[PATCH 0/4] [RESEND] 32/64 bit llseek hashes (v4)

[PATCH 0/4] [RESEND] 32/64 bit llseek hashes (v4)

[Bernd Schubert]

With the ext3/ext4 directory index implementation hashes are used to specify
offsets for llseek(). For compatibility with NFSv2 and 32-bit user space
on 64-bit systems (kernel space) ext3/ext4 currently only return 32-bit 
hashes and therefore the probability of hash collisions for larger directories
is rather high. As recently reported on the NFS mailing list that theoretical
problem also happens on real systems:
http://comments.gmane.org/gmane.linux.nfs/40863

The following series adds two new f_mode flags to tell ext4
to use 32-bit or 64-bit hash values for llseek() calls.
These flags can then used by network file systems, such as NFS, to
request 32-bit or 64-bit offsets (hashes).

Version 4
- Andreas noticed there was HAVE_IS_COMPAT_TASK instead of 
  CONFIG_COMPAT in the 
  "Return 32/64-bit dir name hash according to usage type"
  patch

Version 3:
- remove patch "RFC: Remove check for a 32-bit cookie in nfsd4_readdir()", 
  as it was included upstream separately
- split "nfsd: vfs_llseek() with 32 or 64 bit offsets (hashes)" into two
  two separate patches as suggested by Bruce, one patch to rename 
  'access' to 'may_flags'. And the remainder of the original patch to set 
  FMODE_32BITHASH/FMODE_64BITHASH flags and to introduce the new 
  NFSD_MAY_64BIT_COOKIE flag

Version 2:
- use f_mode instead of O_* flags and also in a separate patch
- introduce EXT4_HTREE_EOF_32BIT and EXT4_HTREE_EOF_64BIT
- fix SEEK_END in ext4_dir_llseek()
- set f_mode flags in NFS code as early as possible and introduce a new
  NFSD_MAY_64BIT_COOKIE flag for that

--
Bernd Schubert
Fraunhofer ITWM

[PATCH 4 1/4] Add new FMODE flags: FMODE_32bithash and FMODE_64bithash

[Bernd Schubert]

Those flags are supposed to be set by NFS readdir() to tell ext3/ext4
to 32bit (NFSv2) or 64bit hash values (offsets) in seekdir().

Signed-off-by: Bernd Schubert <bernd.schubert@itwm.fraunhofer.de>
---
 include/linux/fs.h |    5 +++++
 1 files changed, 5 insertions(+), 0 deletions(-)

diff --git a/include/linux/fs.h b/include/linux/fs.h
index 7aacf31..b158386 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -92,6 +92,11 @@ struct inodes_stat_t {
 /* File is opened using open(.., 3, ..) and is writeable only for ioctls
    (specialy hack for floppy.c) */
 #define FMODE_WRITE_IOCTL	((__force fmode_t)0x100)
+/* 32bit hashes as llseek() offset (for directories) */
+#define FMODE_32BITHASH         ((__force fmode_t)0x200)
+/* 64bit hashes as llseek() offset (for directories) */
+#define FMODE_64BITHASH         ((__force fmode_t)0x400)
+
 
 /*
  * Don't update ctime and mtime.

[PATCH 4 2/4] Return 32/64-bit dir name hash according to usage type

[Bernd Schubert]

From: Fan Yong <yong.fan@whamcloud.com>

Traditionally ext2/3/4 has returned a 32-bit hash value from llseek()
to appease NFSv2, which can only handle a 32-bit cookie for seekdir()
and telldir().  However, this causes problems if there are 32-bit hash
collisions, since the NFSv2 server can get stuck resending the same
entries from the directory repeatedly.

Allow ext4 to return a full 64-bit hash (both major and minor) for
telldir to decrease the chance of hash collisions.  This still needs
integration on the NFS side.

