vfs: trivial __d_lookup_rcu() cleanups
[linux-2.6.git] / fs / nfs / unlink.c
index a2242af..4f9319a 100644 (file)
 #include <linux/nfs_fs.h>
 #include <linux/sched.h>
 #include <linux/wait.h>
+#include <linux/namei.h>
 
 #include "internal.h"
 #include "nfs4_fs.h"
+#include "iostat.h"
+#include "delegation.h"
 
 struct nfs_unlinkdata {
        struct hlist_node list;
@@ -84,7 +87,7 @@ static void nfs_async_unlink_done(struct rpc_task *task, void *calldata)
        struct inode *dir = data->dir;
 
        if (!NFS_PROTO(dir)->unlink_done(task, dir))
-               nfs_restart_rpc(task, NFS_SERVER(dir)->nfs_client);
+               rpc_restart_call_prepare(task);
 }
 
 /**
@@ -110,7 +113,7 @@ void nfs_unlink_prepare(struct rpc_task *task, void *calldata)
        struct nfs_unlinkdata *data = calldata;
        struct nfs_server *server = NFS_SERVER(data->dir);
 
-       if (nfs4_setup_sequence(server->nfs_client, &data->args.seq_args,
+       if (nfs4_setup_sequence(server, &data->args.seq_args,
                                &data->res.seq_res, 1, task))
                return;
        rpc_call_start(task);
@@ -144,23 +147,33 @@ static int nfs_do_call_unlink(struct dentry *parent, struct inode *dir, struct n
 
        alias = d_lookup(parent, &data->args.name);
        if (alias != NULL) {
-               int ret = 0;
+               int ret;
+               void *devname_garbage = NULL;
 
                /*
                 * Hey, we raced with lookup... See if we need to transfer
                 * the sillyrename information to the aliased dentry.
                 */
                nfs_free_dname(data);
+               ret = nfs_copy_dname(alias, data);
                spin_lock(&alias->d_lock);
-               if (alias->d_inode != NULL &&
+               if (ret == 0 && alias->d_inode != NULL &&
                    !(alias->d_flags & DCACHE_NFSFS_RENAMED)) {
+                       devname_garbage = alias->d_fsdata;
                        alias->d_fsdata = data;
                        alias->d_flags |= DCACHE_NFSFS_RENAMED;
                        ret = 1;
-               }
+               } else
+                       ret = 0;
                spin_unlock(&alias->d_lock);
                nfs_dec_sillycount(dir);
                dput(alias);
+               /*
+                * If we'd displaced old cached devname, free it.  At that
+                * point dentry is definitely not a root, so we won't need
+                * that anymore.
+                */
+               kfree(devname_garbage);
                return ret;
        }
        data->dir = igrab(dir);
@@ -177,7 +190,7 @@ static int nfs_do_call_unlink(struct dentry *parent, struct inode *dir, struct n
        task_setup_data.rpc_client = NFS_CLIENT(dir);
        task = rpc_run_task(&task_setup_data);
        if (!IS_ERR(task))
-               rpc_put_task(task);
+               rpc_put_task_async(task);
        return 1;
 }
 
@@ -192,8 +205,6 @@ static int nfs_call_unlink(struct dentry *dentry, struct nfs_unlinkdata *data)
        if (parent == NULL)
                goto out_free;
        dir = parent->d_inode;
-       if (nfs_copy_dname(dentry, data) != 0)
-               goto out_dput;
        /* Non-exclusive lock protects against concurrent lookup() calls */
        spin_lock(&dir->i_lock);
        if (atomic_inc_not_zero(&NFS_I(dir)->silly_count) == 0) {
@@ -244,11 +255,12 @@ void nfs_unblock_sillyrename(struct dentry *dentry)
  * @dir: parent directory of dentry
  * @dentry: dentry to unlink
  */
-int
+static int
 nfs_async_unlink(struct inode *dir, struct dentry *dentry)
 {
        struct nfs_unlinkdata *data;
        int status = -ENOMEM;
+       void *devname_garbage = NULL;
 
