mnt_want_write_file() has problem?
Hi,
While I'm reading some code, I suspected that mnt_want_write_file() may have wrong assumption. I think mnt_want_write_file() is assuming it increments ->mnt_writers if (file->f_mode & FMODE_WRITE). But, if it's special_file(), it is false?
Sorry, I'm still not checking all of those though. E.g. I'm thinking the below.
static inline int __get_file_write_access(struct inode *inode,
struct vfsmount *mnt)
{
[...]
if (!special_file(inode->i_mode)) {
/*
* Balanced in __fput()
*/
error = mnt_want_write(mnt);
if (error)
put_write_access(inode);
}
return error;
}Thanks.
Signature · 24 lines
OGAWA Hirofumi <[email protected]> Signed-off-by: OGAWA Hirofumi <[email protected]> --- fs/namespace.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff -puN fs/namespace.c~mnt_want_write-wrong-assume fs/namespace.c --- linux-2.6/fs/namespace.c~mnt_want_write-wrong-assume 2009-08-03 04:33:35.000000000 +0900 +++ linux-2.6-hirofumi/fs/namespace.c 2009-08-03 04:31:34.000000000 +0900 @@ -316,7 +316,8 @@ EXPORT_SYMBOL_GPL(mnt_clone_write); */ int mnt_want_write_file(struct file *file) { - if (!(file->f_mode & FMODE_WRITE)) + struct inode *inode = file->f_dentry->d_inode; + if (!(file->f_mode & FMODE_WRITE) || special_file(inode->i_mode)) return mnt_want_write(file->f_path.mnt); else return mnt_clone_write(file->f_path.mnt); _
Quoted 12 lines · OGAWA Hirofumi
While I'm reading some code, I suspected that
mnt_want_write_file()may have wrong assumption. I thinkmnt_want_write_file()is assuming it increments ->mnt_writers if (file->f_mode&FMODE_WRITE). But, if it'sspecial_file(), it is false?Sorry, I'm still not checking all of those though. E.g. I'm thinking the below.
static inline int __get_file_write_access(struct inode *inode,
struct vfsmount *mnt)
{
[...]
Quoted 10 lines
if (!special_file(inode->i_mode)) {
/*
* Balanced in __fput()
*/
error = mnt_want_write(mnt);
if (error)
put_write_access(inode);
}
return error;
}In practice I don't think this is an issue. We were never supposed to do mnt_want_write(mnt) for any 'struct file' that was a special_file(), specifically because of what you mention.
Nick's use of mnt_want_write_file() was a 1:1 drop-in for mnt_want_write(). So, if all is well in the world, there should not be any call sites where mnt_want_write_file() gets called on a special_file().
Future users of mnt_want_file_write() may not notice this fact, though. This is probably worth at least a note in the documentation or perhaps a WARN_ON().
-- Dave
Quoted 24 lines · Dave Hansen
Quoted 12 lines · OGAWA Hirofumi
While I'm reading some code, I suspected that
mnt_want_write_file()may have wrong assumption. I thinkmnt_want_write_file()is assuming it increments ->mnt_writers if (file->f_mode&FMODE_WRITE). But, if it'sspecial_file(), it is false?Sorry, I'm still not checking all of those though. E.g. I'm thinking the below.
static inline int __get_file_write_access(struct inode *inode,
struct vfsmount *mnt)
{
[...]Quoted 10 lines
if (!special_file(inode->i_mode)) { /* * Balanced in __fput() */ error = mnt_want_write(mnt); if (error) put_write_access(inode); } return error; }
Quoted 8 lines
In practice I don't think this is an issue. We were never supposed to do
mnt_want_write(mnt)for any 'struct file' that was aspecial_file(), specifically because of what you mention.Nick's use of
mnt_want_write_file()was a 1:1 drop-in formnt_want_write(). So, if all is well in the world, there should not be any call sites wheremnt_want_write_file()gets called on aspecial_file().
void file_update_time(struct file *file)sys_fchmod()sys_fchown()sys_fsetxattr()sys_fremovexattr()
Um..., the users of mnt_want_write_file() seems to be those. I think all of those filp can be special file?
