Within __change_pid(), 'new' may be NULL if it comes from detach_pid(),
and 'link->pid' also may be NULL ("link->pid = new"), so theoretically,\
the original 'link->pid' may be NULL, too.
In real world, at least now, all callers which will call detach_pid()
or change_pid() will not cause issue, but still recommend to check it
in __change_pid() to let itself consistency.
After the modification, it passed a simpe test: "build -> boot up ->
[s/g]et_[p/s/g]id() test by LTP tools".
Signed-off-by: Chen Gang <[email protected]>
---
kernel/pid.c | 3 +++
1 files changed, 3 insertions(+), 0 deletions(-)
diff --git a/kernel/pid.c b/kernel/pid.c
index 9b9a266..15b1b3d 100644
--- a/kernel/pid.c
+++ b/kernel/pid.c
@@ -399,6 +399,9 @@ static void __change_pid(struct task_struct *task, enum pid_type type,
hlist_del_rcu(&link->node);
link->pid = new;
+ if (!pid)
+ return;
+
for (tmp = PIDTYPE_MAX; --tmp >= 0; )
if (!hlist_empty(&pid->tasks[tmp]))
return;
--
1.7.7.6
On 10/07, Chen Gang wrote:
>
> Within __change_pid(), 'new' may be NULL if it comes from detach_pid(),
> and 'link->pid' also may be NULL ("link->pid = new"), so theoretically,\
> the original 'link->pid' may be NULL, too.
I don't really understand this "theoretically",
> In real world, at least now, all callers which will call detach_pid()
> or change_pid() will not cause issue,
Yes,
> but still recommend to check it
> in __change_pid() to let itself consistency.
Why?
Contrary, I think we should not hide the problem. If __change_pid() is
called when task->pids[type].pid is already NULL there is something
seriously wrong.
Oleg.
On 10/07/2013 08:43 PM, Oleg Nesterov wrote:
> On 10/07, Chen Gang wrote:
>>
>> Within __change_pid(), 'new' may be NULL if it comes from detach_pid(),
>> and 'link->pid' also may be NULL ("link->pid = new"), so theoretically,\
>> the original 'link->pid' may be NULL, too.
>
> I don't really understand this "theoretically",
>
>> In real world, at least now, all callers which will call detach_pid()
>> or change_pid() will not cause issue,
>
> Yes,
>
>> but still recommend to check it
>> in __change_pid() to let itself consistency.
>
> Why?
>
> Contrary, I think we should not hide the problem. If __change_pid() is
> called when task->pids[type].pid is already NULL there is something
> seriously wrong.
>
Hmm... In my opinion, it means need BUG_ON() for original 'link->pid'.
--------------------------------patch begin-----------------------------
[PATCH] kernel/pid.c: add BUG_ON() for "!pid" in __change_pid()
Within __change_pid(), 'new' may be NULL if it comes from detach_pid(),
and 'link->pid' also may be NULL ("link->pid = new"), so theoretically,
the original 'link->pid' may be NULL, too.
But in real world, all related extern functions always assume "if
'link->pid' is already NULL, there must be something seriously wrong",
although __change_pid() can accept parameter 'new' as NULL.
So in __change_pid(), need add BUG_ON() for it: "it should not happen,
when it really happen, OS must be continuing blindly, and next will
cause serious issue".
Signed-off-by: Chen Gang <[email protected]>
---
kernel/pid.c | 6 ++++++
1 files changed, 6 insertions(+), 0 deletions(-)
diff --git a/kernel/pid.c b/kernel/pid.c
index 9b9a266..8fc87f1 100644
--- a/kernel/pid.c
+++ b/kernel/pid.c
@@ -396,6 +396,12 @@ static void __change_pid(struct task_struct *task, enum pid_type type,
link = &task->pids[type];
pid = link->pid;
+ /*
+ * If task->pids[type].pid is already NULL, there must be something
+ * seriously wrong
+ */
+ BUG_ON(!pid);
+
hlist_del_rcu(&link->node);
link->pid = new;
--
1.7.7.6
--------------------------------patch end-------------------------------
> Oleg.
>
>
>
--
Chen Gang
On 10/08, Chen Gang wrote:
>
> On 10/07/2013 08:43 PM, Oleg Nesterov wrote:
> >
> >> but still recommend to check it
> >> in __change_pid() to let itself consistency.
> >
> > Why?
> >
> > Contrary, I think we should not hide the problem. If __change_pid() is
> > called when task->pids[type].pid is already NULL there is something
> > seriously wrong.
> >
>
> Hmm... In my opinion, it means need BUG_ON() for original 'link->pid'.
>
> --------------------------------patch begin-----------------------------
>
> [PATCH] kernel/pid.c: add BUG_ON() for "!pid" in __change_pid()
>
> Within __change_pid(), 'new' may be NULL if it comes from detach_pid(),
Yes, this is fine,
> and 'link->pid' also may be NULL ("link->pid = new"),
> ...
> the original 'link->pid' may be NULL, too.
Too? You mean, it becomes NULL after detach_pid().
> But in real world, all related extern functions always assume "if
> 'link->pid' is already NULL, there must be something seriously wrong",
> although __change_pid() can accept parameter 'new' as NULL.
I simply can't understand why you mix "new == NULL" and "link->pid == NULL".
> So in __change_pid(), need add BUG_ON() for it: "it should not happen,
> when it really happen, OS must be continuing blindly,
OS will crash a couple of lines below trying to dereference this pointer.
