Skip to contentSkip to conversation

llseek fixes

PATCH In review Richard Yao 3 participants 11 messages Jun 14, 2013 – Jun 16, 2013

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover llseek fixes

btrfs_file_llseek() and ocfs2_file_llseek() are extremely similar and consequently, contain many of the same flaws. Li Dongyang filed a pull request with ZFSOnLinux for SEEK_HOLE/SEEK_DATA support that included a custom llseek function that appears to have been modelled after the one in ocfs2. The similarity was strong enough that it suffered from many of the same flaws, which I caught during review. I addressed the issues with his patch with one that I wrote. Since a small percentage of Gentoo Linux users are affected by these flaws, I decided to adapt that code that to btrfs and ocfs2.

The specific issues are:

1. Taking the inode->i_mutex lock before calling generic_file_llseek(), which is unnecessary.

2. Failing to take the filp->f_lock spinlock before modifying filp->f_pos and filp->f_version, which differs from generic_file_llseek().

3. Having an offset > inode->i_sb->s_maxbytes check that is dead code in btrfs and active code in ocfs2. In ocfs2, this permits seeking up to the maximum file size possible, even when it is past the end of the file. Seeking beyond that (if possible), would return EINVAL instead of ENXIO.

4. Trying to cover all whence values when in reality such functions should only care about SEEK_HOLE/SEEK_DATA. Any other cases should be passsed to generic_file_llseek().

Richard Yao (2):
ocfs2: Fix llseek() semantics and do some cleanup
btrfs: Cleanup llseek()

 fs/btrfs/file.c | 49 ++++++++++++++++++-------------------------
 fs/ocfs2/file.c | 65 ++++++++++++++++++++++-----------------------------------
 2 files changed, 45 insertions(+), 69 deletions(-)
Signature · 1 line
1.8.1.5
#1 lore raw

PATCH 1/2 ocfs2: Fix llseek() semantics and do some cleanup

There are multiple issues with the custom llseek implemented in ocfs2 for implementing SEEK_HOLE/SEEK_DATA.

1. It takes the inode->i_mutex lock before calling generic_file_llseek(), which is unnecessary.

2. It fails to take the filp->f_lock spinlock before modifying filp->f_pos and filp->f_version, which differs from generic_file_llseek().

3. It does a offset > inode->i_sb->s_maxbytes check that permits seeking up to the maximum file size possible on the ocfs2 filesystem, even when it is past the end of the file. Seeking beyond that (if possible), would return EINVAL instead of ENXIO.

4. The switch statement tries to cover all whence values when in reality it should only care about SEEK_HOLE/SEEK_DATA. Any other cases should be passsed to generic_file_llseek().

btrfs_file_llseek() and ocfs2_file_llseek() are extremely similar and consequently, contain many of the same flaws. Li Dongyang filed a pull request with ZFSOnLinux for SEEK_HOLE/SEEK_DATA support that included a custom llseek function that appears to have been modelled after the one in ocfs2. The similarity was strong enough that it suffered from many of the same flaws, which I caught during review. I addressed the issues with his patch with one that I wrote. Since a small percentage of Gentoo Linux users are affected by these flaws, I decided to adapt that code that to btrfs (separate patch) and ocfs2.

Note that commit 48802c8ae2a9d618ec734a61283d645ad527e06c by Jeff Liu at Oracle mostly addressed #3 in btrfs. The only lingering issue was that the offset > inode->i_sb->s_maxbytes check became dead code. The ocfs2 code was not fortunate enough to have had a similar correction until now.

---
 fs/ocfs2/file.c | 65 ++++++++++++++++++++++-----------------------------------
 1 file changed, 25 insertions(+), 40 deletions(-)
fs/ocfs2/file.c +25 -40
index ff54014..84f8c9c 100644
@@ -2615,54 +2615,39 @@ bail:
 }
 
 /* Refer generic_file_llseek_unlocked() */
-static loff_t ocfs2_file_llseek(struct file *file, loff_t offset, int whence)
+static loff_t ocfs2_file_llseek(struct file *filp, loff_t offset, int whence)
 {
-	struct inode *inode = file->f_mapping->host;
-	int ret = 0;
+	if (whence == SEEK_DATA || whence == SEEK_HOLE) {
+		struct inode *inode = filp->f_mapping->host;
+		int ret;
 
-	mutex_lock(&inode->i_mutex);
+		if (offset < 0 && !(filp->f_mode & FMODE_UNSIGNED_OFFSET))
+			return -EINVAL;
 
