On Mon, Jan 10, 2022 at 06:51:38PM -0800, Suren Baghdasaryan wrote:
> With write operation on psi files replacing old trigger with a new one,
> the lifetime of its waitqueue is totally arbitrary. Overwriting an
> existing trigger causes its waitqueue to be freed and pending poll()
> will stumble on trigger->event_wait which was destroyed.
> Fix this by disallowing to redefine an existing psi trigger. If a write
> operation is used on a file descriptor with an already existing psi
> trigger, the operation will fail with EBUSY error.
> Also bypass a check for psi_disabled in the psi_trigger_destroy as the
> flag can be flipped after the trigger is created, leading to a memory
> leak.
>
> Reported-by: syzbot+cdb5dd11c97cc532efad@xxxxxxxxxxxxxxxxxxxxxxxxx
> Analyzed-by: Eric Biggers <ebiggers@xxxxxxxxxx>
> Suggested-by: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
> Signed-off-by: Suren Baghdasaryan <surenb@xxxxxxxxxx>
Please include Fixes and Cc stable tags.
> diff --git a/kernel/cgroup/cgroup.c b/kernel/cgroup/cgroup.c
> index cafb8c114a21..e6878238fb19 100644
> --- a/kernel/cgroup/cgroup.c
> +++ b/kernel/cgroup/cgroup.c
> @@ -3642,6 +3642,12 @@ static ssize_t cgroup_pressure_write(struct kernfs_open_file *of, char *buf,
> cgroup_get(cgrp);
> cgroup_kn_unlock(of->kn);
>
> + /* Allow only one trigger per file descriptor */
> + if (READ_ONCE(ctx->psi.trigger)) {
> + cgroup_put(cgrp);
> + return -EBUSY;
> + }
> +
Doesn't the task have exclusive access to the file at this point? READ_ONCE()
is only needed instead of a plain load when the field can be concurrently
changed by another thread.
> diff --git a/kernel/sched/psi.c b/kernel/sched/psi.c
> index 1652f2bb54b7..882bf62cc247 100644
> --- a/kernel/sched/psi.c
> +++ b/kernel/sched/psi.c
> @@ -1151,7 +1151,6 @@ struct psi_trigger *psi_trigger_create(struct psi_group *group,
> t->event = 0;
> t->last_event_time = 0;
> init_waitqueue_head(&t->event_wait);
> - kref_init(&t->refcount);
>
> mutex_lock(&group->trigger_lock);
>
> @@ -1180,15 +1179,21 @@ struct psi_trigger *psi_trigger_create(struct psi_group *group,
> return t;
> }
>
> -static void psi_trigger_destroy(struct kref *ref)
> +void psi_trigger_destroy(void **trigger_ptr)
> {
> - struct psi_trigger *t = container_of(ref, struct psi_trigger, refcount);
> - struct psi_group *group = t->group;
> + struct psi_trigger *t;
> + struct psi_group *group;
> struct task_struct *task_to_destroy = NULL;
>
> - if (static_branch_likely(&psi_disabled))
> + /*
> + * We do not check psi_disabled since it might have been disabled after
> + * the trigger got created.
> + */
> + t = xchg(trigger_ptr, NULL);
> + if (!t)
> return;
Likewise, doesn't the task have exclusive access to the file at this point?
This is only called during ->release().
And why does this take a pointer to a pointer instead of just the pointer?
> @@ -1305,14 +1289,24 @@ static ssize_t psi_write(struct file *file, const char __user *user_buf,
>
> buf[buf_size - 1] = '\0';
>
> - new = psi_trigger_create(&psi_system, buf, nbytes, res);
> - if (IS_ERR(new))
> - return PTR_ERR(new);
> -
> seq = file->private_data;
> +
> /* Take seq->lock to protect seq->private from concurrent writes */
> mutex_lock(&seq->lock);
> - psi_trigger_replace(&seq->private, new);
> +
> + /* Allow only one trigger per file descriptor */
> + if (READ_ONCE(seq->private)) {
> + mutex_unlock(&seq->lock);
> + return -EBUSY;
> + }
Likewise, what does this race against that would require the use of READ_ONCE()?
- Eric