Thanks.
Signature · 1 line
OGAWA Hirofumi <[email protected]>
Quoted 35 lines · OGAWA Hirofumi
Quoted 24 lines · Dave Hansen
Quoted 12 lines · OGAWA Hirofumi
While I'm reading some code, I suspected that
mnt_want_write_file()may have wrong assumption. I thinkmnt_want_write_file()is assuming it increments ->mnt_writers if (file->f_mode&FMODE_WRITE). But, if it'sspecial_file(), it is false?Sorry, I'm still not checking all of those though. E.g. I'm thinking the below.
static inline int __get_file_write_access(struct inode *inode,
struct vfsmount *mnt)
{
[...]Quoted 10 lines
if (!special_file(inode->i_mode)) { /* * Balanced in __fput() */ error = mnt_want_write(mnt); if (error) put_write_access(inode); } return error; }Quoted 8 lines
In practice I don't think this is an issue. We were never supposed to do
mnt_want_write(mnt)for any 'struct file' that was aspecial_file(), specifically because of what you mention.Nick's use of
mnt_want_write_file()was a 1:1 drop-in formnt_want_write(). So, if all is well in the world, there should not be any call sites wheremnt_want_write_file()gets called on aspecial_file().
Quoted 8 lines
void
file_update_time(struct file *file)sys_fchmod()sys_fchown()sys_fsetxattr()sys_fremovexattr()Um..., the users of
mnt_want_write_file()seems to be those. I think all of those filp can be special file?
OK, I see where you're going now. I think the race goes like this:
Let's say we have a process with /dev/null opened with FMODE_WRITE. It is the only file open on the filesystem and so the /dev mount has a 0 mnt_writers count. That process goes to f_chmod() its fd to /dev/null. The code checks and notices that (file->f_mode & FMODE_WRITE), and goes to mnt_clone_write().
At the same time, another process tries to 'mount -o remount,ro /dev'. That process never sees mnt_clone_write()'s mnt_writers bump and allows the remount,ro, even though there's an elevated mnt_writers count.
Here's a completely untested/uncompiled patch. I'll see if I can find a test case that triggers this bug with the BUG_ON() in this patch.
fs/namespace.c +19 -5
@@ -294,9 +294,17 @@ EXPORT_SYMBOL_GPL(mnt_want_write);* * After finished, mnt_drop_write must be called as usual to * drop the reference. +* +* Be very careful using this. You must *guarantee* that +* this vfsmount has at least one existing, persistent writer +* that can not possibly go away, before calling this. */ int mnt_clone_write (struct vfsmount *mnt) { +/* This would kill the performance +* optimization in this function +BUG_ON(count_mnt_writers(mnt) > 0); +*/ /* superblock may be r/o */ if (__mnt_is_readonly (mnt))return -EROFS ;
@@ -312,14 +320,20 @@ EXPORT_SYMBOL_GPL(mnt_clone_write);* @file: the file who's mount on which to take a write * * This is like mnt_want_write, but it takes a file and can -* do some optimisations if the file is open for write already +* do some optimisations if the file is open for write already. +* We do not do mnt_want_write() on read-only or special files, +* so we can not use mnt_clone_write() for them. */ int mnt_want_write_file (struct file *file) { -if (!(file->f_mode &FMODE_WRITE )) -return mnt_want_write (file->f_path.mnt); -else -return mnt_clone_write (file->f_path.mnt); +struct path *path = &file->f_path; +struct inode *inode = path->dentry->d_inode; + +if ((file->f_mode &FMODE_WRITE ) && + !special_file (inode)) +return mnt_clone_write (path->mnt); + +return mnt_want_write (path->mnt); }EXPORT_SYMBOL_GPL (mnt_want_write_file);
