On 20/06/2019 08:25, Kishon Vijay Abraham I wrote: > On 14/06/19 6:08 PM, Marc Gonzalez wrote: > >> The issue is usleep_range() being misused ^_^ >> >> Although usleep_range() takes unsigned longs as parameters, it is >> not appropriate over the entire 0-2^64 range. >> >> a) It should not be used with tiny values, because the cost of programming >> the timer interrupt, and processing the resulting IRQ would dominate. >> >> b) It should not be used with large values (above 2000000/HZ) because >> msleep() is more efficient, and is acceptable for these ranges. > > Documentation/timers/timers-howto.txt has all the information on the various > kernel delay/sleep mechanisms. For < ~10us, it recommends to use udelay > (readx_poll_timeout_atomic). Depending on the actual timeout to be used, the > delay mechanism in timers-howto.txt should be used. Hello Kishon, I believe the proposed patch does the right thing: a) polling for the ready bit is not done in atomic context, therefore we don't need to busy-loop b) since we're ultimately calling usleep_range(), we should pass an appropriate parameter, such as max_us = 10 (instead of max_us = 1, which is outside usleep_range spec) Maybe it would help if someone reviewed this patch. Regards.
