Re: __might_sleep in uio_read()?

[Michal Hocko]

On Thu 03-09-15 11:47:34, Andy Grover wrote:
> On 09/03/2015 05:26 AM, Michal Hocko wrote:
> >On Wed 02-09-15 15:45:10, Andy Grover wrote:
> >>Hi Hans and Greg,
> >>
> >>Is this an issue with uio? I swear it didn't used to throw this warning...
> >>
> >>Thanks -- Andy
> >>
> >>[ 5174.883261] ------------[ cut here ]------------
> >>[ 5174.883617] WARNING: CPU: 0 PID: 1532 at
> >>/home/agrover/git/kernel/kernel/sched/core.c:7389 __might_sleep+0x7d/0x90()
> >>[ 5174.884407] do not call blocking ops when !TASK_RUNNING; state=1 set at
> >>[<ffffffffa02a5821>] uio_read+0x91/0x170 [uio]
> >
> >The warning says that the driver is calling copy_to_user with
> >TASK_INTERRUPTIBLE which is wrong in general because this context can
> >sleep and a schedule would destroy the state. It doesn't matter here
> >because the code would break out from the loop regardless of the
> >copy_to_user return value.
> >
> >I assume that TASK_INTERRUPTIBLE is necessary before the event_count
> >check to prevent from wake up races. If that is the case then you can
> >simply do:
> >diff --git a/drivers/uio/uio.c b/drivers/uio/uio.c
> >index 3257d4220d01..7d8959e3833b 100644
> >--- a/drivers/uio/uio.c
> >+++ b/drivers/uio/uio.c
> >@@ -524,6 +524,7 @@ static ssize_t uio_read(struct file *filep, char __user *buf,
> >
> >  		event_count = atomic_read(&idev->event);
> >  		if (event_count != listener->event_count) {
> >+			__set_current_state(TASK_RUNNING);
> >  			if (copy_to_user(buf, &event_count, count))
> >  				retval = -EFAULT;
> >  			else {
> >
> 
> This certainly makes the warning go away. If this looks good to everyone
> else can we get this change in?
---