aufs: dirren 1/6, inum-list of the renamed dir in a branch

This commits brings a list of the inode numbers which indicates the
logically renamed dir into a branch. The list will be referred in
lookup, and its lifetime is equivalent to the branch's, ie. the list is
loaded/created in adding a branch, and stored/deleted in deleting a
branch. The simple storing happens in remounting and unmounting aufs
too.
Signed-off-by: J. R. Okajima <hooanon05g@gmail.com>

fs/aufs/Kconfig

@@ -96,6 +96,19 @@ config AUFS_XATTR
 	branch attributes for EA.
 	See detail in aufs.5.
 
+config AUFS_DIRREN
+	bool "Workaround for rename(2)-ing a directory"
+	help
+	By default, aufs returns EXDEV error in renameing a dir who has
+	his child on the lower branch, since it is a bad idea to issue
+	rename(2) internally for every lower branch. But user may not
+	accept this behaviour. So here is a workaround to allow such
+	rename(2) and store some extra infromation on the writable
+	branch. Obviously this costs high (and I don't like it).
+	To use this feature, you need to enable this configuration AND
+	to specify the mount option `dirren.'
+	See details in aufs.5 and the design documents.
+
 config AUFS_DEBUG
 	bool "Debug aufs"
 	help

fs/aufs/Makefile

@@ -27,4 +27,5 @@ aufs-$(CONFIG_AUFS_HFSNOTIFY) += hfsnotify.o
 aufs-$(CONFIG_AUFS_EXPORT) += export.o
 aufs-$(CONFIG_AUFS_XATTR) += xattr.o
 aufs-$(CONFIG_FS_POSIX_ACL) += posix_acl.o
+aufs-$(CONFIG_AUFS_DIRREN) += dirren.o
 aufs-$(CONFIG_AUFS_DEBUG) += debug.o

fs/aufs/aufs.h

@@ -27,6 +27,7 @@
 #include "dcsub.h"
 #include "dentry.h"
 #include "dir.h"
+#include "dirren.h"
 #include "dynop.h"
 #include "file.h"
 #include "fstype.h"

fs/aufs/branch.c

@@ -20,6 +20,8 @@ static void au_br_do_free(struct au_branch *br)
 	struct au_dykey **key;
 
 	au_hnotify_fin_br(br);
+	/* always, regardless the mount option */
+	au_dr_hino_free(&br->br_dirren);
 	au_xino_put(br);
 
 	AuLCntZero(au_lcnt_read(&br->br_nfiles, /*do_rev*/0));
@@ -361,6 +363,11 @@ static int au_br_init(struct au_branch *br, struct super_block *sb,
 	br->br_id = au_new_br_id(sb);
 	AuDebugOn(br->br_id < 0);
 
+	/* always, regardless the given option */
+	err = au_dr_br_init(sb, br, &add->path);
+	if (unlikely(err))
+		goto out_err;
+
 	if (au_br_writable(add->perm)) {
 		err = au_wbr_init(br, sb, add->perm);
 		if (unlikely(err))
@@ -933,6 +940,9 @@ static void au_br_do_del(struct super_block *sb, aufs_bindex_t bindex,
 	au_br_do_del_hdp(au_di(root), bindex, bbot);
 	au_br_do_del_hip(au_ii(inode), bindex, bbot);
 
+	/* ignore an error */
+	au_dr_br_fin(sb, br); /* always, regardless the mount option */
+
 	dput(h_root);
 	iput(h_inode);
 	au_br_do_free(br);

fs/aufs/branch.h

@@ -13,6 +13,7 @@
 #ifdef __KERNEL__
 
 #include <linux/mount.h>
+#include "dirren.h"
 #include "dynop.h"
 #include "lcnt.h"
 #include "rwsem.h"
@@ -101,6 +102,8 @@ struct au_branch {
 	/* entries under sysfs per mount-point */
 	struct au_brsysfs	br_sysfs[AuBrSysfs_Last];
 #endif
+
+	struct au_dr_br		br_dirren;
 };
 
