Skip to contentSkip to conversation

Use exclusive lock for file_remove_privs

PATCH Superseded Bernd Schubert 4 participants 12 messages Aug 31, 2023 – Sep 7, 2023

Subsystems (2)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

v2 cover Use exclusive lock for file_remove_privs

While adding shared direct IO write locks to fuse Miklos noticed that file_remove_privs() needs an exclusive lock. I then noticed that btrfs actually has the same issue as I had in my patch, it was calling into that function with a shared lock. This series adds a new exported function file_needs_remove_privs(), which used by the follow up btrfs patch and will be used by the DIO code path in fuse as well. If that function returns any mask the shared lock needs to be dropped and replaced by the exclusive variant.

Note: Compilation tested only.

v2:
Already check for IS_NOSEC in btrfs_direct_write before the first lock is taken.
Slight modification to make the code easier to read (boolean pointer is passed to btrfs_write_check, instead of flags).

Cc Miklos Szeredi <[email protected]>
Cc Dharmendra Singh <[email protected]>
Cc Josef Bacik <[email protected]>
Cc Alexander Viro <[email protected]>
Cc Christian Brauner <[email protected]>

Bernd Schubert (2):
fs: Add and export file_needs_remove_privs
btrfs: file_remove_privs needs an exclusive lock

 fs/btrfs/file.c    | 37 +++++++++++++++++++++++++++++--------
 fs/inode.c         |  8 ++++++++
 include/linux/fs.h |  1 +
 3 files changed, 38 insertions(+), 8 deletions(-)
Signature · 1 line
2.39.2
#1 lore raw

PATCH v2 1/2 fs: Add and export file_needs_remove_privs

File systems want to hold a shared lock for DIO writes, but may need to drop file priveliges - that a requires an exclusive lock. The new export function file_needs_remove_privs() is added in order to first check if that is needed.

Cc Miklos Szeredi <[email protected]>
Cc Dharmendra Singh <[email protected]>
Cc Josef Bacik <[email protected]>
Cc Alexander Viro <[email protected]>
Cc Christian Brauner <[email protected]>
---
 fs/inode.c         | 8 ++++++++
 include/linux/fs.h | 1 +
 2 files changed, 9 insertions(+)
fs/inode.c +8 -0
index 67611a360031..9b05db602e41 100644
@@ -2013,6 +2013,14 @@ int dentry_needs_remove_privs(struct mnt_idmap *idmap,
 	return mask;
 }
 
+int file_needs_remove_privs(struct file *file)
+{
+	struct dentry *dentry = file_dentry(file);
+
+	return dentry_needs_remove_privs(file_mnt_idmap(file), dentry);
+}
+EXPORT_SYMBOL_GPL(file_needs_remove_privs);
+
 static int __remove_privs(struct mnt_idmap *idmap,
 			  struct dentry *dentry, int kill)
 {
include/linux/fs.h +1 -0
index 562f2623c9c9..9245f0de00bc 100644
@@ -2721,6 +2721,7 @@ extern struct inode *new_inode_pseudo(struct super_block *sb);
 extern struct inode *new_inode(struct super_block *sb);
 extern void free_inode_nonrcu(struct inode *inode);
 extern int setattr_should_drop_suidgid(struct mnt_idmap *, struct inode *);
+int file_needs_remove_privs(struct file *);
 extern int file_remove_privs(struct file *);
 int setattr_should_drop_sgid(struct mnt_idmap *idmap,
 			     const struct inode *inode);
Signature · 1 line
2.39.2
#2 lore raw

Re: [PATCH 1/2] fs: Add and export file_needs_remove_privs

Quoted 2 lines · Bernd Schubert

File systems want to hold a shared lock for DIO writes, but may need to drop file priveliges - that a requires an

s/priveliges/privileges/
s/that a requires/that requires/

Quoted 50 lines

exclusive lock. The new export function file_needs_remove_privs() is added in order to first check if that is needed.

