Re: sleepy linux self-test

* Ingo Molnar [off-list ref] wrote:

 [   23.893598] Calling initcall 0xc0c518b0: be_sleepy+0x0/0x170()
 [   23.901601] PM: no wakelarm-capable RTC
 [   23.905599] initcall 0xc0c518b0: be_sleepy+0x0/0x170() returned 0.
 [   23.910879] initcall 0xc0c518b0 ran for 3 msecs: be_sleepy+0x0/0x170()

so close, yet so far away :-)

good news: the suspend+resume self-test works perfectly now! :-)

[   24.018541] Calling initcall 0xc0c70a40: be_sleepy+0x0/0x170()
[   24.023780] PM: wakealarm test with Suspend-to-RAM
[   24.027503] PM: Syncing filesystems ... done.
[   24.032177] PM: Preparing system for mem sleep
[   .........]
[   30.317160] initcall 0xc0c70a40: be_sleepy+0x0/0x170() returned 0.
[   30.321593] initcall 0xc0c70a40 ran for 6289 msecs: be_sleepy+0x0/0x170()

that's a successful suspend+resume cycle, all automatic!

now, this was _totally_ non-obvious to set up. (config attached) I had 
the (surely incorrect) impression that the new RTC code does everything 
it can to keep itself not used ;-) It did not complain that it's not the 
default RTC driver, and it did not complain that it's configured in a 
way that makes it functionally inferior.

Please dont be that shy and give us a really prominent option that does 
something like:

   CONFIG_USE_THE_COOL_NEW_RTC_CODE=y

with all its bells and whistles, and punch out the old 
drivers/char/rtc.c RTC driver if this option is enabled. Believe me, 
people _want_ to run your code on all the PC distros too. (and as a 
side-effect all the non-PC platforms are helped too)

Small feature request: could you please measure the duration 
suspend+resume cycle and print out the result to the kernel console like 
this:

   INFO: suspend+resume test succeeded and took 1.289 seconds.

such a message would be useful in flagging suspend+resume delay 
regressions. (and to automate the bisection of such regressions, etc.) 
(Please keep the millioseconds portion too - that way i can follow the 
finer details of suspend+resume performance over time as well.)
 
For example recently someone introduced a 5-10 seconds delay into the 
suspend+resume cycle (yeah, that bloke was me), so it would also be nice 
to add some hard limit on how much it takes:

   WARNING: suspend+resume test took more than 5 seconds.

and print a stack trace via WARN_ON(1), so that kerneloops.org can pick 
up such regressions immediately. This would be dependent on this 
self-test option.

please also rename CONFIG_PM_WAKEALARM_TEST to something more 
descriptive, such as: CONFIG_SUSPEND_SELF_TEST=y.

another set of small usability feature requests: please also add a boot 
option to disable/enable the suspend self-test, and allow the test to be 
built into the kernel but disabled by default. Perhaps even a sysctl to 
activate it runtime. That way distros can build it in by default but can 
keep it disabled.

	Ingo