Missing i_sb NULL pointer check in destroy_inode()
A message that is not in the archive (sent off-list or not yet received): 1068045518.10730.266.camel@socrates
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): 1068066524.10726.289.camel@socrates
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): 1068145132.10735.322.camel@socrates
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): 1068148881.10730.337.camel@socrates
A message that is not in the archive (sent off-list or not yet received): 1068230146.10726.359.camel@socrates
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): 1068419747.687.28.camel@socrates
A message that is not in the archive (sent off-list or not yet received): [email protected]
Hello, Andrew, Marcelo,
destroy_inode() dereferences inode->i_sb without checking if it is NULL. This is inconsistent with its caller: iput() and clear_inode(), both of which check inode->i_sb before dereferencing it. Since iput() calls destroy_inode() after calling file system's .clear_inode method(via clear_inode()), some file systems might choose to clear the i_sb in the .clear_inode super block operation. This results in a crash in destroy_inode().
This issue exists in both 2.6, 2.4 and 2.4 kernel. A simple fix against 2.6.0-test9 is included below. 2.4 based fix should be very similar to this one. Please take a look and consider include it.
Many thanks!!
--Mingming
----------------------------------------------------------
fs/inode.c +1 -1
@@ -160,7 +160,7 @@ void destroy_inode(struct inode *inode)if (inode_has_buffers (inode))BUG ();security_inode_free (inode); -if (inode->i_sb->s_op->destroy_inode) +if (inode->i_sb && inode->i_sb->s_op->destroy_inode) inode->i_sb->s_op->destroy_inode (inode);else kmem_cache_free (inode_cachep, (inode));
Quoted 4 lines · Mingming Cao
destroy_inode()dereferencesinode->i_sbwithout checking if it is NULL. This is inconsistent with its caller:iput()andclear_inode(), both of which checkinode->i_sbbefore dereferencing it.
I assume this has only been observed with an out-of-tree filesystem, but yes, the consistency is good.
Quoted 9 lines · Andrew Morton
Quoted 4 lines · Mingming Cao
destroy_inode()dereferencesinode->i_sbwithout checking if it is NULL. This is inconsistent with its caller:iput()andclear_inode(), both of which checkinode->i_sbbefore dereferencing it.I assume this has only been observed with an out-of-tree filesystem, but yes, the consistency is good.
Yes, the crash happened with an out-of-tree filesystem. Thanks.
-Mingming
Quoted 10 lines · Mingming Cao
Hello, Andrew, Marcelo,
destroy_inode()dereferencesinode->i_sbwithout checking if it is NULL. This is inconsistent with its caller:iput()andclear_inode(), both of which checkinode->i_sbbefore dereferencing it. Sinceiput()callsdestroy_inode()after calling file system's .clear_inodemethod(via clear_inode()), some file systems might choose to clear thei_sbin the .clear_inode super block operation. This results in a crash indestroy_inode().
Quoted 3 lines
This issue exists in both 2.6, 2.4 and 2.4 kernel. A simple fix against 2.6.0-test9 is included below. 2.4 based fix should be very similar to this one. Please take a look and consider include it.
inode->i_sb can't be NULL. We should remove all those checks.
Quoted 11 lines · Christoph Hellwig
Quoted 10 lines · Mingming Cao
Hello, Andrew, Marcelo,
destroy_inode()dereferencesinode->i_sbwithout checking if it is NULL. This is inconsistent with its caller:iput()andclear_inode(), both of which checkinode->i_sbbefore dereferencing it. Sinceiput()callsdestroy_inode()after calling file system's .clear_inodemethod(via clear_inode()), some file systems might choose to clear thei_sbin the .clear_inode super block operation. This results in a crash indestroy_inode().
Quoted 6 lines
Quoted 3 lines
This issue exists in both 2.6, 2.4 and 2.4 kernel. A simple fix against 2.6.0-test9 is included below. 2.4 based fix should be very similar to this one. Please take a look and consider include it.
inode->i_sbcan't be NULL. We should remove all those checks.
Sorry I can not agree with this. Maybe the inode->i_sb should not be NULL, but the kernel still allows the file system to do so. In fact JFS's diReadSpecial() function clears the inode->i_sb to NULL before calling iput().
Acutally iput() in 2.6 is missing the check too.(in 2.4 the check is there). Here is the the incremental fix for 2.6 only:
fs/inode.c +4 -4
@@ -1084,13 +1084,13 @@ static inline void iput_final(struct inovoid iput (struct inode *inode) {if (inode) { -struct super_operations *op = inode->i_sb->s_op; - +struct super_block *sb = inode->i_sb; +if (inode->i_state ==I_CLEAR )BUG (); -if (op && op->put_inode) - op->put_inode (inode); +if (sb && sb->op && sb->op->put_inode) + sb->op->put_inode (inode);if (atomic_dec_and_lock (&inode->i_count, &inode_lock))iput_final (inode);
Quoted 18 lines · Mingming Cao
Quoted 11 lines · Christoph Hellwig
Quoted 10 lines · Mingming Cao
Hello, Andrew, Marcelo,
destroy_inode()dereferencesinode->i_sbwithout checking if it is NULL. This is inconsistent with its caller:iput()andclear_inode(), both of which checkinode->i_sbbefore dereferencing it. Sinceiput()callsdestroy_inode()after calling file system's .clear_inodemethod(via clear_inode()), some file systems might choose to clear thei_sbin the .clear_inode super block operation. This results in a crash indestroy_inode().Quoted 6 lines
Quoted 3 lines
This issue exists in both 2.6, 2.4 and 2.4 kernel. A simple fix against 2.6.0-test9 is included below. 2.4 based fix should be very similar to this one. Please take a look and consider include it.
inode->i_sbcan't be NULL. We should remove all those checks.
Quoted 7 lines
Sorry I can not agree with this. Maybe the
inode->i_sbshould not be NULL, but the kernel still allows the file system to do so. In fact JFS'sdiReadSpecial()function clears theinode->i_sbto NULL before callingiput().Acutally
iput()in 2.6 is missing the check too.(in 2.4 the check is there). Here is the the incremental fix for 2.6 only:
There is a little typo in your patch. The struct member is s_op and not op. The following patch should be right.
Timo
fs/inode.c +4 -4
@@ -1084,13 +1084,13 @@ static inline void iput_final(struct inovoid iput (struct inode *inode) {if (inode) { -struct super_operations *op = inode->i_sb->s_op; - +struct super_block *sb = inode->i_sb; +if (inode->i_state ==I_CLEAR )BUG (); -if (op && op->put_inode) - op->put_inode (inode); +if (sb && sb->s_op && sb->s_op->put_inode) + sb->s_op->put_inode (inode);if (atomic_dec_and_lock (&inode->i_count, &inode_lock))iput_final (inode);