Cc Miklos Szeredi <[email protected]>
Cc Dharmendra Singh <[email protected]>
Cc Josef Bacik <[email protected]>
Cc Alexander Viro <[email protected]>
Cc Christian Brauner <[email protected]>
---
 fs/inode.c         | 8 ++++++++
 include/linux/fs.h | 1 +
 2 files changed, 9 insertions(+)
fs/inode.c +8 -0
index 67611a360031..9b05db602e41 100644
@@ -2013,6 +2013,14 @@ int dentry_needs_remove_privs(struct mnt_idmap *idmap,
 	return mask;
 }
 
+int file_needs_remove_privs(struct file *file)
+{
+	struct dentry *dentry = file_dentry(file);
+
+	return dentry_needs_remove_privs(file_mnt_idmap(file), dentry);
+}
+EXPORT_SYMBOL_GPL(file_needs_remove_privs);
+
 static int __remove_privs(struct mnt_idmap *idmap,
 			  struct dentry *dentry, int kill)
 {
include/linux/fs.h +1 -0
index 562f2623c9c9..9245f0de00bc 100644
@@ -2721,6 +2721,7 @@ extern struct inode *new_inode_pseudo(struct super_block *sb);
 extern struct inode *new_inode(struct super_block *sb);
 extern void free_inode_nonrcu(struct inode *inode);
 extern int setattr_should_drop_suidgid(struct mnt_idmap *, struct inode *);
+int file_needs_remove_privs(struct file *);
 extern int file_remove_privs(struct file *);
 int setattr_should_drop_sgid(struct mnt_idmap *idmap,
 			     const struct inode *inode);
Signature · 1 line
2.39.2
#3 lore raw

Re: [PATCH 1/2] fs: Add and export file_needs_remove_privs

In reply to Bernd Schubert
Quoted 23 lines · Bernd Schubert

File systems want to hold a shared lock for DIO writes, but may need to drop file priveliges - that a requires an exclusive lock. The new export function file_needs_remove_privs() is added in order to first check if that is needed.

Cc Miklos Szeredi <[email protected]>
Cc Dharmendra Singh <[email protected]>
Cc Josef Bacik <[email protected]>
Cc Alexander Viro <[email protected]>
Cc Christian Brauner <[email protected]>
---
 fs/inode.c         | 8 ++++++++
 include/linux/fs.h | 1 +
 2 files changed, 9 insertions(+)
fs/inode.c +0 -0
index 67611a360031..9b05db602e41 100644

fs/inode.c int dentry_needs_remove_privs(struct mnt_idmap *idmap, · line 2020

Quoted 8 lines
fs/inode.c +5 -0
@@ -2013,6 +2013,14 @@ int dentry_needs_remove_privs(struct mnt_idmap *idmap,
 	return mask;
 }
 
+int file_needs_remove_privs(struct file *file)
+{
+	struct dentry *dentry = file_dentry(file);
+
+	return dentry_needs_remove_privs(file_mnt_idmap(file), dentry);

Ugh, I wanted to propose to get rid of this dentry dance but I propsed that before and remembered it's because of __vfs_getxattr() which is called from the capability security hook that we need it...

#4 lore raw

Re: [PATCH 1/2] fs: Add and export file_needs_remove_privs

Quoted 27 lines · Christian Brauner
Quoted 26 lines · Bernd Schubert

File systems want to hold a shared lock for DIO writes, but may need to drop file priveliges - that a requires an exclusive lock. The new export function file_needs_remove_privs() is added in order to first check if that is needed.

Cc Miklos Szeredi <[email protected]>
Cc Dharmendra Singh <[email protected]>
Cc Josef Bacik <[email protected]>
Cc Alexander Viro <[email protected]>
Cc Christian Brauner <[email protected]>
---
  fs/inode.c         | 8 ++++++++
  include/linux/fs.h | 1 +
  2 files changed, 9 insertions(+)
fs/inode.c +0 -0
index 67611a360031..9b05db602e41 100644
@@ -2013,6 +2013,14 @@ int dentry_needs_remove_privs(struct mnt_idmap *idmap,
  	return mask;
  }
 
Quoted 9 lines

fs/inode.c int dentry_needs_remove_privs(struct mnt_idmap *idmap, · line 2020

Quoted 5 lines
fs/inode.c +5 -0
+int file_needs_remove_privs(struct file *file)
+{
+	struct dentry *dentry = file_dentry(file);
+
+	return dentry_needs_remove_privs(file_mnt_idmap(file), dentry);

Ugh, I wanted to propose to get rid of this dentry dance but I propsed that before and remembered it's because of __vfs_getxattr() which is called from the capability security hook that we need it...

Is there anything specific you are suggesting?

And thanks for typo/grammar annotations, I'm going to wait for btrfs feedback before I'm going to send a new version.

