On (03/07/19 13:06), John Ogness wrote:
> On 2019-03-04, Sergey Senozhatsky <sergey.senozhatsky.work@xxxxxxxxx> wrote:
> >> + /* the printk kthread never exits */
> >> + for (;;) {
> >> + ret = prb_iter_wait_next(&iter, buf,
> >> + PRINTK_RECORD_MAX, &master_seq);
> >> + if (ret == -ERESTARTSYS) {
> >> + continue;
> >> + } else if (ret < 0) {
> >> + /* iterator invalid, start over */
> >> + prb_iter_init(&iter, &printk_rb, NULL);
> >> + continue;
> >> + }
> >> +
> >> + msg = (struct printk_log *)buf;
> >> + format_text(msg, master_seq, ext_text, &ext_len, text,
> >> + &len, printk_time);
> >> +
> >> + console_lock();
> >> + if (len > 0 || ext_len > 0) {
> >> + call_console_drivers(ext_text, ext_len, text, len);
> >> + boot_delay_msec(msg->level);
> >> + printk_delay();
> >> + }
> >> + console_unlock();
> >> + }
> >
> > This, theoretically, creates a whole new world of possibilities for
> > console drivers. Now they can do GFP_KERNEL allocations and stall
> > printk_kthread during OOM; or they can explicitly reschedule from
> > ->write() callback (via console_conditional_schedule()) because
> > console_lock() sets console_may_schedule.
>
> This was the intention.
This can stall the entire printing pipeline
OOM -> printk_kthread() -> console_lock() -> con_foo() -> kmalloc(GFP_KERNEL) -> OOM
> Although, as I mentioned in a previous response[0], perhaps we should
> not loosen the requirements on write().
Right. Console drivers better stay restricted; very restricted.
> It is exactly that disable_preempt() that is so harmful for realtime tasks.
I'll reply in another email (later today, or tomorrow).
-ss