        data = kzalloc(sizeof(*data), GFP_KERNEL);
        if (data == NULL)
@@ -259,7 +271,6 @@ nfs_async_unlink(struct inode *dir, struct dentry *dentry)
                status = PTR_ERR(data->cred);
                goto out_free;
        }
-       data->res.seq_res.sr_slotid = NFS4_MAX_SLOT_TABLE;
        data->res.dir_attr = &data->dir_attr;
 
        status = -EBUSY;
@@ -267,8 +278,16 @@ nfs_async_unlink(struct inode *dir, struct dentry *dentry)
        if (dentry->d_flags & DCACHE_NFSFS_RENAMED)
                goto out_unlock;
        dentry->d_flags |= DCACHE_NFSFS_RENAMED;
+       devname_garbage = dentry->d_fsdata;
        dentry->d_fsdata = data;
        spin_unlock(&dentry->d_lock);
+       /*
+        * If we'd displaced old cached devname, free it.  At that
+        * point dentry is definitely not a root, so we won't need
+        * that anymore.
+        */
+       if (devname_garbage)
+               kfree(devname_garbage);
        return 0;
 out_unlock:
        spin_unlock(&dentry->d_lock);
@@ -297,9 +316,282 @@ nfs_complete_unlink(struct dentry *dentry, struct inode *inode)
        if (dentry->d_flags & DCACHE_NFSFS_RENAMED) {
                dentry->d_flags &= ~DCACHE_NFSFS_RENAMED;
                data = dentry->d_fsdata;
+               dentry->d_fsdata = NULL;
        }
        spin_unlock(&dentry->d_lock);
 