Patch-updated-by: Bernd Schubert <bernd.schubert@itwm.fraunhofer.de>
(blame me if something is not correct)

Signed-off-by: Fan Yong <yong.fan@whamcloud.com>
Signed-off-by: Andreas Dilger <adilger@whamcloud.com>
Signed-off-by: Bernd Schubert <bernd.schubert@itwm.fraunhofer.de>
---
 fs/ext4/dir.c  |  185 ++++++++++++++++++++++++++++++++++++++++++++------------
 fs/ext4/ext4.h |    6 ++
 fs/ext4/hash.c |    4 +
 3 files changed, 154 insertions(+), 41 deletions(-)

diff --git a/fs/ext4/dir.c b/fs/ext4/dir.c
index 164c560..cee09f2 100644
--- a/fs/ext4/dir.c
+++ b/fs/ext4/dir.c
@@ -32,24 +32,8 @@ static unsigned char ext4_filetype_table[] = {
 	DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
 };
 
-static int ext4_readdir(struct file *, void *, filldir_t);
 static int ext4_dx_readdir(struct file *filp,
 			   void *dirent, filldir_t filldir);
-static int ext4_release_dir(struct inode *inode,
-				struct file *filp);
-
-const struct file_operations ext4_dir_operations = {
-	.llseek		= ext4_llseek,
-	.read		= generic_read_dir,
-	.readdir	= ext4_readdir,		/* we take BKL. needed?*/
-	.unlocked_ioctl = ext4_ioctl,
-#ifdef CONFIG_COMPAT
-	.compat_ioctl	= ext4_compat_ioctl,
-#endif
-	.fsync		= ext4_sync_file,
-	.release	= ext4_release_dir,
-};
-
 
 static unsigned char get_dtype(struct super_block *sb, int filetype)
 {
@@ -254,22 +238,134 @@ out:
 	return ret;
 }
 
