kernel BUG at fs/locks.c:1723!
Subsystems (2)
There is a race between unshare_files() and the following steal_locks(). As a consequence, steal_locks() may steal some additional FL_POSIX locks that don't belong to the current thread. This triggers a BUG in locks_remove_flock().
In detail, the current thread shares its files struct with other threads. This causes unshare_files() to associate the current thread with a copy of its files_struct. The copy shares all file objects with the original files struct. In the time between unshare_files() and steal_locks(), another thread creates a new file and a FL_POSIX lock on it. The current thread gets into steal_locks() and takes over all FL_POSIX locks that refer to the previous files_struct, including the new lock. We do put_files_struct(original files_struct). This causes the file handle to the new file to be closed. We get into locks_remove_posix() and miss the lock, because its fl_owner field now refers to the new files_struct. Finally we get into locks_remove_flock(), and stumble upon the lock.
While looking into this bug report I gathered the following data with a SUSE kernel (oops and LKCD dump from Chris):
kernel BUG at fs/locks.c:1736!
invalid operand: 0000 [#1]
SMP
CPU: 0 EIP: 0060:[<c01844fb>] Tainted: G U EFLAGS: 00010246 (2.6.5-0-testing)
EIP is at locks_remove_flock+0x8b/0x130
eax: f7b89998 ebx: f61df3fc ecx: f61df354 edx: 00000000
esi: f61df354 edi: f6702b80 ebp: f6179c24 esp: f6179c08
ds: 007b es: 007b ss: 0068
Process owcimomd (pid: 1713, threadinfo=f6178000 task=f66d0d60)
Stack: c1e1fdac c1e1fdac f7fe83c0 00000296 f6702b80 f7fe87c0 f61df354 f6179c3c
c016ce00 f61ddadc f6702b80 00000000 f6703b00 f6179c54 c0168b1f c0000000
0000026f 00000012 f6703b00 f6179c6c c0124ba7 00000001 f6179e5c f6179d88
Call Trace:
[<c016ce00>] __fput+0x30/0x120
[<c0168b1f>] filp_close+0x4f/0x90
[<c0124ba7>] put_files_struct+0x67/0xc0
[<c019d285>] load_elf_binary+0x3f5/0x1596
[<c018a5af>] update_atime+0x9f/0xc0
[<c01478fd>] __generic_file_aio_read+0x1cd/0x200
[<c0145060>] file_read_actor+0x0/0xd0
[<c01784b7>] search_binary_handler+0x97/0x270
[<c017a072>] do_execve+0x172/0x200
[<c0105fb2>] sys_execve+0x32/0x70
[<c0107e21>] sysenter_past_esp+0x52/0x71
Code: 0f 0b c8 06 eb 74 35 c0 eb db b8 00 e0 ff ff 21 e0 8b 10 8bput_files_struct+0x67 is equivalent to fs/binfmt_elf.c:681 in 2.6.6
current->files == fl->fl_ownerfl->fl_file = 0xf6702b80 (a valid struct file)
current->files =
max_fds=32
max_fdset=1024
next_fd=3
Here's a proposed fix. As a side effect, steal_locks no longer walks the global list of locks, but only the locks of all open inodes.
What are the reasons (other than historic ones) for not getting rid of fl_owner and using fl_pid instead, by the way? I think that would clean up the whole mess with file locks a bit.
fs/locks.c +40 -7
@@ -1731,6 +1731,9 @@ void locks_remove_flock(struct file *fillease_modify (before,F_UNLCK );continue ; } +/* FL_POSIX locks of this process have already been +* removed in filp_close->locks_remove_posix. +*/ BUG (); } before = &fl->fl_next;
@@ -1990,22 +1993,52 @@ int lock_may_write(struct inode *inode,EXPORT_SYMBOL (lock_may_write); +static inline void __steal_locks (struct file *file,fl_owner_t from) +{ +struct inode *inode = file->f_dentry->d_inode; +struct file_lock *fl = inode->i_flock; + +while (fl) { +if (fl->fl_file == file && fl->fl_owner == from) + fl->fl_owner = current->files; + fl = fl->fl_next; + } +} + +/* When getting ready for executing a binary, we make sure that current +* has a files_struct on its own. Before dropping the old files_struct, +* we take over ownership of all locks for all file descriptors we own. +* Note that we may accidentally steal a lock for a file that a sibling +* has created since the unshare_files() call. +*/ void steal_locks (fl_owner_t from) { -struct list_head *tmp; +struct files_struct *files = current->files; +int i, j; -if (from == current->files) +if (from == files)return ;lock_kernel (); -list_for_each (tmp, &file_lock_list) { - struct file_lock *fl = list_entry(tmp, struct file_lock, fl_link); -if (fl->fl_owner == from) - fl->fl_owner = current->files; + j =0 ; +for (;;) { +unsigned long set; + i = j * __NFDBITS; +if (i >= files->max_fdset || i >= files->max_fds) +break ; + set = files->open_fds->fds_bits[j++]; +while (set) { +if (set &1 ) { +struct file *file = files->fd[i]; +if (file) +__steal_locks (file, from); + } + i++; + set >>=1 ; + } }unlock_kernel (); } -EXPORT_SYMBOL (steal_locks);static int __init filelock_init (void )
Cheers,
Signature · 2 lines
Andreas Gruenbacher <[email protected]> SUSE Labs, SUSE LINUX AG
På su , 23/05/2004 klokka 17:50, skreiv Andreas Gruenbacher:
Quoted 6 lines
Here's a proposed fix. As a side effect,
steal_locksno longer walks the global list of locks, but only the locks of all open inodes.What are the reasons (other than historic ones) for not getting rid of
fl_ownerand usingfl_pidinstead, by the way? I think that would clean up the whole mess with file locks a bit.
If I understand correctly, the fl_owner was introduced in order to deal with the problem of lockd which has no control over which pids it has to use. You should probably check with Olaf though...
In the end, this made for a horrible "solution", and causes no end of bugs. Look for instance at the code in locks_remove_posix() which breaks POSIX 'cos it gets the whole idea wrong and thinks that it suffices to test the fl_owner instead of doing fl_pid (at least posix_same_owner() & friends get that right).
IMO it would be better to set fl_owner to NULL for ordinary processes (instead of dealing with this mess inside current->files), and then let lockd set current->files in whatever way it needs to in order to distinguish the various "pid spaces" it deals with...
Cheers,
Trond
Quoted 6 lines · Andreas Gruenbacher
There is a race between
unshare_files()and the followingsteal_locks(). As a consequence,steal_locks()may steal some additionalFL_POSIXlocks that don't belong to the current thread. This triggers a BUG inlocks_remove_flock().
Well based on Trond's words it's not clear whether we should be proceeding this way. Although I think he said "the code is crap and someone should rewrite it". Fine, but that doesn't mean we shouldn't be fixing oopses in the current stuff.
I fixed up the big reject which this threw - it's getting a bit optimistic sending out patches against 2.6.5...