RE: [PATCHv3 1/3] gpio: sch: Consolidate similar algorithms

From: Chang, Rebecca Swee Fun
Date: Thu Oct 23 2014 - 03:26:13 EST


Thanks for the review comments. Please check my reply below.

> -----Original Message-----
> From: Alexandre Courbot [mailto:gnurou@xxxxxxxxx]
> Sent: 17 October, 2014 4:44 PM
> To: Chang, Rebecca Swee Fun
> Cc: Linus Walleij; Westerberg, Mika; GPIO Subsystem Mailing List; Linux Kernel
> Mailing List; Denis Turischev
> Subject: Re: [PATCHv3 1/3] gpio: sch: Consolidate similar algorithms
>
> On Thu, Oct 16, 2014 at 12:51 PM, Chang Rebecca Swee Fun
> <rebecca.swee.fun.chang@xxxxxxxxx> wrote:
> > Consolidating similar algorithms into common functions to make GPIO
> > SCH simpler and manageable.
> >
> > Signed-off-by: Chang Rebecca Swee Fun
> > <rebecca.swee.fun.chang@xxxxxxxxx>
> > ---
> > drivers/gpio/gpio-sch.c | 95 ++++++++++++++++++++++++++--------------------
> -
> > 1 file changed, 53 insertions(+), 42 deletions(-)
> >
> > diff --git a/drivers/gpio/gpio-sch.c b/drivers/gpio/gpio-sch.c index
> > 99720c8..6e89be9 100644
> > --- a/drivers/gpio/gpio-sch.c
> > +++ b/drivers/gpio/gpio-sch.c
> > @@ -63,94 +63,105 @@ static unsigned sch_gpio_bit(struct sch_gpio *sch,
> unsigned gpio)
> > return gpio % 8;
> > }
> >
> > -static void sch_gpio_enable(struct sch_gpio *sch, unsigned gpio)
> > +static void sch_gpio_register_set(struct sch_gpio *sch, unsigned gpio,
> > + unsigned reg)
> > {
> > unsigned short offset, bit;
> > u8 enable;
> >
> > spin_lock(&sch->lock);
>
> ...
>
> >
> > - offset = sch_gpio_offset(sch, gpio, GEN);
> > + offset = sch_gpio_offset(sch, gpio, reg);
> > bit = sch_gpio_bit(sch, gpio);
> >
> > enable = inb(sch->iobase + offset);
>
> I see identical blocks of code in sch_gpio_register_clear(),
> sch_gpio_reg_get() and sch_gpio_reg_set() (and possibly other functions?).
> Maybe this could be factorized into an inline function that would return the
> "enable" variable?
>
> Also, "enable" looks like the wrong name here. The exact same result is later
> called "disable" and "curr_dirs" later.

Referring to your comments below, this will be get rid of after the sch_gpio_register_set() and sch_gpio_register_clear() being replaced by sch_gpio_reg_set(gc, gpio, reg, 1) and sch_gpio_reg_set(gc, gpio, reg, 0).
There will be less identical block of "offset", "bit", and "enable" nor "disable". So there is no need to factorize the identical blocks into inline function.