-	switch (whence) {
-	case SEEK_SET:
-		break;
-	case SEEK_END:
-		offset += inode->i_size;
-		break;
-	case SEEK_CUR:
-		if (offset == 0) {
-			offset = file->f_pos;
-			goto out;
+		if (offset >= i_size_read(inode)) {
+			return -ENXIO;
 		}
-		offset += file->f_pos;
-		break;
-	case SEEK_DATA:
-	case SEEK_HOLE:
-		ret = ocfs2_seek_data_hole_offset(file, &offset, whence);
-		if (ret)
-			goto out;
-		break;
-	default:
-		ret = -EINVAL;
-		goto out;
-	}
 
-	if (offset < 0 && !(file->f_mode & FMODE_UNSIGNED_OFFSET))
-		ret = -EINVAL;
-	if (!ret && offset > inode->i_sb->s_maxbytes)
-		ret = -EINVAL;
-	if (ret)
-		goto out;
+		mutex_lock(&inode->i_mutex);
+		ret = ocfs2_seek_data_hole_offset(filp, &offset, whence);
+		mutex_unlock(&inode->i_mutex);
+
+		if (ret) {
+			return ret;
+		}
 
-	if (offset != file->f_pos) {
-		file->f_pos = offset;
-		file->f_version = 0;
+		if (offset != filp->f_pos) {
+			spin_lock(&filp->f_lock);
+			filp->f_pos = offset;
+			filp->f_version = 0;
+			spin_unlock(&filp->f_lock);
+		}
+
+		return offset;
 	}
 
-out:
-	mutex_unlock(&inode->i_mutex);
-	if (ret)
-		return ret;
-	return offset;
+	return generic_file_llseek(filp, offset, whence);
+
 }
 
 const struct inode_operations ocfs2_file_iops = {
Signature · 1 line
1.8.1.5
#2 lore raw

Re: [PATCH 1/2] ocfs2: Fix llseek() semantics and do some cleanup

[Add ocfs2-devel to CC-list]

Hello Richard,

Thanks for your patch.

Quoted 5 lines · Richard Yao

There are multiple issues with the custom llseek implemented in ocfs2 for implementing SEEK_HOLE/SEEK_DATA.

1. It takes the inode->i_mutex lock before calling generic_file_llseek(), which is unnecessary.

Agree, but please see my comments below.

Quoted 4 lines

2. It fails to take the filp->f_lock spinlock before modifying filp->f_pos and filp->f_version, which differs from generic_file_llseek().

Quoted 8 lines

3. It does a offset > inode->i_sb->s_maxbytes check that permits seeking up to the maximum file size possible on the ocfs2 filesystem, even when it is past the end of the file. Seeking beyond that (if possible), would return EINVAL instead of ENXIO.

4. The switch statement tries to cover all whence values when in reality it should only care about SEEK_HOLE/SEEK_DATA. Any other cases should be passsed to generic_file_llseek().

I have another patch set for refactoring ocfs2_file_llseek() but not yet found time to run a comprehensive tests. It can solve the existing issues but also improved the SEEK_DATA/SEEK_HOLE for unwritten extents, i.e. OCFS2_EXT_UNWRITTEN.

With this change, SEEK_DATA/SEEK_HOLE will go into separate function with a little code duplication instead of the current mix-ups in ocfs2_seek_data_hole_offset(), i.e,

loff_t ocfs2_file_llseek()

{
	switch (origin) {
        case SEEK_END:
        case SEEK_CUR:
        case SEEK_SET:
                return generic_file_llseek(file, offset, origin);
        case SEEK_DATA:
                return ocfs2_seek_data(file, offset);
        case SEEK_HOLE:
                return ocfs2_seek_hole(file, offset);
        default:
                return -EINVAL;
        }
}

I personally like keeping SEEK_DATA/SEEK_HOLE in switch...case style rather than dealing with them in a condition check block.

Thanks,
-Jeff

Quoted 106 lines

btrfs_file_llseek() and ocfs2_file_llseek() are extremely similar and consequently, contain many of the same flaws. Li Dongyang filed a pull request with ZFSOnLinux for SEEK_HOLE/SEEK_DATA support that included a custom llseek function that appears to have been modelled after the one in ocfs2. The similarity was strong enough that it suffered from many of the same flaws, which I caught during review. I addressed the issues with his patch with one that I wrote. Since a small percentage of Gentoo Linux users are affected by these flaws, I decided to adapt that code that to btrfs (separate patch) and ocfs2.