 /* ---------------------------------------------------------------------- */

fs/aufs/dirren.c

+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (C) 2017-2019 Junjiro R. Okajima
+ */
+
+/*
+ * special handling in renaming a directory
+ * in order to support looking-up the before-renamed name on the lower readonly
+ * branches
+ */
+
+#include <linux/byteorder/generic.h>
+#include "aufs.h"
+
+/* re-commit later */ __maybe_unused
+static void au_dr_hino_del(struct au_dr_br *dr, struct au_dr_hino *ent)
+{
+	int idx;
+
+	idx = au_dr_ihash(ent->dr_h_ino);
+	au_hbl_del(&ent->dr_hnode, dr->dr_h_ino + idx);
+}
+
+static int au_dr_hino_test_empty(struct au_dr_br *dr)
+{
+	int ret, i;
+	struct hlist_bl_head *hbl;
+
+	ret = 1;
+	for (i = 0; ret && i < AuDirren_NHASH; i++) {
+		hbl = dr->dr_h_ino + i;
+		hlist_bl_lock(hbl);
+		ret &= hlist_bl_empty(hbl);
+		hlist_bl_unlock(hbl);
+	}
+
+	return ret;
+}
+
+/* re-commit later */ __maybe_unused
+static struct au_dr_hino *au_dr_hino_find(struct au_dr_br *dr, ino_t ino)
+{
+	struct au_dr_hino *found, *ent;
+	struct hlist_bl_head *hbl;
+	struct hlist_bl_node *pos;
+	int idx;
+
+	found = NULL;
+	idx = au_dr_ihash(ino);
+	hbl = dr->dr_h_ino + idx;
+	hlist_bl_lock(hbl);
+	hlist_bl_for_each_entry(ent, pos, hbl, dr_hnode)
+		if (ent->dr_h_ino == ino) {
+			found = ent;
+			break;
+		}
+	hlist_bl_unlock(hbl);
+
+	return found;
+}
+
+int au_dr_hino_test_add(struct au_dr_br *dr, ino_t ino,
+			struct au_dr_hino *add_ent)
+{
+	int found, idx;
+	struct hlist_bl_head *hbl;
+	struct hlist_bl_node *pos;
+	struct au_dr_hino *ent;
+
+	found = 0;
+	idx = au_dr_ihash(ino);
+	hbl = dr->dr_h_ino + idx;
+#if 0
+	{
+		struct hlist_bl_node *tmp;
+
+		hlist_bl_for_each_entry_safe(ent, pos, tmp, hbl, dr_hnode)
+			AuDbg("hi%llu\n", (unsigned long long)ent->dr_h_ino);
+	}
+#endif
+	hlist_bl_lock(hbl);
+	hlist_bl_for_each_entry(ent, pos, hbl, dr_hnode)
+		if (ent->dr_h_ino == ino) {
+			found = 1;
+			break;
+		}
+	if (!found && add_ent)
+		hlist_bl_add_head(&add_ent->dr_hnode, hbl);
+	hlist_bl_unlock(hbl);
+
+	if (!found && add_ent)
+		AuDbg("i%llu added\n", (unsigned long long)add_ent->dr_h_ino);
+
+	return found;
+}
+
+void au_dr_hino_free(struct au_dr_br *dr)
+{
+	int i;
+	struct hlist_bl_head *hbl;
+	struct hlist_bl_node *pos, *tmp;
+	struct au_dr_hino *ent;
+
+	/* SiMustWriteLock(sb); */
+
+	for (i = 0; i < AuDirren_NHASH; i++) {
+		hbl = dr->dr_h_ino + i;
+		/* no spinlock since sbinfo must be write-locked */
+		hlist_bl_for_each_entry_safe(ent, pos, tmp, hbl, dr_hnode)
+			au_kfree_rcu(ent);
+		INIT_HLIST_BL_HEAD(hbl);
+	}
+}
+
+/* returns the number of inodes or an error */
+static int au_dr_hino_store(struct super_block *sb, struct au_branch *br,
+			    struct file *hinofile)