        if (data != NULL && (NFS_STALE(inode) || !nfs_call_unlink(dentry, data)))
                nfs_free_unlinkdata(data);
 }
+
+/* Cancel a queued async unlink. Called when a sillyrename run fails. */
+static void
+nfs_cancel_async_unlink(struct dentry *dentry)
+{
+       spin_lock(&dentry->d_lock);
+       if (dentry->d_flags & DCACHE_NFSFS_RENAMED) {
+               struct nfs_unlinkdata *data = dentry->d_fsdata;
+
+               dentry->d_flags &= ~DCACHE_NFSFS_RENAMED;
+               dentry->d_fsdata = NULL;
+               spin_unlock(&dentry->d_lock);
+               nfs_free_unlinkdata(data);
+               return;
+       }
+       spin_unlock(&dentry->d_lock);
+}
+
+struct nfs_renamedata {
+       struct nfs_renameargs   args;
+       struct nfs_renameres    res;
+       struct rpc_cred         *cred;
+       struct inode            *old_dir;
+       struct dentry           *old_dentry;
+       struct nfs_fattr        old_fattr;
+       struct inode            *new_dir;
+       struct dentry           *new_dentry;
+       struct nfs_fattr        new_fattr;
+};
+
+/**
+ * nfs_async_rename_done - Sillyrename post-processing
+ * @task: rpc_task of the sillyrename
+ * @calldata: nfs_renamedata for the sillyrename
+ *
+ * Do the directory attribute updates and the d_move
+ */
+static void nfs_async_rename_done(struct rpc_task *task, void *calldata)
+{
+       struct nfs_renamedata *data = calldata;
+       struct inode *old_dir = data->old_dir;
+       struct inode *new_dir = data->new_dir;
+       struct dentry *old_dentry = data->old_dentry;
+       struct dentry *new_dentry = data->new_dentry;
+
+       if (!NFS_PROTO(old_dir)->rename_done(task, old_dir, new_dir)) {
+               rpc_restart_call_prepare(task);
+               return;
+       }
+
+       if (task->tk_status != 0) {
+               nfs_cancel_async_unlink(old_dentry);
+               return;
+       }
+
+       d_drop(old_dentry);
+       d_drop(new_dentry);
+}
+
+/**
+ * nfs_async_rename_release - Release the sillyrename data.
+ * @calldata: the struct nfs_renamedata to be released
+ */
+static void nfs_async_rename_release(void *calldata)
+{
+       struct nfs_renamedata   *data = calldata;
+       struct super_block *sb = data->old_dir->i_sb;
+
+       if (data->old_dentry->d_inode)
+               nfs_mark_for_revalidate(data->old_dentry->d_inode);
+
+       dput(data->old_dentry);
+       dput(data->new_dentry);
+       iput(data->old_dir);
+       iput(data->new_dir);
+       nfs_sb_deactive(sb);
+       put_rpccred(data->cred);
+       kfree(data);
+}
+
+#if defined(CONFIG_NFS_V4_1)
+static void nfs_rename_prepare(struct rpc_task *task, void *calldata)
+{
+       struct nfs_renamedata *data = calldata;
+       struct nfs_server *server = NFS_SERVER(data->old_dir);
+
+       if (nfs4_setup_sequence(server, &data->args.seq_args,
+                               &data->res.seq_res, 1, task))
+               return;
+       rpc_call_start(task);
+}
+#endif /* CONFIG_NFS_V4_1 */
+
+static const struct rpc_call_ops nfs_rename_ops = {
+       .rpc_call_done = nfs_async_rename_done,
+       .rpc_release = nfs_async_rename_release,
+#if defined(CONFIG_NFS_V4_1)
+       .rpc_call_prepare = nfs_rename_prepare,
+#endif /* CONFIG_NFS_V4_1 */
+};
+
+/**
+ * nfs_async_rename - perform an asynchronous rename operation
+ * @old_dir: directory that currently holds the dentry to be renamed
+ * @new_dir: target directory for the rename
+ * @old_dentry: original dentry to be renamed
+ * @new_dentry: dentry to which the old_dentry should be renamed
+ *
+ * It's expected that valid references to the dentries and inodes are held
+ */
+static struct rpc_task *
+nfs_async_rename(struct inode *old_dir, struct inode *new_dir,
+                struct dentry *old_dentry, struct dentry *new_dentry)
+{
+       struct nfs_renamedata *data;
+       struct rpc_message msg = { };
+       struct rpc_task_setup task_setup_data = {
+               .rpc_message = &msg,
+               .callback_ops = &nfs_rename_ops,
+               .workqueue = nfsiod_workqueue,
+               .rpc_client = NFS_CLIENT(old_dir),
+               .flags = RPC_TASK_ASYNC,
+       };
+
+       data = kzalloc(sizeof(*data), GFP_KERNEL);
+       if (data == NULL)
+               return ERR_PTR(-ENOMEM);
+       task_setup_data.callback_data = data;
+
+       data->cred = rpc_lookup_cred();
+       if (IS_ERR(data->cred)) {
+               struct rpc_task *task = ERR_CAST(data->cred);
+               kfree(data);
+               return task;
+       }
+
+       msg.rpc_argp = &data->args;
+       msg.rpc_resp = &data->res;
+       msg.rpc_cred = data->cred;
+
+       /* set up nfs_renamedata */
+       data->old_dir = old_dir;
