leds: Fix oops about sleeping in led_trigger_blink()
led_trigger_blink() calls led_blink_set() from a RCU read-side critical section so led_blink_set() must not sleep. Note sleeping was not allowed before the switch to RCU either because a spinlock was held before. led_blink_set() does not sleep when sw-blinking is used, but many LED controller drivers with hw blink support have a blink_set function which may sleep, leading to an oops like this one: [ 832.605062] ------------[ cut here ]------------ [ 832.605085] Voluntary context switch within RCU read-side critical section! [ 832.605119] WARNING: CPU: 2 PID: 370 at kernel/rcu/tree_plugin.h:318 rcu_note_context_switch+0x4ee/0x690 <snip> [ 832.606453] Call Trace: [ 832.606466] <TASK> [ 832.606487] __schedule+0x9f/0x1480 [ 832.606527] schedule+0x5d/0xe0 [ 832.606549] schedule_timeout+0x79/0x140 [ 832.606572] ? __pfx_process_timeout+0x10/0x10 [ 832.606599] wait_for_completion_timeout+0x6f/0x140 [ 832.606627] i2c_dw_xfer+0x101/0x460 [ 832.606659] ? psi_group_change+0x168/0x400 [ 832.606680] __i2c_transfer+0x172/0x6d0 [ 832.606709] i2c_smbus_xfer_emulated+0x27d/0x9c0 [ 832.606732] ? __schedule+0x430/0x1480 [ 832.606753] ? preempt_count_add+0x6a/0xa0 [ 832.606778] ? get_nohz_timer_target+0x18/0x190 [ 832.606796] ? lock_timer_base+0x61/0x80 [ 832.606817] ? preempt_count_add+0x6a/0xa0 [ 832.606842] __i2c_smbus_xfer+0xa2/0x3f0 [ 832.606862] i2c_smbus_xfer+0x66/0xf0 [ 832.606882] i2c_smbus_read_byte_data+0x41/0x70 [ 832.606901] ? _raw_spin_unlock_irqrestore+0x23/0x40 [ 832.606922] ? __pm_runtime_suspend+0x46/0xc0 [ 832.606946] cht_wc_byte_reg_read+0x2e/0x60 [ 832.606972] _regmap_read+0x5c/0x120 [ 832.606997] _regmap_update_bits+0x96/0xc0 [ 832.607023] regmap_update_bits_base+0x5b/0x90 [ 832.607053] cht_wc_leds_brightness_get+0x412/0x910 [leds_cht_wcove] [ 832.607094] led_blink_setup+0x28/0x100 [ 832.607119] led_trigger_blink+0x40/0x70 [ 832.607145] power_supply_update_leds+0x1b7/0x1c0 [ 832.607174] power_supply_changed_work+0x67/0xe0 [ 832.607198] process_one_work+0x1c8/0x3c0 [ 832.607222] worker_thread+0x4d/0x380 [ 832.607243] ? __pfx_worker_thread+0x10/0x10 [ 832.607258] kthread+0xe9/0x110 [ 832.607279] ? __pfx_kthread+0x10/0x10 [ 832.607300] ret_from_fork+0x2c/0x50 [ 832.607337] </TASK> [ 832.607344] ---[ end trace 0000000000000000 ]--- Add a new led_blink_set_nosleep() function which defers the actual led_blink_set() call to a workqueue when necessary to fix this. This also fixes an existing race where a pending led_set_brightness() has been deferred to set_brightness_work and might then race with a later led_cdev->blink_set() call. Note this race is only an issue with triggers mixing led_trigger_event() and led_trigger_blink() calls, sysfs API calls and led_trigger_blink_oneshot() are not affected. Note rather then adding a separate blink_set_blocking callback this uses the presence of the already existing brightness_set_blocking callback to detect if the blinking call should be deferred to set_brightness_work. Signed-off-by: Hans de Goede <hdegoede@redhat.com> Reviewed-by: Jacek Anaszewski <jacek.anaszewski@gmail.com> Tested-by: Yauhen Kharuzhy <jekhor@gmail.com> Link: https://lore.kernel.org/r/20230510162234.291439-4-hdegoede@redhat.comSigned-off-by: Lee Jones <lee@kernel.org>
Showing
Please register or sign in to comment