Thanks,
Bernd

#5 lore raw

Re: [PATCH 1/2] fs: Add and export file_needs_remove_privs

Quoted 39 lines · Bernd Schubert
Quoted 19 lines · Christian Brauner
Quoted 18 lines · Bernd Schubert

File systems want to hold a shared lock for DIO writes, but may need to drop file priveliges - that a requires an exclusive lock. The new export function file_needs_remove_privs() is added in order to first check if that is needed.

Cc Miklos Szeredi <[email protected]>
Cc Dharmendra Singh <[email protected]>
Cc Josef Bacik <[email protected]>
Cc Alexander Viro <[email protected]>
Cc Christian Brauner <[email protected]>
---
  fs/inode.c         | 8 ++++++++
  include/linux/fs.h | 1 +
  2 files changed, 9 insertions(+)
Quoted 16 lines

fs/inode.c int dentry_needs_remove_privs(struct mnt_idmap *idmap, · line 2019

Quoted 12 lines
fs/inode.c +5 -0
index 67611a360031..9b05db602e41 100644
@@ -2013,6 +2013,14 @@ int dentry_needs_remove_privs(struct mnt_idmap *idmap,
  	return mask;
  }
+int file_needs_remove_privs(struct file *file)
+{
+	struct dentry *dentry = file_dentry(file);
+
+	return dentry_needs_remove_privs(file_mnt_idmap(file), dentry);

Ugh, I wanted to propose to get rid of this dentry dance but I propsed that before and remembered it's because of __vfs_getxattr() which is called from the capability security hook that we need it...

Quoted 1 line

Is there anything specific you are suggesting?

No, it's not actionable for you here. It would require adding inode methods to set and get filesystem capabilities and then converting it in such a way that we don't need to rely on passing dentries around. That's a separate larger patchset that we would need with surgery across a bunch of filesystems and the vfs - Seth (Forshee) has been working on this.

The callchains are just pointless which I remembered when I saw the patchset:

file_needs_remove_privs(file)
-> dentry_needs_remove_privs(dentry)
-> inode = d_inode(dentry)

      // do inode stuff
      // security_needs_*(dentry)

point is ideally we shouldn't need the dentry in *remove_privs() at all.

#6 lore raw

PATCH v2 2/2 btrfs: file_remove_privs needs an exclusive lock

In reply to Bernd Schubert

file_remove_privs might call into notify_change(), which requires to hold an exclusive lock.
In order to keep the shared lock for most IOs it now first checks if privilege changes are needed, then switches to the exclusive lock, rechecks and only then calls file_remove_privs. This makes usage of the new exported function file_needs_remove_privs().

The file_remove_privs code path is not optimized, under the assumption that it would be a rare call (file_remove_privs calls file_needs_remove_privs a 2nd time).

Fixes e9adabb9712e ("btrfs: use shared lock for direct writes within EOF")
Cc Christoph Hellwig <[email protected]>
Cc Goldwyn Rodrigues <[email protected]>
Cc Miklos Szeredi <[email protected]>
Cc Dharmendra Singh <[email protected]>
Cc Chris Mason <[email protected]>
Cc Josef Bacik <[email protected]>
Cc David Sterba <[email protected]>
---
 fs/btrfs/file.c | 41 ++++++++++++++++++++++++++++++++---------
 1 file changed, 32 insertions(+), 9 deletions(-)
fs/btrfs/file.c +32 -9
index fd03e689a6be..3162ec245d57 100644
@@ -1125,7 +1125,7 @@ static void update_time_for_write(struct inode *inode)
 }
 
 static int btrfs_write_check(struct kiocb *iocb, struct iov_iter *from,
-			     size_t count)
+			     size_t count, bool *shared_lock)
 {
 	struct file *file = iocb->ki_filp;
 	struct inode *inode = file_inode(file);
@@ -1145,9 +1145,17 @@ static int btrfs_write_check(struct kiocb *iocb, struct iov_iter *from,
 	    !(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
 		return -EAGAIN;
 
-	ret = file_remove_privs(file);
-	if (ret)
-		return ret;
+	ret = file_needs_remove_privs(file);
+	if (ret) {
+		if (shared_lock && *shared_lock) {
+			*shared_lock = false;
+			return -EAGAIN;
+		}
+
+		ret = file_remove_privs(file);
+		if (ret)
+			return ret;
+	}
 