Note that commit 48802c8ae2a9d618ec734a61283d645ad527e06c by Jeff Liu at Oracle mostly addressed #3 in btrfs. The only lingering issue was that the offset > inode->i_sb->s_maxbytes check became dead code. The ocfs2 code was not fortunate enough to have had a similar correction until now.

---
 fs/ocfs2/file.c | 65 ++++++++++++++++++++++-----------------------------------
 1 file changed, 25 insertions(+), 40 deletions(-)
fs/ocfs2/file.c +25 -40
index ff54014..84f8c9c 100644
@@ -2615,54 +2615,39 @@ bail:
 }
 
 /* Refer generic_file_llseek_unlocked() */
-static loff_t ocfs2_file_llseek(struct file *file, loff_t offset, int whence)
+static loff_t ocfs2_file_llseek(struct file *filp, loff_t offset, int whence)
 {
-	struct inode *inode = file->f_mapping->host;
-	int ret = 0;
+	if (whence == SEEK_DATA || whence == SEEK_HOLE) {
+		struct inode *inode = filp->f_mapping->host;
+		int ret;
 
-	mutex_lock(&inode->i_mutex);
+		if (offset < 0 && !(filp->f_mode & FMODE_UNSIGNED_OFFSET))
+			return -EINVAL;
 
-	switch (whence) {
-	case SEEK_SET:
-		break;
-	case SEEK_END:
-		offset += inode->i_size;
-		break;
-	case SEEK_CUR:
-		if (offset == 0) {
-			offset = file->f_pos;
-			goto out;
+		if (offset >= i_size_read(inode)) {
+			return -ENXIO;
 		}
-		offset += file->f_pos;
-		break;
-	case SEEK_DATA:
-	case SEEK_HOLE:
-		ret = ocfs2_seek_data_hole_offset(file, &offset, whence);
-		if (ret)
-			goto out;
-		break;
-	default:
-		ret = -EINVAL;
-		goto out;
-	}
 
-	if (offset < 0 && !(file->f_mode & FMODE_UNSIGNED_OFFSET))
-		ret = -EINVAL;
-	if (!ret && offset > inode->i_sb->s_maxbytes)
-		ret = -EINVAL;
-	if (ret)
-		goto out;
+		mutex_lock(&inode->i_mutex);
+		ret = ocfs2_seek_data_hole_offset(filp, &offset, whence);
+		mutex_unlock(&inode->i_mutex);
+
+		if (ret) {
+			return ret;
+		}
 
-	if (offset != file->f_pos) {
-		file->f_pos = offset;
-		file->f_version = 0;
+		if (offset != filp->f_pos) {
+			spin_lock(&filp->f_lock);
+			filp->f_pos = offset;
+			filp->f_version = 0;
+			spin_unlock(&filp->f_lock);
+		}
+
+		return offset;
 	}
 
-out:
-	mutex_unlock(&inode->i_mutex);
-	if (ret)
-		return ret;
-	return offset;
+	return generic_file_llseek(filp, offset, whence);
+
 }
 
 const struct inode_operations ocfs2_file_iops = {
#3 lore raw

Re: [Ocfs2-devel] [PATCH 1/2] ocfs2: Fix llseek() semantics and do some cleanup

Hello, Richard and Jeff,
we found that llseek has another bug when in SEEK_END. it should be
add the inode lock and unlock.
this bug can be reproduce the following scenario:
on one nodeA, open the file and then write some data to file and
close the file .
on another nodeB , open the file and llseek the end of file . the
position of file is old.

Quoted 169 lines · Jeff Liu

[Add ocfs2-devel to CC-list]

Hello Richard,

Thanks for your patch.

Quoted 5 lines · Richard Yao

There are multiple issues with the custom llseek implemented in ocfs2 for implementing SEEK_HOLE/SEEK_DATA.

1. It takes the inode->i_mutex lock before calling generic_file_llseek(), which is unnecessary.

Agree, but please see my comments below.

Quoted 11 lines

2. It fails to take the filp->f_lock spinlock before modifying filp->f_pos and filp->f_version, which differs from generic_file_llseek().

3. It does a offset> inode->i_sb->s_maxbytes check that permits seeking up to the maximum file size possible on the ocfs2 filesystem, even when it is past the end of the file. Seeking beyond that (if possible), would return EINVAL instead of ENXIO.

4. The switch statement tries to cover all whence values when in reality it should only care about SEEK_HOLE/SEEK_DATA. Any other cases should be passsed to generic_file_llseek().

