On 03/19/2018 03:48 PM, Marek Vasut wrote: > On 03/19/2018 02:43 PM, Vladimir Zapolskiy wrote: >> On 03/19/2018 12:56 PM, Marek Vasut wrote: >>> On 03/19/2018 11:53 AM, Geert Uytterhoeven wrote: >>>> Hi Marek, >>>> >>>> On Mon, Mar 19, 2018 at 10:53 AM, Marek Vasut <marek.vasut@xxxxxxxxx> wrote: >>>>> On 03/19/2018 09:38 AM, Simon Horman wrote: >>>>>> On Sun, Mar 18, 2018 at 11:52:52AM +0100, Marek Vasut wrote: >>>>>>> The data link active signal usually takes ~20 uSec to be asserted, >>>>>>> poll the bit more often to avoid useless delays in this function. >>>>>>> >>>>>>> Signed-off-by: Marek Vasut <marek.vasut+renesas@xxxxxxxxx> >>>>>>> Cc: Geert Uytterhoeven <geert+renesas@xxxxxxxxx> >>>>>>> Cc: Phil Edworthy <phil.edworthy@xxxxxxxxxxx> >>>>>>> Cc: Simon Horman <horms+renesas@xxxxxxxxxxxx> >>>>>>> Cc: Wolfram Sang <wsa@xxxxxxxxxxxxx> >>>>>>> Cc: linux-renesas-soc@xxxxxxxxxxxxxxx >>>>>> >>>>>> Unless my eyes deceive me this seems to be quite a lot (100x) more often, >>>>>> but so be it. >>>>> >>>>> It's just a higher frequency to avoid slowdown when bringing the link up. >>>> >>>> No it isn't: you replaced a sleep by a delay, thus making it blocking. >>> >>> For much shorter period of time. >>> >>>> So this can spin for up to 50 ms (+ overhead)? >>> >>> That's what it did before too , it used msleep and now it uses udelay. >>> >> >> msleep() does not spin, it reschedules the process. >> >> Instead to find a balance you may want to play with usleep_range(). > > Which does not work in atomic context, which will be needed when using > this function in suspend/resume later on. > IIRC suspend/resume are never executed in atomic context, and runtime suspend/resume are invoked in atomic context only if pm_runtime_irq_safe() is called (just a few drivers in vanilla uses it at the moment). Nevertheless, the commit message does not say that the change is needed for future suspend/resume add-on. -- With best wishes, Vladimir