+{
+	int err, i;
+	ssize_t ssz;
+	loff_t pos, oldsize;
+	__be64 u64;
+	struct inode *hinoinode;
+	struct hlist_bl_head *hbl;
+	struct hlist_bl_node *n1, *n2;
+	struct au_dr_hino *ent;
+
+	SiMustWriteLock(sb);
+	AuDebugOn(!au_br_writable(br->br_perm));
+
+	hinoinode = file_inode(hinofile);
+	oldsize = i_size_read(hinoinode);
+
+	err = 0;
+	pos = 0;
+	hbl = br->br_dirren.dr_h_ino;
+	for (i = 0; !err && i < AuDirren_NHASH; i++, hbl++) {
+		/* no bit-lock since sbinfo must be write-locked */
+		hlist_bl_for_each_entry_safe(ent, n1, n2, hbl, dr_hnode) {
+			AuDbg("hi%llu, %pD2\n",
+			      (unsigned long long)ent->dr_h_ino, hinofile);
+			u64 = cpu_to_be64(ent->dr_h_ino);
+			ssz = vfsub_write_k(hinofile, &u64, sizeof(u64), &pos);
+			if (ssz == sizeof(u64))
+				continue;
+
+			/* write error */
+			pr_err("ssz %zd, %pD2\n", ssz, hinofile);
+			err = -ENOSPC;
+			if (ssz < 0)
+				err = ssz;
+			break;
+		}
+	}
+	/* regardless the error */
+	if (pos < oldsize) {
+		err = vfsub_trunc(&hinofile->f_path, pos, /*attr*/0, hinofile);
+		AuTraceErr(err);
+	}
+
+	AuTraceErr(err);
+	return err;
+}
+
+static int au_dr_hino_load(struct au_dr_br *dr, struct file *hinofile)
+{
+	int err, hidx;
+	ssize_t ssz;
+	size_t sz, n;
+	loff_t pos;
+	uint64_t u64;
+	struct au_dr_hino *ent;
+	struct inode *hinoinode;
+	struct hlist_bl_head *hbl;
+
+	err = 0;
+	pos = 0;
+	hbl = dr->dr_h_ino;
+	hinoinode = file_inode(hinofile);
+	sz = i_size_read(hinoinode);
+	AuDebugOn(sz % sizeof(u64));
+	n = sz / sizeof(u64);
+	while (n--) {
+		ssz = vfsub_read_k(hinofile, &u64, sizeof(u64), &pos);
+		if (unlikely(ssz != sizeof(u64))) {
+			pr_err("ssz %zd, %pD2\n", ssz, hinofile);
+			err = -EINVAL;
+			if (ssz < 0)
+				err = ssz;
+			goto out_free;
+		}
+
+		ent = kmalloc(sizeof(*ent), GFP_NOFS);
+		if (!ent) {
+			err = -ENOMEM;
+			AuTraceErr(err);
+			goto out_free;
+		}
+		ent->dr_h_ino = be64_to_cpu((__force __be64)u64);
+		AuDbg("hi%llu, %pD2\n",
+		      (unsigned long long)ent->dr_h_ino, hinofile);
+		hidx = au_dr_ihash(ent->dr_h_ino);
+		au_hbl_add(&ent->dr_hnode, hbl + hidx);
+	}
+	goto out; /* success */
+
+out_free:
+	au_dr_hino_free(dr);
+out:
+	AuTraceErr(err);
+	return err;
+}
+
+/*
+ * @bindex/@br is a switch to distinguish whether suspending hnotify or not.
+ * @path is a switch to distinguish load and store.
+ */
+static int au_dr_hino(struct super_block *sb, aufs_bindex_t bindex,
+		      struct au_branch *br, const struct path *path)
+{
+	int err, flags;
+	unsigned char load, suspend;
+	struct file *hinofile;
+	struct au_hinode *hdir;
+	struct inode *dir, *delegated;
+	struct path hinopath;
+	struct qstr hinoname = QSTR_INIT(AUFS_WH_DR_BRHINO,
+					 sizeof(AUFS_WH_DR_BRHINO) - 1);
+
