2021-09-29 18:18:21

by Kees Cook

[permalink] [raw]
Subject: Re: [PATCH 2/5] kernel/fork: allocate task->comm dynamicly

On Wed, Sep 29, 2021 at 11:50:33AM +0000, Yafang Shao wrote:
> task->comm is defined as an array embedded in struct task_struct before.
> This patch changes it to a char pointer. It will be allocated in the fork
> and freed when the task is freed.
>
> Signed-off-by: Yafang Shao <[email protected]>
> ---
> include/linux/sched.h | 2 +-
> kernel/fork.c | 19 +++++++++++++++++++
> 2 files changed, 20 insertions(+), 1 deletion(-)
>
> diff --git a/include/linux/sched.h b/include/linux/sched.h
> index e12b524426b0..b387b5943db4 100644
> --- a/include/linux/sched.h
> +++ b/include/linux/sched.h
> @@ -1051,7 +1051,7 @@ struct task_struct {
> * - access it with [gs]et_task_comm()
> * - lock it with task_lock()
> */
> - char comm[TASK_COMM_LEN];
> + char *comm;

This, I think, is basically a non-starter. It adds another kmalloc to
the fork path without a well-justified reason. TASK_COMM_LEN is small,
yes, but why is growing it valuable enough to slow things down?

(Or, can you prove that this does NOT slow things down? It seems like
it would.)

-Kees

>
> struct nameidata *nameidata;
>
> diff --git a/kernel/fork.c b/kernel/fork.c
> index 38681ad44c76..227aec240501 100644
> --- a/kernel/fork.c
> +++ b/kernel/fork.c
> @@ -721,6 +721,20 @@ static void mmdrop_async(struct mm_struct *mm)
> }
> }
>
> +static int task_comm_alloc(struct task_struct *p)
> +{
> + p->comm = kzalloc(TASK_COMM_LEN, GFP_KERNEL);
> + if (!p->comm)
> + return -ENOMEM;
> +
> + return 0;
> +}
> +
> +static void task_comm_free(struct task_struct *p)
> +{
> + kfree(p->comm);
> +}
> +
> static inline void free_signal_struct(struct signal_struct *sig)
> {
> taskstats_tgid_free(sig);
> @@ -753,6 +767,7 @@ void __put_task_struct(struct task_struct *tsk)
> bpf_task_storage_free(tsk);
> exit_creds(tsk);
> delayacct_tsk_free(tsk);
> + task_comm_free(tsk);
> put_signal_struct(tsk->signal);
> sched_core_free(tsk);
>
> @@ -2076,6 +2091,10 @@ static __latent_entropy struct task_struct *copy_process(
> if (data_race(nr_threads >= max_threads))
> goto bad_fork_cleanup_count;
>
> + retval = task_comm_alloc(p);
> + if (retval)
> + goto bad_fork_cleanup_count;
> +
> delayacct_tsk_init(p); /* Must remain after dup_task_struct() */
> p->flags &= ~(PF_SUPERPRIV | PF_WQ_WORKER | PF_IDLE | PF_NO_SETAFFINITY);
> p->flags |= PF_FORKNOEXEC;
> --
> 2.17.1
>

--
Kees Cook


2021-09-30 13:16:49

by Yafang Shao

[permalink] [raw]
Subject: Re: [PATCH 2/5] kernel/fork: allocate task->comm dynamicly

On Thu, Sep 30, 2021 at 2:11 AM Kees Cook <[email protected]> wrote:
>
> On Wed, Sep 29, 2021 at 11:50:33AM +0000, Yafang Shao wrote:
> > task->comm is defined as an array embedded in struct task_struct before.
> > This patch changes it to a char pointer. It will be allocated in the fork
> > and freed when the task is freed.
> >
> > Signed-off-by: Yafang Shao <[email protected]>
> > ---
> > include/linux/sched.h | 2 +-
> > kernel/fork.c | 19 +++++++++++++++++++
> > 2 files changed, 20 insertions(+), 1 deletion(-)
> >
> > diff --git a/include/linux/sched.h b/include/linux/sched.h
> > index e12b524426b0..b387b5943db4 100644
> > --- a/include/linux/sched.h
> > +++ b/include/linux/sched.h
> > @@ -1051,7 +1051,7 @@ struct task_struct {
> > * - access it with [gs]et_task_comm()
> > * - lock it with task_lock()
> > */
> > - char comm[TASK_COMM_LEN];
> > + char *comm;
>
> This, I think, is basically a non-starter. It adds another kmalloc to
> the fork path without a well-justified reason. TASK_COMM_LEN is small,
> yes, but why is growing it valuable enough to slow things down?
>
> (Or, can you prove that this does NOT slow things down? It seems like
> it would.)
>

Right, the new kmalloc would take some extra latency.
Seems it is not easy to measure which one is more valuable.

