vfs: allow file truncations when both suid and write permissions set
Subsystems (3)
[email protected]
maintainerPATCH 12/12 vfs: allow file truncations when both suid and write permissions set
From: Amerigo Wang <[email protected]>
When suid is set and the non-owner user has write permission, any writing into this file should be allowed and suid should be removed after that.
However, the current kernel only allows writing without truncations, when we do truncations on that file, we get EPERM. This is a bug.
Steps to reproduce this bug:
% ls -l rootdir/file1
-rwsrwsrwx 1 root root 3 Jun 25 15:42 rootdir/file1 % echo h > rootdir/file1
zsh: operation not permitted: rootdir/file1 % ls -l rootdir/file1
-rwsrwsrwx 1 root root 3 Jun 25 15:42 rootdir/file1 % echo h >> rootdir/file1
% ls -l rootdir/file1
-rwxrwxrwx 1 root root 5 Jun 25 16:34 rootdir/file1
This patch fixes it.
---
fs/open.c | 8 ++++++-- include/linux/fs.h | 1 + mm/filemap.c | 12 ++++++++++-- 3 files changed, 17 insertions(+), 4 deletions(-)
fs/open.c renamed from fs/open.c~vfs-allow-file-truncations-when-both-suid-and-write-permissions-set +6 -2
@@ -213,11 +213,15 @@ int do_truncate(struct dentry *dentry, l newattrs.ia_valid |=ATTR_FILE ; } +mutex_lock (&dentry->d_inode->i_mutex);/* Remove suid/sgid on truncate too */ - newattrs.ia_valid |=should_remove_suid (dentry); + err =dentry_remove_suid (dentry); +if (err) + goto unlock; -mutex_lock (&dentry->d_inode->i_mutex); err =notify_change (dentry, &newattrs); + + unlock:mutex_unlock (&dentry->d_inode->i_mutex);return err; }
include/linux/fs.h renamed from include/linux/fs.h~vfs-allow-file-truncations-when-both-suid-and-write-permissions-set +1 -0
@@ -2170,6 +2170,7 @@ extern void destroy_inode(struct inode *extern struct inode *new_inode (struct super_block *);extern int should_remove_suid (struct dentry *);extern int file_remove_suid (struct file *); +extern int dentry_remove_suid (struct dentry *);extern void __insert_inode_hash (struct inode *,unsigned long hashval);extern void remove_inode_hash (struct inode *);
mm/filemap.c renamed from mm/filemap.c~vfs-allow-file-truncations-when-both-suid-and-write-permissions-set +10 -2
@@ -1839,9 +1839,11 @@ static int __remove_suid(struct dentry *return notify_change (dentry, &newattrs); } -int file_remove_suid (struct file *file) +/* +* Note: you need to hold i_mutex before call this. +*/ +int dentry_remove_suid (struct dentry *dentry) { -struct dentry *dentry = file->f_path.dentry;int killsuid =should_remove_suid (dentry);int killpriv =security_inode_need_killpriv (dentry);int error =0 ;
@@ -1855,6 +1857,12 @@ int file_remove_suid(struct file *file)return error; } + +int file_remove_suid (struct file *file) +{ +struct dentry *dentry = file->f_path.dentry; +return dentry_remove_suid (dentry); +}EXPORT_SYMBOL (file_remove_suid); static size_t __iovec_copy_from_user_inatomic(char *vaddr,
_
fs/open.c int do_truncate(struct dentry *dentry, l · line 220
Quoted 13 lines · [email protected] writes:
diff -puN
fs/open.c~vfs-allow-file-truncations-when-both-suid-and-write-permissions-setfs/open.cfs/open.c renamed from fs/open.c~vfs-allow-file-truncations-when-both-suid-and-write-permissions-set +4
-1@@ -213,11 +213,15 @@ int do_truncate(struct dentry *dentry, l newattrs.ia_valid |=ATTR_FILE ; } +mutex_lock (&dentry->d_inode->i_mutex);/* Remove suid/sgid on truncate too */ - newattrs.ia_valid |=should_remove_suid (dentry); + err =dentry_remove_suid (dentry); +if (err) + goto unlock;
Can't we use ATTR_FORCE for this? Because this calls notify_change() twice, and I guess this removes s[ug]id even if vmtruncate() (or in future ->truncate() may return error) or something returned error.
I think it would not be good behavior.
Thanks.
