Whamcloud - gitweb
LU-1994 llite: atomic_open support
[fs/lustre-release.git] / lustre / llite / namei.c
index eca2929..48303a3 100644 (file)
@@ -51,6 +51,9 @@
 #include <lustre_ver.h>
 #include "llite_internal.h"
 
+static int ll_create_it(struct inode *, struct dentry *,
+                       int, struct lookup_intent *);
+
 /*
  * Check if we have something mounted at the named dchild.
  * In such a case there would always be dentry present.
@@ -168,9 +171,10 @@ struct inode *ll_iget(struct super_block *sb, ino_t hash,
 static void ll_invalidate_negative_children(struct inode *dir)
 {
        struct dentry *dentry, *tmp_subdir;
+       struct ll_d_hlist_node *p;
 
        ll_lock_dcache(dir);
-       list_for_each_entry(dentry, &dir->i_dentry, d_alias) {
+       ll_d_hlist_for_each_entry(dentry, p, &dir->i_dentry, d_alias) {
                spin_lock(&dentry->d_lock);
                if (!list_empty(&dentry->d_subdirs)) {
                        struct dentry *child;
@@ -342,14 +346,15 @@ void ll_i2gids(__u32 *suppgids, struct inode *i1, struct inode *i2)
 static struct dentry *ll_find_alias(struct inode *inode, struct dentry *dentry)
 {
        struct dentry *alias, *discon_alias, *invalid_alias;
+       struct ll_d_hlist_node *p;
 
-       if (list_empty(&inode->i_dentry))
+       if (ll_d_hlist_empty(&inode->i_dentry))
                return NULL;
 
        discon_alias = invalid_alias = NULL;
 
        ll_lock_dcache(inode);
-       list_for_each_entry(alias, &inode->i_dentry, d_alias) {
+       ll_d_hlist_for_each_entry(alias, p, &inode->i_dentry, d_alias) {
                LASSERT(alias != dentry);
 
                spin_lock(&alias->d_lock);
@@ -419,6 +424,8 @@ int ll_lookup_it_finish(struct ptlrpc_request *request,
 
        /* NB 1 request reference will be taken away by ll_intent_lock()
         * when I return */
+       CDEBUG(D_DENTRY, "it %p it_disposition %x\n", it,
+              it->d.lustre.it_disposition);
        if (!it_disposition(it, DISP_LOOKUP_NEG)) {
                 rc = ll_prep_inode(&inode, request, (*de)->d_sb);
                 if (rc)
@@ -436,13 +443,21 @@ int ll_lookup_it_finish(struct ptlrpc_request *request,
                    Also see bug 7198. */
        }
 
-       *de = ll_splice_alias(inode, *de);
+       /* Only hash *de if it is unhashed (new dentry).
+        * Atoimc_open may passin hashed dentries for open.
+        */
+       if (d_unhashed(*de))
+               *de = ll_splice_alias(inode, *de);
 
        if (!it_disposition(it, DISP_LOOKUP_NEG)) {
                /* we have lookup look - unhide dentry */
                if (bits & MDS_INODELOCK_LOOKUP)
                        d_lustre_revalidate(*de);
-       } else {
+       } else if (!it_disposition(it, DISP_OPEN_CREATE)) {
+               /* If file created on server, don't depend on parent UPDATE
+                * lock to unhide it. It is left hidden and next lookup can
+                * find it in ll_splice_alias.
+                */
                /* Check that parent has UPDATE lock. */
                struct lookup_intent parent_it = {
                                        .it_op = IT_GETATTR,
@@ -516,7 +531,9 @@ static struct dentry *ll_lookup_it(struct inode *parent, struct dentry *dentry,
         if (IS_ERR(op_data))
                 RETURN((void *)op_data);
 
-        it->it_create_mode &= ~cfs_curproc_umask();
+        /* enforce umask if acl disabled or MDS doesn't support umask */
+        if (!IS_POSIXACL(parent) || !exp_connect_umask(ll_i2mdexp(parent)))
+                it->it_create_mode &= ~cfs_curproc_umask();
 
         rc = md_intent_lock(ll_i2mdexp(parent), op_data, NULL, 0, it,
                             lookup_flags, &req, ll_md_blocking_ast, 0);
@@ -549,6 +566,121 @@ static struct dentry *ll_lookup_it(struct inode *parent, struct dentry *dentry,
         return retval;
 }
 