I have another patch set for refactoring ocfs2_file_llseek() but not yet found time to run a comprehensive tests. It can solve the existing issues but also improved the SEEK_DATA/SEEK_HOLE for unwritten extents, i.e. OCFS2_EXT_UNWRITTEN.

With this change, SEEK_DATA/SEEK_HOLE will go into separate function with a little code duplication instead of the current mix-ups in ocfs2_seek_data_hole_offset(), i.e,

loff_t ocfs2_file_llseek()

{
	switch (origin) {
         case SEEK_END:
         case SEEK_CUR:
         case SEEK_SET:
                 return generic_file_llseek(file, offset, origin);
         case SEEK_DATA:
                 return ocfs2_seek_data(file, offset);
         case SEEK_HOLE:
                 return ocfs2_seek_hole(file, offset);
         default:
                 return -EINVAL;
         }
}

I personally like keeping SEEK_DATA/SEEK_HOLE in switch...case style rather than dealing with them in a condition check block.

Thanks,
-Jeff

Quoted 95 lines

btrfs_file_llseek() and ocfs2_file_llseek() are extremely similar and consequently, contain many of the same flaws. Li Dongyang filed a pull request with ZFSOnLinux for SEEK_HOLE/SEEK_DATA support that included a custom llseek function that appears to have been modelled after the one in ocfs2. The similarity was strong enough that it suffered from many of the same flaws, which I caught during review. I addressed the issues with his patch with one that I wrote. Since a small percentage of Gentoo Linux users are affected by these flaws, I decided to adapt that code that to btrfs (separate patch) and ocfs2.

Note that commit 48802c8ae2a9d618ec734a61283d645ad527e06c by Jeff Liu at Oracle mostly addressed #3 in btrfs. The only lingering issue was that the offset> inode->i_sb->s_maxbytes check became dead code. The ocfs2 code was not fortunate enough to have had a similar correction until now.

---
  fs/ocfs2/file.c | 65 ++++++++++++++++++++++-----------------------------------
  1 file changed, 25 insertions(+), 40 deletions(-)
fs/ocfs2/file.c +23 -35
index ff54014..84f8c9c 100644
@@ -2615,54 +2615,39 @@ bail:
  }
 
  /* Refer generic_file_llseek_unlocked() */
-static loff_t ocfs2_file_llseek(struct file *file, loff_t offset, int whence)
+static loff_t ocfs2_file_llseek(struct file *filp, loff_t offset, int whence)
  {
-	struct inode *inode = file->f_mapping->host;
-	int ret = 0;
+	if (whence == SEEK_DATA || whence == SEEK_HOLE) {
+		struct inode *inode = filp->f_mapping->host;
+		int ret;
 
-	mutex_lock(&inode->i_mutex);
+		if (offset<  0&&  !(filp->f_mode&  FMODE_UNSIGNED_OFFSET))
+			return -EINVAL;
 
-	switch (whence) {
-	case SEEK_SET:
-		break;
-	case SEEK_END:
-		offset += inode->i_size;
-		break;
-	case SEEK_CUR:
-		if (offset == 0) {
-			offset = file->f_pos;
-			goto out;
+		if (offset>= i_size_read(inode)) {
+			return -ENXIO;
  		}
-		offset += file->f_pos;
-		break;
-	case SEEK_DATA:
-	case SEEK_HOLE:
-		ret = ocfs2_seek_data_hole_offset(file,&offset, whence);
-		if (ret)
-			goto out;
-		break;
-	default:
-		ret = -EINVAL;
-		goto out;
-	}
 
-	if (offset<  0&&  !(file->f_mode&  FMODE_UNSIGNED_OFFSET))
-		ret = -EINVAL;
-	if (!ret&&  offset>  inode->i_sb->s_maxbytes)
-		ret = -EINVAL;
-	if (ret)
-		goto out;
+		mutex_lock(&inode->i_mutex);
+		ret = ocfs2_seek_data_hole_offset(filp,&offset, whence);
+		mutex_unlock(&inode->i_mutex);
+
+		if (ret) {
+			return ret;
+		}
 
-	if (offset != file->f_pos) {
-		file->f_pos = offset;
-		file->f_version = 0;
+		if (offset != filp->f_pos) {
+			spin_lock(&filp->f_lock);
+			filp->f_pos = offset;
+			filp->f_version = 0;
+			spin_unlock(&filp->f_lock);
+		}
+
+		return offset;
  	}
 

fs/ocfs2/file.c bail: · line 2653

Quoted 10 lines
fs/ocfs2/file.c +2 -5
-out:
-	mutex_unlock(&inode->i_mutex);
-	if (ret)
-		return ret;
-	return offset;
+	return generic_file_llseek(filp, offset, whence);
+
  }
 
  const struct inode_operations ocfs2_file_iops = {

_______________________________________________
Ocfs2-devel mailing list
Ocfs2-devel at oss.oracle.com
https://oss.oracle.com/mailman/listinfo/ocfs2-devel

.