Quoted 7 lines
fs/open.c renamed from fs/open.c~vfs-allow-file-truncations-when-both-suid-and-write-permissions-set +2
-1-mutex_lock (&dentry->d_inode->i_mutex); err =notify_change (dentry, &newattrs); + + unlock:mutex_unlock (&dentry->d_inode->i_mutex);return err; }
Signature · 1 line
OGAWA Hirofumi <[email protected]>
Quoted 22 lines · OGAWA Hirofumi
Quoted 4 lines · [email protected] writes:
diff -puN
fs/open.c~vfs-allow-file-truncations-when-both-suid-and-write-permissions-setfs/open.cfs/open.c int do_truncate(struct dentry *dentry, l · line 221
Quoted 10 lines
fs/open.c renamed from fs/open.c~vfs-allow-file-truncations-when-both-suid-and-write-permissions-set +4
-1@@ -213,11 +213,15 @@ int do_truncate(struct dentry *dentry, l newattrs.ia_valid |=ATTR_FILE ; } +mutex_lock (&dentry->d_inode->i_mutex);/* Remove suid/sgid on truncate too */ - newattrs.ia_valid |=should_remove_suid (dentry); + err =dentry_remove_suid (dentry); +if (err) + goto unlock;Can't we use
ATTR_FORCEfor this? Because this callsnotify_change()twice, and I guess this removes s[ug]id even ifvmtruncate()(or in future ->truncate() may return error) or something returned error.
Quoted 2 lines
I think it would not be good behavior.
Hi, please check:
http://lkml.org/lkml/2009/7/1/459
Quoted 23 lines · Amerigo Wang
Quoted 22 lines · OGAWA Hirofumi
Quoted 4 lines · [email protected] writes:
diff -puN
fs/open.c~vfs-allow-file-truncations-when-both-suid-and-write-permissions-setfs/open.cfs/open.c int do_truncate(struct dentry *dentry, l · line 221
Quoted 10 lines
fs/open.c renamed from fs/open.c~vfs-allow-file-truncations-when-both-suid-and-write-permissions-set +4
-1@@ -213,11 +213,15 @@ int do_truncate(struct dentry *dentry, l newattrs.ia_valid |=ATTR_FILE ; } +mutex_lock (&dentry->d_inode->i_mutex);/* Remove suid/sgid on truncate too */ - newattrs.ia_valid |=should_remove_suid (dentry); + err =dentry_remove_suid (dentry); +if (err) + goto unlock;Can't we use
ATTR_FORCEfor this? Because this callsnotify_change()twice, and I guess this removes s[ug]id even ifvmtruncate()(or in future ->truncate() may return error) or something returned error.
Quoted 5 lines
Quoted 2 lines
I think it would not be good behavior.
Hi, please check:
http://lkml.org/lkml/2009/7/1/459
Sorry for same argument. I see. However, um...
I found this piece in security/selinux/hooks.c
static int selinux_inode_setattr(struct dentry *dentry, struct iattr *iattr)
{
const struct cred *cred = current_cred();
if (iattr->ia_valid & ATTR_FORCE)
return 0;
if (iattr->ia_valid & (ATTR_MODE | ATTR_UID | ATTR_GID |
ATTR_ATIME_SET | ATTR_MTIME_SET))
return dentry_has_perm(cred, NULL, dentry, FILE__SETATTR);
return dentry_has_perm(cred, NULL, dentry, FILE__WRITE);
}I guess it's assuming the ia_valid doesn't have (ATTR_MODE | ATTR_SIZE), but truncate() already does it, I don't know whether it's ok. The definition of ATTR_FORCE is unclear at all, it would be problem. But, I'm not sure though, I suspect the above code also has problem...
Thanks.
Signature · 1 line
OGAWA Hirofumi <[email protected]>
Quoted 18 lines · OGAWA Hirofumi
Sorry for same argument. I see. However, um...
I found this piece in
security/selinux/hooks.cstatic int
selinux_inode_setattr(struct dentry *dentry, struct iattr *iattr){ const struct cred *cred = current_cred(); if (iattr->ia_valid & ATTR_FORCE) return 0; if (iattr->ia_valid & (ATTR_MODE | ATTR_UID | ATTR_GID | ATTR_ATIME_SET | ATTR_MTIME_SET)) return dentry_has_perm(cred, NULL, dentry, FILE__SETATTR); return dentry_has_perm(cred, NULL, dentry, FILE__WRITE); }
Quoted 4 lines
I guess it's assuming the
ia_validdoesn't have (ATTR_MODE|ATTR_SIZE), buttruncate()already does it, I don't know whether it's ok. The definition ofATTR_FORCEis unclear at all, it would be problem. But, I'm not sure though, I suspect the above code also has problem...
