Am 07.04.24 um 17:32 schrieb Hans de Goede:
Hi,
On 4/6/24 8:57 PM, Armin Wolf wrote:
Am 27.03.24 um 22:45 schrieb Armin Wolf:
Since commit e2ffcda16290 ("ACPI: OSL: Allow Notify () handlers to run
on all CPUs"), the ACPI core allows multiple notify calls to be active
at the same time. This means that two instances of quickstart_notify()
running at the same time can mess which each others input sequences.
Fix this by protecting the input sequence with a mutex.
Compile-tested only.
Any thoughts on this?
I wonder if we need this at all ?
The input_event() / input_report_key() / input_sync() functions
which underpin sparse_keymap_report_event() all are safe to be called
from multiple threads at the same time AFAIK.
The only thing which can then still go "wrong" if we have
2 sparse_keymap_report_event() functions racing for the same
quickstart button and thus for the same keycode is that we may
end up with:
input_report_key(dev, keycode, 1);
input_report_key(dev, keycode, 1); /* This is a no-op */
input_sync(); /* + another input_sync() somewhere which is a no-op */
input_report_key(dev, keycode, 0);
input_report_key(dev, keycode, 0); /* This is a no-op */
input_sync(); /* + another input_sync() somewhere which is a no-op */
IOW if 2 racing notifiers hit the perfect race conditions then
only 1 key press is reported, instead of 2 which seems like
it is not a problem since arguably if the same event gets
reported twice at the exact same time it probably really
is only a single button press.
Also I think it is highly unlikely we will actually see
2 notifiers for this racing in practice.
So I don't think we need this at all. But if others feel strongly
about adding this I can still merge it... ?
Regards,
Hans
Hi,
the locking issue was originally brought up by Ilpo Jarvinen during the review of the lenovo-wmi-camera driver.
Also the race condition can cause an invalid input sequence to be emitted:
input_report_key(dev, keycode, 1);
input_sync();
input_report_key(dev, keycode, 0); // Possible invalid sequence?
input_report_key(dev, keycode, 1);
input_sync();
input_sync();
input_report_key(dev, keycode, 0);
input_sync();
I think this input sequence would be invalid, so we need the locking.
Thanks,
Armin Wolf
Fixes: afd66f2a739e ("platform/x86: Add ACPI quickstart button (PNP0C32) driver")
Signed-off-by: Armin Wolf <W_Armin@xxxxxx>
---
This applies on the branch "review-hans". Maybe we could somehow
document the concurrency rules for ACPI notify handlers?
---
drivers/platform/x86/quickstart.c | 17 +++++++++++++++++
1 file changed, 17 insertions(+)
diff --git a/drivers/platform/x86/quickstart.c b/drivers/platform/x86/quickstart.c
index ba3a7a25dda7..e40f852d42c1 100644
--- a/drivers/platform/x86/quickstart.c
+++ b/drivers/platform/x86/quickstart.c
@@ -18,6 +18,7 @@
#include <linux/input/sparse-keymap.h>
#include <linux/kernel.h>
#include <linux/module.h>
+#include <linux/mutex.h>
#include <linux/platform_device.h>
#include <linux/sysfs.h>
#include <linux/types.h>
@@ -35,6 +36,7 @@
struct quickstart_data {
struct device *dev;
+ struct mutex input_lock; /* Protects input sequence during notify */
struct input_dev *input_device;
char input_name[32];
char phys[32];
@@ -73,7 +75,10 @@ static void quickstart_notify(acpi_handle handle, u32 event, void *context)
switch (event) {
case QUICKSTART_EVENT_RUNTIME:
+ mutex_lock(&data->input_lock);
sparse_keymap_report_event(data->input_device, 0x1, 1, true);
+ mutex_unlock(&data->input_lock);
+
acpi_bus_generate_netlink_event(DRIVER_NAME, dev_name(data->dev), event, 0);
break;
default:
@@ -147,6 +152,13 @@ static void quickstart_notify_remove(void *context)
acpi_remove_notify_handler(handle, ACPI_DEVICE_NOTIFY, quickstart_notify);
}
+static void quickstart_mutex_destroy(void *data)
+{
+ struct mutex *lock = data;
+
+ mutex_destroy(lock);
+}
+
static int quickstart_probe(struct platform_device *pdev)
{
struct quickstart_data *data;
@@ -165,6 +177,11 @@ static int quickstart_probe(struct platform_device *pdev)
data->dev = &pdev->dev;
dev_set_drvdata(&pdev->dev, data);
+ mutex_init(&data->input_lock);
+ ret = devm_add_action_or_reset(&pdev->dev, quickstart_mutex_destroy, &data->input_lock);
+ if (ret < 0)
+ return ret;
+
/* We have to initialize the device wakeup before evaluating GHID because
* doing so will notify the device if the button was used to wake the machine
* from S5.
--
2.39.2