+#ifdef HAVE_IOP_ATOMIC_OPEN
+static struct dentry *ll_lookup_nd(struct inode *parent, struct dentry *dentry,
+                                  unsigned int flags)
+{
+       struct lookup_intent *itp, it = { .it_op = IT_GETATTR };
+       struct dentry *de;
+
+       CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s,dir=%lu/%u(%p),flags=%u\n",
+              dentry->d_name.len, dentry->d_name.name, parent->i_ino,
+              parent->i_generation, parent, flags);
+
+       /* Optimize away (CREATE && !OPEN). Let .create handle the race. */
+       if ((flags & LOOKUP_CREATE ) && !(flags & LOOKUP_OPEN)) {
+               ll_dops_init(dentry, 1, 1);
+               __d_lustre_invalidate(dentry);
+               d_add(dentry, NULL);
+               return NULL;
+       }
+
+       if (flags & (LOOKUP_PARENT|LOOKUP_OPEN|LOOKUP_CREATE))
+               itp = NULL;
+       else
+               itp = &it;
+       de = ll_lookup_it(parent, dentry, itp, 0);
+
+       if (itp != NULL)
+               ll_intent_release(itp);
+
+       return de;
+}
+
+/*
+ * For cached negative dentry and new dentry, handle lookup/create/open
+ * together.
+ */
+static int ll_atomic_open(struct inode *dir, struct dentry *dentry,
+                         struct file *file, unsigned open_flags,
+                         umode_t mode, int *opened)
+{
+       struct lookup_intent *it;
+       struct dentry *de;
+       long long lookup_flags = LOOKUP_OPEN;
+       int rc = 0;
+       ENTRY;
+
+       CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s,dir=%lu/%u(%p),file %p,"
+                          "open_flags %x,mode %x opened %d\n",
+              dentry->d_name.len, dentry->d_name.name, dir->i_ino,
+              dir->i_generation, dir, file, open_flags, mode, *opened);
+
+       OBD_ALLOC(it, sizeof(*it));
+       if (!it)
+               RETURN(-ENOMEM);
+
+       it->it_op = IT_OPEN;
+       if (mode) {
+               it->it_op |= IT_CREAT;
+               lookup_flags |= LOOKUP_CREATE;
+       }
+       it->it_create_mode = (mode & S_IALLUGO) | S_IFREG;
+       it->it_flags = (open_flags & ~O_ACCMODE) | OPEN_FMODE(open_flags);
+
+       /* Dentry added to dcache tree in ll_lookup_it */
+       de = ll_lookup_it(dir, dentry, it, lookup_flags);
+       if (IS_ERR(de))
+               rc = PTR_ERR(de);
+       else if (de != NULL)
+               dentry = de;
+
+       if (!rc) {
+               if (it_disposition(it, DISP_OPEN_CREATE)) {
+                       /* Dentry instantiated in ll_create_it. */
+                       rc = ll_create_it(dir, dentry, mode, it);
+                       if (rc) {
+                               /* We dget in ll_splice_alias. */
+                               if (de != NULL)
+                                       dput(de);
+                               goto out_release;
+                       }
+
+                       *opened |= FILE_CREATED;
+               }
+               if (dentry->d_inode && it_disposition(it, DISP_OPEN_OPEN)) {
+                       /* Open dentry. */
+                       if (S_ISFIFO(dentry->d_inode->i_mode)) {
+                               /* We cannot call open here as it would
+                                * deadlock.
+                                */
+                               if (it_disposition(it, DISP_ENQ_OPEN_REF))
+                                       ptlrpc_req_finished(
+                                                      (struct ptlrpc_request *)
+                                                         it->d.lustre.it_data);
+                               rc = finish_no_open(file, de);
+                       } else {
+                               file->private_data = it;
+                               rc = finish_open(file, dentry, NULL, opened);
+                               /* We dget in ll_splice_alias. finish_open takes
+                                * care of dget for fd open.
+                                */
+                               if (de != NULL)
+                                       dput(de);
+                       }
+               } else {
+                       rc = finish_no_open(file, de);
+               }
+       }
+
+out_release:
+       ll_intent_release(it);
+       OBD_FREE(it, sizeof(*it));
+
+       RETURN(rc);
+}
+
+#else /* !HAVE_IOP_ATOMIC_OPEN */
 struct lookup_intent *ll_convert_intent(struct open_intent *oit,
                                         int lookup_flags)
 {
@@ -643,6 +775,7 @@ static struct dentry *ll_lookup_nd(struct inode *parent, struct dentry *dentry,
 
         RETURN(de);
 }
+#endif /* HAVE_IOP_ATOMIC_OPEN */
 
 /* We depend on "mode" being set with the proper file type/umask by now */
 static struct inode *ll_create_node(struct inode *dir, const char *name,
@@ -665,7 +798,7 @@ static struct inode *ll_create_node(struct inode *dir, const char *name,
         if (rc)
                 GOTO(out, inode = ERR_PTR(rc));
 
-        LASSERT(list_empty(&inode->i_dentry));
+       LASSERT(ll_d_hlist_empty(&inode->i_dentry));
 
         /* We asked for a lock on the directory, but were granted a
          * lock on the inode.  Since we finally have an inode pointer,
@@ -788,7 +921,8 @@ static int ll_mknod_generic(struct inode *dir, struct qstr *name, int mode,
                name->len, name->name, dir->i_ino, dir->i_generation, dir,
                mode, rdev);
 
-        mode &= ~cfs_curproc_umask();
+        if (!IS_POSIXACL(dir) || !exp_connect_umask(ll_i2mdexp(dir)))
+                mode &= ~cfs_curproc_umask();
 
         switch (mode & S_IFMT) {
         case 0:
@@ -814,8 +948,32 @@ static int ll_mknod_generic(struct inode *dir, struct qstr *name, int mode,
         RETURN(err);
 }
 
