[PATCH] coredump: elf_core_dump: skip kernel threads

[PATCH] coredump: elf_core_dump: skip kernel threads

[Oleg Nesterov]

linux_binfmt->core_dump() runs before the process does exit_aio(), this means
that we can hit the kernel thread which shares the same ->mm. Afaics, nothing
really bad can happen, but perhaps it makes sense to fix this minor bug.

It is sad we have to iterate over all threads in system and use GFP_ATOMIC.
Hopefully we can kill theses ugly do_each_thread()s, but this needs some
nontrivial changes in mm_struct and do_coredump.

Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>

 binfmt_elf.c       |    6 ++++++
 binfmt_elf_fdpic.c |    3 +++
 2 files changed, 9 insertions(+)

--- 26-rc2/fs/binfmt_elf.c~CD_SKIP_KTHREAD	2008-06-02 20:46:54.000000000 +0400
+++ 26-rc2/fs/binfmt_elf.c	2008-06-16 20:00:00.000000000 +0400
@@ -1522,6 +1522,9 @@ static int fill_note_info(struct elfhdr 
 	rcu_read_lock();
 	do_each_thread(g, p)
 		if (p->mm == dump_task->mm) {
+			if (p->flags & PF_KTHREAD)
+				continue;
+
 			t = kzalloc(offsetof(struct elf_thread_core_info,
 					     notes[info->thread_notes]),
 				    GFP_ATOMIC);
@@ -1726,6 +1729,9 @@ static int fill_note_info(struct elfhdr 
 		rcu_read_lock();
 		do_each_thread(g, p)
 			if (current->mm == p->mm && current != p) {
+				if (p->flags & PF_KTHREAD)
+					continue;
+
 				ets = kzalloc(sizeof(*ets), GFP_ATOMIC);
 				if (!ets) {
 					rcu_read_unlock();
--- 26-rc2/fs/binfmt_elf_fdpic.c~CD_SKIP_KTHREAD	2008-06-02 20:46:54.000000000 +0400
+++ 26-rc2/fs/binfmt_elf_fdpic.c	2008-06-16 18:57:46.000000000 +0400
@@ -1626,6 +1626,9 @@ static int elf_fdpic_core_dump(long sign
 		rcu_read_lock();
 		do_each_thread(g,p)
 			if (current->mm == p->mm && current != p) {
+				if (p->flags & PF_KTHREAD)
+					continue;
+
 				tmp = kzalloc(sizeof(*tmp), GFP_ATOMIC);
 				if (!tmp) {
 					rcu_read_unlock();

Re: [PATCH] coredump: elf_core_dump: skip kernel threads

[Roland McGrath]

Perhaps a little prettier like this:

--- a/fs/binfmt_elf.c
+++ b/fs/binfmt_elf.c
@@ -1469,6 +1469,12 @@ static int fill_thread_core_info(struct 
 	return 1;
 }
 
+static bool is_dump_thread(struct task_struct *dump_task,
+			   struct task_struct *thread)
+{
+	return !(thread->flags & PF_KTHREAD) && thread->mm == dump_task->mm;
+}
+
 static int fill_note_info(struct elfhdr *elf, int phdrs,
 			  struct elf_note_info *info,
 			  long signr, struct pt_regs *regs)
@@ -1518,7 +1524,7 @@ static int fill_note_info(struct elfhdr 
 	 */
 	rcu_read_lock();
 	do_each_thread(g, p)
-		if (p->mm == dump_task->mm) {
+		if (is_dump_thread(dump_task, p)) {
 			t = kzalloc(offsetof(struct elf_thread_core_info,
 					     notes[info->thread_notes]),
 				    GFP_ATOMIC);
@@ -1722,7 +1728,7 @@ static int fill_note_info(struct elfhdr 
 		struct elf_thread_status *ets;
 		rcu_read_lock();
 		do_each_thread(g, p)
-			if (current->mm == p->mm && current != p) {
+			if (p != current && is_dump_thread(current, p)) {
 				ets = kzalloc(sizeof(*ets), GFP_ATOMIC);
 				if (!ets) {
 					rcu_read_unlock();

Or maybe you'll have other reasons to add a task_user_mm() that returns
NULL for kthreads, and then just use task_user_mm(p) == dump_task->mm here.

> It is sad we have to iterate over all threads in system and use GFP_ATOMIC.
> Hopefully we can kill theses ugly do_each_thread()s, but this needs some
> nontrivial changes in mm_struct and do_coredump.

Agreed, and twice at that (coredump_wait).  But this is something more to
consider for the future, and I wouldn't worry about it right now.


Thanks,
Roland

Re: [PATCH] coredump: elf_core_dump: skip kernel threads

[Oleg Nesterov]

On 06/19, Roland McGrath wrote:
>
> Perhaps a little prettier like this:
> 
> --- a/fs/binfmt_elf.c
> +++ b/fs/binfmt_elf.c
> @@ -1469,6 +1469,12 @@ static int fill_thread_core_info(struct 
>  	return 1;
>  }
>  
> +static bool is_dump_thread(struct task_struct *dump_task,
> +			   struct task_struct *thread)
> +{
> +	return !(thread->flags & PF_KTHREAD) && thread->mm == dump_task->mm;
> +}
> +
>  static int fill_note_info(struct elfhdr *elf, int phdrs,
>  			  struct elf_note_info *info,
>  			  long signr, struct pt_regs *regs)
> @@ -1518,7 +1524,7 @@ static int fill_note_info(struct elfhdr 
>  	 */
>  	rcu_read_lock();
>  	do_each_thread(g, p)
> -		if (p->mm == dump_task->mm) {
> +		if (is_dump_thread(dump_task, p)) {
>  			t = kzalloc(offsetof(struct elf_thread_core_info,
>  					     notes[info->thread_notes]),
>  				    GFP_ATOMIC);

Agreed.

> > It is sad we have to iterate over all threads in system and use GFP_ATOMIC.
> > Hopefully we can kill theses ugly do_each_thread()s, but this needs some
> > nontrivial changes in mm_struct and do_coredump.
> 
> Agreed, and twice at that (coredump_wait).  But this is something more to
> consider for the future, and I wouldn't worry about it right now.

In most cases coredump_wait() doesn't need to do for_each_process().

And I think there are other reasons to kill these do_each_thread()s.
We can simplify the code and make the coredumping process visible to
oom_kill. I sent the preparatory patches.

Oleg.