kernel/pid.c: check pid whether be NULL in __change_pid()
PATCH 1/1 kernel/pid.c: check pid whether be NULL 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.
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".
--- kernel/pid.c | 3 +++ 1 files changed, 3 insertions(+), 0 deletions(-)
kernel/pid.c +3 -0
@@ -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 ;
Signature · 1 line
1.7.7.6
Quoted 4 lines · Chen Gang
Within
__change_pid(), 'new' may be NULL if it comes fromdetach_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",
Quoted 2 lines
In real world, at least now, all callers which will call
detach_pid()orchange_pid()will not cause issue,
Yes,
Quoted 2 lines
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.
Quoted 13 lines · Oleg Nesterov
Quoted 4 lines · Chen Gang
Within
__change_pid(), 'new' may be NULL if it comes fromdetach_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",
Quoted 2 lines
In real world, at least now, all callers which will call
detach_pid()orchange_pid()will not cause issue,Yes,
Quoted 9 lines
Quoted 2 lines
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".
--- kernel/pid.c | 6 ++++++ 1 files changed, 6 insertions(+), 0 deletions(-)
kernel/pid.c +6 -0
@@ -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;
Signature · 3 lines
1.7.7.6 --------------------------------patch end-------------------------------
Quoted 4 lines
Oleg.
Signature · 1 line
Chen Gang
Quoted 13 lines · Chen Gang
Quoted 10 lines · Oleg Nesterov
Quoted 3 lines
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.
Quoted 7 lines
Hmm... In my opinion, it means need
BUG_ON()for original 'link->pid'.--------------------------------patch begin-----------------------------
[PATCH]
kernel/pid.c: addBUG_ON()for "!pid" in__change_pid()Within
__change_pid(), 'new' may be NULL if it comes fromdetach_pid(),
Yes, this is fine,
Quoted 3 lines
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().
Quoted 3 lines
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".
Quoted 2 lines
So in
__change_pid(), need addBUG_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.
kernel/pid.c static void __change_pid(struct task_struct *task, enum pid_type type, · line 403
Quoted 11 lines
kernel/pid.c +5
-0@@ -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.
Quoted 60 lines · Oleg Nesterov
Quoted 13 lines · Chen Gang
Quoted 10 lines · Oleg Nesterov
Quoted 3 lines
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.Quoted 7 lines
Hmm... In my opinion, it means need
BUG_ON()for original 'link->pid'.--------------------------------patch begin-----------------------------
[PATCH]
kernel/pid.c: addBUG_ON()for "!pid" in__change_pid()Within
__change_pid(), 'new' may be NULL if it comes fromdetach_pid(),Yes, this is fine,
Quoted 3 lines
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().Quoted 3 lines
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".Quoted 2 lines
So in
__change_pid(), need addBUG_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.
kernel/pid.c static void __change_pid(struct task_struct *task, enum pid_type type, · line 403
Quoted 11 lines
kernel/pid.c +5
-0@@ -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); ...
Quoted 5 lines
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 addBUG_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.
--- kernel/pid.c | 6 +++--- 1 files changed, 3 insertions(+), 3 deletions(-)
kernel/pid.c +3 -3
@@ -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); }
Signature · 7 lines
1.7.7.6 --------------------------------patch end------------------------------- Thanks. -- Chen Gang