2020-03-04 00:44:10

by Qian Cai

Subject: [PATCH -next] posix-cpu-timers: fix memory leaks for task_struct

The recent commit removed put_task_struct() in posix_cpu_timer_del()
results in many memory leaks like this,

unreferenced object 0xc0000016d9b44480 (size 8192):
comm "timer_create01", pid 57749, jiffies 4295163733 (age 6159.670s)
hex dump (first 32 bytes):
02 00 00 00 00 00 00 00 10 00 00 00 00 00 00 00 ................
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
backtrace:
[<0000000056aca129>] copy_process+0x26c/0x18e0
alloc_task_struct_node at kernel/fork.c:169
(inlined by) dup_task_struct at kernel/fork.c:877
(inlined by) copy_process at kernel/fork.c:1929
[<00000000bdbbf9f8>] _do_fork+0xac/0xb20
[<00000000dcb1c445>] __do_sys_clone+0x98/0xe0
__do_sys_clone at kernel/fork.c:2591
[<000000006c059205>] ppc_clone+0x8/0xc
ppc_clone at arch/powerpc/kernel/entry_64.S:479

Fixes: 672ebe8eb017a5 ("posix-cpu-timers: Store a reference to a pid not a task")
Signed-off-by: Qian Cai <[email protected]>
---
kernel/time/posix-cpu-timers.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/kernel/time/posix-cpu-timers.c b/kernel/time/posix-cpu-timers.c
index afd1e959a282..e0b580deb61a 100644
--- a/kernel/time/posix-cpu-timers.c
+++ b/kernel/time/posix-cpu-timers.c
@@ -446,8 +446,10 @@ static int posix_cpu_timer_del(struct k_itimer *timer)

out:
rcu_read_unlock();
- if (!ret)
+ if (!ret) {
put_pid(ctmr->pid);
+ put_task_struct(p);
+ }

return ret;
}
--
2.21.0 (Apple Git-122.2)

2020-03-04 05:19:34

by Eric W. Biederman

Subject: Re: [PATCH -next] posix-cpu-timers: fix memory leaks for task_struct

Qian Cai <[email protected]> writes:

> The recent commit removed put_task_struct() in posix_cpu_timer_del()
> results in many memory leaks like this,

Good spotting but no. The leak is in posix_cpu_timer_create.
There is a strong likely hood but no guarantee that the task
in posix_cpu_timer_del is the same as the task in
posix_cpu_timer_create.

Plus the point of it all is to use pid references instead of task
references.

Thank you very much for catching my braino.

Eric


> unreferenced object 0xc0000016d9b44480 (size 8192):
> comm "timer_create01", pid 57749, jiffies 4295163733 (age 6159.670s)
> hex dump (first 32 bytes):
> 02 00 00 00 00 00 00 00 10 00 00 00 00 00 00 00 ................
> 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
> backtrace:
> [<0000000056aca129>] copy_process+0x26c/0x18e0
> alloc_task_struct_node at kernel/fork.c:169
> (inlined by) dup_task_struct at kernel/fork.c:877
> (inlined by) copy_process at kernel/fork.c:1929
> [<00000000bdbbf9f8>] _do_fork+0xac/0xb20
> [<00000000dcb1c445>] __do_sys_clone+0x98/0xe0
> __do_sys_clone at kernel/fork.c:2591
> [<000000006c059205>] ppc_clone+0x8/0xc
> ppc_clone at arch/powerpc/kernel/entry_64.S:479
>
> Fixes: 672ebe8eb017a5 ("posix-cpu-timers: Store a reference to a pid not a task")
> Signed-off-by: Qian Cai <[email protected]>
> ---
> kernel/time/posix-cpu-timers.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/kernel/time/posix-cpu-timers.c b/kernel/time/posix-cpu-timers.c
> index afd1e959a282..e0b580deb61a 100644
> --- a/kernel/time/posix-cpu-timers.c
> +++ b/kernel/time/posix-cpu-timers.c
> @@ -446,8 +446,10 @@ static int posix_cpu_timer_del(struct k_itimer *timer)
>
> out:
> rcu_read_unlock();
> - if (!ret)
> + if (!ret) {
> put_pid(ctmr->pid);
> + put_task_struct(p);
> + }
>
> return ret;
> }

2020-03-04 05:25:12

by Eric W. Biederman

Subject: [PATCH timers/core] posix-cpu-timers: Put the task_struct in posix_cpu_timers_create

Qian Cai <[email protected]> writes:
> The recent commit removed put_task_struct() in posix_cpu_timer_del()
> results in many memory leaks like this,
>
> unreferenced object 0xc0000016d9b44480 (size 8192):
> comm "timer_create01", pid 57749, jiffies 4295163733 (age 6159.670s)
> hex dump (first 32 bytes):
> 02 00 00 00 00 00 00 00 10 00 00 00 00 00 00 00 ................
> 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
> backtrace:
> [<0000000056aca129>] copy_process+0x26c/0x18e0
> alloc_task_struct_node at kernel/fork.c:169
> (inlined by) dup_task_struct at kernel/fork.c:877
> (inlined by) copy_process at kernel/fork.c:1929
> [<00000000bdbbf9f8>] _do_fork+0xac/0xb20
> [<00000000dcb1c445>] __do_sys_clone+0x98/0xe0
> __do_sys_clone at kernel/fork.c:2591
> [<000000006c059205>] ppc_clone+0x8/0xc
> ppc_clone at arch/powerpc/kernel/entry_64.S:479
>

I forgot that get_task_for_clock called by posix_cpu_timer_create
returns a reference to a task_struct. Put that reference
to avoid the leak.