+static inline int is_32bit_api(void)
+{
+#ifdef CONFIG_COMPAT
+	return is_compat_task();
+#else
+	return (BITS_PER_LONG == 32);
+#endif
+}
+
 /*
  * These functions convert from the major/minor hash to an f_pos
- * value.
+ * value for dx directories
  *
- * Currently we only use major hash numer.  This is unfortunate, but
- * on 32-bit machines, the same VFS interface is used for lseek and
- * llseek, so if we use the 64 bit offset, then the 32-bit versions of
- * lseek/telldir/seekdir will blow out spectacularly, and from within
- * the ext2 low-level routine, we don't know if we're being called by
- * a 64-bit version of the system call or the 32-bit version of the
- * system call.  Worse yet, NFSv2 only allows for a 32-bit readdir
- * cookie.  Sigh.
+ * Upper layer (for example NFS) should specify FMODE_32BITHASH or
+ * FMODE_64BITHASH explicitly. On the other hand, we allow ext4 to be mounted
+ * directly on both 32-bit and 64-bit nodes, under such case, neither
+ * FMODE_32BITHASH nor FMODE_64BITHASH is specified.
  */
-#define hash2pos(major, minor)	(major >> 1)
-#define pos2maj_hash(pos)	((pos << 1) & 0xffffffff)
-#define pos2min_hash(pos)	(0)
+static inline loff_t hash2pos(struct file *filp, __u32 major, __u32 minor)
+{
+	if ((filp->f_flags & FMODE_32BITHASH) ||
+	    (!(filp->f_flags & FMODE_64BITHASH) && is_32bit_api()))
+		return major >> 1;
+	else
+		return ((__u64)(major >> 1) << 32) | (__u64)minor;
+}
+
+static inline __u32 pos2maj_hash(struct file *filp, loff_t pos)
+{
+	if ((filp->f_flags & FMODE_32BITHASH) ||
+	    (!(filp->f_mode & FMODE_64BITHASH) && is_32bit_api()))
+		return (pos << 1) & 0xffffffff;
+	else
+		return ((pos >> 32) << 1) & 0xffffffff;
+}
+
+static inline __u32 pos2min_hash(struct file *filp, loff_t pos)
+{
+	if ((filp->f_flags & FMODE_32BITHASH) ||
+	    (!(filp->f_flags & FMODE_64BITHASH) && is_32bit_api()))
+		return 0;
+	else
+		return pos & 0xffffffff;
+}
+
+/*
+ * Return 32- or 64-bit end-of-file for dx directories
+ */
+static inline loff_t ext4_get_htree_eof(struct file *filp)
+{
+	if ((filp->f_mode & FMODE_32BITHASH) ||
+	    (!(filp->f_mode & FMODE_64BITHASH) && is_32bit_api()))
+		return EXT4_HTREE_EOF_32BIT;
+	else
+		return EXT4_HTREE_EOF_64BIT;
+}
+
+
+/*
+ * ext4_dir_llseek() based on generic_file_llseek() to handle both
+ * non-htree and htree directories, where the "offset" is in terms
+ * of the filename hash value instead of the byte offset.
+ *
+ * NOTE: offsets obtained *before* ext4_set_inode_flag(dir, EXT4_INODE_INDEX)
+ *       will be invalid once the directory was converted into a dx directory
+ */
+loff_t ext4_dir_llseek(struct file *file, loff_t offset, int origin)
+{
+	struct inode *inode = file->f_mapping->host;
+	loff_t ret = -EINVAL;
+	int is_dx_dir = ext4_test_inode_flag(inode, EXT4_INODE_INDEX);
+
+	mutex_lock(&inode->i_mutex);
+
+	/* NOTE: relative offsets with dx directories might not work
+	 *       as expected, as it is difficult to figure out the
+	 *       correct offset between dx hashes */
+
+	switch (origin) {
+	case SEEK_END:
+		if (unlikely(offset > 0))
+			goto out_err; /* not supported for directories */
+
+		/* so only negative offsets are left, does that have a
+		 * meaning for directories at all? */
+		if (is_dx_dir)
+			offset += ext4_get_htree_eof(file);
+		else
+			offset += inode->i_size;
+		break;
+	case SEEK_CUR:
+		/*
+		 * Here we special-case the lseek(fd, 0, SEEK_CUR)
+		 * position-querying operation.  Avoid rewriting the "same"
+		 * f_pos value back to the file because a concurrent read(),
+		 * write() or lseek() might have altered it
+		 */
+		if (offset == 0) {
+			offset = file->f_pos;
+			goto out_ok;
+		}
+
+		offset += file->f_pos;
+		break;
+	}
+
+	if (unlikely(offset < 0))
+		goto out_err;
+
+	if (!is_dx_dir) {
+		if (offset > inode->i_sb->s_maxbytes)
+			goto out_err;
+	} else if (offset > ext4_get_htree_eof(file))
+		goto out_err;
+
+	/* Special lock needed here? */
+	if (offset != file->f_pos) {
+		file->f_pos = offset;
+		file->f_version = 0;
+	}
+
+out_ok:
+	ret = offset;
+out_err:
+	mutex_unlock(&inode->i_mutex);
+
+	return ret;
+}
 
 /*
  * This structure holds the nodes of the red-black tree used to store
@@ -330,15 +426,16 @@ static void free_rb_tree_fname(struct rb_root *root)
 }
 
 
-static struct dir_private_info *ext4_htree_create_dir_info(loff_t pos)
+static struct dir_private_info *ext4_htree_create_dir_info(struct file *filp,
+							   loff_t pos)
 {
 	struct dir_private_info *p;
 
 	p = kzalloc(sizeof(struct dir_private_info), GFP_KERNEL);
 	if (!p)
 		return NULL;
-	p->curr_hash = pos2maj_hash(pos);
-	p->curr_minor_hash = pos2min_hash(pos);
+	p->curr_hash = pos2maj_hash(filp, pos);
+	p->curr_minor_hash = pos2min_hash(filp, pos);
 	return p;
 }
 
