aufs: inode op, del, unlink

Implement i_op->unlink() with supporting logical deletion by whiteout.
Signed-off-by: J. R. Okajima <hooanon05g@gmail.com>

fs/aufs/Makefile

@@ -17,7 +17,7 @@ aufs-y := module.o sbinfo.o super.o branch.o xino.o sysaufs.o opts.o \
 	dynop.o \
 	finfo.o file.o \
 	dir.o vdir.o \
-	iinfo.o inode.o i_op.o i_op_add.o
+	iinfo.o inode.o i_op.o i_op_add.o i_op_del.o
 
 # all are boolean
 aufs-$(CONFIG_PROC_FS) += procfs.o plink.o

fs/aufs/i_op.c

@@ -647,6 +647,7 @@ struct inode_operations aufs_iop[] = {
 	[AuIop_DIR] = {
 		.create		= aufs_create,
 		.lookup		= aufs_lookup,
+		.unlink		= aufs_unlink,
 		.symlink	= aufs_symlink,
 		.mkdir		= aufs_mkdir,
 		.mknod		= aufs_mknod,

fs/aufs/i_op_del.c

+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (C) 2005-2019 Junjiro R. Okajima
+ */
+
+/*
+ * inode operations (del entry)
+ */
+
+#include <linux/iversion.h>
+#include "aufs.h"
+
+/*
+ * decide if a new whiteout for @dentry is necessary or not.
+ * when it is necessary, prepare the parent dir for the upper branch whose
+ * branch index is @bcpup for creation. the actual creation of the whiteout will
+ * be done by caller.
+ * return value:
+ * 0: wh is unnecessary
+ * plus: wh is necessary
+ * minus: error
+ */
+int au_wr_dir_need_wh(struct dentry *dentry, int isdir, aufs_bindex_t *bcpup)
+{
+	int need_wh, err;
+	aufs_bindex_t btop;
+	struct super_block *sb;
+
+	sb = dentry->d_sb;
+	btop = au_dbtop(dentry);
+	if (*bcpup < 0) {
+		*bcpup = btop;
+		if (au_test_ro(sb, btop, d_inode(dentry))) {
+			err = AuWbrCopyup(au_sbi(sb), dentry);
+			*bcpup = err;
+			if (unlikely(err < 0))
+				goto out;
+		}
+	} else
+		AuDebugOn(btop < *bcpup
+			  || au_test_ro(sb, *bcpup, d_inode(dentry)));
+	AuDbg("bcpup %d, btop %d\n", *bcpup, btop);
+
+	if (*bcpup != btop) {
+		err = au_cpup_dirs(dentry, *bcpup);
+		if (unlikely(err))
+			goto out;
+		need_wh = 1;
+	} else {
+		struct au_dinfo *dinfo, *tmp;
+
+		need_wh = -ENOMEM;
+		dinfo = au_di(dentry);
+		tmp = au_di_alloc(sb, AuLsc_DI_TMP);
+		if (tmp) {
+			au_di_cp(tmp, dinfo);
+			au_di_swap(tmp, dinfo);
+			/* returns the number of positive dentries */
+			need_wh = au_lkup_dentry(dentry, btop + 1,
+						 /* AuLkup_IGNORE_PERM */ 0);
+			au_di_swap(tmp, dinfo);
+			au_rw_write_unlock(&tmp->di_rwsem);
+			au_di_free(tmp);
+		}
+	}
+	AuDbg("need_wh %d\n", need_wh);
+	err = need_wh;
+
+out:
+	return err;
+}
+
+/*
+ * simple tests for the del-entry operations.
+ * following the checks in vfs, plus the parent-child relationship.
+ */
+int au_may_del(struct dentry *dentry, aufs_bindex_t bindex,
+	       struct dentry *h_parent, int isdir)
+{
+	int err;
+	umode_t h_mode;
+	struct dentry *h_dentry, *h_latest;
+	struct inode *h_inode;
+
+	h_dentry = au_h_dptr(dentry, bindex);
+	if (d_really_is_positive(dentry)) {
+		err = -ENOENT;
+		if (unlikely(d_is_negative(h_dentry)))
+			goto out;
+		h_inode = d_inode(h_dentry);
+		if (unlikely(!h_inode->i_nlink))
+			goto out;
+
+		h_mode = h_inode->i_mode;
+		if (!isdir) {
+			err = -EISDIR;
+			if (unlikely(S_ISDIR(h_mode)))
+				goto out;
+		} else if (unlikely(!S_ISDIR(h_mode))) {
+			err = -ENOTDIR;
+			goto out;
+		}
+	} else {
+		/* rename(2) case */
+		err = -EIO;
+		if (unlikely(d_is_positive(h_dentry)))