Link: https://lore.kernel.org/lkml/[email protected]/
Fixes: 672ebe8eb017a5 ("posix-cpu-timers: Store a reference to a pid not a task")
Reported-by: Qian Cai <[email protected]>
Signed-off-by: "Eric W. Biederman" <[email protected]>
---
kernel/time/posix-cpu-timers.c | 1 +
1 file changed, 1 insertion(+)

diff --git a/kernel/time/posix-cpu-timers.c b/kernel/time/posix-cpu-timers.c
index 1c21f2fd3d9b..cd88c1217224 100644
--- a/kernel/time/posix-cpu-timers.c
+++ b/kernel/time/posix-cpu-timers.c
@@ -405,6 +405,7 @@ static int posix_cpu_timer_create(struct k_itimer *new_timer)
new_timer->kclock = &clock_posix_cpu;
timerqueue_init(&new_timer->it.cpu.node);
new_timer->it.cpu.pid = get_task_pid(p, cpu_timer_pid_type(new_timer));
+ put_task_struct(p);
return 0;
}

--
2.20.1

2020-03-04 08:58:09

by Thomas Gleixner

Subject: Re: [PATCH timers/core] posix-cpu-timers: Put the task_struct in posix_cpu_timers_create

[email protected] (Eric W. Biederman) writes:

> Qian Cai <[email protected]> writes:
>> The recent commit removed put_task_struct() in posix_cpu_timer_del()
>> results in many memory leaks like this,
>>
>> unreferenced object 0xc0000016d9b44480 (size 8192):
>> comm "timer_create01", pid 57749, jiffies 4295163733 (age 6159.670s)
>> hex dump (first 32 bytes):
>> 02 00 00 00 00 00 00 00 10 00 00 00 00 00 00 00 ................
>> 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
>> backtrace:
>> [<0000000056aca129>] copy_process+0x26c/0x18e0
>> alloc_task_struct_node at kernel/fork.c:169
>> (inlined by) dup_task_struct at kernel/fork.c:877
>> (inlined by) copy_process at kernel/fork.c:1929
>> [<00000000bdbbf9f8>] _do_fork+0xac/0xb20
>> [<00000000dcb1c445>] __do_sys_clone+0x98/0xe0
>> __do_sys_clone at kernel/fork.c:2591
>> [<000000006c059205>] ppc_clone+0x8/0xc
>> ppc_clone at arch/powerpc/kernel/entry_64.S:479
>>
>
> I forgot that get_task_for_clock called by posix_cpu_timer_create
> returns a reference to a task_struct. Put that reference
> to avoid the leak.

I took the liberty to fold this back into the affected commit and add a
comment why this put_task_struct() is actually required.

Thanks,

tglx

2020-03-04 14:36:33

by Eric W. Biederman

Subject: Re: [PATCH timers/core] posix-cpu-timers: Put the task_struct in posix_cpu_timers_create

Thomas Gleixner <[email protected]> writes:

> [email protected] (Eric W. Biederman) writes:
>
>> Qian Cai <[email protected]> writes:
>>> The recent commit removed put_task_struct() in posix_cpu_timer_del()
>>> results in many memory leaks like this,
>>>
>>> unreferenced object 0xc0000016d9b44480 (size 8192):
>>> comm "timer_create01", pid 57749, jiffies 4295163733 (age 6159.670s)
>>> hex dump (first 32 bytes):
>>> 02 00 00 00 00 00 00 00 10 00 00 00 00 00 00 00 ................
>>> 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
>>> backtrace:
>>> [<0000000056aca129>] copy_process+0x26c/0x18e0
>>> alloc_task_struct_node at kernel/fork.c:169
>>> (inlined by) dup_task_struct at kernel/fork.c:877
>>> (inlined by) copy_process at kernel/fork.c:1929
>>> [<00000000bdbbf9f8>] _do_fork+0xac/0xb20
>>> [<00000000dcb1c445>] __do_sys_clone+0x98/0xe0
>>> __do_sys_clone at kernel/fork.c:2591
>>> [<000000006c059205>] ppc_clone+0x8/0xc
>>> ppc_clone at arch/powerpc/kernel/entry_64.S:479
>>>
>>
>> I forgot that get_task_for_clock called by posix_cpu_timer_create
>> returns a reference to a task_struct. Put that reference
>> to avoid the leak.
>
> I took the liberty to fold this back into the affected commit and add a
> comment why this put_task_struct() is actually required.

Good enough.

We should be able to use rcu and remove the reference entirely.
But that is a change for another day.

Eric