@@ -429,7 +526,7 @@ static int call_filldir(struct file *filp, void *dirent,
 		       "null fname?!?\n");
 		return 0;
 	}
-	curr_pos = hash2pos(fname->hash, fname->minor_hash);
+	curr_pos = hash2pos(filp, fname->hash, fname->minor_hash);
 	while (fname) {
 		error = filldir(dirent, fname->name,
 				fname->name_len, curr_pos,
@@ -454,13 +551,13 @@ static int ext4_dx_readdir(struct file *filp,
 	int	ret;
 
 	if (!info) {
-		info = ext4_htree_create_dir_info(filp->f_pos);
+		info = ext4_htree_create_dir_info(filp, filp->f_pos);
 		if (!info)
 			return -ENOMEM;
 		filp->private_data = info;
 	}
 
-	if (filp->f_pos == EXT4_HTREE_EOF)
+	if (filp->f_pos == ext4_get_htree_eof(filp))
 		return 0;	/* EOF */
 
 	/* Some one has messed with f_pos; reset the world */
@@ -468,8 +565,8 @@ static int ext4_dx_readdir(struct file *filp,
 		free_rb_tree_fname(&info->root);
 		info->curr_node = NULL;
 		info->extra_fname = NULL;
-		info->curr_hash = pos2maj_hash(filp->f_pos);
-		info->curr_minor_hash = pos2min_hash(filp->f_pos);
+		info->curr_hash = pos2maj_hash(filp, filp->f_pos);
+		info->curr_minor_hash = pos2min_hash(filp, filp->f_pos);
 	}
 
 	/*
@@ -501,7 +598,7 @@ static int ext4_dx_readdir(struct file *filp,
 			if (ret < 0)
 				return ret;
 			if (ret == 0) {
-				filp->f_pos = EXT4_HTREE_EOF;
+				filp->f_pos = ext4_get_htree_eof(filp);
 				break;
 			}
 			info->curr_node = rb_first(&info->root);
@@ -521,7 +618,7 @@ static int ext4_dx_readdir(struct file *filp,
 			info->curr_minor_hash = fname->minor_hash;
 		} else {
 			if (info->next_hash == ~0) {
-				filp->f_pos = EXT4_HTREE_EOF;
+				filp->f_pos = ext4_get_htree_eof(filp);
 				break;
 			}
 			info->curr_hash = info->next_hash;
@@ -540,3 +637,15 @@ static int ext4_release_dir(struct inode *inode, struct file *filp)
 
 	return 0;
 }
+
+const struct file_operations ext4_dir_operations = {
+	.llseek		= ext4_dir_llseek,
+	.read		= generic_read_dir,
+	.readdir	= ext4_readdir,
+	.unlocked_ioctl = ext4_ioctl,
+#ifdef CONFIG_COMPAT
+	.compat_ioctl	= ext4_compat_ioctl,
+#endif
+	.fsync		= ext4_sync_file,
+	.release	= ext4_release_dir,
+};
diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
index 1554b15..d3fe1ea 100644
--- a/fs/ext4/ext4.h
+++ b/fs/ext4/ext4.h
@@ -1599,7 +1599,11 @@ struct dx_hash_info
 	u32		*seed;
 };
 
-#define EXT4_HTREE_EOF	0x7fffffff
+
+/* 32 and 64 bit signed EOF for dx directories */
+#define EXT4_HTREE_EOF_32BIT   ((1UL  << (32 - 1)) - 1)
+#define EXT4_HTREE_EOF_64BIT   ((1ULL << (64 - 1)) - 1)
+
 
 /*
  * Control parameters used by ext4_htree_next_block
diff --git a/fs/ext4/hash.c b/fs/ext4/hash.c
index ac8f168..fa8e491 100644
--- a/fs/ext4/hash.c
+++ b/fs/ext4/hash.c
@@ -200,8 +200,8 @@ int ext4fs_dirhash(const char *name, int len, struct dx_hash_info *hinfo)
 		return -1;
 	}
 	hash = hash & ~1;
-	if (hash == (EXT4_HTREE_EOF << 1))
-		hash = (EXT4_HTREE_EOF-1) << 1;
+	if (hash == (EXT4_HTREE_EOF_32BIT << 1))
+		hash = (EXT4_HTREE_EOF_32BIT - 1) << 1;
 	hinfo->hash = hash;
 	hinfo->minor_hash = minor_hash;
 	return 0;

[PATCH 4 3/4] nfsd_open(): rename 'int access' to 'int may_flags' in nfsd_open()

[Bernd Schubert]

Just rename this variable, as the next patch will add a flag and
'access' as variable name would not be correct any more.


