leds: avoid flush_work in atomic context

It turns out that various triggers use led_blink_setup() from atomic
context, so we can't do a flush_work there. Flush is still needed for
slow LEDs, but we can move it to sysfs code where it is safe.

WARNING: inconsistent lock state
5.2.0-rc1 #1 Tainted: G W
--------------------------------
inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W} usage.
swapper/1/0 [HC0[0]:SC1[1]:HE1:SE0] takes:
000000006e30541b
((work_completion)(&led_cdev->set_brightness_work)){+.?.}, at:
+__flush_work+0x3b/0x38a
{SOFTIRQ-ON-W} state was registered at:
lock_acquire+0x146/0x1a1
__flush_work+0x5b/0x38a
flush_work+0xb/0xd
led_blink_setup+0x1e/0xd3
led_blink_set+0x3f/0x44
tpt_trig_timer+0xdb/0x106
ieee80211_mod_tpt_led_trig+0xed/0x112

Fixes: 0db37915d912 ("leds: avoid races with workqueue")
Signed-off-by: Pavel Machek <[e-mail]>
Tested-by: Hugh Dickins <[e-mail]>
Signed-off-by: Jacek Anaszewski <[e-mail]>

drivers/leds/led-core.c

···
 		     unsigned long *delay_on,
 		     unsigned long *delay_off)
 {
-	/*
-	 * If "set brightness to 0" is pending in workqueue, we don't
-	 * want that to be reordered after blink_set()
-	 */
-	flush_work(&led_cdev->set_brightness_work);
 	if (!test_bit(LED_BLINK_ONESHOT, &led_cdev->work_flags) &&
 	    led_cdev->blink_set &&
 	    !led_cdev->blink_set(led_cdev, delay_on, delay_off))

drivers/leds/trigger/ledtrig-timer.c

···
 		led_cdev->flags &= ~LED_INIT_DEFAULT_TRIGGER;
 	}
 
+	/*
+	 * If "set brightness to 0" is pending in workqueue, we don't
+	 * want that to be reordered after blink_set()
+	 */
+	flush_work(&led_cdev->set_brightness_work);
 	led_blink_set(led_cdev, &led_cdev->blink_delay_on,
 		      &led_cdev->blink_delay_off);