+       ihold(old_dir);
+       data->new_dir = new_dir;
+       ihold(new_dir);
+       data->old_dentry = dget(old_dentry);
+       data->new_dentry = dget(new_dentry);
+       nfs_fattr_init(&data->old_fattr);
+       nfs_fattr_init(&data->new_fattr);
+
+       /* set up nfs_renameargs */
+       data->args.old_dir = NFS_FH(old_dir);
+       data->args.old_name = &old_dentry->d_name;
+       data->args.new_dir = NFS_FH(new_dir);
+       data->args.new_name = &new_dentry->d_name;
+
+       /* set up nfs_renameres */
+       data->res.old_fattr = &data->old_fattr;
+       data->res.new_fattr = &data->new_fattr;
+
+       nfs_sb_active(old_dir->i_sb);
+
+       NFS_PROTO(data->old_dir)->rename_setup(&msg, old_dir);
+
+       return rpc_run_task(&task_setup_data);
+}
+
+/**
+ * nfs_sillyrename - Perform a silly-rename of a dentry
+ * @dir: inode of directory that contains dentry
+ * @dentry: dentry to be sillyrenamed
+ *
+ * NFSv2/3 is stateless and the server doesn't know when the client is
+ * holding a file open. To prevent application problems when a file is
+ * unlinked while it's still open, the client performs a "silly-rename".
+ * That is, it renames the file to a hidden file in the same directory,
+ * and only performs the unlink once the last reference to it is put.
+ *
+ * The final cleanup is done during dentry_iput.
+ *
+ * (Note: NFSv4 is stateful, and has opens, so in theory an NFSv4 server
+ * could take responsibility for keeping open files referenced.  The server
+ * would also need to ensure that opened-but-deleted files were kept over
+ * reboots.  However, we may not assume a server does so.  (RFC 5661
+ * does provide an OPEN4_RESULT_PRESERVE_UNLINKED flag that a server can
+ * use to advertise that it does this; some day we may take advantage of
+ * it.))
+ */
+int
+nfs_sillyrename(struct inode *dir, struct dentry *dentry)
+{
+       static unsigned int sillycounter;
+       const int      fileidsize  = sizeof(NFS_FILEID(dentry->d_inode))*2;
+       const int      countersize = sizeof(sillycounter)*2;
+       const int      slen        = sizeof(".nfs")+fileidsize+countersize-1;
+       char           silly[slen+1];
+       struct dentry *sdentry;
+       struct rpc_task *task;
+       int            error = -EIO;
+
+       dfprintk(VFS, "NFS: silly-rename(%s/%s, ct=%d)\n",
+               dentry->d_parent->d_name.name, dentry->d_name.name,
+               dentry->d_count);
+       nfs_inc_stats(dir, NFSIOS_SILLYRENAME);
+
+       /*
+        * We don't allow a dentry to be silly-renamed twice.
+        */
+       error = -EBUSY;
+       if (dentry->d_flags & DCACHE_NFSFS_RENAMED)
+               goto out;
+
+       sprintf(silly, ".nfs%*.*Lx",
+               fileidsize, fileidsize,
+               (unsigned long long)NFS_FILEID(dentry->d_inode));
+
+       /* Return delegation in anticipation of the rename */
+       nfs_inode_return_delegation(dentry->d_inode);
+
+       sdentry = NULL;
+       do {
+               char *suffix = silly + slen - countersize;
+
+               dput(sdentry);
+               sillycounter++;
+               sprintf(suffix, "%*.*x", countersize, countersize, sillycounter);
+
+               dfprintk(VFS, "NFS: trying to rename %s to %s\n",
+                               dentry->d_name.name, silly);
+
+               sdentry = lookup_one_len(silly, dentry->d_parent, slen);
+               /*
+                * N.B. Better to return EBUSY here ... it could be
+                * dangerous to delete the file while it's in use.
+                */
+               if (IS_ERR(sdentry))
+                       goto out;
+       } while (sdentry->d_inode != NULL); /* need negative lookup */
+
+       /* queue unlink first. Can't do this from rpc_release as it
+        * has to allocate memory
+        */
+       error = nfs_async_unlink(dir, dentry);
+       if (error)
+               goto out_dput;
+
+       /* populate unlinkdata with the right dname */
+       error = nfs_copy_dname(sdentry,
+                               (struct nfs_unlinkdata *)dentry->d_fsdata);
+       if (error) {
+               nfs_cancel_async_unlink(dentry);
+               goto out_dput;
+       }
+
+       /* run the rename task, undo unlink if it fails */
+       task = nfs_async_rename(dir, dir, dentry, sdentry);
+       if (IS_ERR(task)) {
+               error = -EBUSY;
+               nfs_cancel_async_unlink(dentry);
+               goto out_dput;
+       }
+
+       /* wait for the RPC task to complete, unless a SIGKILL intervenes */
+       error = rpc_wait_for_completion_task(task);
+       if (error == 0)
+               error = task->tk_status;
+       rpc_put_task(task);
+out_dput:
+       dput(sdentry);
+out:
+       return error;
+}