 	/*
 	 * We reserve space for updating the inode when we reserve space for the
@@ -1204,7 +1212,7 @@ static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
 	if (ret <= 0)
 		goto out;
 
-	ret = btrfs_write_check(iocb, i, ret);
+	ret = btrfs_write_check(iocb, i, ret, NULL);
 	if (ret < 0)
 		goto out;
 
@@ -1462,13 +1470,20 @@ static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
 	ssize_t err;
 	unsigned int ilock_flags = 0;
 	struct iomap_dio *dio;
+	bool shared_lock;
 
 	if (iocb->ki_flags & IOCB_NOWAIT)
 		ilock_flags |= BTRFS_ILOCK_TRY;
 
-	/* If the write DIO is within EOF, use a shared lock */
-	if (iocb->ki_pos + iov_iter_count(from) <= i_size_read(inode))
+	/* If the write DIO is within EOF, use a shared lock and also only
+	 * if security bits will likely not be dropped. Either will need
+	 * to be rechecked after the lock was acquired.
+	 */
+	if (iocb->ki_pos + iov_iter_count(from) <= i_size_read(inode) &&
+	    IS_NOSEC(inode)) {
 		ilock_flags |= BTRFS_ILOCK_SHARED;
+		shared_lock = true;
+	}
 
 relock:
 	err = btrfs_inode_lock(BTRFS_I(inode), ilock_flags);
@@ -1481,8 +1496,15 @@ static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
 		return err;
 	}
 
-	err = btrfs_write_check(iocb, from, err);
+	err = btrfs_write_check(iocb, from, err, &shared_lock);
 	if (err < 0) {
+		if (err == -EAGAIN && ilock_flags & BTRFS_ILOCK_SHARED &&
+		    !shared_lock) {
+			btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
+			ilock_flags &= ~BTRFS_ILOCK_SHARED;
+			goto relock;
+		}
+
 		btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
 		goto out;
 	}
@@ -1496,6 +1518,7 @@ static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
 	    pos + iov_iter_count(from) > i_size_read(inode)) {
 		btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
 		ilock_flags &= ~BTRFS_ILOCK_SHARED;
+		shared_lock = false;
 		goto relock;
 	}
 
@@ -1632,7 +1655,7 @@ static ssize_t btrfs_encoded_write(struct kiocb *iocb, struct iov_iter *from,
 	if (ret || encoded->len == 0)
 		goto out;
 
-	ret = btrfs_write_check(iocb, from, encoded->len);
+	ret = btrfs_write_check(iocb, from, encoded->len, NULL);
 	if (ret < 0)
 		goto out;
 
Signature · 1 line
2.39.2
#7 lore raw
In reply to Bernd Schubert
Quoted 11 lines · Bernd Schubert

While adding shared direct IO write locks to fuse Miklos noticed that file_remove_privs() needs an exclusive lock. I then noticed that btrfs actually has the same issue as I had in my patch, it was calling into that function with a shared lock. This series adds a new exported function file_needs_remove_privs(), which used by the follow up btrfs patch and will be used by the DIO code path in fuse as well. If that function returns any mask the shared lock needs to be dropped and replaced by the exclusive variant.

Note: Compilation tested only.

The fix makes sense, there should be no noticeable performance impact, basically the same check is done in the newly exported helper for the IS_NOSEC bit. I can give it a test locally for the default case, I'm not sure if we have specific tests for the security layers in fstests.

Regarding merge, I can take the two patches via btrfs tree or can wait until the export is present in Linus' tree in case FUSE needs it independently.

#8 lore raw
Quoted 13 lines · David Sterba
Quoted 11 lines · Bernd Schubert

While adding shared direct IO write locks to fuse Miklos noticed that file_remove_privs() needs an exclusive lock. I then noticed that btrfs actually has the same issue as I had in my patch, it was calling into that function with a shared lock. This series adds a new exported function file_needs_remove_privs(), which used by the follow up btrfs patch and will be used by the DIO code path in fuse as well. If that function returns any mask the shared lock needs to be dropped and replaced by the exclusive variant.

Note: Compilation tested only.

Quoted 8 lines

The fix makes sense, there should be no noticeable performance impact, basically the same check is done in the newly exported helper for the IS_NOSEC bit. I can give it a test locally for the default case, I'm not sure if we have specific tests for the security layers in fstests.