#4 lore raw

Re: [Ocfs2-devel] [PATCH 1/2] ocfs2: Fix llseek() semantics and do some cleanup

Quoted 8 lines · shencanquan

Hello, Richard and Jeff,
we found that llseek has another bug when in SEEK_END. it should be
add the inode lock and unlock.
this bug can be reproduce the following scenario:
on one nodeA, open the file and then write some data to file and
close the file .
on another nodeB , open the file and llseek the end of file . the
position of file is old.

Did these operations occur sequentially or did they occur concurrently?

If you meant the former, the inode cache is not being invalidated. That should be a bug because Oracle claims OCFS2 is cache-coherent. However, it is possible that this case was left out of the cache-coherence protocol for performance purposes. If that is the case, then this would be by design. someone who works for Oracle would need to comment on that though.

If you meant the latter, you should ask yourself what would happen when you run two separate programs on the same file in a local filesystem. There should be no way to avoid a race without some kind of a locking mechanism.

#5 lore raw

Re: [Ocfs2-devel] [PATCH 1/2] ocfs2: Fix llseek() semantics and do some cleanup

Quoted 9 lines · Richard Yao
Quoted 8 lines · shencanquan

Hello, Richard and Jeff,
we found that llseek has another bug when in SEEK_END. it should be
add the inode lock and unlock.
this bug can be reproduce the following scenario:
on one nodeA, open the file and then write some data to file and
close the file .
on another nodeB , open the file and llseek the end of file . the
position of file is old.

Quoted 8 lines

Did these operations occur sequentially or did they occur concurrently?

If you meant the former, the inode cache is not being invalidated. That should be a bug because Oracle claims OCFS2 is cache-coherent. However, it is possible that this case was left out of the cache-coherence protocol for performance purposes. If that is the case, then this would be by design. someone who works for Oracle would need to comment on that though.

it is a occur sequentially. after close the file on NodeA , on
nodeB and then open the file and llseek the end of file.

Quoted 6 lines

If you meant the latter, you should ask yourself what would happen when you run two separate programs on the same file in a local filesystem. There should be no way to avoid a race without some kind of a locking mechanism.

#6 lore raw

Re: [Ocfs2-devel] [PATCH 1/2] ocfs2: Fix llseek() semantics and do some cleanup

Remove kernel-dev from the CC-list as this is particular to OCFS2.

Quoted 10 lines · shencanquan
Quoted 9 lines · Richard Yao
Quoted 8 lines · shencanquan

Hello, Richard and Jeff,
we found that llseek has another bug when in SEEK_END. it should be
add the inode lock and unlock.
this bug can be reproduce the following scenario:
on one nodeA, open the file and then write some data to file and
close the file .
on another nodeB , open the file and llseek the end of file . the
position of file is old.

This sounds like a bug because SEEK_END references the file size, hence it requires the OCFS2 specified inode lock protection.

So patch is always welcome.

Thanks,
-Jeff

Quoted 20 lines
Quoted 8 lines

Did these operations occur sequentially or did they occur concurrently?

If you meant the former, the inode cache is not being invalidated. That should be a bug because Oracle claims OCFS2 is cache-coherent. However, it is possible that this case was left out of the cache-coherence protocol for performance purposes. If that is the case, then this would be by design. someone who works for Oracle would need to comment on that though.

it is a occur sequentially. after close the file on NodeA , on nodeB
and then open the file and llseek the end of file.

Quoted 6 lines

If you meant the latter, you should ask yourself what would happen when you run two separate programs on the same file in a local filesystem. There should be no way to avoid a race without some kind of a locking mechanism.

#7 lore raw

Re: [PATCH 1/2] ocfs2: Fix llseek() semantics and do some cleanup

In reply to Jeff Liu
Quoted 52 lines · Jeff Liu

[Add ocfs2-devel to CC-list]

Hello Richard,

Thanks for your patch.

