Re: [RFC PATCH] mtd: tests: Fix eraseblock read speed miscalculation for lower partition sizes

[Miquel Raynal <miquel.raynal@bootlin.com>]

Hi Amit,

+Cc: David, who's maintaining the tools. Please keep him in the
recipients list!

amit.kumar-mahapatra@xilinx.com wrote on Thu, 3 Feb 2022 18:54:34 +0530:

> While calculating speed during  mtd_speedtest, the time interval
> (i.e., start - finish) is rounded off to the nearest milliseconds by
> ignoring the fractional part. This leads to miscalculation of speed.
> The miscalculation is more visible while running speed test on small
> partition sizes(i.e., when partition size is equal to eraseblock size or
> twice the eraseblock size) at higher spi frequencies.
> 
> For e.g., while calculating eraseblock read speed for a mtd partition with
> size equal to the eraseblock size(i.e., 64KiB) the eraseblock read time
> interval comes out to be 966490 nanosecond. This is then converted to
> millisecond(i.e., 0.966 msec.). The integer part (i.e., 0 msec) of the
> value is considered and the fractional part (i.e., 0.966) is ignored,for
> calculating the eraseblock read speed. So the reported eraseblock read
> speed is 0 KiB/s, which is incorrect.
> 
> There are two approaches to fix this issue.
> 
> First approach will be to keep the time interval in millisecond. and round
> up the integer value, with this approach the 0.966msec time interval in the
> above example will be rounded up to 1msec and this value is used for
> calculating the speed. Downside of this approach is that the reported speed
> is still not accurate.
> 
> Second approach will be to convert the time interval to microseconds
> instead of milliseconds, with this approach the 966490 nanosecond time
> interval in the above example will be converted t0 966.490usec and this
> value is used for calculating the speed. As compared to the current
> implementation and the suggested First approach, this approach will report
> a more accurate speed. Downside of this approach is that, in future if the
> mtd size is too large then the u64 variable, that holds the number of
> bytes, might overflow.
> 
> In this patch we have gone with the second approach as this reports a more
> accurate speed. With this approach the eraseblock read speed in the above
> example comes out to be 132505 KiB/s when the spi clock is configured at
> 150Mhz.
> 
> Signed-off-by: Amit Kumar Mahapatra <amit.kumar-mahapatra@xilinx.com>
> ---
> BRANCH: mtd/next
> ---
>  drivers/mtd/tests/speedtest.c | 10 +++++-----
>  1 file changed, 5 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/mtd/tests/speedtest.c b/drivers/mtd/tests/speedtest.c
> index 93e76648f676..2b76e7750c68 100644
> --- a/drivers/mtd/tests/speedtest.c
> +++ b/drivers/mtd/tests/speedtest.c
> @@ -161,13 +161,13 @@ static inline void stop_timing(void)
>  static long calc_speed(void)
>  {
>  	uint64_t k;
> -	long ms;
> +	long us;

Should this be an explicit 64-bit value? And unsigned?
unsigned long long int or uint64_t? I believe we are now 1000x closer
to the 4GiB limit so we might need to enlarge this variable.

>  
> -	ms = ktime_ms_delta(finish, start);
> -	if (ms == 0)
> +	us = ktime_us_delta(finish, start);
> +	if (us == 0)
>  		return 0;
> -	k = (uint64_t)goodebcnt * (mtd->erasesize / 1024) * 1000;
> -	do_div(k, ms);
> +	k = (uint64_t)goodebcnt * (mtd->erasesize / 1024) * 1000000;
> +	do_div(k, us);
>  	return k;
>  }
>  

Otherwise lgtm!

Reviewed-by: Miquel Raynal <miquel.raynal@bootlin.com>


Thanks,
Miquèl