On Wed, Mar 21, 2018 at 07:45:29PM +0200, Andy Shevchenko wrote:

On Sat, Mar 17, 2018 at 5:50 PM, William Breathitt Gray
[off-list ref] wrote:

quoted

The ACCES I/O PCI-IDIO-16 series of devices provides 16
optically-isolated digital inputs accessed via two 8-bit ports. Since
eight input lines are acquired on a single port input read, the
PCI-IDIO-16 GPIO driver may improve multiple input reads by utilizing a
get_multiple callback. This patch implements the
idio_16_gpio_get_multiple function which serves as the respective
get_multiple callback.

quoted

+static int idio_16_gpio_get_multiple(struct gpio_chip *chip,
+       unsigned long *mask, unsigned long *bits)
+{
+       struct idio_16_gpio *const idio16gpio = gpiochip_get_data(chip);
+       size_t i;
+       const unsigned int gpio_reg_size = 8;
+       unsigned int bits_offset;
+       size_t word_index;
+       unsigned int word_offset;
+       unsigned long word_mask;

quoted

+       const unsigned long port_mask = GENMASK(gpio_reg_size, 0);

gpio_reg_size - 1?

Oops, looks like I made an off-by-one error here so I'll make sure to
fix that up.

Though I would prefer not to have that variable at all, just use 8 or
7 respectively.

This device is simple enough that throughout this function I could
inline gpio_reg_size and port_mask to 8 and 0xFF respectively, but I
would like to keep the code generic enough for reuse in other drivers.
In addition, I believe the variable names help keep the intention of the
code clear, so I'll stick with dedicated const variables for now if
there are no other objections.

quoted

+       unsigned long port_state;

quoted

+       u8 __iomem ports[] = {
+               idio16gpio->reg->out0_7, idio16gpio->reg->out8_15,

quoted

+               idio16gpio->reg->in0_7, idio16gpio->reg->in8_15

I would leave comma even here.

Will do.

quoted

+       };

quoted

+       /* get bits are evaluated a gpio port register at a time */
+       for (i = 0; i < ARRAY_SIZE(ports); i++) {
+               /* gpio offset in bits array */
+               bits_offset = i * gpio_reg_size;
+
+               /* word index for bits array */
+               word_index = BIT_WORD(bits_offset);
+
+               /* gpio offset within current word of bits array */
+               word_offset = bits_offset % BITS_PER_LONG;
+
+               /* mask of get bits for current gpio within current word */
+               word_mask = mask[word_index] & (port_mask << word_offset);
+               if (!word_mask) {
+                       /* no get bits in this port so skip to next one */
+                       continue;
+               }
+
+               /* read bits from current gpio port */
+               port_state = ioread8(ports + i);
+
+               /* store acquired bits at respective bits array offset */
+               bits[word_index] |= port_state << word_offset;
+       }

I would propose to do other way around, i.e.
read all ports to the bitmap array and call bitmap_and() after.

Further optimization can be something like introduction of generic

bitmap_copy_uXX_off(unsigned long *dst, u8 src, unsigned int offset);

It can be done using macros, though it's another story not quite
related to the topic.

Port I/O is significantly more costly to perform than the bitmask
evaluations for each port. Despite the increased complexity of the loop
logic, I believe the latency improvements of skipping unnecessary I/O
port reads are worth the trouble.

I do like the idea of a bitmap_copy_uXX_off macro as that could be quite
useful in general. Even if not for this particular patchset, I would be
interested in seeing that functionality added to the bitmap API. Perhaps
I might implement it as a standlone patch when I have some free time.

William Breathitt Gray

quoted

+}

-- 
With Best Regards,
Andy Shevchenko