Quoted 5 lines · Richard Yao

There are multiple issues with the custom llseek implemented in ocfs2 for implementing SEEK_HOLE/SEEK_DATA.

1. It takes the inode->i_mutex lock before calling generic_file_llseek(), which is unnecessary.

Agree, but please see my comments below.

Quoted 4 lines

2. It fails to take the filp->f_lock spinlock before modifying filp->f_pos and filp->f_version, which differs from generic_file_llseek().

Quoted 8 lines

3. It does a offset > inode->i_sb->s_maxbytes check that permits seeking up to the maximum file size possible on the ocfs2 filesystem, even when it is past the end of the file. Seeking beyond that (if possible), would return EINVAL instead of ENXIO.

4. The switch statement tries to cover all whence values when in reality it should only care about SEEK_HOLE/SEEK_DATA. Any other cases should be passsed to generic_file_llseek().

I have another patch set for refactoring ocfs2_file_llseek() but not yet found time to run a comprehensive tests. It can solve the existing issues but also improved the SEEK_DATA/SEEK_HOLE for unwritten extents, i.e. OCFS2_EXT_UNWRITTEN.

With this change, SEEK_DATA/SEEK_HOLE will go into separate function with a little code duplication instead of the current mix-ups in ocfs2_seek_data_hole_offset(), i.e,

loff_t ocfs2_file_llseek()

{
	switch (origin) {
        case SEEK_END:
        case SEEK_CUR:
        case SEEK_SET:
                return generic_file_llseek(file, offset, origin);
        case SEEK_DATA:
                return ocfs2_seek_data(file, offset);
        case SEEK_HOLE:
                return ocfs2_seek_hole(file, offset);
        default:
                return -EINVAL;
        }
}
Quoted 2 lines

I personally like keeping SEEK_DATA/SEEK_HOLE in switch...case style rather than dealing with them in a condition check block.

I would prefer to see the code structured like this:

loff_t ocfs2_file_llseek()

{
	switch (origin) {
        case SEEK_DATA:
                return ocfs2_seek_data(file, offset);
        case SEEK_HOLE:
                return ocfs2_seek_hole(file, offset);
        default:
                return generic_file_llseek(file, offset, origin);
        }
}

Unfortunately, I just noticed that this code has a problem. In specific, generic_file_llseek() calls generic_file_llseek_size(), which has a switch statement for whence that fails to distinguish between SEEK_SET and invalid whence values. Invalid whence values are mapped to SEEK_SET instead of returning EINVAL, which is wrong. That issue affects all filesystems that do not specify a custom llseek() function and it would affect ocfs2 if my version of the function is used.

#8 lore raw

Re: [PATCH 1/2] ocfs2: Fix llseek() semantics and do some cleanup

Quoted 70 lines · Richard Yao
Quoted 52 lines · Jeff Liu

[Add ocfs2-devel to CC-list]

Hello Richard,

Thanks for your patch.

Quoted 5 lines · Richard Yao

There are multiple issues with the custom llseek implemented in ocfs2 for implementing SEEK_HOLE/SEEK_DATA.

1. It takes the inode->i_mutex lock before calling generic_file_llseek(), which is unnecessary.

Agree, but please see my comments below.

Quoted 4 lines

2. It fails to take the filp->f_lock spinlock before modifying filp->f_pos and filp->f_version, which differs from generic_file_llseek().

Quoted 8 lines

3. It does a offset > inode->i_sb->s_maxbytes check that permits seeking up to the maximum file size possible on the ocfs2 filesystem, even when it is past the end of the file. Seeking beyond that (if possible), would return EINVAL instead of ENXIO.

4. The switch statement tries to cover all whence values when in reality it should only care about SEEK_HOLE/SEEK_DATA. Any other cases should be passsed to generic_file_llseek().

I have another patch set for refactoring ocfs2_file_llseek() but not yet found time to run a comprehensive tests. It can solve the existing issues but also improved the SEEK_DATA/SEEK_HOLE for unwritten extents, i.e. OCFS2_EXT_UNWRITTEN.

With this change, SEEK_DATA/SEEK_HOLE will go into separate function with a little code duplication instead of the current mix-ups in ocfs2_seek_data_hole_offset(), i.e,

loff_t ocfs2_file_llseek()

{
	switch (origin) {
        case SEEK_END:
        case SEEK_CUR:
        case SEEK_SET:
                return generic_file_llseek(file, offset, origin);
        case SEEK_DATA:
                return ocfs2_seek_data(file, offset);
        case SEEK_HOLE:
                return ocfs2_seek_hole(file, offset);
        default:
                return -EINVAL;
        }
}
Quoted 2 lines

I personally like keeping SEEK_DATA/SEEK_HOLE in switch...case style rather than dealing with them in a condition check block.