>
> > - if (!(enable & (1 << bit)))
> > - outb(enable | (1 << bit), sch->iobase + offset);
> > + if (!(enable & BIT(bit)))
> > + outb(enable | BIT(bit), sch->iobase + offset);
> >
> > spin_unlock(&sch->lock);
> > }
> >
> > -static int sch_gpio_direction_in(struct gpio_chip *gc, unsigned
> > gpio_num)
> > +static void sch_gpio_register_clear(struct sch_gpio *sch, unsigned gpio,
> > + unsigned reg)
> > {
> > - struct sch_gpio *sch = to_sch_gpio(gc);
> > - u8 curr_dirs;
> > unsigned short offset, bit;
> > + u8 disable;
> >
> > spin_lock(&sch->lock);
> >
> > - offset = sch_gpio_offset(sch, gpio_num, GIO);
> > - bit = sch_gpio_bit(sch, gpio_num);
> > -
> > - curr_dirs = inb(sch->iobase + offset);
> > + offset = sch_gpio_offset(sch, gpio, reg);
> > + bit = sch_gpio_bit(sch, gpio);
> >
> > - if (!(curr_dirs & (1 << bit)))
> > - outb(curr_dirs | (1 << bit), sch->iobase + offset);
> > + disable = inb(sch->iobase + offset);
> > + if (disable & BIT(bit))
> > + outb(disable & ~BIT(bit), sch->iobase + offset);
> >
> > spin_unlock(&sch->lock);
> > - return 0;
> > }
> >
> > -static int sch_gpio_get(struct gpio_chip *gc, unsigned gpio_num)
> > +static int sch_gpio_reg_get(struct gpio_chip *gc, unsigned gpio,
> > +unsigned reg)
> > {
> > struct sch_gpio *sch = to_sch_gpio(gc);
> > - int res;
> > unsigned short offset, bit;
> > + u8 curr_dirs;
> >
> > - offset = sch_gpio_offset(sch, gpio_num, GLV);
> > - bit = sch_gpio_bit(sch, gpio_num);
> > + offset = sch_gpio_offset(sch, gpio, reg);
> > + bit = sch_gpio_bit(sch, gpio);
> >
> > - res = !!(inb(sch->iobase + offset) & (1 << bit));
> > + curr_dirs = !!(inb(sch->iobase + offset) & BIT(bit));
> >
> > - return res;
> > + return curr_dirs;
> > }
> >
> > -static void sch_gpio_set(struct gpio_chip *gc, unsigned gpio_num, int
> > val)
> > +static void sch_gpio_reg_set(struct gpio_chip *gc, unsigned gpio, unsigned
> reg,
> > + int val)
> > {
> > struct sch_gpio *sch = to_sch_gpio(gc);
> > - u8 curr_vals;
> > unsigned short offset, bit;
> > + u8 curr_dirs;
> >
> > - spin_lock(&sch->lock);
> > -
> > - offset = sch_gpio_offset(sch, gpio_num, GLV);
> > - bit = sch_gpio_bit(sch, gpio_num);
> > + offset = sch_gpio_offset(sch, gpio, reg);
> > + bit = sch_gpio_bit(sch, gpio);
> >
> > - curr_vals = inb(sch->iobase + offset);
> > + curr_dirs = inb(sch->iobase + offset);
> >
> > if (val)
> > - outb(curr_vals | (1 << bit), sch->iobase + offset);
> > + outb(curr_dirs | BIT(bit), sch->iobase + offset);
> > else
> > - outb((curr_vals & ~(1 << bit)), sch->iobase + offset);
> > + outb((curr_dirs & ~BIT(bit)), sch->iobase + offset); }
>
> Mmmm this really looks like sch_gpio_register_set() and
> sch_gpio_register_clear() could just call sch_gpio_reg_set() right after
> acquiring the spinlock. Also the names are very similar and it is not clear why
> you need two different functions here. Couldn't you call sch_gpio_reg_set(gc,
> gpio, reg, 1) in place of
> sch_gpio_register_set() and sch_gpio_reg_set(gc, gpio, reg, 0) for
> sch_gpio_register_clear()? You may need to acquire the spinlock before some
> call sites, but that's preferable to having very similar functions bearing a very
> similar name IMHO.

Thanks for pointing that out. I think I can replaced all sch_gpio_register_set() and sch_gpio_register_clear() with sch_gpio_reg_set(gc, gpio, reg, 0/1).
Regarding the spinlock, in fact, I have tested using the GPIO driver through sysfs. I didn't encounter any deadlock issue, but the double locking is really an issue.
This double lock problem can also be fix by calling sch_gpio_reg_set(gc, gpio, reg, 0/1) in place of sch_gpio_register_set() and sch_gpio_register_clear().

>
> >
> > +static int sch_gpio_direction_in(struct gpio_chip *gc, unsigned
> > +gpio_num) {
> > + struct sch_gpio *sch = to_sch_gpio(gc);
> > +
> > + spin_lock(&sch->lock);
> > + sch_gpio_register_set(sch, gpio_num, GIO);
>
> So here you are acquiring sch->lock, then calling
> sch_gpio_register_set() which will try to acquire the same spinlock first thing.
> Wouldn't that deadlock? Have you tested changing the direction of a GPIO? This
> again speaks in favor or reducing the number of similar functions in this file.
>
> Unless I misunderstood something there are still some issues that make this file
> harder to understand than it should, and which can sometimes bork the system
> altogether. It is a good idea to cleanup this file, but please try to go one step
> further - this should simplify locking and ultimately get rid of the locking issues.
>
> And hopefully I will also take less time to review v4. :P

Thanks for the review. I will take note on the locking part and resend later or next week. Appreciate your review comments. Thank you!

Rebecca