I found the SuSv3 says for ftruncate():
Upon successful completion, if fildes refers to a regular file, ftruncate() shall mark for update the last data modification and last file status change timestamps of the file and the S_ISUID and S_ISGID bits of the file mode may be cleared. If the ftruncate() function is unsuccessful, the file is unaffected.
And vmtruncate() can return error easily with RLIMIT_FSIZE or ->s_maxbytes. So, I think clearing s[ug]id first may be bad behavior without good reason.
Thanks.
Signature · 1 line
OGAWA Hirofumi <[email protected]>
Quoted 8 lines · OGAWA Hirofumi
I found the SuSv3 says for
ftruncate():Upon successful completion, if fildes refers to a regular file,
ftruncate()shall mark for update the last data modification and last file status change timestamps of the file and theS_ISUIDandS_ISGIDbits of the file mode may be cleared. If theftruncate()function is unsuccessful, the file is unaffected.
Quoted 4 lines
And
vmtruncate()can return error easily withRLIMIT_FSIZEor ->s_maxbytes. So, I think clearing s[ug]id first may be bad behavior without good reason.
Hmm, you mean we should clear suid after we do notify_change()? Good point!
I will update the patch now...
Thanks for your review!
Quoted 9 lines · Amerigo Wang
Quoted 8 lines · OGAWA Hirofumi
I found the SuSv3 says for
ftruncate():Upon successful completion, if fildes refers to a regular file,
ftruncate()shall mark for update the last data modification and last file status change timestamps of the file and theS_ISUIDandS_ISGIDbits of the file mode may be cleared. If theftruncate()function is unsuccessful, the file is unaffected.
Quoted 7 lines
Quoted 4 lines
And
vmtruncate()can return error easily withRLIMIT_FSIZEor ->s_maxbytes. So, I think clearing s[ug]id first may be bad behavior without good reason.Hmm, you mean we should clear suid after we do
notify_change()? Good point!
I guess it would be better than current patch in practice, although it's not perfect.
Thanks.
Signature · 1 line
OGAWA Hirofumi <[email protected]>
Quoted 53 lines · OGAWA Hirofumi
Quoted 25 lines · Amerigo Wang
Quoted 23 lines · OGAWA Hirofumi
Quoted 5 lines · [email protected] writes:
diff -puN
fs/open.c~vfs-allow-file-truncations-when-both-suid-and-write-permissions-setfs/open.cfs/open.c renamed from fs/open.c~vfs-allow-file-truncations-when-both-suid-and-write-permissions-set +0
-0@@ -213,11 +213,15 @@ int do_truncate(struct dentry *dentry, l newattrs.ia_valid |=ATTR_FILE ;fs/open.c int do_truncate(struct dentry *dentry, l · line 222
Quoted 10 lines
fs/open.c renamed from fs/open.c~vfs-allow-file-truncations-when-both-suid-and-write-permissions-set +4
-1} +mutex_lock (&dentry->d_inode->i_mutex);/* Remove suid/sgid on truncate too */ - newattrs.ia_valid |=should_remove_suid (dentry); + err =dentry_remove_suid (dentry); +if (err) + goto unlock;Can't we use
ATTR_FORCEfor this? Because this callsnotify_change()twice, and I guess this removes s[ug]id even ifvmtruncate()(or in future ->truncate() may return error) or something returned error.Quoted 6 lines
Quoted 3 lines
I think it would not be good behavior.
Hi, please check:
http://lkml.org/lkml/2009/7/1/459Sorry for same argument. I see. However, um...
I found this piece in
security/selinux/hooks.cstatic int
selinux_inode_setattr(struct dentry *dentry, struct iattr *iattr){ const struct cred *cred = current_cred(); if (iattr->ia_valid & ATTR_FORCE) return 0; if (iattr->ia_valid & (ATTR_MODE | ATTR_UID | ATTR_GID | ATTR_ATIME_SET | ATTR_MTIME_SET)) return dentry_has_perm(cred, NULL, dentry, FILE__SETATTR); return dentry_has_perm(cred, NULL, dentry, FILE__WRITE); }
Quoted 2 lines
I guess it's assuming the
ia_validdoesn't have (ATTR_MODE|ATTR_SIZE), buttruncate()already does it, I don't know whether it's ok.
No, here we should only force ATTR_KILL_SUID and/or ATTR_KILL_SGID. do_truncate() has ATTR_SIZE and ATTR_FILE.
