Serge E. Hallyn wrote:
> Quoting Pierre Peiffer (pierre.peiffer@bull.net):
>>
>> Serge E. Hallyn wrote:
>>> Quoting pierre.peiffer@bull.net (pierre.peiffer@bull.net):
>>>> From: Pierre Peiffer <pierre.peiffer@bull.net>
>>>>
>>>> In order to modify the semundo-list of a task from procfs, we must be able to
>>>> work on any target task.
>>>> But all the existing code playing with the semundo-list, currently works
>>>> only on the 'current' task, and does not allow to specify any target task.
>>>>
>>>> This patch changes all these routines to allow them to work on a specified
>>>> task, passed in parameter, instead of current.
>>>>
>>>> This is mainly a preparation for the semundo_write() operation, on the
>>>> /proc/<pid>/semundo file, as provided in the next patch.
>>>>
>>>> Signed-off-by: Pierre Peiffer <pierre.peiffer@bull.net>
>>>> ---
>>>>
>>>>  ipc/sem.c |   90 ++++++++++++++++++++++++++++++++++++++++++++++----------------
>>>>  1 file changed, 68 insertions(+), 22 deletions(-)
>>>>
>>>> Index: b/ipc/sem.c
>>>> ===================================================================
>>>> --- a/ipc/sem.c
>>>> +++ b/ipc/sem.c
>>>> @@ -1017,8 +1017,9 @@ asmlinkage long sys_semctl (int semid, i
>>>>  }
>>>>
>>>>  /* If the task doesn't already have a undo_list, then allocate one
>>>> - * here.  We guarantee there is only one thread using this undo list,
>>>> - * and current is THE ONE
>>>> + * here.
>>>> + * The target task (tsk) is current in the general case, except when
>>>> + * accessed from the procfs (ie when writting to /proc/<pid>/semundo)
>>>>   *
>>>>   * If this allocation and assignment succeeds, but later
>>>>   * portions of this code fail, there is no need to free the sem_undo_list.
>>>> @@ -1026,22 +1027,60 @@ asmlinkage long sys_semctl (int semid, i
>>>>   * at exit time.
>>>>   *
>>>>   * This can block, so callers must hold no locks.
>>>> + *
>>>> + * Note: task_lock is used to synchronize 1. several possible concurrent
>>>> + * creations and 2. the free of the undo_list (done when the task using it
>>>> + * exits). In the second case, we check the PF_EXITING flag to not create
>>>> + * an undo_list for a task which has exited.
>>>> + * If there already is an undo_list for this task, there is no need
>>>> + * to held the task-lock to retrieve it, as the pointer can not change
>>>> + * afterwards.
>>>>   */
>>>> -static inline int get_undo_list(struct sem_undo_list **undo_listp)
>>>> +static inline int get_undo_list(struct task_struct *tsk,
>>>> +				struct sem_undo_list **ulp)
>>>>  {
>>>> -	struct sem_undo_list *undo_list;
>>>> +	if (tsk->sysvsem.undo_list == NULL) {
>>>> +		struct sem_undo_list *undo_list;
>>> Hmm, this is weird.  If there was no undo_list and
>>> tsk!=current, you set the refcnt to 2.  But if there was an
>>> undo list and tsk!=current, where do you inc the refcnt?
>>>
>> I inc it  outside this function, as I don't call get_undo_list() if there is an
>> undo_list.
>> This appears most clearly in the next patch, in semundo_open() for example.
> 
> Ok, so however unlikely, there is a flow that could cause you a problem:
> T2 calls semundo_open() for T1.  T1 does not yet have a semundolist.
> T2.semundo_open() calls get_undo_list, just then T1 creats its own
> semundo_list.  T2 comes to top of get_undo_list() and see
> tsk->sysvsem.undo_list != NULL, simply returns a pointer to the
> undo_list.  Now you never increment the count.
> 
Right.

And yesterday, with more testing in the corners, I've found another issue: if I
use /proc/self/semundo, I don't have tsk != current and the refcnt is wrong too.

Thanks for finding this !

P.