Regarding merge, I can take the two patches via btrfs tree or can wait until the export is present in Linus' tree in case FUSE needs it independently.

Both fuse and btrfs need it afaict. We can grab it and provide a tag post -rc1? Whatever works best.

#9 lore raw
Quoted 19 lines · Christian Brauner
Quoted 18 lines · David Sterba
Quoted 11 lines · Bernd Schubert

While adding shared direct IO write locks to fuse Miklos noticed that file_remove_privs() needs an exclusive lock. I then noticed that btrfs actually has the same issue as I had in my patch, it was calling into that function with a shared lock. This series adds a new exported function file_needs_remove_privs(), which used by the follow up btrfs patch and will be used by the DIO code path in fuse as well. If that function returns any mask the shared lock needs to be dropped and replaced by the exclusive variant.

Note: Compilation tested only.

The fix makes sense, there should be no noticeable performance impact, basically the same check is done in the newly exported helper for the IS_NOSEC bit. I can give it a test locally for the default case, I'm not sure if we have specific tests for the security layers in fstests.

Quoted 6 lines
Quoted 3 lines

Regarding merge, I can take the two patches via btrfs tree or can wait until the export is present in Linus' tree in case FUSE needs it independently.

Both fuse and btrfs need it afaict. We can grab it and provide a tag post -rc1? Whatever works best.

fuse will need it for my direct IO patches - hopefully in 6.7. For btrfs it is a bug fix, should go in asap?

Christoph has some objections for to use the new exported helper (file_needs_remove_privs). Maybe I better send a version for btrfs that only uses S_NOSEC? For fuse we cannot use it, unfortunately.

Thanks,
Bernd

#10 lore raw
Quoted 29 lines · Bernd Schubert
Quoted 19 lines · Christian Brauner
Quoted 18 lines · David Sterba
Quoted 11 lines · Bernd Schubert

While adding shared direct IO write locks to fuse Miklos noticed that file_remove_privs() needs an exclusive lock. I then noticed that btrfs actually has the same issue as I had in my patch, it was calling into that function with a shared lock. This series adds a new exported function file_needs_remove_privs(), which used by the follow up btrfs patch and will be used by the DIO code path in fuse as well. If that function returns any mask the shared lock needs to be dropped and replaced by the exclusive variant.

Note: Compilation tested only.

The fix makes sense, there should be no noticeable performance impact, basically the same check is done in the newly exported helper for the IS_NOSEC bit. I can give it a test locally for the default case, I'm not sure if we have specific tests for the security layers in fstests.

Quoted 6 lines
Quoted 3 lines

Regarding merge, I can take the two patches via btrfs tree or can wait until the export is present in Linus' tree in case FUSE needs it independently.

Both fuse and btrfs need it afaict. We can grab it and provide a tag post -rc1? Whatever works best.

Quoted 6 lines

fuse will need it for my direct IO patches - hopefully in 6.7. For btrfs it is a bug fix, should go in asap?

Christoph has some objections for to use the new exported helper (file_needs_remove_privs). Maybe I better send a version for btrfs that only uses S_NOSEC? For fuse we cannot use it, unfortunately.

Sure.

#11 lore raw
In reply to Christian Brauner
Quoted 19 lines · Christian Brauner
Quoted 18 lines · David Sterba
Quoted 11 lines · Bernd Schubert

While adding shared direct IO write locks to fuse Miklos noticed that file_remove_privs() needs an exclusive lock. I then noticed that btrfs actually has the same issue as I had in my patch, it was calling into that function with a shared lock. This series adds a new exported function file_needs_remove_privs(), which used by the follow up btrfs patch and will be used by the DIO code path in fuse as well. If that function returns any mask the shared lock needs to be dropped and replaced by the exclusive variant.

Note: Compilation tested only.

The fix makes sense, there should be no noticeable performance impact, basically the same check is done in the newly exported helper for the IS_NOSEC bit. I can give it a test locally for the default case, I'm not sure if we have specific tests for the security layers in fstests.

Quoted 6 lines
Quoted 3 lines

Regarding merge, I can take the two patches via btrfs tree or can wait until the export is present in Linus' tree in case FUSE needs it independently.

Both fuse and btrfs need it afaict. We can grab it and provide a tag post -rc1? Whatever works best.

Git tree sync won't be needed, Bernd sent the fix within btrfs code.