Hi, On Tue, Sep 1, 2020 at 6:47 AM Jonathan Cameron <Jonathan.Cameron@xxxxxxxxxx> wrote: > > On Tue, 1 Sep 2020 00:03:55 -0700 > Stephen Boyd <swboyd@xxxxxxxxxxxx> wrote: > > > Quoting Doug Anderson (2020-08-31 11:59:37) > > > On Sat, Aug 29, 2020 at 10:18 AM Jonathan Cameron <jic23@xxxxxxxxxx> wrote: > > > > Agreed. However, I'd like a comment next to the place we turn it on > > > > saying what delays we are trying to mitigate by enabling it in this > > > > driver. > > > > > > OK, I used function graph to get a trace. Some of the time is simply > > > i2c transfers, which we do A LOT of during initialization (and i2c > > > transfers aren't known for being instant). However, that's not the > > > big problem. Check out the regmap_read_poll_timeout() in > > > sx9310_init_compensation(). 20 ms per sleep with a 2 second timeout. > > > This seems to eat over around 220 ms of the time on my system. > > > > > > So how do you want to do this? Do you want me to re-post the patch > > > and mention the regmap_read_poll_timeout() in the commit message? Do > > > you want to just add that to the commit message yourself? > > > > > > > I think the suggestion is to put a comment next to the > > PROBE_PREFER_ASYNCHRONOUS flag indicating that > > sx9310_init_compensation() can take a while. > > Yup. That's what I meant. Sorry for lack of clarity. > A comment in the code stops others copying it without making sure they > understand whether it is relevant in their case. No one is going > to look in commit messages. OK, posted. I expect it'll show up soon at: https://lore.kernel.org/r/20200901081920.v2.1.Id02b2f451b3eed71ddd580f4b8b44b3e33e84970@changeid > > Should the compensation be > > delayed even further though to some time when userspace opens the device > > node? > > It's tricky because we may hit it on an individual sysfs read. > We could spin off a thread to do it I suppose, though that would be > more or less the same as allowing async probe. Also: even if we delayed the compensation further we'd still end up with a somewhat slow probe just due to the metric buttload of i2c transfers we do at probe time. -Doug