+	AuDebugOn(bindex < 0 && !br);
+	AuDebugOn(bindex >= 0 && br);
+
+	err = -EINVAL;
+	suspend = !br;
+	if (suspend)
+		br = au_sbr(sb, bindex);
+	load = !!path;
+	if (!load) {
+		path = &br->br_path;
+		AuDebugOn(!au_br_writable(br->br_perm));
+		if (unlikely(!au_br_writable(br->br_perm)))
+			goto out;
+	}
+
+	hdir = NULL;
+	if (suspend) {
+		dir = d_inode(sb->s_root);
+		hdir = au_hinode(au_ii(dir), bindex);
+		dir = hdir->hi_inode;
+		au_hn_inode_lock_nested(hdir, AuLsc_I_CHILD);
+	} else {
+		dir = d_inode(path->dentry);
+		inode_lock_nested(dir, AuLsc_I_CHILD);
+	}
+	hinopath.dentry = vfsub_lkup_one(&hinoname, path->dentry);
+	err = PTR_ERR(hinopath.dentry);
+	if (IS_ERR(hinopath.dentry))
+		goto out_unlock;
+
+	err = 0;
+	flags = O_RDONLY;
+	if (load) {
+		if (d_is_negative(hinopath.dentry))
+			goto out_dput; /* success */
+	} else {
+		if (au_dr_hino_test_empty(&br->br_dirren)) {
+			if (d_is_positive(hinopath.dentry)) {
+				delegated = NULL;
+				err = vfsub_unlink(dir, &hinopath, &delegated,
+						   /*force*/0);
+				AuTraceErr(err);
+				if (unlikely(err))
+					pr_err("ignored err %d, %pd2\n",
+					       err, hinopath.dentry);
+				if (unlikely(err == -EWOULDBLOCK))
+					iput(delegated);
+				err = 0;
+			}
+			goto out_dput;
+		} else if (!d_is_positive(hinopath.dentry)) {
+			err = vfsub_create(dir, &hinopath, 0600,
+					   /*want_excl*/false);
+			AuTraceErr(err);
+			if (unlikely(err))
+				goto out_dput;
+		}
+		flags = O_WRONLY;
+	}
+	hinopath.mnt = path->mnt;
+	hinofile = vfsub_dentry_open(&hinopath, flags);
+	if (suspend)
+		au_hn_inode_unlock(hdir);
+	else
+		inode_unlock(dir);
+	dput(hinopath.dentry);
+	AuTraceErrPtr(hinofile);
+	if (IS_ERR(hinofile)) {
+		err = PTR_ERR(hinofile);
+		goto out;
+	}
+
+	if (load)
+		err = au_dr_hino_load(&br->br_dirren, hinofile);
+	else
+		err = au_dr_hino_store(sb, br, hinofile);
+	fput(hinofile);
+	goto out;
+
+out_dput:
+	dput(hinopath.dentry);
+out_unlock:
+	if (suspend)
+		au_hn_inode_unlock(hdir);
+	else
+		inode_unlock(dir);
+out:
+	AuTraceErr(err);
+	return err;
+}
+
+int au_dr_br_init(struct super_block *sb, struct au_branch *br,
+		  const struct path *path)
+{
+	int err, i;
+	struct au_dr_br *dr;
+	struct hlist_bl_head *hbl;
+
+	dr = &br->br_dirren;
+	hbl = dr->dr_h_ino;
+	for (i = 0; i < AuDirren_NHASH; i++, hbl++)
+		INIT_HLIST_BL_HEAD(hbl);
+
+	err = 0;
+	if (au_opt_test(au_mntflags(sb), DIRREN))
+		err = au_dr_hino(sb, /*bindex*/-1, br, path);
+
+	AuTraceErr(err);
+	return err;
+}
+
+int au_dr_br_fin(struct super_block *sb, struct au_branch *br)
+{
+	int err;
+
+	err = 0;
+	if (au_br_writable(br->br_perm))
+		err = au_dr_hino(sb, /*bindex*/-1, br, /*path*/NULL);
+	if (!err)
+		au_dr_hino_free(&br->br_dirren);
+
+	return err;
+}

