On 06/30/2017 08:54 PM, Stephen Boyd wrote:

On 06/30, Gabriel FERNANDEZ wrote:

quoted

On 06/30/2017 02:20 AM, Stephen Boyd wrote:

quoted

On 06/29, Gabriel FERNANDEZ wrote:

quoted

On 06/28/2017 05:59 PM, Stephen Boyd wrote:

quoted

On 06/27, Gabriel FERNANDEZ wrote:

quoted

On 06/22/2017 12:07 AM, Stephen Boyd wrote:

quoted

readl_poll_timeout?

if i use readl_poll_timeout (wich use 'ktime_get()') it can be
operational only after the selection of clocksource ? (device_initcall).
And then if a driver turn on a clock before, it could blocked the linux
console ?

Ok. I wonder if we could add some sort of starting check to
readl_poll_timeout() that tests system_state for booting vs.
scheduling? That should be sufficient to handle this case?

Oops i think i understood my problem...
i used readl_poll_timeout in atomic context.
I have to move my code in the .prepare ops.

If you are ok with that i will send a v5

There's readl_poll_timeout_atomic() for those modes.

yes it's exactly the test i made (use 'readl_poll_timeout()_atomic' in
.enable ops) but i'm blocked.

if i do the same in .prepare ops with 'readl_poll_timeout()' it's ok.

I'm still confused. readl_poll_timeout_atomic() uses ktime_get(),
and so does readl_poll_timeout(), so how does moving to the
prepare op fix the problem? What's the actual problem?

Yes both use ktime_get().

The issue concerns internal/external oscillator clocks (time 
stabilization could be long) and
if a driver wants to enable one these clocks before device_initcall().

By default the clocksource is the jiffies until the end of the boot 
(fs_initcall)
Then after, the best clocksource is selected (arm_system_timer or stm32 
timer, etc...)
There is no problem after because the counter is a hardware register.

But the jiffies counter is incremented by interruption and enable op
does not allow to be interrupted. (we can with prepare op).