentry_data(struct rcu_head *head)
+{
+       struct ll_dentry_data *lld;
+
+       lld = container_of(head, struct ll_dentry_data, lld_rcu_head);
+       OBD_FREE_PTR(lld);
+}
+
 /* should NOT be called with the dcache lock, see fs/dcache.c */
 static void ll_release(struct dentry *de)
 {
@@ -62,15 +69,16 @@ static void ll_release(struct dentry *de)
                 ll_intent_release(lld->lld_it);
                 OBD_FREE(lld->lld_it, sizeof(*lld->lld_it));
         }
-        LASSERT(lld->lld_cwd_count == 0);
-        LASSERT(lld->lld_mnt_count == 0);
-        OBD_FREE(de->d_fsdata, sizeof(*lld));
+       LASSERT(lld->lld_cwd_count == 0);
+       LASSERT(lld->lld_mnt_count == 0);
+       de->d_fsdata = NULL;
+       call_rcu(&lld->lld_rcu_head, free_dentry_data);
 
-        EXIT;
+       EXIT;
 }
 
 /* Compare if two dentries are the same.  Don't match if the existing dentry
- * is marked DCACHE_LUSTRE_INVALID.  Returns 1 if different, 0 if the same.
+ * is marked invalid.  Returns 1 if different, 0 if the same.
  *
  * This avoids a race where ll_lookup_it() instantiates a dentry, but we get
  * an AST before calling d_revalidate_it().  The dentry still exists (marked
@@ -135,8 +143,8 @@ static inline int return_if_equal(struct ldlm_lock *lock, void *data)
  *      < 0    error */
 static int find_cbdata(struct inode *inode)
 {
-        struct ll_inode_info *lli = ll_i2info(inode);
-        struct ll_sb_info *sbi = ll_i2sbi(inode);
+       struct ll_sb_info *sbi = ll_i2sbi(inode);
+       struct lov_stripe_md *lsm;
         int rc = 0;
         ENTRY;
 
@@ -146,11 +154,14 @@ static int find_cbdata(struct inode *inode)
         if (rc != 0)
                  RETURN(rc);
 
-        if (lli->lli_smd)
-                rc = obd_find_cbdata(sbi->ll_dt_exp, lli->lli_smd,
-                                     return_if_equal, NULL);
+       lsm = ccc_inode_lsm_get(inode);
+       if (lsm == NULL)
+               RETURN(rc);
 
-        RETURN(rc);
+       rc = obd_find_cbdata(sbi->ll_dt_exp, lsm, return_if_equal, NULL);
+       ccc_inode_lsm_put(inode, lsm);
+
+       RETURN(rc);
 }
 
 /**
@@ -171,31 +182,35 @@ static int ll_ddelete(HAVE_D_DELETE_CONST struct dentry *de)
               d_unhashed((struct dentry *)de) ? "" : "hashed,",
               list_empty(&de->d_subdirs) ? "" : "subdirs");
 
+#ifdef HAVE_DCACHE_LOCK
+       LASSERT(d_refcount(de) == 0);
+#else
        /* kernel >= 2.6.38 last refcount is decreased after this function. */
-#ifdef DCACHE_OP_DELETE
        LASSERT(d_refcount(de) == 1);
-#else
-       LASSERT(d_refcount(de) == 0);
 #endif
 
+       /* Disable this piece of code temproarily because this is called
+        * inside dcache_lock so it's not appropriate to do lots of work
+        * here. */
+#if 0
        /* if not ldlm lock for this inode, set i_nlink to 0 so that
         * this inode can be recycled later b=20433 */
        if (de->d_inode && !find_cbdata(de->d_inode))
-               de->d_inode->i_nlink = 0;
+               clear_nlink(de->d_inode);
+#endif
 
        if (d_lustre_invalid((struct dentry *)de))
-                RETURN(1);
-
-        RETURN(0);
+               RETURN(1);
+       RETURN(0);
 }
 
 static int ll_set_dd(struct dentry *de)
 {
-        ENTRY;
-        LASSERT(de != NULL);
+       ENTRY;
+       LASSERT(de != NULL);
 
