Re: [PATCH v2 21/72] ncr5380: Sleep when polling, if possible

From: Geert Uytterhoeven
Date: Sun Dec 06 2015 - 05:18:12 EST


Hi Finn,

On Sun, Dec 6, 2015 at 2:31 AM, Finn Thain <fthain@xxxxxxxxxxxxxxxxxxx> wrote:
> When in process context, sleep during polling if doing so won't add
> significant latency. In interrupt context or if the lock is held, poll
> briefly then give up. Keep both core drivers in sync.
>
> Calibrate busy-wait iterations to allow for variation in chip register
> access times between different 5380 hardware implementations.
>
> Signed-off-by: Finn Thain <fthain@xxxxxxxxxxxxxxxxxxx>
>
> ---
>
> Changed since v1:
> - Don't rely on loops_per_jiffy to estimate register access speed, measure
> it instead.
>
> ---
> drivers/scsi/NCR5380.c | 77 +++++++++++++++++++++++++------------------
> drivers/scsi/NCR5380.h | 1
> drivers/scsi/atari_NCR5380.c | 59 ++++++++++++++++++++------------
> 3 files changed, 84 insertions(+), 53 deletions(-)
>
> Index: linux/drivers/scsi/NCR5380.c
> ===================================================================
> --- linux.orig/drivers/scsi/NCR5380.c 2015-12-06 12:29:51.000000000 +1100
> +++ linux/drivers/scsi/NCR5380.c 2015-12-06 12:29:54.000000000 +1100
> @@ -293,44 +293,48 @@ static inline void initialize_SCp(struct
> }
>
> /**
> - * NCR5380_poll_politely - wait for NCR5380 status bits
> - * @instance: controller to poll
> - * @reg: 5380 register to poll
> - * @bit: Bitmask to check
> - * @val: Value required to exit
> - *
> - * Polls the NCR5380 in a reasonably efficient manner waiting for
> - * an event to occur, after a short quick poll we begin giving the
> - * CPU back in non IRQ contexts
> + * NCR5380_poll_politely - wait for chip register value
> + * @instance: controller to poll
> + * @reg: 5380 register to poll
> + * @bit: Bitmask to check
> + * @val: Value required to exit
> + * @wait: Time-out in jiffies
> + *
> + * Polls the chip in a reasonably efficient manner waiting for an
> + * event to occur. After a short quick poll we begin to yield the CPU
> + * (if possible). In irq contexts the time-out is arbitrarily limited.
> + * Callers may hold locks as long as they are held in irq mode.
> *
> - * Returns the value of the register or a negative error code.
> + * Returns 0 if event occurred otherwise -ETIMEDOUT.
> */
> -
> -static int NCR5380_poll_politely(struct Scsi_Host *instance, int reg, int bit, int val, int t)
> +
> +static int NCR5380_poll_politely(struct Scsi_Host *instance,
> + int reg, int bit, int val, int wait)
> {
> - int n = 500; /* At about 8uS a cycle for the cpu access */
> - unsigned long end = jiffies + t;
> - int r;
> -
> - while( n-- > 0)
> - {
> - r = NCR5380_read(reg);
> - if((r & bit) == val)
> + struct NCR5380_hostdata *hostdata = shost_priv(instance);
> + unsigned long deadline = jiffies + wait;
> + unsigned long n;
> +
> + /* Busy-wait for up to 10 ms */
> + n = min(10000U, jiffies_to_usecs(wait));
> + n *= hostdata->accesses_per_ms;
> + n /= 1000;
> + do {
> + if ((NCR5380_read(reg) & bit) == val)
> return 0;
> cpu_relax();
> - }

> @@ -773,6 +777,7 @@ static int NCR5380_init(struct Scsi_Host
> {
> int i;
> struct NCR5380_hostdata *hostdata = (struct NCR5380_hostdata *) instance->hostdata;
> + unsigned long deadline;
>
> if(in_interrupt())
> printk(KERN_ERR "NCR5380_init called with interrupts off!\n");
> @@ -812,6 +817,16 @@ static int NCR5380_init(struct Scsi_Host
> NCR5380_write(MODE_REG, MR_BASE);
> NCR5380_write(TARGET_COMMAND_REG, 0);
> NCR5380_write(SELECT_ENABLE_REG, 0);
> +
> + /* Calibrate register polling loop */
> + i = 0;
> + deadline = jiffies + msecs_to_jiffies(100) + 1;
> + do {
> + NCR5380_read(STATUS_REG);
> + ++i;
> + } while (time_is_after_jiffies(deadline));
> + hostdata->accesses_per_ms = i / 100;

As the caller of NCR5380_poll_politely() passes a timeout value in jiffies,
calculations may become simpler if you store the number of accesses per jiffy
instead of per ms.

Unlike the historical calibrating-delay-loop code, you don't wait for a jiffy
change before starting the calibration. At first I thought that was OK, but on
some platforms, HZ can be as low as 24, which means the result can vary by 33%
(based on 100 ms -> 3 jiffies).

The same change is made to atari_NCR5380.c.
I guess you plan to deduplicate this code when merging the drivers, i.e.
after this series?

Gr{oetje,eeting}s,

Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/