-- Dave
Quoted 7 lines · OGAWA Hirofumi
diff -puN
fs/namespace.c~mnt_want_write-wrong-assumefs/namespace.c
---
linux-2.6/fs/namespace.c~mnt_want_write-wrong-assume 2009-08-03
04:33:35.000000000 +0900
+++linux-2.6-hirofumi/fs/namespace.c2009-08-03 04:31:34.000000000
+0900
Review EXPORT_SYMBOL_GPL(mnt_clone_write); · line 323
Quoted 9 lines
(quoted diff) +2
-1@@ -316,7 +316,8 @@ EXPORT_SYMBOL_GPL(mnt_clone_write); */ int mnt_want_write_file(struct file *file) { - if (!(file->f_mode & FMODE_WRITE)) + struct inode *inode = file->f_dentry->d_inode; + if (!(file->f_mode & FMODE_WRITE) || special_file(inode->i_mode)) return mnt_want_write(file->f_path.mnt); else return mnt_clone_write(file->f_path.mnt);
I'm fine with this. I'd like a debugging check in mnt_clone_write() since this bug is easy to detect, but such a check will also cost all of the performance gains that Nick added. So, we can't have it unconditionally.
Signature · 3 lines
Signed-off-by: Dave Hansen <[email protected]> -- Dave
Quoted 14 lines · Dave Hansen
Quoted 13 lines · OGAWA Hirofumi
diff -puN
fs/namespace.c~mnt_want_write-wrong-assumefs/namespace.c
---
linux-2.6/fs/namespace.c~mnt_want_write-wrong-assume 2009-08-03
04:33:35.000000000 +0900
+++linux-2.6-hirofumi/fs/namespace.c2009-08-03 04:31:34.000000000
+0900(quoted diff) +2
-1@@ -316,7 +316,8 @@ EXPORT_SYMBOL_GPL(mnt_clone_write); */ int mnt_want_write_file(struct file *file) { - if (!(file->f_mode & FMODE_WRITE)) + struct inode *inode = file->f_dentry->d_inode; + if (!(file->f_mode & FMODE_WRITE) || special_file(inode->i_mode))
Quoted 8 lines
Review EXPORT_SYMBOL_GPL(mnt_clone_write); · line 323
Quoted 3 lines
(quoted diff) +0
-0return mnt_want_write(file->f_path.mnt); else return mnt_clone_write(file->f_path.mnt);I'm fine with this. I'd like a debugging check in
mnt_clone_write()since this bug is easy to detect, but such a check will also cost all of the performance gains that Nick added. So, we can't have it unconditionally.
Yeah, good catch, thanks for this.
Al Viro
maintainerQuoted 14 lines · Dave Hansen
Quoted 13 lines · OGAWA Hirofumi
diff -puN
fs/namespace.c~mnt_want_write-wrong-assumefs/namespace.c
---
linux-2.6/fs/namespace.c~mnt_want_write-wrong-assume 2009-08-03
04:33:35.000000000 +0900
+++linux-2.6-hirofumi/fs/namespace.c2009-08-03 04:31:34.000000000
+0900(quoted diff) +2
-1@@ -316,7 +316,8 @@ EXPORT_SYMBOL_GPL(mnt_clone_write); */ int mnt_want_write_file(struct file *file) { - if (!(file->f_mode & FMODE_WRITE)) + struct inode *inode = file->f_dentry->d_inode; + if (!(file->f_mode & FMODE_WRITE) || special_file(inode->i_mode))
Quoted 8 lines
Review EXPORT_SYMBOL_GPL(mnt_clone_write); · line 323
Quoted 3 lines
(quoted diff) +0
-0return mnt_want_write(file->f_path.mnt); else return mnt_clone_write(file->f_path.mnt);I'm fine with this. I'd like a debugging check in
mnt_clone_write()since this bug is easy to detect, but such a check will also cost all of the performance gains that Nick added. So, we can't have it unconditionally.
[Very belated] ACK.