-        CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
-               de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
+       CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
+               de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
                d_refcount(de));
 
        if (de->d_fsdata == NULL) {
@@ -209,12 +224,12 @@ static int ll_set_dd(struct dentry *de)
                        else
                                OBD_FREE_PTR(lld);
                        spin_unlock(&de->d_lock);
-                } else {
-                        RETURN(-ENOMEM);
-                }
-        }
+               } else {
+                       RETURN(-ENOMEM);
+               }
+       }
 
-        RETURN(0);
+       RETURN(0);
 }
 
 int ll_dops_init(struct dentry *de, int block, int init_sa)
@@ -233,10 +248,11 @@ int ll_dops_init(struct dentry *de, int block, int init_sa)
         if (lld != NULL && init_sa != 0)
                 lld->lld_sa_generation = 0;
 
-#ifdef DCACHE_OP_HASH
-       LASSERT(de->d_op == &ll_d_ops);
-#else
+#ifdef HAVE_DCACHE_LOCK
        de->d_op = &ll_d_ops;
+#else
+       /* kernel >= 2.6.38 d_op is set in d_alloc() */
+       LASSERT(de->d_op == &ll_d_ops);
 #endif
         return rc;
 }
@@ -246,7 +262,20 @@ void ll_intent_drop_lock(struct lookup_intent *it)
         struct lustre_handle *handle;
 
         if (it->it_op && it->d.lustre.it_lock_mode) {
-                handle = (struct lustre_handle *)&it->d.lustre.it_lock_handle;
+               struct ldlm_lock *lock;
+
+               handle = (struct lustre_handle *)&it->d.lustre.it_lock_handle;
+               lock = ldlm_handle2lock(handle);
+               if (lock != NULL) {
+                       /* it can only be allowed to match after layout is
+                        * applied to inode otherwise false layout would be
+                        * seen. Applying layout shoud happen before dropping
+                        * the intent lock. */
+                       if (it->d.lustre.it_lock_bits & MDS_INODELOCK_LAYOUT)
+                               ldlm_lock_allow_match(lock);
+                       LDLM_LOCK_PUT(lock);
+               }
+
                 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
                        " from it %p\n", handle->cookie, it);
                 ldlm_lock_decref(handle, it->d.lustre.it_lock_mode);
@@ -280,6 +309,7 @@ void ll_intent_release(struct lookup_intent *it)
 void ll_invalidate_aliases(struct inode *inode)
 {
        struct dentry *dentry;
+       struct ll_d_hlist_node *p;
        ENTRY;
 
        LASSERT(inode != NULL);
@@ -288,11 +318,11 @@ void ll_invalidate_aliases(struct inode *inode)
               inode->i_ino, inode->i_generation, inode);
 
        ll_lock_dcache(inode);
-       cfs_list_for_each_entry(dentry, &inode->i_dentry, d_alias) {
-                CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
-                       "inode %p flags %d\n", dentry->d_name.len,
-                       dentry->d_name.name, dentry, dentry->d_parent,
-                       dentry->d_inode, dentry->d_flags);
+       ll_d_hlist_for_each_entry(dentry, p, &inode->i_dentry, d_alias) {
+               CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
+                      "inode %p flags %d\n", dentry->d_name.len,
+                      dentry->d_name.name, dentry, dentry->d_parent,
+                      dentry->d_inode, dentry->d_flags);
 
                 if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') {
                         CERROR("called on root (?) dentry=%p, inode=%p "
@@ -413,12 +443,12 @@ int ll_revalidate_it(struct dentry *de, int lookup_flags,
        if (it->it_op == IT_LOOKUP && !d_lustre_invalid(de))
                 RETURN(1);
 
-        if ((it->it_op == IT_OPEN) && de->d_inode) {
-                struct inode *inode = de->d_inode;
-                struct ll_inode_info *lli = ll_i2info(inode);
-                struct obd_client_handle **och_p;
-                __u64 *och_usecount;
-                __u64 ibits;
+       if (it->it_op == IT_OPEN) {
+               struct inode *inode = de->d_inode;
+               struct ll_inode_info *lli = ll_i2info(inode);
+               struct obd_client_handle **och_p;
+               __u64 *och_usecount;
+               __u64 ibits;
 