Signed-off-by: Bernd Schubert <bernd.schubert@itwm.fraunhofer.de>
Acked-by: J. Bruce Fields<bfields@redhat.com>
---
 fs/nfsd/vfs.c |   18 ++++++++++--------
 1 files changed, 10 insertions(+), 8 deletions(-)

diff --git a/fs/nfsd/vfs.c b/fs/nfsd/vfs.c
index d25a723..20fa252 100644
--- a/fs/nfsd/vfs.c
+++ b/fs/nfsd/vfs.c
@@ -726,12 +726,13 @@ static int nfsd_open_break_lease(struct inode *inode, int access)
 
 /*
  * Open an existing file or directory.
- * The access argument indicates the type of open (read/write/lock)
+ * The may_flags argument indicates the type of open (read/write/lock)
+ * and additional flags.
  * N.B. After this call fhp needs an fh_put
  */
 __be32
 nfsd_open(struct svc_rqst *rqstp, struct svc_fh *fhp, umode_t type,
-			int access, struct file **filp)
+			int may_flags, struct file **filp)
 {
 	struct dentry	*dentry;
 	struct inode	*inode;
@@ -746,7 +747,7 @@ nfsd_open(struct svc_rqst *rqstp, struct svc_fh *fhp, umode_t type,
 	 * and (hopefully) checked permission - so allow OWNER_OVERRIDE
 	 * in case a chmod has now revoked permission.
 	 */
-	err = fh_verify(rqstp, fhp, type, access | NFSD_MAY_OWNER_OVERRIDE);
+	err = fh_verify(rqstp, fhp, type, may_flags | NFSD_MAY_OWNER_OVERRIDE);
 	if (err)
 		goto out;
 
@@ -757,7 +758,7 @@ nfsd_open(struct svc_rqst *rqstp, struct svc_fh *fhp, umode_t type,
 	 * or any access when mandatory locking enabled
 	 */
 	err = nfserr_perm;
-	if (IS_APPEND(inode) && (access & NFSD_MAY_WRITE))
+	if (IS_APPEND(inode) && (may_flags & NFSD_MAY_WRITE))
 		goto out;
 	/*
 	 * We must ignore files (but only files) which might have mandatory
@@ -770,12 +771,12 @@ nfsd_open(struct svc_rqst *rqstp, struct svc_fh *fhp, umode_t type,
 	if (!inode->i_fop)
 		goto out;
 
-	host_err = nfsd_open_break_lease(inode, access);
+	host_err = nfsd_open_break_lease(inode, may_flags);
 	if (host_err) /* NOMEM or WOULDBLOCK */
 		goto out_nfserr;
 
-	if (access & NFSD_MAY_WRITE) {
-		if (access & NFSD_MAY_READ)
+	if (may_flags & NFSD_MAY_WRITE) {
+		if (may_flags & NFSD_MAY_READ)
 			flags = O_RDWR|O_LARGEFILE;
 		else
 			flags = O_WRONLY|O_LARGEFILE;
@@ -785,7 +786,8 @@ nfsd_open(struct svc_rqst *rqstp, struct svc_fh *fhp, umode_t type,
 	if (IS_ERR(*filp))
 		host_err = PTR_ERR(*filp);
 	else
-		host_err = ima_file_check(*filp, access);
+		host_err = ima_file_check(*filp, may_flags);
+
 out_nfserr:
 	err = nfserrno(host_err);
 out:

[PATCH 4 4/4] nfsd: vfs_llseek() with 32 or 64 bit offsets (hashes)

[Bernd Schubert]

Use 32-bit or 64-bit llseek() hashes for directory offsets depending on
the NFS version. NFSv2 gets 32-bit hashes only.