Quoted 11 lines · Amerigo Wang
Quoted 11 lines
static int
selinux_inode_setattr(struct dentry *dentry, struct iattr *iattr){ const struct cred *cred = current_cred(); if (iattr->ia_valid & ATTR_FORCE) return 0; if (iattr->ia_valid & (ATTR_MODE | ATTR_UID | ATTR_GID | ATTR_ATIME_SET | ATTR_MTIME_SET)) return dentry_has_perm(cred, NULL, dentry, FILE__SETATTR);
Quoted 8 lines
Quoted 5 lines
return dentry_has_perm(cred, NULL, dentry, FILE__WRITE); }I guess it's assuming the
ia_validdoesn't have (ATTR_MODE|ATTR_SIZE), buttruncate()already does it, I don't know whether it's ok.No, here we should only force
ATTR_KILL_SUIDand/orATTR_KILL_SGID.do_truncate()hasATTR_SIZEandATTR_FILE.
I guess security module should do,
ia_valid = iattr->ia_valid;
if (!(ia_valid & ATTR_FORCE) && (ia_valid & ATTR_FORCE_MASK)) {
err = dentry_has_perm(cred, NULL, dentry, FILE__SETATTR);
if (err)
return err;
ia_valid &= ~ATTR_FORCE_MASK;
}
if (ia_valid & ATTR_NOT_FORCE_MASK)
err = dentry_has_perm(cred, NULL, dentry, FILE__WRITE);
return err;or something. Because do_truncate() already do (ATTR_MODE | ATTR_SIZE) without ATTR_FORCE.
Thanks.
Signature · 1 line
OGAWA Hirofumi <[email protected]>
Quoted 35 lines · OGAWA Hirofumi
Quoted 11 lines · Amerigo Wang
Quoted 11 lines
static int
selinux_inode_setattr(struct dentry *dentry, struct iattr *iattr){ const struct cred *cred = current_cred(); if (iattr->ia_valid & ATTR_FORCE) return 0; if (iattr->ia_valid & (ATTR_MODE | ATTR_UID | ATTR_GID | ATTR_ATIME_SET | ATTR_MTIME_SET)) return dentry_has_perm(cred, NULL, dentry, FILE__SETATTR);Quoted 8 lines
Quoted 5 lines
return dentry_has_perm(cred, NULL, dentry, FILE__WRITE); }I guess it's assuming the
ia_validdoesn't have (ATTR_MODE|ATTR_SIZE), buttruncate()already does it, I don't know whether it's ok.No, here we should only force
ATTR_KILL_SUIDand/orATTR_KILL_SGID.do_truncate()hasATTR_SIZEandATTR_FILE.I guess security module should do,
ia_valid = iattr->ia_valid; if (!(ia_valid & ATTR_FORCE) && (ia_valid & ATTR_FORCE_MASK)) { err = dentry_has_perm(cred, NULL, dentry, FILE__SETATTR); if (err) return err; ia_valid &= ~ATTR_FORCE_MASK; } if (ia_valid & ATTR_NOT_FORCE_MASK) err = dentry_has_perm(cred, NULL, dentry, FILE__WRITE); return err;
Quoted 2 lines
or something. Because
do_truncate()already do (ATTR_MODE|ATTR_SIZE) withoutATTR_FORCE.
BTW, it seems original code doesn't check ATTR_SIZE if (ATTR_MODE | ATTR_SIZE), right? So, ATTR_FORCE is just forcing ATTR_MODE, but I guess that's problem itself.
Thanks.
Signature · 1 line
OGAWA Hirofumi <[email protected]>
Quoted 38 lines · OGAWA Hirofumi
Quoted 35 lines · OGAWA Hirofumi
Quoted 11 lines · Amerigo Wang
Quoted 11 lines
static int
selinux_inode_setattr(struct dentry *dentry, struct iattr *iattr){ const struct cred *cred = current_cred(); if (iattr->ia_valid & ATTR_FORCE) return 0; if (iattr->ia_valid & (ATTR_MODE | ATTR_UID | ATTR_GID | ATTR_ATIME_SET | ATTR_MTIME_SET)) return dentry_has_perm(cred, NULL, dentry, FILE__SETATTR);Quoted 8 lines
Quoted 6 lines
return dentry_has_perm(cred, NULL, dentry, FILE__WRITE); }I guess it's assuming the
ia_validdoesn't have (ATTR_MODE|ATTR_SIZE), buttruncate()already does it, I don't know whether it's ok.No, here we should only force
ATTR_KILL_SUIDand/orATTR_KILL_SGID.do_truncate()hasATTR_SIZEandATTR_FILE.I guess security module should do,
ia_valid = iattr->ia_valid; if (!(ia_valid & ATTR_FORCE) && (ia_valid & ATTR_FORCE_MASK)) { err = dentry_has_perm(cred, NULL, dentry, FILE__SETATTR); if (err) return err; ia_valid &= ~ATTR_FORCE_MASK; } if (ia_valid & ATTR_NOT_FORCE_MASK) err = dentry_has_perm(cred, NULL, dentry, FILE__WRITE); return err;
Quoted 8 lines
Quoted 3 lines
or something. Because
do_truncate()already do (ATTR_MODE|ATTR_SIZE) withoutATTR_FORCE.BTW, it seems original code doesn't check
ATTR_SIZEif (ATTR_MODE|ATTR_SIZE), right? So,ATTR_FORCEis just forcingATTR_MODE, but I guess that's problem itself.