fs/aufs/dirren.h

+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Copyright (C) 2017-2019 Junjiro R. Okajima
+ */
+
+/*
+ * renamed dir info
+ */
+
+#ifndef __AUFS_DIRREN_H__
+#define __AUFS_DIRREN_H__
+
+#ifdef __KERNEL__
+
+#include "hbl.h"
+
+#define AuDirren_NHASH 100
+
+#ifdef CONFIG_AUFS_DIRREN
+struct au_dr_hino {
+	struct hlist_bl_node	dr_hnode;
+	ino_t			dr_h_ino;
+};
+
+struct au_dr_br {
+	struct hlist_bl_head	dr_h_ino[AuDirren_NHASH];
+};
+#else
+struct au_dr_hino;
+/* empty */
+struct au_dr_br { };
+#endif
+
+/* ---------------------------------------------------------------------- */
+
+struct au_branch;
+struct au_hinode;
+struct path;
+struct super_block;
+#ifdef CONFIG_AUFS_DIRREN
+int au_dr_hino_test_add(struct au_dr_br *dr, ino_t h_ino,
+			struct au_dr_hino *add_ent);
+void au_dr_hino_free(struct au_dr_br *dr);
+int au_dr_br_init(struct super_block *sb, struct au_branch *br,
+		  const struct path *path);
+int au_dr_br_fin(struct super_block *sb, struct au_branch *br);
+#else
+AuStubInt0(au_dr_hino_test_add, struct au_dr_br *dr, ino_t h_ino,
+	   struct au_dr_hino *add_ent);
+AuStubVoid(au_dr_hino_free, struct au_dr_br *dr);
+AuStubInt0(au_dr_br_init, struct super_block *sb, struct au_branch *br,
+	   const struct path *path);
+AuStubInt0(au_dr_br_fin, struct super_block *sb, struct au_branch *br);
+#endif
+
+/* ---------------------------------------------------------------------- */
+
+#ifdef CONFIG_AUFS_DIRREN
+static inline int au_dr_ihash(ino_t h_ino)
+{
+	return h_ino % AuDirren_NHASH;
+}
+#else
+AuStubInt0(au_dr_ihash, ino_t h_ino);
+#endif
+
+#endif /* __KERNEL__ */
+#endif /* __AUFS_DIRREN_H__ */

fs/aufs/opts.h

@@ -32,11 +32,16 @@ struct file;
 #define AuOpt_SUM_W		(1 << 11)	/* unimplemented */
 #define AuOpt_VERBOSE		(1 << 13)	/* print the cause of error */
 #define AuOpt_DIO		(1 << 14)	/* direct io */
+#define AuOpt_DIRREN		(1 << 15)	/* directory rename */
 
 #ifndef CONFIG_AUFS_HNOTIFY
 #undef AuOpt_UDBA_HNOTIFY
 #define AuOpt_UDBA_HNOTIFY	0
 #endif
+#ifndef CONFIG_AUFS_DIRREN
+#undef AuOpt_DIRREN
+#define AuOpt_DIRREN		0
+#endif
 
 #define AuOpt_Def	(AuOpt_XINO \
 			 | AuOpt_UDBA_REVAL \

include/uapi/linux/aufs_type.h

@@ -88,6 +88,10 @@ typedef int16_t aufs_bindex_t;
 #define AUFS_PLINK_MAINT_DIR	"fs/" AUFS_NAME
 #define AUFS_PLINK_MAINT_PATH	AUFS_PLINK_MAINT_DIR "/" AUFS_PLINK_MAINT_NAME
 
+/* whiteouted doubly */
+#define AUFS_DR_BRHINO_NAME	AUFS_WH_PFX "hino"
+#define AUFS_WH_DR_BRHINO	AUFS_WH_PFX AUFS_DR_BRHINO_NAME
+
 #define AUFS_DIROPQ_NAME	AUFS_WH_PFX ".opq" /* whiteouted doubly */
 #define AUFS_WH_DIROPQ		AUFS_WH_PFX AUFS_DIROPQ_NAME