NOTE: This patch got rather complex as Christoph asked to set the
filp->f_mode flag in the open call or immediatly after dentry_open()
in nfsd_open() to avoid races.
Personally I still do not see a reason for that and in my opinion
FMODE_32BITHASH/FMODE_64BITHASH flags could be set nfsd_readdir(), as it
follows directly after nfsd_open() without a chance of races.


Signed-off-by: Bernd Schubert <bernd.schubert@itwm.fraunhofer.de>
Acked-by: J. Bruce Fields<bfields@redhat.com>
---
 fs/nfsd/vfs.c |   15 +++++++++++++--
 fs/nfsd/vfs.h |    2 ++
 2 files changed, 15 insertions(+), 2 deletions(-)

diff --git a/fs/nfsd/vfs.c b/fs/nfsd/vfs.c
index 20fa252..f162dc3 100644
--- a/fs/nfsd/vfs.c
+++ b/fs/nfsd/vfs.c
@@ -785,9 +785,15 @@ nfsd_open(struct svc_rqst *rqstp, struct svc_fh *fhp, umode_t type,
 			    flags, current_cred());
 	if (IS_ERR(*filp))
 		host_err = PTR_ERR(*filp);
-	else
+	else {
 		host_err = ima_file_check(*filp, may_flags);
 
+		if (may_flags & NFSD_MAY_64BIT_COOKIE)
+			(*filp)->f_mode |= FMODE_64BITHASH;
+		else
+			(*filp)->f_mode |= FMODE_32BITHASH;
+	}
+
 out_nfserr:
 	err = nfserrno(host_err);
 out:
@@ -2011,8 +2017,13 @@ nfsd_readdir(struct svc_rqst *rqstp, struct svc_fh *fhp, loff_t *offsetp,
 	__be32		err;
 	struct file	*file;
 	loff_t		offset = *offsetp;
+	int             may_flags = NFSD_MAY_READ;
+
+	/* NFSv2 only supports 32 bit cookies */
+	if (rqstp->rq_vers > 2)
+		may_flags |= NFSD_MAY_64BIT_COOKIE;
 
-	err = nfsd_open(rqstp, fhp, S_IFDIR, NFSD_MAY_READ, &file);
+	err = nfsd_open(rqstp, fhp, S_IFDIR, may_flags, &file);
 	if (err)
 		goto out;
 
diff --git a/fs/nfsd/vfs.h b/fs/nfsd/vfs.h
index 1dcd238..2ac2f3e 100644
--- a/fs/nfsd/vfs.h
+++ b/fs/nfsd/vfs.h
@@ -27,6 +27,8 @@
 #define NFSD_MAY_BYPASS_GSS		0x400
 #define NFSD_MAY_READ_IF_EXEC		0x800
 
+#define NFSD_MAY_64BIT_COOKIE	2048 /* 64 bit readdir cookies for >= NFSv3 */
+
 #define NFSD_MAY_CREATE		(NFSD_MAY_EXEC|NFSD_MAY_WRITE)
 #define NFSD_MAY_REMOVE		(NFSD_MAY_EXEC|NFSD_MAY_WRITE|NFSD_MAY_TRUNC)
 

Re: [PATCH 4 4/4] nfsd: vfs_llseek() with 32 or 64 bit offsets (hashes)

[Bernd Schubert]

On 01/09/2012 11:16 AM, Bernd Schubert wrote:
>   #define NFSD_MAY_BYPASS_GSS		0x400
>   #define NFSD_MAY_READ_IF_EXEC		0x800
>
> +#define NFSD_MAY_64BIT_COOKIE	2048 /* 64 bit readdir cookies for>= NFSv3 */
> +

Oh, I just notice the format of the flags has been updated in the mean 
time and the value for NFSD_MAY_64BIT_COOKIE needs to be increased to 
avoid an conflict.
Will send a new series in a moment.

Sorry for not noticing it earlier,
Bernd

--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html