+			goto out;
+	}
+
+	err = -ENOENT;
+	/* expected parent dir is locked */
+	if (unlikely(h_parent != h_dentry->d_parent))
+		goto out;
+	err = 0;
+
+	/*
+	 * rmdir a dir may break the consistency on some filesystem.
+	 * let's try heavy test.
+	 */
+	err = -EACCES;
+	if (unlikely(au_test_h_perm(d_inode(h_parent), MAY_EXEC | MAY_WRITE)))
+		goto out;
+
+	h_latest = au_sio_lkup_one(&dentry->d_name, h_parent);
+	err = -EIO;
+	if (IS_ERR(h_latest))
+		goto out;
+	if (h_latest == h_dentry)
+		err = 0;
+	dput(h_latest);
+
+out:
+	return err;
+}
+
+/*
+ * decide the branch where we operate for @dentry. the branch index will be set
+ * @rbcpup. after deciding it, 'pin' it and store the timestamps of the parent
+ * dir for reverting.
+ * when a new whiteout is necessary, create it.
+ */
+static struct dentry*
+lock_hdir_create_wh(struct dentry *dentry, int isdir, aufs_bindex_t *rbcpup,
+		    struct au_dtime *dt, struct au_pin *pin)
+{
+	struct dentry *wh_dentry;
+	struct super_block *sb;
+	struct path h_path;
+	int err, need_wh;
+	unsigned int udba;
+	aufs_bindex_t bcpup;
+
+	need_wh = au_wr_dir_need_wh(dentry, isdir, rbcpup);
+	wh_dentry = ERR_PTR(need_wh);
+	if (unlikely(need_wh < 0))
+		goto out;
+
+	sb = dentry->d_sb;
+	udba = au_opt_udba(sb);
+	bcpup = *rbcpup;
+	err = au_pin(pin, dentry, bcpup, udba,
+		     AuPin_DI_LOCKED | AuPin_MNT_WRITE);
+	wh_dentry = ERR_PTR(err);
+	if (unlikely(err))
+		goto out;
+
+	h_path.dentry = au_pinned_h_parent(pin);
+	if (udba != AuOpt_UDBA_NONE
+	    && au_dbtop(dentry) == bcpup) {
+		err = au_may_del(dentry, bcpup, h_path.dentry, isdir);
+		wh_dentry = ERR_PTR(err);
+		if (unlikely(err))
+			goto out_unpin;
+	}
+
+	h_path.mnt = au_sbr_mnt(sb, bcpup);
+	au_dtime_store(dt, au_pinned_parent(pin), &h_path);
+	wh_dentry = NULL;
+	if (!need_wh)
+		goto out; /* success, no need to create whiteout */
+
+	wh_dentry = au_wh_create(dentry, bcpup, h_path.dentry);
+	if (IS_ERR(wh_dentry))
+		goto out_unpin;
+
+	/* returns with the parent is locked and wh_dentry is dget-ed */
+	goto out; /* success */
+
+out_unpin:
+	au_unpin(pin);
+out:
+	return wh_dentry;
+}
+
+/*
+ * final procedure for deleting a entry.
+ * maintain dentry and iattr.
+ */
+static void epilog(struct inode *dir, struct dentry *dentry,
+		   aufs_bindex_t bindex)
+{
+	struct inode *inode;
+
+	inode = d_inode(dentry);
+	d_drop(dentry);
+	inode->i_ctime = dir->i_ctime;
+
+	au_dir_ts(dir, bindex);
+	inode_inc_iversion(dir);
+}
+
+/*
+ * when an error happened, remove the created whiteout and revert everything.
+ */
+static int do_revert(int err, struct inode *dir, aufs_bindex_t bindex,
+		     aufs_bindex_t bwh, struct dentry *wh_dentry,
+		     struct dentry *dentry, struct au_dtime *dt)
+{
+	int rerr;
+	struct path h_path = {
+		.dentry	= wh_dentry,
+		.mnt	= au_sbr_mnt(dir->i_sb, bindex)
+	};
+
+	rerr = au_wh_unlink_dentry(au_h_iptr(dir, bindex), &h_path, dentry);
+	if (!rerr) {
+		au_set_dbwh(dentry, bwh);
+		au_dtime_revert(dt);
+		return 0;
+	}
+
+	AuIOErr("%pd reverting whiteout failed(%d, %d)\n", dentry, err, rerr);
+	return -EIO;
+}
+
+/* ---------------------------------------------------------------------- */
+
+int aufs_unlink(struct inode *dir, struct dentry *dentry)
+{
+	int err;
