Re: [PATCH] vt: keyboard: serialize KDGETLED with keyboard_tasklet

[Greg KH]

On Mon, May 11, 2026 at 12:43:26AM +0800, Cen Zhang wrote:
> Hi Greg,
> 
> > On Sun, May 10, 2026 at 06:22:xxPM +0200, Greg KH wrote:
> 
> > Why is that an issue?
> 
> Thanks, that's a fair point. The issue I was trying to describe is not
> that KDGETLED can race with a normal LED state transition and return
> either the old or the new value. That part is just the usual snapshot
> semantics, and the value can of course change immediately after the ioctl
> returns.
> 
> The specific case I was concerned about is the VT-switch force-update
> path. vt_set_leds_compute_shiftstate() sets vt_switch and schedules
> keyboard_tasklet, and then kbd_bh() does:
> 
>     if (vt_switch) {
>             ledstate = ~leds;
>             vt_switch = false;
>     }
> 
>     if (leds != ledstate) {
>             kbd_propagate_led_state(ledstate, leds);
>             ledstate = leds;
>     }
> 
> That first assignment is only an internal old-state sentinel to force LED
> propagation when switching VTs. It is not a committed LED state intended
> to be returned to userspace. Since KDGETLED currently returns
> ledstate & 0xff without quiescing the tasklet, it can expose that
> temporary complemented value. For example, if the real LED state is
> LED_NUM only, a read in that window can return the complemented low byte
> instead of 0x02.
> 
> > stable for "what" exactly? This is a snapshot in time, be it before or
> > after it changes is not always going to really matter here, as it can
> > change right after you "enable" the tasklet, right?
> 
> Yes, I agree. "stable snapshot" was too broad a description.
> 
> By "stable" I only meant "not while kbd_bh() is in the middle of using
> that internal force-update value". tasklet_disable() would make KDGETLED
> observe either the old value before the tasklet runs, or the committed
> value after it finishes. It does not make the value stable after
> tasklet_enable(), and the changelog should not imply that.
> 
> This is a small userspace-visible consistency issue, not a memory-safety
> problem. If this behavior is considered acceptable for KDGETLED, I can
> drop the patch; otherwise I can send a v2 with a narrower changelog.

In the past decades of use, has anyone every noticed or complained about
this consitency issue?  If not, I would recommend not worrying about it :)

thanks,

greg k-h