Re: [Patch] vfs: increase scope of critical locked path in fget_light to avoid race

[Al Viro <viro@parcelfarce.linux.theplanet.co.uk>]

On Fri, May 20, 2005 at 11:25:50AM -0400, Neil Horman wrote:
> I don't have the complete race scenario, just a stack that suggests that
> files->fd was corrupted.  This problem isn't recreatable at will (yet), so this
> is really based on a thought experiment more than anything else.  The conditions
> that I was envisioning was a multithreaded application in which two threads
> modified  the same file descriptor at the same time.  Its my understanding, from
> the way I read the code that the ref count on a file_struct will still be one
> for a multithreaded application, and as such it would be possible, using the
> fget_light routine for one thread to be be preforming an operation on an
> descriptor in the fd array, while another thread preformed another operation

Incorrect.  References to files_struct are held by task_struct.  Kernel
stack is determined by task_struct.  So your two threads would have
to share task_struct (i.e. be purely userland ones) and could not
run in the kernel at the same time for very obvious reasons.

The rules are simple:
	* all access to files_struct is done from upper-half (i.e. is
process-synchronous).
	* the only files_struct you can modify is *(current->files)
	* each task_struct that has ->files pointing to given files_struct
contributes 1 to ->count of that files_struct.  There might be other holders
of temporary references and they also contribute to ->count.
	* all changes of task->files itself are process-synchronous.  Only
two kinds of changes are possible:
	1) current->files can be set to NULL.  That drops a reference to
original files_struct.
	2) current->files can be replaced with a pointer to a new copy of
previous files_struct.  This operations drops a reference to old one and
sets the refcount on a copy to 1.  It could either be done explicitly (when
unshare(2) gets merged into Linus' tree) or implicitly at the clone()/fork()
time.  In the latter case that's done by parent to child before the child
gets a chance to run.
	* at task creation time, child inherits ->files from parent; that
acquires a new reference to it.  That might be followed by implicit unshare()
(see above).  fork(2) always unshares ->files, clone(2) does that unless
CLONE_FILES had been passed to it in flags.

IOW, the only way for two tasks to have ->files pointing to the same object
is to have it unchanged all the way back to common ancestor.  In particular,
if current->files->count is 1, we know that no other task has ->files pointing
to our files_struct and that will remain true until we call clone().  It does
*not* mean that current->files->count will remain 1; somebody might acquire
a temporary reference to our files_struct.  However, we are guaranteed that
all such references will be used only for read-only access (that happens,
e.g., when somebody does ls /proc/<our_pid>/fd - they will grab a reference
to our ->files and go looking at the descriptor table; they are not allowed
to change it, though).