+	aufs_bindex_t bwh, bindex, btop;
+	struct inode *inode, *h_dir, *delegated;
+	struct dentry *parent, *wh_dentry;
+	/* to reduce stack size */
+	struct {
+		struct au_dtime dt;
+		struct au_pin pin;
+		struct path h_path;
+	} *a;
+
+	IMustLock(dir);
+
+	err = -ENOMEM;
+	a = kmalloc(sizeof(*a), GFP_NOFS);
+	if (unlikely(!a))
+		goto out;
+
+	err = aufs_read_lock(dentry, AuLock_DW | AuLock_GEN);
+	if (unlikely(err))
+		goto out_free;
+	err = au_d_hashed_positive(dentry);
+	if (unlikely(err))
+		goto out_unlock;
+	inode = d_inode(dentry);
+	IMustLock(inode);
+	err = -EISDIR;
+	if (unlikely(d_is_dir(dentry)))
+		goto out_unlock; /* possible? */
+
+	btop = au_dbtop(dentry);
+	bwh = au_dbwh(dentry);
+	bindex = -1;
+	parent = dentry->d_parent; /* dir inode is locked */
+	di_write_lock_parent(parent);
+	wh_dentry = lock_hdir_create_wh(dentry, /*isdir*/0, &bindex, &a->dt,
+					&a->pin);
+	err = PTR_ERR(wh_dentry);
+	if (IS_ERR(wh_dentry))
+		goto out_parent;
+
+	a->h_path.mnt = au_sbr_mnt(dentry->d_sb, btop);
+	a->h_path.dentry = au_h_dptr(dentry, btop);
+	dget(a->h_path.dentry);
+	if (bindex == btop) {
+		h_dir = au_pinned_h_dir(&a->pin);
+		delegated = NULL;
+		err = vfsub_unlink(h_dir, &a->h_path, &delegated, /*force*/0);
+		if (unlikely(err == -EWOULDBLOCK)) {
+			pr_warn("cannot retry for NFSv4 delegation"
+				" for an internal unlink\n");
+			iput(delegated);
+		}
+	} else {
+		/* dir inode is locked */
+		h_dir = d_inode(wh_dentry->d_parent);
+		IMustLock(h_dir);
+		err = 0;
+	}
+
+	if (!err) {
+		vfsub_drop_nlink(inode);
+		epilog(dir, dentry, bindex);
+
+		/* update target timestamps */
+		if (bindex == btop)
+			inode->i_ctime = d_inode(a->h_path.dentry)->i_ctime;
+		else
+			/* todo: this timestamp may be reverted later */
+			inode->i_ctime = h_dir->i_ctime;
+		goto out_unpin; /* success */
+	}
+
+	/* revert */
+	if (wh_dentry) {
+		int rerr;
+
+		rerr = do_revert(err, dir, bindex, bwh, wh_dentry, dentry,
+				 &a->dt);
+		if (rerr)
+			err = rerr;
+	}
+
+out_unpin:
+	au_unpin(&a->pin);
+	dput(wh_dentry);
+	dput(a->h_path.dentry);
+out_parent:
+	di_write_unlock(parent);
+out_unlock:
+	aufs_read_unlock(dentry, AuLock_DW);
+out_free:
+	au_kfree_rcu(a);
+out:
+	return err;
+}

fs/aufs/inode.h

@@ -172,6 +172,12 @@ int aufs_create(struct inode *dir, struct dentry *dentry, umode_t mode,
 int aufs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode);
 int aufs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode);
 
+/* i_op_del.c */
+int au_wr_dir_need_wh(struct dentry *dentry, int isdir, aufs_bindex_t *bcpup);
+int au_may_del(struct dentry *dentry, aufs_bindex_t bindex,
+	       struct dentry *h_parent, int isdir);
+int aufs_unlink(struct inode *dir, struct dentry *dentry);
+
 /* iinfo.c */
 struct inode *au_h_iptr(struct inode *inode, aufs_bindex_t bindex);
 void au_hiput(struct au_hinode *hinode);

fs/aufs/vfsub.h

@@ -37,6 +37,12 @@ enum {
 
 /* ---------------------------------------------------------------------- */
 
+static inline void vfsub_drop_nlink(struct inode *inode)
+{
+	AuDebugOn(!inode->i_nlink);
+	drop_nlink(inode);
+}
+
 int vfsub_sync_filesystem(struct super_block *h_sb, int wait);
 
 /* ---------------------------------------------------------------------- */