Re: [PATCH v9 04/20] gpio: uapi: define uAPI v2

From: Arnd Bergmann
Date: Tue Sep 22 2020 - 03:41:34 EST


On Tue, Sep 22, 2020 at 4:34 AM Kent Gibson <warthog618@xxxxxxxxx> wrote:
> +/**
> + * struct gpio_v2_line_attribute - a configurable attribute of a line
> + * @id: attribute identifier with value from &enum gpio_v2_line_attr_id
> + * @padding: reserved for future use and must be zero filled
> + * @flags: if id is GPIO_V2_LINE_ATTR_ID_FLAGS, the flags for the GPIO
> + * line, with values from enum gpio_v2_line_flag, such as
> + * GPIO_V2_LINE_FLAG_ACTIVE_LOW, GPIO_V2_LINE_FLAG_OUTPUT etc, OR:ed
> + * together. This overrides the default flags contained in the &struct
> + * gpio_v2_line_config for the associated line.
> + * @values: if id is GPIO_V2_LINE_ATTR_ID_OUTPUT_VALUES, a bitmap
> + * containing the values to which the lines will be set, with each bit
> + * number corresponding to the index into &struct
> + * gpio_v2_line_request.offsets.
> + * @debounce_period_us: if id is GPIO_V2_LINE_ATTR_ID_DEBOUNCE, the desired
> + * debounce period, in microseconds
> + */
> +struct gpio_v2_line_attribute {
> + __u32 id;
> + __u32 padding;
> + union {
> + __aligned_u64 flags;
> + __aligned_u64 values;
> + __u32 debounce_period_us;
> + };
> +};

Having different-sized members in the union makes it hard for
something like strace to print the contents. How about just making
them all __aligned_u64 even when 32 bits are sufficient?

> +struct gpio_v2_line_request {
> + __u32 offsets[GPIO_V2_LINES_MAX];
> + char consumer[GPIO_MAX_NAME_SIZE];
> + struct gpio_v2_line_config config;
> + __u32 num_lines;
> + __u32 event_buffer_size;
> + /* Pad to fill implicit padding and reserve space for future use. */
> + __u32 padding[5];
> + __s32 fd;
> +};

> +struct gpio_v2_line_info {
> + char name[GPIO_MAX_NAME_SIZE];
> + char consumer[GPIO_MAX_NAME_SIZE];
> + __u32 offset;
> + __u32 num_attrs;
> + __aligned_u64 flags;
> + struct gpio_v2_line_attribute attrs[GPIO_V2_LINE_NUM_ATTRS_MAX];
> + /* Space reserved for future use. */
> + __u32 padding[4];
> +};

These are both several hundred bytes long, requiring a lot of data
to be copied to the stack and take up space there. I see this is not
actually much different for the v1 API, but I wonder if there has been
any analysis of whether this has a noticeable effect on application
runtime.

Arnd