>
> >
> > struct nameidata *nameidata;
> >
> > diff --git a/kernel/fork.c b/kernel/fork.c
> > index 38681ad44c76..227aec240501 100644
> > --- a/kernel/fork.c
> > +++ b/kernel/fork.c
> > @@ -721,6 +721,20 @@ static void mmdrop_async(struct mm_struct *mm)
> > }
> > }
> >
> > +static int task_comm_alloc(struct task_struct *p)
> > +{
> > + p->comm = kzalloc(TASK_COMM_LEN, GFP_KERNEL);
> > + if (!p->comm)
> > + return -ENOMEM;
> > +
> > + return 0;
> > +}
> > +
> > +static void task_comm_free(struct task_struct *p)
> > +{
> > + kfree(p->comm);
> > +}
> > +
> > static inline void free_signal_struct(struct signal_struct *sig)
> > {
> > taskstats_tgid_free(sig);
> > @@ -753,6 +767,7 @@ void __put_task_struct(struct task_struct *tsk)
> > bpf_task_storage_free(tsk);
> > exit_creds(tsk);
> > delayacct_tsk_free(tsk);
> > + task_comm_free(tsk);
> > put_signal_struct(tsk->signal);
> > sched_core_free(tsk);
> >
> > @@ -2076,6 +2091,10 @@ static __latent_entropy struct task_struct *copy_process(
> > if (data_race(nr_threads >= max_threads))
> > goto bad_fork_cleanup_count;
> >
> > + retval = task_comm_alloc(p);
> > + if (retval)
> > + goto bad_fork_cleanup_count;
> > +
> > delayacct_tsk_init(p); /* Must remain after dup_task_struct() */
> > p->flags &= ~(PF_SUPERPRIV | PF_WQ_WORKER | PF_IDLE | PF_NO_SETAFFINITY);
> > p->flags |= PF_FORKNOEXEC;
> > --
> > 2.17.1
> >
>
> --
> Kees Cook



--
Thanks
Yafang

2021-09-30 20:36:04

by Petr Mladek

[permalink] [raw]
Subject: Re: [PATCH 2/5] kernel/fork: allocate task->comm dynamicly

On Thu 2021-09-30 20:41:40, Yafang Shao wrote:
> On Thu, Sep 30, 2021 at 2:11 AM Kees Cook <[email protected]> wrote:
> >
> > On Wed, Sep 29, 2021 at 11:50:33AM +0000, Yafang Shao wrote:
> > > task->comm is defined as an array embedded in struct task_struct before.
> > > This patch changes it to a char pointer. It will be allocated in the fork
> > > and freed when the task is freed.
> > >
> > > Signed-off-by: Yafang Shao <[email protected]>
> > > ---
> > > include/linux/sched.h | 2 +-
> > > kernel/fork.c | 19 +++++++++++++++++++
> > > 2 files changed, 20 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/include/linux/sched.h b/include/linux/sched.h
> > > index e12b524426b0..b387b5943db4 100644
> > > --- a/include/linux/sched.h
> > > +++ b/include/linux/sched.h
> > > @@ -1051,7 +1051,7 @@ struct task_struct {
> > > * - access it with [gs]et_task_comm()
> > > * - lock it with task_lock()
> > > */
> > > - char comm[TASK_COMM_LEN];
> > > + char *comm;
> >
> > This, I think, is basically a non-starter. It adds another kmalloc to
> > the fork path without a well-justified reason. TASK_COMM_LEN is small,
> > yes, but why is growing it valuable enough to slow things down?
> >
> > (Or, can you prove that this does NOT slow things down? It seems like
> > it would.)
> >
>
> Right, the new kmalloc would take some extra latency.
> Seems it is not easy to measure which one is more valuable.

Honestly, I do not think that this exercise is worth it. The patchset
adds a lot of complexity and potential problems just to extend
comm from 16 to 24 for kthreads.

Is the problem real or just cosmetic?

If you really want it then it would be much easier to increase
TASK_COMM_LEN. task_struct is growing rather regularly. Extra
8 bytes should be acceptable.

If you want to make it more acceptable then keep 16 for
CONFIG_BASE_SMALL.


> > > diff --git a/kernel/fork.c b/kernel/fork.c
> > > index 38681ad44c76..227aec240501 100644
> > > --- a/kernel/fork.c
> > > +++ b/kernel/fork.c
> > > @@ -753,6 +767,7 @@ void __put_task_struct(struct task_struct *tsk)
> > > bpf_task_storage_free(tsk);
> > > exit_creds(tsk);
> > > delayacct_tsk_free(tsk);
> > > + task_comm_free(tsk);

Just one example of the potential problems. Are you sure that nobody
will access tsk->comm after this point?

task->comm is widely used to describe the affected task_struct because
it is user friendly.

Also __put_task_struct() later calls also profile_handoff_task() that might
get registered even by some external module.

Best Regards,
Petr

PS: I think that the fork performance is important. It is tested by
benchmarks, for example, lmbench. But for me, the reliability is even
more important and any pointer/alloc/free just adds another weak
point.