I am not sure if I understand you correctly... You must be referring notify_change(), it seems to do what you said.
But clearly ATTR_FORCE is the way to bypass the security module on purpose. I agree that we perhaps should have some wrapper function to do this (instead of calling notify_change() twice), but currently this is fine.
Quoted 44 lines · Amerigo Wang
Quoted 43 lines · OGAWA Hirofumi
Quoted 39 lines · OGAWA Hirofumi
Quoted 11 lines · Amerigo Wang
Quoted 11 lines
static int
selinux_inode_setattr(struct dentry *dentry, struct iattr *iattr){ const struct cred *cred = current_cred(); if (iattr->ia_valid & ATTR_FORCE) return 0; if (iattr->ia_valid & (ATTR_MODE | ATTR_UID | ATTR_GID | ATTR_ATIME_SET | ATTR_MTIME_SET)) return dentry_has_perm(cred, NULL, dentry, FILE__SETATTR);Quoted 9 lines
Quoted 6 lines
return dentry_has_perm(cred, NULL, dentry, FILE__WRITE); }I guess it's assuming the
ia_validdoesn't have (ATTR_MODE|ATTR_SIZE), buttruncate()already does it, I don't know whether it's ok.No, here we should only force
ATTR_KILL_SUIDand/orATTR_KILL_SGID.do_truncate()hasATTR_SIZEandATTR_FILE.I guess security module should do,
ia_valid = iattr->ia_valid; if (!(ia_valid & ATTR_FORCE) && (ia_valid & ATTR_FORCE_MASK)) { err = dentry_has_perm(cred, NULL, dentry, FILE__SETATTR); if (err) return err; ia_valid &= ~ATTR_FORCE_MASK; } if (ia_valid & ATTR_NOT_FORCE_MASK) err = dentry_has_perm(cred, NULL, dentry, FILE__WRITE); return err;or something. Because
do_truncate()already do (ATTR_MODE|ATTR_SIZE) withoutATTR_FORCE.
Quoted 7 lines
Quoted 4 lines
BTW, it seems original code doesn't check
ATTR_SIZEif (ATTR_MODE|ATTR_SIZE), right? So,ATTR_FORCEis just forcingATTR_MODE, but I guess that's problem itself.I am not sure if I understand you correctly... You must be referring
notify_change(), it seems to do what you said.
As far as I can see,
do_truncate pass ATTR_SIZE
security_inode_setattr() dentry_has_perm(cred, NULL, dentry, FILE__WRITE);
do_truncate pass (ATTR_SIZE | ATTR_MODE) [ATTR_MODE is set by ATTR_KILL_*.
This should works fine for file owner without patch.]
security_inode_setattr()
dentry_has_perm(cred, NULL, dentry, FILE__SETATTR);
Doesn't security module really have to call dentry_has_perm() with FILE__WRITE if (ATTR_SIZE | ATTR_MODE)?
Quoted 2 lines
But clearly
ATTR_FORCEis the way to bypass the security module on purpose.
Even if it's true, why don't we change security modules? I'm saying always bypass was wrong way, if it doesn't want. AFAIK, it's not binary driver, and it has good source codes.
Quoted 3 lines
I agree that we perhaps should have some wrapper function to do this (instead of calling
notify_change()twice), but currently this is fine.
I don't object to that patch though, but it's fine? Really? The error handling is wrong, and note to call fs-driver twice means it can be cause of I/O (network, storage) twice, fsnotify() is also called twice, and fs-driver can't fix those.
I don't think those are fine at all.
Signature · 1 line
OGAWA Hirofumi <[email protected]>