                 /*
                  * We used to check for MDS_INODELOCK_OPEN here, but in fact
@@ -445,7 +475,7 @@ int ll_revalidate_it(struct dentry *de, int lookup_flags,
                 ibits = MDS_INODELOCK_LOOKUP;
                 if (!ll_have_md_lock(inode, &ibits, LCK_MINMODE))
                         goto do_lock;
-                cfs_mutex_lock(&lli->lli_och_mutex);
+               mutex_lock(&lli->lli_och_mutex);
                 if (*och_p) { /* Everything is open already, do nothing */
                         /*(*och_usecount)++;  Do not let them steal our open
                           handle from under us */
@@ -457,10 +487,10 @@ int ll_revalidate_it(struct dentry *de, int lookup_flags,
                            hope the lock won't be invalidated in between. But
                            if it would be, we'll reopen the open request to
                            MDS later during file open path */
-                        cfs_mutex_unlock(&lli->lli_och_mutex);
+                       mutex_unlock(&lli->lli_och_mutex);
                         RETURN(1);
                 } else {
-                        cfs_mutex_unlock(&lli->lli_och_mutex);
+                       mutex_unlock(&lli->lli_och_mutex);
                 }
         }
 
@@ -479,7 +509,8 @@ do_lock:
         if (IS_ERR(op_data))
                 RETURN(PTR_ERR(op_data));
 
-        it->it_create_mode &= ~cfs_curproc_umask();
+        if (!IS_POSIXACL(parent) || !exp_connect_umask(exp))
+                it->it_create_mode &= ~cfs_curproc_umask();
         it->it_create_mode |= M_CHECK_STALE;
         rc = md_intent_lock(exp, op_data, NULL, 0, it,
                             lookup_flags,
@@ -526,6 +557,9 @@ out:
         if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
                 ptlrpc_req_finished(req);
         if (rc == 0) {
+               /* mdt may grant layout lock for the newly created file, so
+                * release the lock to avoid leaking */
+               ll_intent_drop_lock(it);
                ll_invalidate_aliases(de->d_inode);
        } else {
                __u64 bits = 0;
@@ -608,12 +642,41 @@ out_sa:
         goto mark;
 }
 
+#ifdef HAVE_IOP_ATOMIC_OPEN
+/*
+ * Always trust cached dentries. Update statahead window if necessary.
+ */
+int ll_revalidate_nd(struct dentry *dentry, unsigned int flags)
+{
+       struct inode *parent = dentry->d_parent->d_inode;
+       int unplug = 0;
+
+       ENTRY;
+       CDEBUG(D_VFSTRACE, "VFS Op:name=%s,flags=%u\n",
+              dentry->d_name.name, flags);
+
+       if (!(flags & (LOOKUP_PARENT|LOOKUP_OPEN|LOOKUP_CREATE)) &&
+           ll_need_statahead(parent, dentry) > 0) {
+               if (flags & LOOKUP_RCU)
+                       RETURN(-ECHILD);
+
+               if (dentry->d_inode == NULL)
+                       unplug = 1;
+               do_statahead_enter(parent, &dentry, unplug);
+               ll_statahead_mark(parent, dentry);
+       }
+
+       RETURN(1);
+}
+
+#else /* !HAVE_IOP_ATOMIC_OPEN */
 int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
 {
         int rc;
         ENTRY;
 
-#ifdef LOOKUP_RCU
+#ifndef HAVE_DCACHE_LOCK
+       /* kernel >= 2.6.38 supports rcu-walk, but lustre doesn't. */
        if (nd->flags & LOOKUP_RCU)
                return -ECHILD;
 #endif
@@ -674,13 +737,14 @@ out_it:
 
         RETURN(rc);
 }
+#endif /* HAVE_IOP_ATOMIC_OPEN */
 
 void ll_d_iput(struct dentry *de, struct inode *inode)
 {
-        LASSERT(inode);
-        if (!find_cbdata(inode))
-                inode->i_nlink = 0;
-        iput(inode);
+       LASSERT(inode);
+       if (!find_cbdata(inode))
+               clear_nlink(inode);
+       iput(inode);
 }
 
 struct dentry_operations ll_d_ops = {