I would prefer to see the code structured like this:

loff_t ocfs2_file_llseek()

{
	switch (origin) {
        case SEEK_DATA:
                return ocfs2_seek_data(file, offset);
        case SEEK_HOLE:
                return ocfs2_seek_hole(file, offset);
        default:
                return generic_file_llseek(file, offset, origin);
        }
}
Quoted 7 lines

Unfortunately, I just noticed that this code has a problem. In specific, generic_file_llseek() calls generic_file_llseek_size(), which has a switch statement for whence that fails to distinguish between SEEK_SET and invalid whence values. Invalid whence values are mapped to SEEK_SET instead of returning EINVAL, which is wrong. That issue affects all filesystems that do not specify a custom llseek() function and it would affect ocfs2 if my version of the function is used.

Hmm?? Did you mean to say that an invalid whence(i.e, whence > SEEK_MAX) can be passed into generic_file_llseek()?
If so, I don't think that is a problem because any invalid whence should be rejected at the entrance of VFS lseek(2) with EINVAL.

Thanks,
-Jeff

#9 lore raw

Re: [PATCH 1/2] ocfs2: Fix llseek() semantics and do some cleanup

Quoted 80 lines · Jeff Liu
Quoted 70 lines · Richard Yao
Quoted 52 lines · Jeff Liu

[Add ocfs2-devel to CC-list]

Hello Richard,

Thanks for your patch.

Quoted 5 lines · Richard Yao

There are multiple issues with the custom llseek implemented in ocfs2 for implementing SEEK_HOLE/SEEK_DATA.

1. It takes the inode->i_mutex lock before calling generic_file_llseek(), which is unnecessary.

Agree, but please see my comments below.

Quoted 4 lines

2. It fails to take the filp->f_lock spinlock before modifying filp->f_pos and filp->f_version, which differs from generic_file_llseek().

Quoted 8 lines

3. It does a offset > inode->i_sb->s_maxbytes check that permits seeking up to the maximum file size possible on the ocfs2 filesystem, even when it is past the end of the file. Seeking beyond that (if possible), would return EINVAL instead of ENXIO.

4. The switch statement tries to cover all whence values when in reality it should only care about SEEK_HOLE/SEEK_DATA. Any other cases should be passsed to generic_file_llseek().

I have another patch set for refactoring ocfs2_file_llseek() but not yet found time to run a comprehensive tests. It can solve the existing issues but also improved the SEEK_DATA/SEEK_HOLE for unwritten extents, i.e. OCFS2_EXT_UNWRITTEN.

With this change, SEEK_DATA/SEEK_HOLE will go into separate function with a little code duplication instead of the current mix-ups in ocfs2_seek_data_hole_offset(), i.e,

loff_t ocfs2_file_llseek()

{
	switch (origin) {
        case SEEK_END:
        case SEEK_CUR:
        case SEEK_SET:
                return generic_file_llseek(file, offset, origin);
        case SEEK_DATA:
                return ocfs2_seek_data(file, offset);
        case SEEK_HOLE:
                return ocfs2_seek_hole(file, offset);
        default:
                return -EINVAL;
        }
}
Quoted 2 lines

I personally like keeping SEEK_DATA/SEEK_HOLE in switch...case style rather than dealing with them in a condition check block.

I would prefer to see the code structured like this:

loff_t ocfs2_file_llseek()

{
	switch (origin) {
        case SEEK_DATA:
                return ocfs2_seek_data(file, offset);
        case SEEK_HOLE:
                return ocfs2_seek_hole(file, offset);
        default:
                return generic_file_llseek(file, offset, origin);
        }
}
Quoted 7 lines

Unfortunately, I just noticed that this code has a problem. In specific, generic_file_llseek() calls generic_file_llseek_size(), which has a switch statement for whence that fails to distinguish between SEEK_SET and invalid whence values. Invalid whence values are mapped to SEEK_SET instead of returning EINVAL, which is wrong. That issue affects all filesystems that do not specify a custom llseek() function and it would affect ocfs2 if my version of the function is used.

Quoted 8 lines

Hmm?? Did you mean to say that an invalid whence(i.e, whence > SEEK_MAX) can be passed into generic_file_llseek()?
If so, I don't think that is a problem because any invalid whence should be rejected at the entrance of VFS lseek(2) with EINVAL.

