2023-05-02 01:03:05

by Dmitry Torokhov

[permalink] [raw]
Subject: [PATCH] Input: avoid calling input_set_abs_val() in the event handling core

input_abs_set_val() can nominally call input_alloc_absinfo() which may
allocate memory with GFP_KERNEL flag. This does not happen when
input_abs_set_val() is called by the input core to set current MT slot when
handling a new input event, but it trips certain static analyzers.

Rearrange the code to access the relevant structures directly.

Reported-by: Teng Qi <[email protected]>
Signed-off-by: Dmitry Torokhov <[email protected]>
---
drivers/input/input.c | 6 ++++--
1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/drivers/input/input.c b/drivers/input/input.c
index 37e876d45eb9..f791d14ecf23 100644
--- a/drivers/input/input.c
+++ b/drivers/input/input.c
@@ -190,6 +190,7 @@ static int input_handle_abs_event(struct input_dev *dev,
unsigned int code, int *pval)
{
struct input_mt *mt = dev->mt;
+ bool is_new_slot = false;
bool is_mt_event;
int *pold;

@@ -210,6 +211,7 @@ static int input_handle_abs_event(struct input_dev *dev,
pold = &dev->absinfo[code].value;
} else if (mt) {
pold = &mt->slots[mt->slot].abs[code - ABS_MT_FIRST];
+ is_new_slot = mt->slot != dev->absinfo[ABS_MT_SLOT].value;
} else {
/*
* Bypass filtering for multi-touch events when
@@ -228,8 +230,8 @@ static int input_handle_abs_event(struct input_dev *dev,
}

/* Flush pending "slot" event */
- if (is_mt_event && mt && mt->slot != input_abs_get_val(dev, ABS_MT_SLOT)) {
- input_abs_set_val(dev, ABS_MT_SLOT, mt->slot);
+ if (is_new_slot) {
+ dev->absinfo[ABS_MT_SLOT].value = mt->slot;
return INPUT_PASS_TO_HANDLERS | INPUT_SLOT;
}

--
2.40.1.495.gc816e09b53d-goog


--
Dmitry


2023-05-02 06:19:53

by Peter Hutterer

[permalink] [raw]
Subject: Re: [PATCH] Input: avoid calling input_set_abs_val() in the event handling core

On Mon, May 01, 2023 at 06:01:19PM -0700, Dmitry Torokhov wrote:
> input_abs_set_val() can nominally call input_alloc_absinfo() which may
> allocate memory with GFP_KERNEL flag. This does not happen when
> input_abs_set_val() is called by the input core to set current MT slot when
> handling a new input event, but it trips certain static analyzers.
>
> Rearrange the code to access the relevant structures directly.
>
> Reported-by: Teng Qi <[email protected]>
> Signed-off-by: Dmitry Torokhov <[email protected]>

Reviewed-by: Peter Hutterer <[email protected]>

If git grep is to be believed, this is the only use of
input_abs_set_val. Maybe removing that function is an option?

Are there external drivers that rely on this (instead of sending an
event)?

Cheers,
Peter

> ---
> drivers/input/input.c | 6 ++++--
> 1 file changed, 4 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/input/input.c b/drivers/input/input.c
> index 37e876d45eb9..f791d14ecf23 100644
> --- a/drivers/input/input.c
> +++ b/drivers/input/input.c
> @@ -190,6 +190,7 @@ static int input_handle_abs_event(struct input_dev *dev,
> unsigned int code, int *pval)
> {
> struct input_mt *mt = dev->mt;
> + bool is_new_slot = false;
> bool is_mt_event;
> int *pold;
>
> @@ -210,6 +211,7 @@ static int input_handle_abs_event(struct input_dev *dev,
> pold = &dev->absinfo[code].value;
> } else if (mt) {
> pold = &mt->slots[mt->slot].abs[code - ABS_MT_FIRST];
> + is_new_slot = mt->slot != dev->absinfo[ABS_MT_SLOT].value;
> } else {
> /*
> * Bypass filtering for multi-touch events when
> @@ -228,8 +230,8 @@ static int input_handle_abs_event(struct input_dev *dev,
> }
>
> /* Flush pending "slot" event */
> - if (is_mt_event && mt && mt->slot != input_abs_get_val(dev, ABS_MT_SLOT)) {
> - input_abs_set_val(dev, ABS_MT_SLOT, mt->slot);
> + if (is_new_slot) {
> + dev->absinfo[ABS_MT_SLOT].value = mt->slot;
> return INPUT_PASS_TO_HANDLERS | INPUT_SLOT;
> }
>
> --
> 2.40.1.495.gc816e09b53d-goog
>
>
> --
> Dmitry

2023-05-02 20:28:14

by Dmitry Torokhov

[permalink] [raw]
Subject: Re: [PATCH] Input: avoid calling input_set_abs_val() in the event handling core

On Tue, May 02, 2023 at 04:05:31PM +1000, Peter Hutterer wrote:
> On Mon, May 01, 2023 at 06:01:19PM -0700, Dmitry Torokhov wrote:
> > input_abs_set_val() can nominally call input_alloc_absinfo() which may
> > allocate memory with GFP_KERNEL flag. This does not happen when
> > input_abs_set_val() is called by the input core to set current MT slot when
> > handling a new input event, but it trips certain static analyzers.
> >
> > Rearrange the code to access the relevant structures directly.
> >
> > Reported-by: Teng Qi <[email protected]>
> > Signed-off-by: Dmitry Torokhov <[email protected]>
>
> Reviewed-by: Peter Hutterer <[email protected]>

Thanks for the review.

>
> If git grep is to be believed, this is the only use of
> input_abs_set_val. Maybe removing that function is an option?

We generate input_abs_[get|set]_<item>() accessors and setters, and
there is a couple of users of input_abs_get_val() in the code, so
dropping just input_abs_set_val() is kind of hard. It might be useful
elsewhere although I have no idea if anyone is actually using it
anywhere...

Thanks.

--
Dmitry