+#ifdef HAVE_IOP_ATOMIC_OPEN
+/*
+ * Plain create. Intent create is handled in atomic_open.
+ */
+static int ll_create_nd(struct inode *dir, struct dentry *dentry,
+                       umode_t mode, bool want_excl)
+{
+       int rc;
+
+       CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s,dir=%lu/%u(%p),"
+                          "flags=%u, excl=%d\n",
+              dentry->d_name.len, dentry->d_name.name, dir->i_ino,
+              dir->i_generation, dir, mode, want_excl);
+
+       rc = ll_mknod_generic(dir, &dentry->d_name, mode, 0, dentry);
+
+       ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_CREATE, 1);
+
+       CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, unhashed %d\n",
+              dentry->d_name.len, dentry->d_name.name, d_unhashed(dentry));
+
+       return rc;
+}
+#else /* !HAVE_IOP_ATOMIC_OPEN */
 static int ll_create_nd(struct inode *dir, struct dentry *dentry,
-                        int mode, struct nameidata *nd)
+                       ll_umode_t mode, struct nameidata *nd)
 {
         struct lookup_intent *it = ll_d2d(dentry)->lld_it;
         int rc;
@@ -850,6 +1008,7 @@ out:
 
         return rc;
 }
+#endif /* HAVE_IOP_ATOMIC_OPEN */
 
 static int ll_symlink_generic(struct inode *dir, struct qstr *name,
                               const char *tgt, struct dentry *dchild)
@@ -912,7 +1071,9 @@ static int ll_mkdir_generic(struct inode *dir, struct qstr *name,
         CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s,dir=%lu/%u(%p)\n",
                name->len, name->name, dir->i_ino, dir->i_generation, dir);
 
-        mode = (mode & (S_IRWXUGO|S_ISVTX) & ~cfs_curproc_umask()) | S_IFDIR;
+        if (!IS_POSIXACL(dir) || !exp_connect_umask(ll_i2mdexp(dir)))
+                mode &= ~cfs_curproc_umask();
+        mode = (mode & (S_IRWXUGO|S_ISVTX)) | S_IFDIR;
         err = ll_new_node(dir, name, NULL, mode, 0, dchild, LUSTRE_OPC_MKDIR);
 
         if (!err)
@@ -926,10 +1087,10 @@ static int ll_mkdir_generic(struct inode *dir, struct qstr *name,
 static void ll_get_child_fid(struct inode * dir, struct qstr *name,
                              struct lu_fid *fid)
 {
-        struct dentry *parent, *child;
+       struct dentry *parent, *child;
 
-        parent = list_entry(dir->i_dentry.next, struct dentry, d_alias);
-        child = d_lookup(parent, name);
+       parent = ll_d_hlist_entry(dir->i_dentry, struct dentry, d_alias);
+       child = d_lookup(parent, name);
         if (child) {
                 if (child->d_inode)
                         *fid = *ll_inode2fid(child->d_inode);
@@ -1128,7 +1289,7 @@ static int ll_rename_generic(struct inode *src, struct dentry *src_dparent,
         RETURN(err);
 }
 
-static int ll_mknod(struct inode *dir, struct dentry *dchild, int mode,
+static int ll_mknod(struct inode *dir, struct dentry *dchild, ll_umode_t mode,
                    dev_t rdev)
 {
        return ll_mknod_generic(dir, &dchild->d_name, mode,
@@ -1139,25 +1300,30 @@ static int ll_unlink(struct inode * dir, struct dentry *dentry)
 {
         return ll_unlink_generic(dir, NULL, dentry, &dentry->d_name);
 }
-static int ll_mkdir(struct inode *dir, struct dentry *dentry, int mode)
+
+static int ll_mkdir(struct inode *dir, struct dentry *dentry, ll_umode_t mode)
 {
         return ll_mkdir_generic(dir, &dentry->d_name, mode, dentry);
 }
+
 static int ll_rmdir(struct inode *dir, struct dentry *dentry)
 {
         return ll_rmdir_generic(dir, NULL, dentry, &dentry->d_name);
 }
+
 static int ll_symlink(struct inode *dir, struct dentry *dentry,
                       const char *oldname)
 {
         return ll_symlink_generic(dir, &dentry->d_name, oldname, dentry);
 }
+
 static int ll_link(struct dentry *old_dentry, struct inode *dir,
                    struct dentry *new_dentry)
 {
         return ll_link_generic(old_dentry->d_inode, dir, &new_dentry->d_name,
                                new_dentry);
 }
+
 static int ll_rename(struct inode *old_dir, struct dentry *old_dentry,
                      struct inode *new_dir, struct dentry *new_dentry)
 {
@@ -1177,6 +1343,9 @@ static int ll_rename(struct inode *old_dir, struct dentry *old_dentry,
 
 struct inode_operations ll_dir_inode_operations = {
        .mknod              = ll_mknod,
+#ifdef HAVE_IOP_ATOMIC_OPEN
+       .atomic_open        = ll_atomic_open,
+#endif
        .lookup             = ll_lookup_nd,
        .create             = ll_create_nd,
        /* We need all these non-raw things for NFSD, to not patch it. */