>>>> -	undo_list = current->sysvsem.undo_list;
>>>> -	if (!undo_list) {
>>>> -		undo_list = kzalloc(sizeof(*undo_list), GFP_KERNEL);
>>>> +		/* we must alloc a new one */
>>>> +		undo_list = kmalloc(sizeof(*undo_list), GFP_KERNEL);
>>>>  		if (undo_list == NULL)
>>>>  			return -ENOMEM;
>>>> +
>>>> +		task_lock(tsk);
>>>> +
>>>> +		/* check again if there is an undo_list for this task */
>>>> +		if (tsk->sysvsem.undo_list) {
>>>> +			if (tsk != current)
>>>> +				atomic_inc(&tsk->sysvsem.undo_list->refcnt);
>>>> +			task_unlock(tsk);
>>>> +			kfree(undo_list);
>>>> +			goto out;
>>>> +		}
>>>> +
>>>>  		spin_lock_init(&undo_list->lock);
>>>> -		atomic_set(&undo_list->refcnt, 1);
>>>> -		undo_list->ns = get_ipc_ns(current->nsproxy->ipc_ns);
>>>> -		current->sysvsem.undo_list = undo_list;
>>>> +		/*
>>>> +		 * If tsk is not current (meaning that current is creating
>>>> +		 * a semundo_list for a target task through procfs), and if
>>>> +		 * it's not being exited then refcnt must be 2: the target
>>>> +		 * task tsk + current.
>>>> +		 */
>>>> +		if (tsk == current)
>>>> +			atomic_set(&undo_list->refcnt, 1);
>>>> +		else if (!(tsk->flags & PF_EXITING))
>>>> +			atomic_set(&undo_list->refcnt, 2);
>>>> +		else {
>>>> +			task_unlock(tsk);
>>>> +			kfree(undo_list);
>>>> +			return -EINVAL;
>>>> +		}
>>>> +		undo_list->ns = get_ipc_ns(tsk->nsproxy->ipc_ns);
>>>> +		undo_list->proc_list = NULL;
>>>> +		tsk->sysvsem.undo_list = undo_list;
>>>> +		task_unlock(tsk);
>>>>  	}
>>>> -	*undo_listp = undo_list;
>>>> +out:
>>>> +	*ulp = tsk->sysvsem.undo_list;
>>>>  	return 0;
>>>>  }
>>>>
>>>> @@ -1065,17 +1104,12 @@ static struct sem_undo *lookup_undo(stru
>>>>  	return un;
>>>>  }
>>>>
>>>> -static struct sem_undo *find_undo(struct ipc_namespace *ns, int semid)
>>>> +static struct sem_undo *find_undo(struct sem_undo_list *ulp, int semid)
>>>>  {
>>>>  	struct sem_array *sma;
>>>> -	struct sem_undo_list *ulp;
>>>>  	struct sem_undo *un, *new;
>>>> +	struct ipc_namespace *ns;
>>>>  	int nsems;
>>>> -	int error;
>>>> -
>>>> -	error = get_undo_list(&ulp);
>>>> -	if (error)
>>>> -		return ERR_PTR(error);
>>>>
>>>>  	spin_lock(&ulp->lock);
>>>>  	un = lookup_undo(ulp, semid);
>>>> @@ -1083,6 +1117,8 @@ static struct sem_undo *find_undo(struct
>>>>  	if (likely(un!=NULL))
>>>>  		goto out;
>>>>
>>>> +	ns = ulp->ns;
>>>> +
>>>>  	/* no undo structure around - allocate one. */
>>>>  	sma = sem_lock_check(ns, semid);
>>>>  	if (IS_ERR(sma))
>>>> @@ -1133,6 +1169,7 @@ asmlinkage long sys_semtimedop(int semid
>>>>  	struct sem_array *sma;
>>>>  	struct sembuf fast_sops[SEMOPM_FAST];
>>>>  	struct sembuf* sops = fast_sops, *sop;
>>>> +	struct sem_undo_list *ulp;
>>>>  	struct sem_undo *un;
>>>>  	int undos = 0, alter = 0, max;
>>>>  	struct sem_queue queue;
>>>> @@ -1177,9 +1214,13 @@ asmlinkage long sys_semtimedop(int semid
>>>>  			alter = 1;
>>>>  	}
>>>>
>>>> +	error = get_undo_list(current, &ulp);
>>>> +	if (error)
>>>> +		goto out_free;
>>>> +
>>>>  retry_undos:
>>>>  	if (undos) {
>>>> -		un = find_undo(ns, semid);
>>>> +		un = find_undo(ulp, semid);
>>>>  		if (IS_ERR(un)) {
>>>>  			error = PTR_ERR(un);
>>>>  			goto out_free;
>>>> @@ -1305,7 +1346,7 @@ int copy_semundo(unsigned long clone_fla
>>>>  	int error;
>>>>
>>>>  	if (clone_flags & CLONE_SYSVSEM) {
>>>> -		error = get_undo_list(&undo_list);
>>>> +		error = get_undo_list(current, &undo_list);
>>>>  		if (error)
>>>>  			return error;
>>>>  		atomic_inc(&undo_list->refcnt);
>>>> @@ -1405,10 +1446,15 @@ next_entry:
>>>>  	kfree(undo_list);
>>>>  }
>>>>
>>>> -/* called from do_exit() */
>>>> +/* exit_sem: called from do_exit()
>>>> + * task_lock is used to synchronize with get_undo_list()
>>> Ok I had to think about this again.  I'd like the comment
>>> here to point out that the task_lock here acts as a barrier
>>> between the prior setting of PF_EXITING and the undo_list
>>> being freed here, so that get_undo_list() will either see
>>> PF_EXITING is NOT in the tsk->flags, in which case it will