Re: BUG: sleeping function called from ras_epow_interrupt context

On 07/16/2015 01:23 AM, Thomas Huth wrote:

On 07/15/2015 09:58 PM, Nathan Fontenot wrote:

quoted

On 07/15/2015 09:35 AM, Thomas Huth wrote:

quoted

On 07/14/2015 11:22 PM, Benjamin Herrenschmidt wrote:

quoted

On Tue, 2015-07-14 at 20:43 +0200, Thomas Huth wrote:

quoted

Any suggestions how to fix this? Simply revert 587f83e8dd50d? Use
mdelay() instead of msleep() in rtas_busy_delay()? Something more
fancy?

A proper fix would be more fancy, the get_sensor should happen in a
kernel thread instead.

I'm not very familiar with this stuff, but isn't the EPOW interrupt
something that is very time-critical? Moving parts of the handler into a
kernel thread then does not sound like a very good idea to me...

Another question: Can it happen at all that this get-sensor call results
in a sleep condition? Looking at commit ID
81b73dd92b97423b8f5324a59044da478c04f4c4 ("Fix might-sleep warning on
removing cpus"), which apparently fixed a similar issue for CPU
hot-plugging, indicates that at least some of the rtas calls are never
returning the busy code? In that case we could fix this by introducing a
similar rtas_get_sensor_fast() function? (or simply revert 587f83e8dd50d
which would be quite similar, I think)

Looking at the PAPR, the get-sensor-state rtas call for the EPOW sensor
is listed as a fast call and should not return a busy indication.

Great, good to know, thanks for looking that up! So IMHO we should
either introduce a rtas_get_sensor_fast() function or revert
587f83e8dd50d ... any preferences? Shall I come up with a patch?

A quick look at the kernel, I only find three places that rtas_get_sensor
is called. The instance you point out here for the EPOW sensor is the
only time I find it called for a sensor that should not return a busy
indication.

Reverting commit 587f83e8dd50d would solve the issue but not fix any future
users of a fast get-sensor call. I don't have an issue with a patch for a
rtas_get_sensor_fast().

-Nathan
 

quoted

I'm curious as to why we're getting a busy return indication when
making this call.

Looking at the code again, rtas_busy_delay() likely never slept ... it's
likely just the "might_sleep()" annotation in that function that causes
the BUG.

 Thomas