Skip to contentSkip to conversation

introduce get_task_pid() to fix unsafe get_pid()

PATCH In review Oleg Nesterov 2 participants 4 messages Sep 11, 2006

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 introduce get_task_pid() to fix unsafe get_pid()

(COMPILE TESTED, needs an ack from Eric)

proc_pid_make_inode:

	ei->pid = get_pid(task_pid(task));

I think this is not safe. get_pid() can be preempted after checking "pid != NULL". Then the task exits, does detach_pid(), and RCU frees the pid.

include/linux/pid.h renamed from include/linux/pid.h~1_tgp +2 -0
@@ -68,6 +68,8 @@ extern struct task_struct *FASTCALL(pid_
 extern struct task_struct *FASTCALL(get_pid_task(struct pid *pid,
 						enum pid_type));
 
+extern struct pid *get_task_pid(struct task_struct *task, enum pid_type type);
+
 /*
  * attach_pid() and detach_pid() must be called with the tasklist_lock
  * write-held.
kernel/pid.c renamed from kernel/pid.c~1_tgp +9 -0
@@ -305,6 +305,15 @@ struct task_struct *find_task_by_pid_typ
 
 EXPORT_SYMBOL(find_task_by_pid_type);
 
+struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
+{
+	struct pid *pid;
+	rcu_read_lock();
+	pid = get_pid(task->pids[type].pid);
+	rcu_read_unlock();
+	return pid;
+}
+
 struct task_struct *fastcall get_pid_task(struct pid *pid, enum pid_type type)
 {
 	struct task_struct *result;
fs/proc/base.c renamed from fs/proc/base.c~1_tgp +2 -2
@@ -958,7 +958,7 @@ static struct inode *proc_pid_make_inode
 	/*
 	 * grab the reference to task.
 	 */
-	ei->pid = get_pid(task_pid(task));
+	ei->pid = get_task_pid(task, PIDTYPE_PID);
 	if (!ei->pid)
 		goto out_unlock;
 
@@ -1665,7 +1665,7 @@ static struct dentry *proc_base_instanti
 	/*
 	 * grab the reference to the task.
 	 */
-	ei->pid = get_pid(task_pid(task));
+	ei->pid = get_task_pid(task, PIDTYPE_PID);
 	if (!ei->pid)
 		goto out_iput;
 
#1 lore raw
Quoted 9 lines · Oleg Nesterov

(COMPILE TESTED, needs an ack from Eric)

proc_pid_make_inode:

	ei->pid = get_pid(task_pid(task));

I think this is not safe. get_pid() can be preempted after checking "pid != NULL". Then the task exits, does detach_pid(), and RCU frees the pid.

Ugh. I had forgotten that the pid of a task gets freed even if you hold a reference to the task struct. So the preemption case looks possible.

Your technique to handle this problem looks fine.

As for the functions can we build them in all 4 varieties. struct pid *get_task_pid(struct task *);
struct pid *get_task_tgid(struct task *);
struct pid *get_task_pgrp(struct task *);
struct pid *get_task_session(struct task *);

Functions without a flag are less error prone to use, and clearer to read.

Either that or we can just drop in some rcu_read_lock() rcu_read_unlock() into the call sites.

Eric

#2 lore raw
Quoted 6 lines · Eric W. Biederman

As for the functions can we build them in all 4 varieties. struct pid *get_task_pid(struct task *);
struct pid *get_task_tgid(struct task *);
struct pid *get_task_pgrp(struct task *);
struct pid *get_task_session(struct task *);

Something like the patch below?

Quoted 2 lines

Either that or we can just drop in some rcu_read_lock() rcu_read_unlock() into the call sites.

Possible. I don't have a strong opinion, please feel free to send a different patch.

[PATCH] introduce get_task_pid() to fix unsafe get_pid()

proc_pid_make_inode:

	ei->pid = get_pid(task_pid(task));

I think this is not safe. get_pid() can be preempted after checking "pid != NULL". Then the task exits, does detach_pid(), and RCU frees the pid.

include/linux/pid.h renamed from include/linux/pid.h~1_tgp +2 -0
@@ -68,6 +68,8 @@ extern struct task_struct *FASTCALL(pid_
 extern struct task_struct *FASTCALL(get_pid_task(struct pid *pid,
 						enum pid_type));
 
+extern struct pid *__get_task_pid(struct task_struct *task, enum pid_type type);
+
 /*
  * attach_pid() and detach_pid() must be called with the tasklist_lock
  * write-held.
kernel/pid.c renamed from kernel/pid.c~1_tgp +9 -0
@@ -305,6 +305,15 @@ struct task_struct *find_task_by_pid_typ
 
 EXPORT_SYMBOL(find_task_by_pid_type);
 
+struct pid *__get_task_pid(struct task_struct *task, enum pid_type type)
+{
+	struct pid *pid;
+	rcu_read_lock();
+	pid = get_pid(task->pids[type].pid);
+	rcu_read_unlock();
+	return pid;
+}
+
 struct task_struct *fastcall get_pid_task(struct pid *pid, enum pid_type type)
 {
 	struct task_struct *result;
include/linux/sched.h renamed from include/linux/sched.h~1_tgp +5 -0
@@ -1073,6 +1073,11 @@ static inline struct pid *task_session(s
 	return task->group_leader->pids[PIDTYPE_SID].pid;
 }
 
+static inline struct pid *get_task_pid(struct task_struct *task)
+{
+	return __get_task_pid(task, PIDTYPE_PID);
+}
+
 /**
  * pid_alive - check that a task structure is not stale
  * @p: Task structure to be checked.
fs/proc/base.c renamed from fs/proc/base.c~1_tgp +2 -2
@@ -958,7 +958,7 @@ static struct inode *proc_pid_make_inode
 	/*
 	 * grab the reference to task.
 	 */
-	ei->pid = get_pid(task_pid(task));
+	ei->pid = get_task_pid(task);
 	if (!ei->pid)
 		goto out_unlock;
 
@@ -1665,7 +1665,7 @@ static struct dentry *proc_base_instanti
 	/*
 	 * grab the reference to the task.
 	 */
-	ei->pid = get_pid(task_pid(task));
+	ei->pid = get_task_pid(task);
 	if (!ei->pid)
 		goto out_iput;
 
#3 lore raw
Quoted 9 lines · Oleg Nesterov
Quoted 6 lines · Eric W. Biederman

As for the functions can we build them in all 4 varieties. struct pid *get_task_pid(struct task *);
struct pid *get_task_tgid(struct task *);
struct pid *get_task_pgrp(struct task *);
struct pid *get_task_session(struct task *);

Something like the patch below?

Yes something like that. Although it doesn't provide for the get_task_tgid case, and your patch only get_task_pid.

Quoted 5 lines
Quoted 2 lines

Either that or we can just drop in some rcu_read_lock() rcu_read_unlock() into the call sites.

Possible. I don't have a strong opinion, please feel free to send a different patch.

I just might. Coming up with an idiom that is hard to get wrong, is desirable here, or at least with an idiom that is consistent.

I need to sleep on it before I can answer which way we handle that. The pain with a new idiom is that I will have to update all of the users so all of the examples in the kernel are consistent.

I might just need to do that anyway, but...

Eric