> --- a/kernel/pid.c
> +++ b/kernel/pid.c
> @@ -396,6 +396,12 @@ static void __change_pid(struct task_struct *task, enum pid_type type,
> link = &task->pids[type];
> pid = link->pid;
>
> + /*
> + * If task->pids[type].pid is already NULL, there must be something
> + * seriously wrong
> + */
> + BUG_ON(!pid);
Following this logic you should also add
BUG_ON(!task);
BUG_ON(!link->node.next);
BUG_ON(!link->node.prev || link->node.prev == LIST_POISON2);
...
Seriously, I do not understand the point. Yes, detach_pid() should not
be called twice. And it has a single caller. And this caller will crash
too if it is called twice. So you can also add BUG_ON() into
__unhash_process(). And so on.
Oleg.
On 10/09/2013 01:56 AM, Oleg Nesterov wrote:
> On 10/08, Chen Gang wrote:
>>
>> On 10/07/2013 08:43 PM, Oleg Nesterov wrote:
>>>
>>>> but still recommend to check it
>>>> in __change_pid() to let itself consistency.
>>>
>>> Why?
>>>
>>> Contrary, I think we should not hide the problem. If __change_pid() is
>>> called when task->pids[type].pid is already NULL there is something
>>> seriously wrong.
>>>
>>
>> Hmm... In my opinion, it means need BUG_ON() for original 'link->pid'.
>>
>> --------------------------------patch begin-----------------------------
>>
>> [PATCH] kernel/pid.c: add BUG_ON() for "!pid" in __change_pid()
>>
>> Within __change_pid(), 'new' may be NULL if it comes from detach_pid(),
>
> Yes, this is fine,
>
>> and 'link->pid' also may be NULL ("link->pid = new"),
>> ...
>> the original 'link->pid' may be NULL, too.
>
> Too? You mean, it becomes NULL after detach_pid().
>
>> But in real world, all related extern functions always assume "if
>> 'link->pid' is already NULL, there must be something seriously wrong",
>> although __change_pid() can accept parameter 'new' as NULL.
>
> I simply can't understand why you mix "new == NULL" and "link->pid == NULL".
>
>> So in __change_pid(), need add BUG_ON() for it: "it should not happen,
>> when it really happen, OS must be continuing blindly,
>
> OS will crash a couple of lines below trying to dereference this pointer.
>
>> --- a/kernel/pid.c
>> +++ b/kernel/pid.c
>> @@ -396,6 +396,12 @@ static void __change_pid(struct task_struct *task, enum pid_type type,
>> link = &task->pids[type];
>> pid = link->pid;
>>
>> + /*
>> + * If task->pids[type].pid is already NULL, there must be something
>> + * seriously wrong
>> + */
>> + BUG_ON(!pid);
>
> Following this logic you should also add
>
> BUG_ON(!task);
> BUG_ON(!link->node.next);
> BUG_ON(!link->node.prev || link->node.prev == LIST_POISON2);
> ...
>
> Seriously, I do not understand the point. Yes, detach_pid() should not
> be called twice. And it has a single caller. And this caller will crash
> too if it is called twice. So you can also add BUG_ON() into
> __unhash_process(). And so on.
>
In my opinion, for using BUG_ON(), it has 3 requirements:
- OS is just continuing blindly.
- next, will cause real issue (or need use WARN_ON instead of).
- Can let the related code self consitency (or will add many wastes).
Your demo are match 2 requrements, but not match the 3rd one: "it is
reasonable to assume 'task', 'link', and 'link->node' are valid in
__change_pid()".
But for link->pid, the function name '__change_pid' tells us it is only
for changing pid, if 'new' can be NULL, 'link->pid' also can be NULL,
so the original 'link-pid' can be NULL, too.
So for self consistency, we also can change the function name from
'__change_pid' to another one (e.g. 'change_orig_valid_pid'), to let
itself consistency (so don't need BUG_ON)
The related patch is below, please check, thanks.
--------------------------------patch begin-----------------------------
kernel/pid.c: use 'change_orig_valid_pid' instead of '__change_pid' for function name
For function name '__change_pid' is only for changing pid. In fact, it
always assumes the original pid is valid, but new pid can be NULL, so
recommend to use 'change_orig_valid_pid' instead of.
Signed-off-by: Chen Gang <[email protected]>
---
kernel/pid.c | 6 +++---
1 files changed, 3 insertions(+), 3 deletions(-)
diff --git a/kernel/pid.c b/kernel/pid.c
index 9b9a266..408a3b5 100644
--- a/kernel/pid.c
+++ b/kernel/pid.c
@@ -386,7 +386,7 @@ void attach_pid(struct task_struct *task, enum pid_type type)
hlist_add_head_rcu(&link->node, &link->pid->tasks[type]);
}
-static void __change_pid(struct task_struct *task, enum pid_type type,
+static void change_orig_valid_pid(struct task_struct *task, enum pid_type type,
struct pid *new)
{
struct pid_link *link;
@@ -408,13 +408,13 @@ static void __change_pid(struct task_struct *task, enum pid_type type,
void detach_pid(struct task_struct *task, enum pid_type type)
{
- __change_pid(task, type, NULL);
+ change_orig_valid_pid(task, type, NULL);
}
void change_pid(struct task_struct *task, enum pid_type type,
struct pid *pid)
{
- __change_pid(task, type, pid);
+ change_orig_valid_pid(task, type, pid);
attach_pid(task, type);
}
--
1.7.7.6
--------------------------------patch end-------------------------------
Thanks.
--
Chen Gang