On Wed, Sep 04, 2024 at 08:41:30PM +0200, Javier Carrasco wrote:
> On 04/09/2024 20:21, Dmitry Torokhov wrote:
> > Hi Javier,
> >
> > On Wed, Sep 04, 2024 at 03:53:40PM +0200, Javier Carrasco wrote:
> >> On 04/09/2024 06:47, Dmitry Torokhov wrote:
> >>> Using guard notation makes the code more compact and error handling
> >>> more robust by ensuring that mutexes are released in all code paths
> >>> when control leaves critical section.
> >>>
> >>> Signed-off-by: Dmitry Torokhov <dmitry.torokhov@xxxxxxxxx>
> >>> ---
> >>> drivers/input/misc/iqs269a.c | 46 +++++++++++++-----------------------
> >>> 1 file changed, 16 insertions(+), 30 deletions(-)
> >>>
> >>> diff --git a/drivers/input/misc/iqs269a.c b/drivers/input/misc/iqs269a.c
> >>> index 843f8a3f3410..c34d847fa4af 100644
> >>> --- a/drivers/input/misc/iqs269a.c
> >>> +++ b/drivers/input/misc/iqs269a.c
> >>
> >> ...
> >>
> >>> @@ -453,9 +449,9 @@ static int iqs269_ati_base_get(struct iqs269_private *iqs269,
> >>> if (ch_num >= IQS269_NUM_CH)
> >>> return -EINVAL;
> >>>
> >>> - mutex_lock(&iqs269->lock);
> >>> + guard(mutex)(&iqs269->lock);
> >>> +
> >>> engine_b = be16_to_cpu(ch_reg[ch_num].engine_b);
> >>
> >> maybe scoped_guard() to keep the scope of the mutex as it used to be?
> >
> > Thank you for looking over patches.
> >
> > It is just a few computations extra, so I decided not to use
> > scoped_guard(). Note that original code was forced to release mutex
> > early to avoid having to unlock it in all switch branches.
> >
> >>
> >>> - mutex_unlock(&iqs269->lock);
> >>>
> >>> switch (engine_b & IQS269_CHx_ENG_B_ATI_BASE_MASK) {
> >>> case IQS269_CHx_ENG_B_ATI_BASE_75:
> >>> @@ -491,7 +487,7 @@ static int iqs269_ati_target_set(struct iqs269_private *iqs269,
> >>> if (target > IQS269_CHx_ENG_B_ATI_TARGET_MAX)
> >>> return -EINVAL;
> >>>
> >>> - mutex_lock(&iqs269->lock);
> >>> + guard(mutex)(&iqs269->lock);
> >>>
> >>> engine_b = be16_to_cpu(ch_reg[ch_num].engine_b);
> >>>
> >>> @@ -501,8 +497,6 @@ static int iqs269_ati_target_set(struct iqs269_private *iqs269,
> >>> ch_reg[ch_num].engine_b = cpu_to_be16(engine_b);
> >>> iqs269->ati_current = false;
> >>>
> >>> - mutex_unlock(&iqs269->lock);
> >>> -
> >>> return 0;
> >>> }
> >>>
> >>> @@ -515,10 +509,9 @@ static int iqs269_ati_target_get(struct iqs269_private *iqs269,
> >>> if (ch_num >= IQS269_NUM_CH)
> >>> return -EINVAL;
> >>>
> >>> - mutex_lock(&iqs269->lock);
> >>> - engine_b = be16_to_cpu(ch_reg[ch_num].engine_b);
> >>> - mutex_unlock(&iqs269->lock);
> >>> + guard(mutex)(&iqs269->lock);
> >>
> >> same here?
> >>
> >>>
> >>> + engine_b = be16_to_cpu(ch_reg[ch_num].engine_b);
> >>> *target = (engine_b & IQS269_CHx_ENG_B_ATI_TARGET_MASK) * 32;
> >
> > Same here, calculating the line above will take no time at all...
> >
> > Thanks.
> >
>
> As you pointed out, in reality the extra locked instructions will not
> make any difference. But as the conversion added instructions to be
> locked by the mutex without mentioning it, I thought it should be either
> left as it used to be with scoped_guard(), or explicitly mentioned in
> the description.
>
> No strong feelings against it, but out of curiosity, why would you
> rather use guard()? I think scoped_guard() is a better way to
> self-document what has to be accessed via mutex, and what not.
Simply less indentation ;) and in this driver uniformity with for example
iqs269_ati_target_set() where critical section does indeed extend to the
whole function.
Not super strong arguments either.
--
Dmitry
