I2C_F_POLL (poll for completion -- no sleeping) is used a lot in sys/dev/i2c.
In some places, it's used conditionally based on the value of 'cold', which
seems fine, and I am going to centralize that logic in i2c_exec.c so it doesn't
have to be replicated.
But there's a number of places using I2C_F_POLL outside of cold auto
configuration, and I can't believe that it's actually necessary in all of them.
Each i2c device has exclusive access to the i2c bus while doing its operation,
and I don't believe that's allowed in any interrupt context at all. Certainly,
I don't think the various i2c real-time clock drivers need to use I2C_F_POLL --
inittodr() is called very late in the startup process, and resettodr() is not
called in the crash or shutdown process. That said, resettodr() *is* called
from kern_time.c:settime1(), which wraps the entire function body in a
splclock() / splx() pair. This seems ... unnecessary. Seems like the
following diff is safe / appropriate, though I'm still suspicious of the whole
splclock() thing.
Comments / thoughts / rotten produce?
diff --git a/sys/kern/kern_time.c b/sys/kern/kern_time.c
index 6ff70ad74be9..8226372404f8 100644
--- a/sys/kern/kern_time.c
+++ b/sys/kern/kern_time.c
@@ -160,9 +160,9 @@ settime1(struct proc *p, const struct timespec *ts, bool
check_kauth)
tc_setclock(ts);
timespecadd(&boottime, &delta, &boottime);
+ splx(s);
resettodr();
- splx(s);
return (0);
}
-- thorpej
