On Wed, Sep 29, 2021 at 11:50:35AM +0000, Yafang Shao wrote:
> This patch increases the size of ktread's comm from 16 to 24, which is
> the same with workqueue's, to improve this situation. After this cahnge,
> [...]
> Because there're only a few of kthreads, so it won't increase too much
> memory.
Even without the performance impact changes, the math here doesn't hold
either, since using kmalloc means there are slabs being allocated to hold
the task "comm"s now (which comes with overhead), and every task added
a pointer to those 16 bytes (i.e. 8 more bytes on 64-bit systems). So
this change, even if there was 0 overhead in using slabs, would be
identical to having just raised TASK_COMM_LEN to 24. 8 byte pointer,
16 byte allocation == 24 bytes.
-Kees
--
Kees Cook
On Thu, Sep 30, 2021 at 2:20 AM Kees Cook <[email protected]> wrote:
>
> On Wed, Sep 29, 2021 at 11:50:35AM +0000, Yafang Shao wrote:
> > This patch increases the size of ktread's comm from 16 to 24, which is
> > the same with workqueue's, to improve this situation. After this cahnge,
> > [...]
> > Because there're only a few of kthreads, so it won't increase too much
> > memory.
>
> Even without the performance impact changes, the math here doesn't hold
> either, since using kmalloc means there are slabs being allocated to hold
> the task "comm"s now (which comes with overhead), and every task added
> a pointer to those 16 bytes (i.e. 8 more bytes on 64-bit systems). So
> this change, even if there was 0 overhead in using slabs, would be
> identical to having just raised TASK_COMM_LEN to 24. 8 byte pointer,
> 16 byte allocation == 24 bytes.
>
Right, thanks for the explanation. I missed the pointer before.
What about reusing the kthread_data() to store the the comm if the
kthread is not a kworker?
struct kthread {
...
void *data; // reuse this pointer
...
}
The logic will be something as follows,
if (kthread_is_kworker) {
store_worker_desc_into_kthread_data(); // already did in the kernel
} else {
store_comm_into_kthread_data(); // that is what we should change
}
And then we modify the proc_task_name() correspondingly.
--
Thanks
Yafang