Thanks,
-Jeff

You are correct. I had not looked there.

#10 lore raw

PATCH 2/2 btrfs: Cleanup llseek()

In reply to Richard Yao

There are multiple issues with the custom llseek implemented in btrfs for implementing SEEK_HOLE/SEEK_DATA.

1. It takes the inode->i_mutex lock before calling generic_file_llseek(), which is unnecessary.

2. It fails to take the filp->f_lock spinlock before modifying filp->f_pos and filp->f_version, which differs from generic_file_llseek().

3. It does a offset > inode->i_sb->s_maxbytes check that is dead code because the the offset >= i_size_read(inode) will evaluate to true and return ENXIO whenever the former comparison evaluates to true.

4. The switch statement tries to cover all whence values when in reality it should only care about SEEK_HOLE/SEEK_DATA. Any other cases should be passsed to generic_file_llseek().

btrfs_file_llseek() and ocfs2_file_llseek() are extremely similar and consequently, contain many of the same flaws. Li Dongyang filed a pull request with ZFSOnLinux for SEEK_HOLE/SEEK_DATA support that included a custom llseek function that appears to have been modelled after the one in ocfs2. The similarity was strong enough that it suffered from many of the same flaws, which I caught during review. I addressed the issues with his patch with one that I wrote. I decided to adapt that code to both btrfs and ocfs2 (separate patch) because a small percentage of Gentoo Linux users are affected by these flaws.

Note that #3 would have been worse had it not been for commit 48802c8ae2a9d618ec734a61283d645ad527e06c by Jeff Liu at Oracle. Prior to his commit, btrfs llseek() would permit seeking up to the maximum file size possible on the btrfs filesystem, even when it is past the end of the file. Seeking beyond that (if possible), would return EINVAL instead of ENXIO. That was the same behavior that the ocfs2 llseek had. corrected that issue. However, the ocfs2 code was not fortunate enough to have had this corrected at that time.

---
 fs/btrfs/file.c | 49 ++++++++++++++++++++-----------------------------
 1 file changed, 20 insertions(+), 29 deletions(-)
fs/btrfs/file.c +20 -29
index 4205ba7..1f437d8 100644
@@ -2400,48 +2400,39 @@ out:
 	return ret;
 }
 
-static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
+static loff_t btrfs_file_llseek(struct file *filp, loff_t offset, int whence)
 {
-	struct inode *inode = file->f_mapping->host;
-	int ret;
+	if (whence == SEEK_DATA || whence == SEEK_HOLE) {
+		struct inode *inode = filp->f_mapping->host;
+		int ret;
+
+		if (offset < 0 && !(filp->f_mode & FMODE_UNSIGNED_OFFSET))
+			return -EINVAL;
 
-	mutex_lock(&inode->i_mutex);
-	switch (whence) {
-	case SEEK_END:
-	case SEEK_CUR:
-		offset = generic_file_llseek(file, offset, whence);
-		goto out;
-	case SEEK_DATA:
-	case SEEK_HOLE:
 		if (offset >= i_size_read(inode)) {
-			mutex_unlock(&inode->i_mutex);
 			return -ENXIO;
 		}
 
+		mutex_lock(&inode->i_mutex);
 		ret = find_desired_extent(inode, &offset, whence);
+		mutex_unlock(&inode->i_mutex);
+
 		if (ret) {
-			mutex_unlock(&inode->i_mutex);
 			return ret;
 		}
-	}
 
-	if (offset < 0 && !(file->f_mode & FMODE_UNSIGNED_OFFSET)) {
-		offset = -EINVAL;
-		goto out;
-	}
-	if (offset > inode->i_sb->s_maxbytes) {
-		offset = -EINVAL;
-		goto out;
-	}
+		if (offset != filp->f_pos) {
+			spin_lock(&filp->f_lock);
+			filp->f_pos = offset;
+			filp->f_version = 0;
+			spin_unlock(&filp->f_lock);
+		}
 
-	/* Special lock needed here? */
-	if (offset != file->f_pos) {
-		file->f_pos = offset;
-		file->f_version = 0;
+		return offset;
 	}
-out:
-	mutex_unlock(&inode->i_mutex);
-	return offset;
+
+	return generic_file_llseek(filp, offset, whence);
+
 }
 
 const struct file_operations btrfs_file_operations = {
Signature · 1 line
1.8.1.5