On Sun, Jun 19, 2022 at 01:31:42PM +0100, Jonathan Cameron wrote: > On Thu, 16 Jun 2022 17:02:08 +0000 > Dmitry Rokosov <DDRokosov@xxxxxxxxxxxxxx> wrote: > > > > > > > + err = -EINVAL; > > > > + mutex_lock(&msa311->lock); > > > > + for (odr = 0; odr < ARRAY_SIZE(msa311_odr_table); ++odr) > > > > + if (val == msa311_odr_table[odr].val && > > > > + val2 == msa311_odr_table[odr].val2) { > > > > + err = msa311_set_odr(msa311, odr); > > > > > > > + if (err) { > > > > + dev_err(dev, "cannot update freq (%d)\n", err); > > > > + goto failed; > > > > + } > > > > > > Why is this inside the loop and more important under lock? Also you > > > may cover the initial error code by this message when moving it out of > > > the loop and lock. > > > > > > Ditto for other code snippets in other function(s) where applicable. > > > > > > > Yes, I can move dev_err() outside of loop. But all ODR search loop > > should be under lock fully, because other msa311 operations should not > > be executed when we search proper ODR place. > > I don't see why? The search itself is for a match of the input to const data. > That can occur before taking the lock to do the actual write. > > I don't see any additional race beyond the one that is always there of > a thread updating ODR whilst another is accessing the device. Which order > those events happen in is not controlled by the driver, but the output > will be consistent with one or other order of those two accesses. > > Jonathan Agreed, will be changed in the v4. -- Thank you, Dmitry