Calling spi_write periodically in a linux driver

问题

I am writing a driver for an LCD display. According to the application note, I need to write a dummy SPI write to the command periodically to maximize its contrast. To accomplish this, I set up a timer and attempt to write the contrast-maximizing 2-byte dummy command from the timer handler.

However, something goes wrong because the spi_write function causes a complete kernel crash with the following error:

BUG: scheduling while atomic: swapper/1/0/0x00000102

Based on the following post: How to solve "BUG: scheduling while atomic: swapper /0x00000103/0, CPU#0"? in TSC2007 Driver?

"Scheduling while atomic" indicates that you've tried to sleep somewhere that you shouldn't - like within a spinlock-protected critical section or an interrupt handler.

Maybe the call to spi_write triggers some sort of sleep behavior. It would make sense to disallow sleeping here, because based on the stack trace, I see that the code is in a soft IRQ state:

[<404ec600>] (schedule_timeout) from [<404eac3c>] (wait_for_common+0x114/0x15c)
[<404eac3c>] (wait_for_common) from [<4031c7a4>] (spi_sync+0x70/0x88)
[<4031c7a4>] (spi_sync) from [<3f08a6b0>] (plt_lcd_send_toggle_comin_cmd+0x7c/0x84 [plt_lcd_spi])
[<3f08a6b0>] (plt_lcd_send_toggle_comin_cmd [plt_lcd_spi]) from [<3f08a6c4>] (plt_lcd_timer_handler+0xc/0x2c [plt_lcd_spi])
[<3f08a6c4>] (plt_lcd_timer_handler [plt_lcd_spi]) from [<40058818>] (call_timer_fn.isra.26+0x20/0x30)
[<40058818>] (call_timer_fn.isra.26) from [<40058f30>] (run_timer_softirq+0x1ec/0x21c)
[<40058f30>] (run_timer_softirq) from [<40023414>] (__do_softirq+0xe0/0x1c8)
[<40023414>] (__do_softirq) from [<400236f0>] (irq_exit+0x58/0xac)
[<400236f0>] (irq_exit) from [<4004ee4c>] (__handle_domain_irq+0x80/0xa0)
[<4004ee4c>] (__handle_domain_irq) from [<400085ac>] (gic_handle_irq+0x38/0x5c)
[<400085ac>] (gic_handle_irq) from [<40011740>] (__irq_svc+0x40/0x74)

My question is: what is the right way to implement such periodic behavior, where an SPI transaction needs to occur periodically?

The following is a summary of the timer handler (albeit with some manual modifications to make the names more generic -- I might have inserted some typos in the process)

static void lcd_timer_handler(unsigned long data)
{
    // priv is a private structure that contains private info for the 
    // driver: timer structure, timer timeout, context for the dummy command
    lcd_priv * const priv = (memlcd_priv *) data;

    unsigned char dummy[2];
    dummy[0] = get_dummy_command_code(priv);
    dummy[1] = 0; // command must be terminated by a 0.

    // This is the call that causes the failure.
    // priv->spi is a struct spi_device *
    spi_write(priv->spi, ((const void *) dummy), 2);

    // Re-arm the timer
    mod_timer(&priv->timer, jiffies + priv->timer_timeout);
}

Thanks!

EDIT: Here is what I came up with after implementing the recommendations from the answer below. Works nicely, but using delayed_work involved having to jump through a few hoops.

typedef struct lcd_priv {
    /* private stuff: */
    /* ... */

    /* workqueue stuff: */
    struct workqueue_struct * wq;
    struct delayed_work periodic_work;
} lcd_priv;


void lcd_periodic_work(struct work_struct * work_struct_ptr)
{
    /*
     * Old documentation refers to a "data" pointer, but the API
     * no longer supports it. The developer is invited to put the work_struct
     * inside what would have been pointed to by "data" and to use container_of()
     * to recover this master struct.
     * See http://lwn.net/Articles/211279/ for more info.
    */

    struct delayed_work * delayed = container_of(work_struct_ptr, struct delayed_work, work);
    lcd_priv * priv = container_of(delayed, lcd_priv, periodic_work);

    /* (prepare spi buffer in priv->spi_buf) */
    /* ... */

    /* This could be any activity that goes to sleep: */
    spi_write(priv->spi, ((const void *) &priv->spi_buf[0]), 2);

    queue_delayed_work(priv->wq, &priv->periodic_work, TOGGLE_FREQUENCY);
}

static void lcd_start_workqueue(lcd_priv * const priv) {
    priv->wq = create_singlethread_workqueue("lcd_periodic_st_wq");

    INIT_DELAYED_WORK(&priv->periodic_work, lcd_periodic_work);
    queue_delayed_work(priv->wq, &priv->periodic_work, TOGGLE_FREQUENCY);
}

static void lcd_stop_workqueue(lcd_priv * const priv) {
    destroy_workqueue(priv->wq);
}

回答1:

If look at spi_write source code, it calls spi_sync, and if look at first lines of spi_sync -> mutex_lock, so spi_write can not be run inside interrupt, and it can not be fixed via .config or sysfs.

My question is: what is the right way to implement such periodic behavior, where > an SPI transaction needs to occur periodically?

Answer depend on your hardware, how often you want send data via SPI, what latency you accept etc.

you can use spi_write inside workqueue callback, see https://www.safaribooksonline.com/library/view/understanding-the-linux/0596005652/ch04s08.html

workqueue specially designed for such kind of things (running something that can not be run in interrupt context),

also you can use spi_async to schedule write via spi. spy_async can be called inside interrupt handler.

also you move things to userspace if latency not matter, and write to SPI via spidev interface.

来源：https://stackoverflow.com/questions/37149212/calling-spi-write-periodically-in-a-linux-driver

标签

Linux

driver

atomic

periodic-task